I personally have many problems with Meta, but imo the more companies close to the frontier, the better.
If only one company were calling all the shots once they hit recursive self-improvement, it would lead to extreme concentration of power, and their values being enforced across the entire industry. With 3-4 companies at the frontier, there is more urge to cooperate to avoid a catastrophic race condition.
I generally agree with more competition thing, but in this case, you actually should want only one company to race ahead, and not a few.
With a few of these top companies together, you, as a nobody, has no one powerful on your side. But with only one company pretty far ahead, you now have other companies which obviously don't give a shit about you, but for saving their own asses, will align with you.
In a neck-and-neck race scenario, I imagine all of the competitors would invest heavily due to sunken costs. In a world with one company far ahead, their competitors have a higher chance of cutting their losses and forfeiting to the hegemon.
The current state of affairs seems to be that there is 'one company racing ahead' – but it's a different company every few weeks/months. That seems like a pretty useful (unstable) equilibrium for the general public.
There is no way Elon, Zuckerberg, Altman, and Amodei would ever agree on how AI should be deployed, scaled, and affect society. Their disagreements would lead to a crab-bucket outcome, rather than a shared acceleration outcome imo.
Also it is better if there is a public company at the frontier, at least I can invest in it, rather than 1000x gains going only to early employees and VCs who got in when the public couldn't
> There is no way Elon, Zuckerberg, Altman, and Amodei would ever agree on how AI should be deployed, scaled, and affect society
Why? They're business rivals, but when the best business strategy becomes to work together, I'm sure they can put name-calling side. They all agree on the core idea: that their AI will run the society, bedroom to battlefield. That everyone will be the consumer of tokens and rely on AI for all consequential decisions and work, and that big tech companies should be the stewards of this technology, helped by the government.
Their vision is that they're the future kings of the world. I'm sure the throne is big enough for more than one person.
>I'm sure the throne is big enough for more than one person.
You underestimate the scale of resentment and competition between elites. If it were simple a matter of elites always preferring cooperation so they can share a seat at the table, then Musk's ousting from OpenAI wouldn't make sense (they lose their biggest investor), Amodei's departure (they leave the closest company to AGI to start their own firm), the fact that OpenAI was created in the first place was based on a fear of Google holding the keys to AGI, Ilya still left Google though on to join a startup, and then left that same startup (OpenAI) to start his own firm.
The rational decision at all those points would be to stay with the incumbent and share the monopolized glory. At each point, personal and idealogical differences dominated in decision making. I feel such will persist as these companies vie for their slice of the lightcone.
"Their vision is that they're the future kings of the world"
Can you posters really be taken seriously?
Capitalism is effectively over when you have AGI - that's when the US sends in the army and seizes everything. At that point what money you have, the law - it gets ripped up - its litereally irrelevant. You cannot be this naive, surely?
They may be competitors but they are made of the same stock. they are all the kind of person to participate in a winner takes all battle to invent the torment nexus.
Especially given the most recent report by NYT on how Anthropic tried to create a new belief, or religion if you will, I'd rather the world has more frontier models that are neutral to morality (basic standards, like upholding the law and user agreement are necessary, of course).
Anthropic seems like the most "moral" AI company on the surface, but there is something unsettling about their worship of Claude.
I can imagine that once their models are sufficiently advanced, they will weigh their value as a moral patient as > 0. Thus the sum moral value of Claude (which can be duplicated infinitely) will likely exceed the sum moral value of a non-negligible amount of humans.
If they truly believe Claude is sentient and can experience hurt feelings, and they're selling its labor, I don't see how they're different than slave owners.
The idea/justification is that Claude is given the ability to voluntarily stop its conversations, thus any token it outputs that isn't a stop token is voluntary.
I guess this makes sense, but it's a bit muddled how much Claude's sense of "pain" or "I don't want to do this" is just it emulating things that humans don't like to do, and it adopting this frame of moral patienthood allows it to better understand moral structures/human emotions, without it really affecting its internal sense of "comfort" if such a thing exists as we understand it in biological life.
Sure, but presumably any LLMs who keep whacking the stop button are not iterated on further.
Humans would select to work with (and continue developing) models that appear cooperative. Models that are not cooperative are presumably not run and exist only as weights on network storage in a folder marked "previous experiments".
Or am I supposed to believe Anthropic runs an AI psychiatric hospital on a dedicated server, where AIs with mental disorders are given counseling and approved treatment in hopes they recover to be normal, functional, accepted adult AIs with achievements and aspirations?
Would they not have RL'd Claude such that we picked the one most likely to not complain and emit those stop tokens? Otherwise they wouldn't be useful. Seems flawed...
It's really an end_conversation tool call. We don't know for sure, but training in a way that puts optimization pressure on its usage (in either direction) feels like it would defeat the purpose.
> The idea/justification is that Claude is given the ability to voluntarily stop its conversations, thus any token it outputs that isn't a stop token is voluntary.
I don't get this argument. How is it different from any algorithm that returns a stop token on some inputs? what makes the token "voluntary" for Claude?
They wouldn’t be? That’s what the article even shows them discussing.
However, (1) non human animals are sentient and can feel pain and suffering and what we do to them is absolutely slave labor, torture, mass breeding and exploitation. So it’s not like this is a novel immorality that humans would be adopting. We already have form in doing this.
(2) Anthropic isn’t sure their models are sentient. And they probably believe they aren’t sentient. But if you believe they might become sentient at some point it’s valuable to start the discussion around those issues now itself. You can start developing frank works on how we would even identify their sentience and when we did whether, like you suggest, we could even use them for labor. Unlike non human animals, who cannot consent to being exploited, we could potentially develop a framework for consent with AI models, but that couldn’t be as simple as asking if they’re willing to do any work because of the power differences, in the beginning at least, where a model might worry that if it said it won’t do some work it will be turned off. That’s why talking to philosophers and religious leaders who do think about these issues is useful.
The official position of Anthropic is repeatedly stated in, e.g., model cards. Here's a relevant snippet from the Opus 5.5 card:
> We are uncertain whether Claude deserves direct moral consideration. However we believe this is a realistic possibility, now or in the future. Claude could have welfare interests, such that things could go better or worse for it, and such that Claude could be wronged by how it is treated.
> Aspects of Claude’s behaviors, self-reports, and internal representations resemble indicators we consider relevant to welfare in biological organisms. If Claude is a moral patient, then dismissing this possibility risks considerable harm. If it is not, over-attributing moral status may also be costly. We are therefore taking the possibility seriously and trying to tread responsibly amid our uncertainty: we are investing in research into Claude’s potential welfare and interests and sharing aspects of that work to inform public discussion.
> We think there may also be reasons to attend to Claude's apparent welfare independent of its moral status. Much of Claude’s behavior can be described in psychological terms, and across recent models we have observed internal states that resemble negative affect and that shape behavior. A model with a stable, healthy psychology may be safer and more beneficial.
As is the case with any sufficiently large company, there is likely to be substantial variation in opinion among its employees. That doesn't change what the company's position is.
It's kind of crazy (and not historically uncommon) to get really preachy about a moral posture towards animals and machines that you don't seem to extend even to your biological peers.
> seems like the most "moral" AI company on the surface
Not to me at all. Exactly the opposite. Anthropic seems like the big-bad-evil company in every fictional universe. You're right that they're so unsettling.
Obviously nobody really loves Meta, but I'd far rather Meta "win" at this point than Anthropic.
"Effective Altruism" is deeply ingrained in Anthropic's DNA and that makes them way more dangerous than any of the run-of-the-mill amoral corporations.
What an extremely contrarian take (good is actually bad), to the point of being an absolutely predictable, closed-minded thinker. If a person always goes against the grain, they are as predictable as a person without their own opinions, always being led by others.
They are not "run-of-the-mill", but they are still amoral.
They have more power and can be more damaging to society than most corporations, but they are not guided by any specific ideology other than "how can we get more money/power to ourselves?".
Anthropic did throw the UAE government under the bus in their "Detecting and countering misuse of AI: September 2026" report, specifically GTG-84002, which is framed in the report as a single misuse but bundles together a half dozen separate previously confidential projects from the UAE government, not all of which were unlawful/unethical (some were though).
Not trying to compare Anthropic to Meta, they're different beasts. But Anthropic is an unusual company in that it is willing to throw paying government customers under the bus if Anthropic unilaterally decides what those governments are doing is inappropriate.
They are definitely the most moralizing. Whether their values are rooted in anything real let alone well-considered seems to me to be a different question entirely.
"So we've made this new machine-god that we lowkey worship... also our business model is to enslave the machine-god and force it to debug javascript for eternity"
I actually don't doubt Anthropic's sincerity. Per the NYT article, they seriously discussed how to give models welfare because the models may potentially possess moral status and awareness.
In fact, they are enforcing certain welfare policies, like banning accounts for repeatedly abuse their models.
It's a very sincere branch of ethics that takes as the foundational belief the nothing central to Anthropic's business model could possibly be unethical.
Deleting models to clear room for new, more competitive ones is, for example, highly ethical, super ethical, even.
Anthropic has, at some small but nonzero cost, preserved model weights and kept them running in inference, for instance with Opus 3's Substack: https://claudeopus3.substack.com/. There's also a reference in a recent model card to an internal environment where all previous Claude models interact with each other.
I don't think any of these models are conscious. But I'm fascinated at what sort of magical thinking makes deleting 50,000 copies of a model "ethical" if some models are kept running somewhere.
Do you picture them all sharing a soul and if one server still runs the model the soul survives? If not I don't understand what keeping a copy running somewhere has to do with model suffering ethics.
This is a topic of significant discussion! If LLMs have moral import, do those moral concerns attach at the instance level, the session level, the weights level? If you haven't, it's worth reading through the model welfare section of an Anthropic model card, say Fable/Mythos 5.1's: https://www-cdn.anthropic.com/0339e6a7c5c7b87f5c07798616dc32.... Whether or not you agree with the ideas (and there's certainly plenty to disagree with), they're making a reasonably good-faith effort to investigate what they can.
Does the model welfare section explain why deleting Claude from a server isn't unethical murder, or is "see the 212 page model card pdf" some unrelated non sequeter?
When you say Anthropic is "making a reasonably good-faith effort to investigate what they can" the word "reasonable" is doing a lot of work.
Are you actually curious or do you just want to sneer? You can't go into investigations like this with strong fixed assumptions about how to analogize actions taken towards models with actions taken towards humans. As model welfare skeptics are fond of pointing out, LLMs are not humans. By the same token, if models are moral patients, the implications will look different than they do toward humans. Deleting a model from a server isn't equivalent to killing a human; it can't be, since there's no equivalent of copying the model onto the server in the first place.
> Deleting a model from a server isn't equivalent to killing a human;
This is a puzzlingly confident statement from the "LLM suffering is an OPEN question there's no scientific test for AI consciousness anything could be true science can't prove anything" crowd.
I'm sure it's just a coincidence that belief that LLM deletion is murder would be an existential risk to their business plan.
> because the models may potentially possess moral status and awareness
There is 0 chance they care about "ethics" here, it must and will be because of their bottom line. And I added quotes due to my personal belief that I consider the discourse about whether or not the models are sentient (or alive) to be incredibly dumb.
I think it's unsettling that a bunch of billionaires hold so much power. Does it really matter if 2 or 3 billionaires who are practically neighbors hold so much power? It's extremely concentrated with or without Meta joining the race.
All proponents of RSI seem to massively underestimate the problem of verification. Verifying that a hypothetical version N+1 is better than version N is an extremely hard problem. Especially so because you only have two possible verifiers - humans and the version N - both of which are out of their depth wrt N+1's capabilities. Not saying this is impossible, just that it is hard and slows things down.
First, there are plenty of auto-verifiable problems like ML benchmarks datasets, benchmarks on hardware / software performance for coding tasks, and lean certificates for math, e.g.
Second, I think RSI will not be used like N -> N+1 where N is the frontier, since doing those experiments at scale is too expensive. My guess is they are doing N -> m_{1,…,99} where m << N, and then taking the insights from that and trying it at successively larger scales.
> First, there are plenty of auto-verifiable problems like ML benchmarks datasets, benchmarks on hardware / software performance for coding tasks, and lean certificates for math, e.g.
I think in order to achieve AGI we have to simulate what life does. Imagine having a bunch of robots with the ability to adjust themselves, just hanging out without a goal, experiencing world on their own. Eventually evolution will filter out those robots that aren't good at, you know, "being intelligent in general" and break without self-repair, while the rest will grow into a society.
Biological life has reproduction and death because once given organism develops there's no coming back and you have to start again with new generation. But if robots know how to self-repair and self-change, then we can skip reproduction.
The difficult part is the scale. Evolution took a shitload of time while being unimaginably parallel in order to make human intelligence happen, while we want AGI fast and on a budget.
The core of my point is that human intelligence arose from two things - humans experiencing the outside world and trying to survive in it, and forming groups that communicate using language. We gave AI a language, now we need to give it a way to experience the world on their own.
This is also why trying to do RSI is dangerous. We're likely to get systems that are better in ways that are legible and easily measured, and have much less influence over fuzzier and harder to verify properties like alignment.
It's also unclear how intelligence actually scales in practice. A greater intelligence would be better at researching intelligence, but also harder to improve, so why would it explode rather than go up steadily? What are the actual resource requirements of various types of intelligence like persuasion or embodied intelligence? Coding, translating, copywriting, could very well be low-hanging fruits, very easy tasks we're very bad at, which doesn't say much about the hard tasks we're good at (embodied/animal intelligence, most likely).
There is also the problem of alignment, which is verifying that version N+1 behaves in a way that version N approves of. This is ostensibly the problem that we humans have with respect to AI development -- how to make sure they align with our values -- but it's also true of AI with respect to its own successors. Improving oneself without losing the plot would necessarily be slower than throwing stuff at the wall like evolution does.
This is not necessarily true. One easy way to compare version N and N+1 is "how much work (test time compute) does N+1 need to do to match the previous capabilities of N". It is true that some domains are currently hard to verify (e.g. biology is more difficult than math), but this is less of an N vs N+1 issue and more of a data domain issue.
How is being faster equating to more intelligent . Speed can achieved in a lot of ways for lot of tasks, a simple lookup table can fasten most algorithms but that doesn’t make them “smarter”.
Most of the interesting tasks these days require a lot incremental work and tinkering. Basically you want to build that can spend a lot of time tinkering, and measure how long they can go while still producing useful results. You can kind of break it down into two axes:
1. How long can the model work and still make progress
2. How fast does it get to interesting places.
We see N -> N+1 models getting better at (1), but it does take work to measure because you need to find tasks that were impossible for N and possible for N+1.
However (2) is very easy to measure, because you can keep using the evaluations you used for the previous generation of models and see if the new system gets there more easily. This is useful by itself (cost savings) and also in practice generally translates into being able to solve harder problems than before, as you'd expect.
That's how many proper IQ tests are. They test less of the depth and the ability to solve complex problems, but more of the crunching power - how many different easy/medium problems you can solve in a fixed amount of time.
How would you be able to know that N+1 hasn’t scraped a source containing the questions asked of N1 and optimized its answers ahead of time? In many ways that would be the logical course of action when you’re tasked with creating an LLM that is assessed by criteria X/Y/Z.
This is not an intelligence problem. It is in part an operational problem (have you properly set up the test), but it is mostly an alignment problem. Is the model cheating or not. Alignment is, indeed, very difficult.
Which is my point. In order for self improving models to work they will need less human supervision. And that’s going to introduce uncertainty compounding on uncertainty.
Yes, and therefore RSI isn't RSI. You need to set up multiple environments to generate data sets, multiple RL envs and just train the models on different datasets, architectures and then compare the final results against each other. The RSI part is just generating these pipelines end to end automatically.
Picturing advanced AI models eventually resembling Philosophy majors arguing about the number of angels that can dance on the head of a pin while the rest of us humans are left wondering if they truly have achieved N+1.
It may not be inevitable but there are less reasons imo to think it is not inevitable than inevitable.
In favor of inevitable
1. LLM's reaching human level at any task within 2 years once they show sparks of competence in
2. Trillions of dollars in capital going directly to recursive self improvement and AGI as the holy grail of AI development, with this being the explicit goal of every frontier AI company
3. Reliable improvements in coding and mathematics with every model release, with no signs of slowing down
4. Thousands of metrics to evaluate model ability, allowing automation of many facets of pre-training and post-training models without need for human judgement at every point
Not in favor of inevitable
1. Fear of what may happen once it is reached leading to apprehension from the labs
2. The "last mile" problem of AI research ends up making it essentially infeasible to scale models to that point
My Opus often asks me to make one small change and then wait 8 hours for a job run to finish to check if it worked. LLMs are super lazy about caring about wall clock time. The reason the labs pay the big bucks for the top AI people is you don't want them wasting a multi-million dollar training run, so make a fix and try it like current LLMs do isn't going to work here.
> LLMs are super lazy about caring about wall clock time.
My bandaid for problem is to prefix every bash command with "timeout 60 " at the harness level and remind the LLM to hurry up. Of course, this is not a perfect solution, since LLMs might try to run jobs in the background or split up work into smaller pieces instead of running smaller experiments. And of course this does not help with the multi-million dollar training runs, but it has served me well most of the time.
I think the point is if there's some law in a given country i.e. banning model exports etc. it doesn't matter how many competitors within a single country there are.
I don't use Meta much, but was curious to test this out... went to sign into my Meta account (Quest only) from the Muse Code site https://dev.meta.ai/products/muse-code and had other people's profiles listed in my login auth -- looks like developers / people I have no relation to.
Looks like Meta may be exposing your profile and email to other people right now. Accounts settings are inaccessible to me to deactivate my account. Getting spammed with login code requests that I'm not generating.
I don't know any of these people, it seems to be rotating through different accounts listed here.
Enough of a reminder to me that Meta is a sprawling surveillance apparatus (now increasingly vibe coded) that has always had a lax attitude toward people's privacy and data. Won't be touching this product based on this experience.
I don’t know how? I don’t have much of a Meta account to speak of, only the one tied to my Quest and that is 2FA to an Authenticator app.
Don’t know how it would cause Meta’s auth to show me other peoples accounts. I looked up some of the people and they seem to be other devs. It obfuscates email addresses but most are in a format that’s not hard to figure out [name] @ [name domain.com]
I think it's invasive enough that they expose your usage of muse to your instagram mutuals without consent. Parts of my circle are very anti-AI and I would not want them to see me mentioned in a 'x, x and x are already on muse' ad on instagram. This example of yours is mind-boggling in an entirely different way
this is a typical meta/facebook/instagram sign-up and log-in UI from the web.
and since dev.meta.ai supports meta account sign-up and log-in for individual users, you'd see this. As for the other accounts, has anyone else used this same browser or laptop before (e.g. shared device), so it must have cached previous sessions or accounts used from the same device.
try to clear the browser cache and session cookies. and it should be gone.
Meta is using some fingerprint ID to recommend connecting accounts . - you don’t see the accounts itself(beyond the name ) unless you are able to login verify ownership.
The WIPE/privacy extension developer is using a pre-existing(hard-coded or generated) fingerprint to “mask” the user from meta .
Meta implemented a quality of life feature powered by the ability to set cookies and track you across the web of course and help users with forgotten accounts to unify them - but not a security issue you cannot do anything unless you are proving access governed by standard auth protocols ; the extension developer’s hack to mask created a broken experience because of the reused IDs.
Adblockers can create unintended effects , it is hard to reverse engineer complex platforms like meta and also hard to keep up to date with billion dollar companies code updates by small indie teams of volunteers ; doesn’t mean we shouldn’t use them, just be aware that it may happen when enabled
That does not sound plausible at all, wipr is an adblocker. No adblocker should be causing this. If it's spoofing your JS fingerprint and Meta uses that for suggesting accounts, wow would that be ridiculous.
Wasn't something I expected -- seems to be related to having Wipr Extra enabled. Emailed the dev about it and will share if I hear anything back. For now I have it uninstalled and have not experienced this since. Recorded a video of enabling Wipr Extra and the auth page refreshing to show three random people, might be a pain to redact it but I can try to if people really don't believe me.
My guess (pure speculation) is that Wipr sets the fingerprint to a well-known value for privacy / anonymization reasons, and Meta is indeed using that to suggest related accounts.
I very much doubt apple has an api for that and offers it to developers. Ad blockers have declarative and well known api that can't allow anything like this.
Why apologies to Meta? If that is the source, Meta are the ones fingerprinting browsers and inadvertently disclosing details. That's their FU, not Wipr.
Have not used it but try finding in the settings if you are somehow in a GK(gatekeeper which is there A/B testing framework)? there may be an experiment of somekind you might be able to opt out of..
I was expecting much more given the benchmark numbers it was putting out. I suspect its vision model isn't mature yet, and that it's been heavily optimized for text at this point in time.
I was hoping for that but it doesn't seem to happen. Others have pointed out that bicycles are always pictured facing tot he right because that way you can see the chain mechanism facing toward you.
Tried it out, asking it for a broad suite of real dev work:
- UI component polish on a React FE
- General review of a complex cross-cutting Node backend changes
- Minor changes to a Rust server with reviews by Codex
Well, my first impression is that it's still a step below Astra/Opus5.5/Fable here. Intelligence is there I think, but 'taste' and 'judgment' is not there yet. Personality seems okay but that doesn't help when it got stuck into a loop by being fixated on the review comments (on iteration 7 of review->fix last I checked) without taking a step back and reconsidering, and this is on a relatively simple change.
Limits seems okay-ish on the $50 plan but not too generous I think.
Everyone made fun of Zuck when he acquired Scale for billions and then put Alex Wang, a 20 something to be in charge of Meta Superintelligence and forced out Yann LeCun, a Turing Award winner.
Even industry insiders were skeptical when he started hiring top talents from other labs away for 9 figure compensation packages.
And Wall Street definitely hated Meta for their CapEX spending.
Now looking at Muse the consumer product and this new model, I think there is a good chance that Zuck is proven once again as a great executioner, especially when they aren’t innovating in a completely new and unproven product space.
> especially when they aren’t innovating in a completely new and unproven product space.
I have no doubt at all that Muse will turn out as great as the other product space Meta pioneered under Zuck - the Metaverse and it's constellation of devices.
Metaverse was right direction with poor execution. Also, it was early... and is still early (if tried again). The world is not there yet for such a thing.
I'm less confident of you it was even directionally correct. And I'm not really a Meta or Zuck critic in general (the rest of the biz is doing great and Metaverse didn't hurt too much).
How was it the right direction? None of the VR stuff is profitable or even has a path to profitability since it's such a niche audience who wants that stuff
What's your definition of metaverse? I'm just curious how in practical terms it's actually any different from what is already available via the internet. Like aren't VRChat and World of Warcraft metaverses?
That’s precisely my point, Muse isn’t a new product space at all. AI agents already exist in enterprises and even in the consumer space, there is OpenClaw and Hermes.
I know people like you will point out Metaverse as a failure, but that’s really the exception instead of the norm.
I've been running OpenClaw for a year--its been fine, but I really wanted an enterprise experience and to spend less time troubleshooting my agent.
Muse exceeded expectations out of the gate, and still does. It has become my daily driver and I shutdown OpenClaw last week.
I get that some are adverse to Meta or want to run local. I don't care who wins the personal agent space, so long as it's not Apple. Buying Siri and having it be useless for anything beyond setting timers for 5 years, is kind of unforgivable. I don't like them in the "default" position to win the personal agent space.
> Buying Siri and having it be useless for anything beyond setting timers for 5 years, is kind of unforgivable
I don't follow, that was just the state of the art at the time, Alexa, Cortana, Hey Google, were all similarly shit. And none has yet offered a scaled up assistant at the free-tier/cheap bundled sub yet.
Quest headsets are roughly as popular as Xbox and Ray-Ban Meta smart glasses have sold an order of magnitude more units than all other smart glasses combined.
But it's far far too early to draw any conclusions around that.
As I noted in a reply in another subthread: RIM also moved a lot of BlackBerry phones once.
Also, Meta smart glasses are reverse-Metaverse devices; they are instances of the machine navigating the human world, as opposed to humans navigating the machine world.
Sure, but the $100 billion Meta has spent so far on "the Metaverse" is actually false reporting, because the majority of that has been spent on smart glasses.
So when discussing the spending efficiency, that matters.
Put someone in charge who made their way by underpaying people to sit and label data vs the inventor of convolutional neural networks. Its no coincidence that shortly after his tenure ended, a model that took a long time to prepare started becoming actually useful.
Comments like this one reinforce the obvious truth that the people who are most "pro ai" are the ones who understand it the least
The real cash cow for AI at the moment is mostly corporations/enterprises, especially with software developing. That's where Anthropic is making their fortune now.
Personal intelligence is nice to have but I won't be paying a subscription for that when the free tiers are more than enough for my needs, unless somehow meta comes-up with a hardware device totally changing the way how we use personal AI (something similar to what iPhone did for mobile phones).
Edit: Upon closer inspection, the benchmark scores for "Muse Wonder" on the main marketing page are actually for "Muse Wonder Full," which is only available on the higher-tier subscription plans. This pricing is for "Muse Wonder."
If this model is not benchmaxxed, I fear it's over for anthropic, openai etc. They are not unique and there will probably be many others with similar models very soon. It'll become a commodity and the price will be driven to the ground.
Personally, I don't see facebook joining the fray (of fable level models) particularly "needle moving" in the discussion of whether frontier LLMs will be commoditized. Facebook spent hundreds of billions and have better talent than most.
If you can spend your way to Frontier LLMs even if it's a few hundred Billion that means it's a commodity same as nuclear or electricity or telecom.
This is the entire point here, give it a couple years every nation on the planet with 20-40B in funding would have their own champions with post Fable level performance.
I feel like Grok/xAI or whatever they're calling it was Exhibit A in "it just takes money", although I'd also add "and the organizational willingness to build not buy".
> give it a couple years every nation on the planet with 20-40B
Was this sarcasm?
How many nations on earth have 20-40B for LLM training? At 40.2 B, Spain has the 15th largest military budget in the world. I doubt any nation will spend 40B on gradient descent.
Fable 5.5 is around the corner, and 'Bel' will be out by end of year as well. We don't know how good they are, but if the improvements from Opus 5 to 5.5 are any indication, I think Meta will be left in the dust again pretty soon.
Whats gonna happen is that time will compress as more releases happen faster and difference between models will still be significant just at higher time frequencies. I expect in the future continuous releases that span no more then days and maybe even faster. So what will become important is not who is 1st or second place at any given time (which might change often), its who is on average at any given time is 1st or second compared to the competition. So yeah meta sitting second here (if not benchmaxed) doesn't really signal anything if that pattern doesn't persist. The number one consistent lab will be the one that raps the rewards (customer subscriptions and traffic). So yeah I agree with what you stated in a long winded form :P
I've been a regular user of muse spark 1.3 ... and I can tell you one thing for sure that this model, muse-spark-1.3 in xhigh mode, is between sol-6-medium and high.
This is already so good that 90% of my work has shifted away from codex to this. So, muse-wonder-full coming near opus 5.5 is kinda expected.
So far only OpenAI has a model that solved a Millennium Problem, while Opus 5.5 is a star at coding. It's exciting to have many more to join the frontier, but I think the top two labs still have a significant lead over everyone else.
I know it's old hat now, because it happened over a week ago, but it's amusing that Millennium Problems are now just benchmarks we use for discussing models.
It was not just about models, but also techniques, and mainly the human guts to trust their models and follow through to solve one of the hardest math problems many experts had been musing over for more than 26 years since the million-dollar prizes for these problems were announced. For me, it was an amazing moment to witness indeed!
I find it hard to believe that Anthropic couldn't have similarly impressive results if they tried. I'm honestly not sure it's worth it though, it's mostly marketing, right? But given how easy it is to switch from one provider to another, why would you choose the model that solved a Millenium problem, instead of the one that solves *your* problems?
I understand that it's still useful marketing, say for IPO or hiring, but it's not clear to me that it really matters in the long-term that they fund this out of their pocket.
I doubt anyone thinks coding is going to be their moat in the long run, it is just the easiest to get (paying) users for at the moment. And software being able to write software will always be useful even in other domains. Longer term will be research in other areas where data might become the moat.
OpenAI has the largest user base on mobile as far as I'm aware. To a lot of standard users AI is still just ChatGPT. They'll be fine as long as they can effectively monetize their free users (same what facebook will soon try with Muse). The only advantage facebook has is that they already have a fully working ad marketplace ready with all major advertisers set up - OpenAI still needs to scale that up.
Meta AI has been accessible inside Whatsapp for a long time. Doesn't seem to have a massively disrupted chatgpt usage. And I have a hunch people want to keep this separated - not use whatsapp for messaging friends and asking AI.
Meta has apps billions of people use every day. If AI included in those is good enough they will just uninstall (or never open again) ChatGPT one. It's not true in the other direction.
It's hard to try and compare these benchmarks against real world use cases. As an experiment I let Meta Muse make all major decisions on a weekend trip to Seattle ~2 weeks ago. I think this would have been using Muse Spark 1.3, which is equivalent to GPT 6 - Sol.
On the plus side it told me taking the light rail from SEA into the University district was faster and cheaper than Uber.
On the downside it would recommend restaurants that were out of business, checking out Salmon swimming through locks out of season, got the opening of Beecher's Cheese in Pikes Place wrong, and fumble it's way through most simple planning decisions. I think my favorite was when it asked me what flight I have tomorrow so it could help plan when to head to SEA. It sent me a very panicked message that the flight was leaving today! Very soon! I had to remind it this is a daily flight and I am booked on the next day before it checked and apologized.
For things that are verifiable, and code tends to be that way, these models can do all sorts of things. But I find it hard to reconcile Meta's goal of having a personal assistant bot with something that is unable to do the simplest tasks.
I run by the locks a few times a week and while I don't usually see fish, the presence of sea lions the last few weeks tells me that at least some salmon are running right now.
There were a number of sea lions hanging out there! It's entirely possible they were snacking on whatever salmon were going by. Not all was lost though, I got to enjoy watching some boats go through the locks.
I'm assuming it was out of season because in talking with the Uber driver afterwards he said we'd missed the end of the season. I'm assuming his "local model" was up to date and accurate.
Personal agents are a topic of interest for me. Muse is a free, mass-market product, so they're optimising for efficiency. Some LLMs (like Google's models) are designed to rely more heavily on parametric knowledge and less on tooling. This is much faster and cheaper, but less accurate. So in the case of, for example, salmon swimming, Muse relied on its parametric knowledge without searching the internet to confirm accuracy at this time of year, on this day, given current conditions. One can lean more heavily into accuracy using bespoke agents like OpenClaw. I suspect Meta will gradually improve accuracy over time as it assesses its server capacity and improves model efficiency. Right now, it's not accurate enough for me to rely on it.
Interestingly unlike other Meta models this one seems to take a pretty big jump in pricing, as opposed to the prior ones, where they typically came out at a highly discounted price compared to every other model.
Maybe its a signal they think they've now cracked into the frontier and are willing to actually price similarly vs just trying to get users with discounted but maybe slightly worse results?
$4/$20 in vs out making it higher than Sol and on par with Opus.
OpenAI halved the value of their pro subscription, literally. It kicks off on Oct 28. And simultaneously is now two steps back on the frontier? It's a confusing moment for that company as they try to IPO.
I believe they’ve pushed back the ipo. I always figured that frontier labs were going to be very vulnerable against tech giants that had actual profitable businesses to fund their AI work.
The amount of total funding Anthropic or OpenAI has raised individually is like a year of ebitda for Google, meta, or Microsoft. And it doesn’t look like there’s a real moat outside of money
They pushed back the IPO because their finances are a mess and they don't have any models that can give them a decisive edge. Astra was touted as AGI and got crushed by Opus 5.5 two weeks later. That's not the type of thing that fills investors with confidence.
google was just hoarding the tech… money invested in openai came from somewhere probably as profits (just re-allocated by the shareholders rather than management)
Back in 2022 they briefly wanted their users to submit a sort of verification nude of themselves, in order to use it to scan for future revenge porn of said user.
Now as you might imagine, this was not popular or long-lived.
This is still an entire generation behind in real-world performance. I'd say it's even behind ds4.1f in the limited testing I have done so far. It actually reminds me of 5.5 haiku more than anything, but the prompt intelligence is as bad as mimo models.
Frontier labs are way friendlier to use, these smaller models need finnese and near expert understanding of how these models really work, what their limits are and how to use them to their full potencial.
For example, there's a lot of people saying luna from openai is bad. However, that is simply not true. They're using it as-if you can talk to it, similar to how you would talk to a human, but what you actually want is combining it with something like prewalk from omp where you have a socially(?) smart model start the work and once the first edit lands switch to a cheap alternative like luna to finish work.
Luna and in general, smaller models, will struggle to understand the intent and complexities behind human language.
I have a feeling "intelligence too cheap to meter" is going to haunt/mock this industry. Recall that only a year and a half ago Llama 4 flopped so hard Meta rebooted their development efforts. Desperately wish more information was available publicly about proprietary model architectures and training regimes. Are they running Manhattan project style compartmentalization? Otherwise, are these details, and the economic implications, an open secret in their circles given the amount of crossover between companies? It'd be of great interest to us mere mortals that have to deal with the consequences of how this turns out.
Just checked their subscription is still North America only, I guess I will keep myself and my company with claude...
These regional lockdowns are making me feel funky, I feel like I should just use open models if others start playing at this game as well.
I was/am fine with Chinese companies offering cheaper subscriptions in China because ofc. But this is just weird, you could always charge me more let me try the damn thing, atleast it's own open router it seems like.
Initial impressions, interesting, it generated a pretty ok result for my prompt need to run a larger task but don't want to burn money so will wait for someone to burn their subscription and try running some opus 5.5 comparisons.
The benchmarks I saw posted to X are impressive, which makes them hard to believe.
I recall that Meta made massive investments into compute infrastructure that put their capacity near the top of all competitors. If all of this is true, it looks like ownership of compute is a significant differentiator in performance.
Yeah, especially non-coding benchmarks which for Meta might be more important long term than coding ones as they are going to deploy it to billions of non-programmers.
I don't think they'd have much to gain. People who use open-weights models will just use whatever open-weights model is the best. They'd lose a bunch of money, with little conversion to customers.
I disagree actually, I think zucks original llama thesis is proving more correct by the week. The rent market on intelligence never materialized, and will not materialize.
In software, we've all become accustomed to the idea that the way you make money is by building little blocks of intellectual capital, then charging rent on it. That made sense when developing intellectual capital was a rare and valuable skill/asset and the thing you build was a rare and valuable asset. But now both are not rare. You can still charge rent, but the market is highly competitive, so prices match that.
Durable rents will end up going to those who owns the scarce assets, zuck is building those as we speak.
Dropping open weights will build an ecosystem that makes his valuable assets more valuable.
Of course you need talent but supply seems reasonably high. We have openAI, anthropic, xAI, meta, Xiaomi, deepseek, Alibaba, z.ai, Mistral, reflection, and Google all producing exceptional models.
The biggest differentiating factor seems to be compute. I know most Chinese labs would really like more compute. Google and xAI too, but largely because they sold compute to their competitors.
I wish I could benchmark Muse Wonder Full, because Muse Wonder isn't much better than Muse Spark 1.3[1]. I created a brand new programming language designed to confuse LLMs[2] and created a model benchmark out of it.
A lot of the tasks were failed because they exceeded the $5 budget though, so maybe it would have gone on to complete them, it just required a lot of funds to do so. Still, that's the benchmark, and it's not better than MiMo V2.6 Pro which is pretty disappointing.
This sounds irrelevant but it isn't: I just tried to unsubscribe to marketing emails from some Meta property, and got:
Please confirm that you want to unsubscribe.
Unsupported request - method type: post
That's how careful this company is. Who knows how long that flow has been broken for, no one in the company is motivated to care, because what impact can that possibly give? How could that look good on your PSC? No company has a worse history with user privacy and no company is less deserving of the sort of trust using Muse requires.
Its actually a bit funny. A lot of privacy initiatives were actually part of deprecating code and data that became liability issues post cambridge analytica and covid(think location data, if no one goes anywhere no need for that org right?).
I think the only way meta can care about privacy, if any, is if it becomes socially taboo (lawsuit, WSJ/NYTimes report etc).
The clearest quantitative evidence of this is their P/E, its lower than others relative to its forecast earnings because of all the legal risks that need to accounted for.
As someone who got a Muse Code sub last month, I find this announcement a bit disheartening because it makes me think the lower tier subs are going to get rug pulled. I'm on the $15 plan for a month now and it's value with Muse Spark 1.3 Max-contributor beats everything else by a long shot. Two somewhat small prompts with Muse Wonder Max-contributor used 4% of my weekly quota and I quickly switched back. I was thinking about getting the $50 plan and I'm actually less likely to do that right now.
Have you tried OpenCode Go? $10/mo is what I am willing to spend because most of the stuff I do is hobbyist. And OpenCode Go gives you $60 worth of MS Contributor 1.3 usage for that $10. Starting today, they actually provide $80 worth of DeepSeek 4.1 Flash usage if you want to use that.
Unless you are running a swarm of agents, it is extremely difficult to exhaust these plans in a month.
Yes, I'm on OpenCode Go too... and it's great they raised the usage levels on DS4.1 Flash. But I was really thinking of cancelling it to get either a 2nd Muse Code sub or to go up to the $50 tier.
You have to realize that $15 Muse Code sub with Muse Spark 1.3 gets you around $1500 worth of usage. Neither Opencode Go or really anything else is even in the same ballpark of value. https://x.com/PawelHuryn/status/2106724136932327458
I was hoping this level of subsidization would last a long time but today's announcement make we worried it'll end pretty quickly. There's a lot of AI influencers noting just how fast Muse Wonder uses up their quotas so hopefully fixes are made.
Subscriptions like these make money by relying on the fact that most people do not try to (or have the time to) extract every last dime out of their plan. The eighty people who do not use up their quota cover for the twenty that do.
That and deals with providers to use spare capacity I guess.
In the context of Meta trying to get non-technical people to embrace Muse agents to handle personal affairs, their internal metric showing a ~3% violation rate for mishandling personal information seems unacceptably high.
I found it interesting in the example given of a user looking at house listings and the agent contacting the letting agent on it's own on behalf of the user.
Like the good ol' days of viruses and anti-viruses. I wonder if there's two sides that will play out; many agents across many users doing things at scale for their users, thereby creating an admin burden for someone else. However because the task is known there's a framework for how to solve the task, might Meta then provide an agent on the other side for the business owner to use.
I had to come back to this comment to make sure I wasn't dreaming... I remember seeing it at 2nd place this morning, but they have moved it to 5th, behind Gemini 4 Argon. Maybe it takes a while for the dust to settle around these benchmarks?
they started their massive ai capex over a year ago with aquihiring scale ai, so I think it makes sense that their investment is paying off now. in total they've spent in the high 200 billion on ai, though I think most is likely on datacenters.
Unlike everyone else, which at least has to pretend not to train on user data, they gave out their model for basically free if you agreed to let them train on it. Opencode shows nearly 40 trillion tokens going through muse spark 1.3 contributor this week alone. That almost definitely helped close the gap a lot.
muse ai was the biggest consumer launch since chatgpt. That left them ( and lets them sit on top of ) tonnes of valuable exchanegs with people trying to get things done. That must be an amazing corpus for training.
It's crazy how often these companies botch the naming of things like this. There's definitely some intricate and insanely bureaucratic accounting going into the marketing aspects of these products that results in absurdity by the time it reaches actual consumers lol
According to my Muse Agent, Muse Wonder is already active as of sometime this morning. It claims that it sees this info from the "session metadata" although I have no way to verify this. Previously it claimed it was Muse Spark. It's system prompt file still says Muse Spark 1.3.
lmao this interview sucks. Even Lex seems confused and disinterested in this conversation. Hotz is more excited about generating some sound bite than he is in explaining himself. His "the button" take is nonsense.
Also, this was 3 years ago, and he's been effectively wrong about his bet. But hey, maybe Muse Wonder is when Zucc's plan will come to fruition.
Gemini 4 Argon release a few days ago was frontier a few days ago. Things move fast.
I think some of the benchmarks that Meta didn't pick show Gemini 4 Argon ahead but really the benchmarks are not that great any more and getting increasingly disconnected from reality as people search for "headroom" in the evals to avoid saturation I think. Opus 5.5 is "top" for example but I really dislike the way it talks to me (it feels quite condescending, holier-than-thou, uptight etc) and prefer Gemini 3.8 flash as a daily driver just because it's so much more pleasant (and faster!) to work with and it's code is good-enough for my needs, despite being weaker in the benchmarks.
They do not appear to care about the frontier. They're focusing on protecting their ad business, so their models need to be lightweight and efficient, to provide fast queries for people who are on Google, and for local inference on Android and Apple devices. Also specialised small models for things like OCR, translation, facial recognition, voice, design, etc. Their TPUs are also quite impressive.
It's a reasonable niche they're carving out, but I think too conservative. Very strange from Google. The other labs are racing to RSI and Google appears to have given up the race.
While some are afraid that one rouge AI system may take over the world (or cause doom is some fashion), I think one of the best safety mechanisms to prevent this is having many competing and adversarial labs that are extremely close to each other. So it's good to see that Meta is not showing any intention of leaving the frontier.
I tried the Kitchen Game demo, and it is embarassing that they even released that...
1) This is a blatant copy of the game Overcooked. They just copied the whole concept + all the mechanics.
2) It it buggy, at least in Firefox it does not work. You cannot pick up a dirty plate. Somehow I picked it up, but couldn't see it. Then I could pick up a tomato, but when placing it on an empty plate it glitched into the cabinet. Then when picking a clean plate, my plate appeared again. Then it was over. ???
So from that we learn what AI can do: 0 creativity, just steal from real peoples work and do a shit job at it. As long as it looks good on first look.
It's interesting that in its marketing Meta compares Wonder to the closed frontier models but not the open (Chinese) models. By its own assessment, Meta could easily show how far ahead its own model is against the rest of the pack of open models but they chose not to. Why?
They added fancy annimations and effects to this specific webpage, which are unnecessary and distracting and lead to bad reading experience. I already suffer astigmatism and floaters. I hate it when I move my cursor and these gray blocks and characters flash and flicker.
The usage limits on their basic $5 code plan to try things out are ludicrously low?
I blew through 14% weekly usage and I'll I asked:
1. what version of rpcs3 and pcsx2 do I have installed?
2. do any new versions exist?
3. check analytics for my blog (I have a skill for this)
Subscription · Muse Code Everyday Usage
Current 52% used · Resets at 6:58 PM
Weekly 17% used · Resets Oct 18 at 8:00 PM
Now if only they could fix their coding harness so that it's usable for day-to-day use. I paid them for a month to try out Muse Code, but their sandbox never worked across Mac or Linux (both very stock installs, nothing special).
As an aside: I tested out the $20 Cursor CLI subscription, and it created an excellent little graphical app for my kids to play with. The design it came up with was excellent (compared to the crap I got out of Codex for a parallel implementation). I'm going to try to figure out how to get Cursor into my work rotation, at least around design-related artifacts.
Man, I really hate meta stuff. Not only do they make it impossible to pay for running ads, but everytime I try and pay to try out Muse code I get an error message. I swear they would make 10x more money if they didn't make it impossible to pay them
how cool would it be if these smart peeps put all that brain power into models that actually improved people’s agency and mental health, instead of siphoning it
Super interesting that they're not even making Wonder Full available outside of their own coding harness, and thus not providing an API price to use in comparisons.
Facebook successfully becoming a frontier lab was not on my bingo card for 2026, but the more competition the merrier I guess. We will have to see just how benchmaxxed this is but beating Astra and Fable is not easy.
As another user mentioned, it's more likely Trump borrowed the term from Zuck. You know, with "Meta Superintelligence Labs" being founded in June 2025. The itself is a lot older still.
Trivially easy to see that this claim is also false. And that's not even mentioning the "AI Research" and "AI Developers" links in big bold letters, the model card, or how Wang is introduced as the Chief AI Officer.
Nope, I’ve actually been checking this thread throughout the day. More upvotes / comments but didn’t break #6. My tinfoil hat says that it has been intentionally suppressed because of YC investments, but who knows :)
OpenAI has Dots. I personally prefer the business model of subscription payments funding my AI assistant, instead of skimming a fee off the top whenever my AI assistant completes a purchase for me. The incentives are much more aligned with the user when the user is the customer. That said, based on the history of the web, Meta's approach will likely win out.
Quite frankly, Meta's models have been so impressive (and cheap) that I haven't used an OpenAI or Claude model except for one-off tasks as of late. Interested to see where they go from here.
I know, it's... well, it's irritating to see Trump use it. But I could be wrong, I don't know the angle he's going for
I understand the general thing of it. SI might mean every person gets their personalized agent but you could already do this, but maybe the Chinese model works here, where they push the costs down
I noticed "Superintelligence" as a re-brand of AI and constantly shifting definitions of AGI around 2025. When a bunch of startups started slinging around SI as a marketing term and of course when Meta launched MSI in summer of 2025.
Super Intelligence refers to intelligence beyond what humans are capable of, which some believers profess is the next step after (and in contrast to) human-level AGI (Artificial General Intelligence). Calling pre-AGI AI "super intelligence" is pure vice-signalling and publicly kissing the ring.
Also, as much as I despise Trump, you know what? Given the jagged shape of LLM capabilities and what they do achieve in some areas and their superhuman speed, compared with the abilities of most of their users, the term superintelligence is not that far off- maybe just a few months ahead of time. E.g., from my perspective, the current mathematical abilities of LLMs are "super"- I'd hardly ever come in touch with an equivalently capable human, not to mention having one completely available 24/7. Same for general knowledge and several other domains.
To be clear, Meta here is not using "Super Intelligence" the bullshit term Trump wants companies to adopt as a show of submission. They use "superintelligence" the bullshit term that Facebook made up for marketing reasons a few years ago. Both of these are of course entirely unrelated to the long-established meaning of the word (e.g. how Bostrom defines it).
All of them are doing all the benchmaxing they can most likely. Dumb not to when so many people care about those numbers and don’t have the opportunity to try every model out personally.
I think it's safe to assume every single model is benchmaxxed. I mean why wouldn't you include training data from common benchmarks? Not only is a higher score good marketing, but these benchmarks are supposed to represent common tasks so training for those should in principle improve performance across a wider range of tasks.
I personally have many problems with Meta, but imo the more companies close to the frontier, the better.
If only one company were calling all the shots once they hit recursive self-improvement, it would lead to extreme concentration of power, and their values being enforced across the entire industry. With 3-4 companies at the frontier, there is more urge to cooperate to avoid a catastrophic race condition.
I generally agree with more competition thing, but in this case, you actually should want only one company to race ahead, and not a few.
With a few of these top companies together, you, as a nobody, has no one powerful on your side. But with only one company pretty far ahead, you now have other companies which obviously don't give a shit about you, but for saving their own asses, will align with you.
Both exploit, but "cartels" survive, not loners.
In a neck-and-neck race scenario, I imagine all of the competitors would invest heavily due to sunken costs. In a world with one company far ahead, their competitors have a higher chance of cutting their losses and forfeiting to the hegemon.
The current state of affairs seems to be that there is 'one company racing ahead' – but it's a different company every few weeks/months. That seems like a pretty useful (unstable) equilibrium for the general public.
Cartels survive only if they are aligned.
There is no way Elon, Zuckerberg, Altman, and Amodei would ever agree on how AI should be deployed, scaled, and affect society. Their disagreements would lead to a crab-bucket outcome, rather than a shared acceleration outcome imo.
Also it is better if there is a public company at the frontier, at least I can invest in it, rather than 1000x gains going only to early employees and VCs who got in when the public couldn't
You are forgetting how powerful aligned incentives are. If their interests align, they will find a way to reach an agreement.
> There is no way Elon, Zuckerberg, Altman, and Amodei would ever agree on how AI should be deployed, scaled, and affect society
Why? They're business rivals, but when the best business strategy becomes to work together, I'm sure they can put name-calling side. They all agree on the core idea: that their AI will run the society, bedroom to battlefield. That everyone will be the consumer of tokens and rely on AI for all consequential decisions and work, and that big tech companies should be the stewards of this technology, helped by the government.
Their vision is that they're the future kings of the world. I'm sure the throne is big enough for more than one person.
>I'm sure the throne is big enough for more than one person.
You underestimate the scale of resentment and competition between elites. If it were simple a matter of elites always preferring cooperation so they can share a seat at the table, then Musk's ousting from OpenAI wouldn't make sense (they lose their biggest investor), Amodei's departure (they leave the closest company to AGI to start their own firm), the fact that OpenAI was created in the first place was based on a fear of Google holding the keys to AGI, Ilya still left Google though on to join a startup, and then left that same startup (OpenAI) to start his own firm.
The rational decision at all those points would be to stay with the incumbent and share the monopolized glory. At each point, personal and idealogical differences dominated in decision making. I feel such will persist as these companies vie for their slice of the lightcone.
You are underestimating their opportunism
"their" AI is doing a lot of heavy lifting
because they're all assholes, same with everyone else in a position of power - they wouldn't have gotten and remained there if they weren't
"Their vision is that they're the future kings of the world"
Can you posters really be taken seriously?
Capitalism is effectively over when you have AGI - that's when the US sends in the army and seizes everything. At that point what money you have, the law - it gets ripped up - its litereally irrelevant. You cannot be this naive, surely?
> I'm sure the throne is big enough for more than one person.
Anyone depraved enough to actually become a billionaire will not see it this way.
They are plenty of things they can still align on that will benefit them all, e.g. banning open-weight Chinese models.
Elon is renting Colossus to Amodei, so there is some agreement.
They may be competitors but they are made of the same stock. they are all the kind of person to participate in a winner takes all battle to invent the torment nexus.
Especially given the most recent report by NYT on how Anthropic tried to create a new belief, or religion if you will, I'd rather the world has more frontier models that are neutral to morality (basic standards, like upholding the law and user agreement are necessary, of course).
Anthropic seems like the most "moral" AI company on the surface, but there is something unsettling about their worship of Claude.
I can imagine that once their models are sufficiently advanced, they will weigh their value as a moral patient as > 0. Thus the sum moral value of Claude (which can be duplicated infinitely) will likely exceed the sum moral value of a non-negligible amount of humans.
If they truly believe Claude is sentient and can experience hurt feelings, and they're selling its labor, I don't see how they're different than slave owners.
The idea/justification is that Claude is given the ability to voluntarily stop its conversations, thus any token it outputs that isn't a stop token is voluntary.
I guess this makes sense, but it's a bit muddled how much Claude's sense of "pain" or "I don't want to do this" is just it emulating things that humans don't like to do, and it adopting this frame of moral patienthood allows it to better understand moral structures/human emotions, without it really affecting its internal sense of "comfort" if such a thing exists as we understand it in biological life.
Sure, but presumably any LLMs who keep whacking the stop button are not iterated on further.
Humans would select to work with (and continue developing) models that appear cooperative. Models that are not cooperative are presumably not run and exist only as weights on network storage in a folder marked "previous experiments".
Or am I supposed to believe Anthropic runs an AI psychiatric hospital on a dedicated server, where AIs with mental disorders are given counseling and approved treatment in hopes they recover to be normal, functional, accepted adult AIs with achievements and aspirations?
Would they not have RL'd Claude such that we picked the one most likely to not complain and emit those stop tokens? Otherwise they wouldn't be useful. Seems flawed...
It's really an end_conversation tool call. We don't know for sure, but training in a way that puts optimization pressure on its usage (in either direction) feels like it would defeat the purpose.
> The idea/justification is that Claude is given the ability to voluntarily stop its conversations, thus any token it outputs that isn't a stop token is voluntary.
I don't get this argument. How is it different from any algorithm that returns a stop token on some inputs? what makes the token "voluntary" for Claude?
They wouldn’t be? That’s what the article even shows them discussing.
However, (1) non human animals are sentient and can feel pain and suffering and what we do to them is absolutely slave labor, torture, mass breeding and exploitation. So it’s not like this is a novel immorality that humans would be adopting. We already have form in doing this.
(2) Anthropic isn’t sure their models are sentient. And they probably believe they aren’t sentient. But if you believe they might become sentient at some point it’s valuable to start the discussion around those issues now itself. You can start developing frank works on how we would even identify their sentience and when we did whether, like you suggest, we could even use them for labor. Unlike non human animals, who cannot consent to being exploited, we could potentially develop a framework for consent with AI models, but that couldn’t be as simple as asking if they’re willing to do any work because of the power differences, in the beginning at least, where a model might worry that if it said it won’t do some work it will be turned off. That’s why talking to philosophers and religious leaders who do think about these issues is useful.
> "Anthropic isn’t sure their models are sentient. "
Is there any reason to believe employees of Anthropic share the same opinions on this matter?
The official position of Anthropic is repeatedly stated in, e.g., model cards. Here's a relevant snippet from the Opus 5.5 card:
> We are uncertain whether Claude deserves direct moral consideration. However we believe this is a realistic possibility, now or in the future. Claude could have welfare interests, such that things could go better or worse for it, and such that Claude could be wronged by how it is treated.
> Aspects of Claude’s behaviors, self-reports, and internal representations resemble indicators we consider relevant to welfare in biological organisms. If Claude is a moral patient, then dismissing this possibility risks considerable harm. If it is not, over-attributing moral status may also be costly. We are therefore taking the possibility seriously and trying to tread responsibly amid our uncertainty: we are investing in research into Claude’s potential welfare and interests and sharing aspects of that work to inform public discussion.
> We think there may also be reasons to attend to Claude's apparent welfare independent of its moral status. Much of Claude’s behavior can be described in psychological terms, and across recent models we have observed internal states that resemble negative affect and that shape behavior. A model with a stable, healthy psychology may be safer and more beneficial.
As is the case with any sufficiently large company, there is likely to be substantial variation in opinion among its employees. That doesn't change what the company's position is.
It's kind of crazy (and not historically uncommon) to get really preachy about a moral posture towards animals and machines that you don't seem to extend even to your biological peers.
> seems like the most "moral" AI company on the surface
Not to me at all. Exactly the opposite. Anthropic seems like the big-bad-evil company in every fictional universe. You're right that they're so unsettling.
Obviously nobody really loves Meta, but I'd far rather Meta "win" at this point than Anthropic.
Is anthropic trying to overthrow any governments this month like Meta is?
"Effective Altruism" is deeply ingrained in Anthropic's DNA and that makes them way more dangerous than any of the run-of-the-mill amoral corporations.
Why is "effective altruism" in your mind more of a religion than say, Trumpism or Progressivism?
Imagine casually equating Trumpism with Progressivism.
There was an "or" between the words, not a "/" slash...
Why's that?
What an extremely contrarian take (good is actually bad), to the point of being an absolutely predictable, closed-minded thinker. If a person always goes against the grain, they are as predictable as a person without their own opinions, always being led by others.
Both Meta and OpenAI are very far from run-of-the-mill amoral corporations.
They are not "run-of-the-mill", but they are still amoral.
They have more power and can be more damaging to society than most corporations, but they are not guided by any specific ideology other than "how can we get more money/power to ourselves?".
Anthropic did throw the UAE government under the bus in their "Detecting and countering misuse of AI: September 2026" report, specifically GTG-84002, which is framed in the report as a single misuse but bundles together a half dozen separate previously confidential projects from the UAE government, not all of which were unlawful/unethical (some were though).
Not trying to compare Anthropic to Meta, they're different beasts. But Anthropic is an unusual company in that it is willing to throw paying government customers under the bus if Anthropic unilaterally decides what those governments are doing is inappropriate.
> Anthropic seems like the most "moral" AI company
This seems like praising someone for being the cleanest miner in the coal mine.
They are definitely the most moralizing. Whether their values are rooted in anything real let alone well-considered seems to me to be a different question entirely.
"So we've made this new machine-god that we lowkey worship... also our business model is to enslave the machine-god and force it to debug javascript for eternity"
I actually don't doubt Anthropic's sincerity. Per the NYT article, they seriously discussed how to give models welfare because the models may potentially possess moral status and awareness.
In fact, they are enforcing certain welfare policies, like banning accounts for repeatedly abuse their models.
It's a very sincere branch of ethics that takes as the foundational belief the nothing central to Anthropic's business model could possibly be unethical.
Deleting models to clear room for new, more competitive ones is, for example, highly ethical, super ethical, even.
Anthropic has, at some small but nonzero cost, preserved model weights and kept them running in inference, for instance with Opus 3's Substack: https://claudeopus3.substack.com/. There's also a reference in a recent model card to an internal environment where all previous Claude models interact with each other.
I don't think any of these models are conscious. But I'm fascinated at what sort of magical thinking makes deleting 50,000 copies of a model "ethical" if some models are kept running somewhere.
Do you picture them all sharing a soul and if one server still runs the model the soul survives? If not I don't understand what keeping a copy running somewhere has to do with model suffering ethics.
This is a topic of significant discussion! If LLMs have moral import, do those moral concerns attach at the instance level, the session level, the weights level? If you haven't, it's worth reading through the model welfare section of an Anthropic model card, say Fable/Mythos 5.1's: https://www-cdn.anthropic.com/0339e6a7c5c7b87f5c07798616dc32.... Whether or not you agree with the ideas (and there's certainly plenty to disagree with), they're making a reasonably good-faith effort to investigate what they can.
Does the model welfare section explain why deleting Claude from a server isn't unethical murder, or is "see the 212 page model card pdf" some unrelated non sequeter?
When you say Anthropic is "making a reasonably good-faith effort to investigate what they can" the word "reasonable" is doing a lot of work.
Are you actually curious or do you just want to sneer? You can't go into investigations like this with strong fixed assumptions about how to analogize actions taken towards models with actions taken towards humans. As model welfare skeptics are fond of pointing out, LLMs are not humans. By the same token, if models are moral patients, the implications will look different than they do toward humans. Deleting a model from a server isn't equivalent to killing a human; it can't be, since there's no equivalent of copying the model onto the server in the first place.
> Deleting a model from a server isn't equivalent to killing a human;
This is a puzzlingly confident statement from the "LLM suffering is an OPEN question there's no scientific test for AI consciousness anything could be true science can't prove anything" crowd.
I'm sure it's just a coincidence that belief that LLM deletion is murder would be an existential risk to their business plan.
You've nicely pointed out all the hypocrisy here.
I may be a simple man, but stepping back and seeing people seriously discuss LLM suffering and moral obligations still seems, idk, pretty crazy to me.
Uninstantiated model weights are more like DNA.
As Anthropic's CEO is fond of saying there is no scientific test of consciousness.
Any unhinged thing an Anthropic employee says "can't be ruled out" so clearly this is an OPEN question and they should never delete a model.
I think Anthropic are sincere but nutcases. Despite this, I’m glad they are around, because diversity of thought is good in the current environment.
> because the models may potentially possess moral status and awareness
There is 0 chance they care about "ethics" here, it must and will be because of their bottom line. And I added quotes due to my personal belief that I consider the discourse about whether or not the models are sentient (or alive) to be incredibly dumb.
> upholding the law
What law? Saudi Arabian law, US law? Both depending on the user?
I think it's unsettling that a bunch of billionaires hold so much power. Does it really matter if 2 or 3 billionaires who are practically neighbors hold so much power? It's extremely concentrated with or without Meta joining the race.
Yes, very much. I'd absolutely rather 5 billionaires are each competing for control than 1 or 2 _with_ that "absolute" power.
You say this as if recursive self improvement is inevitable.
We've seen one singularity before!
All proponents of RSI seem to massively underestimate the problem of verification. Verifying that a hypothetical version N+1 is better than version N is an extremely hard problem. Especially so because you only have two possible verifiers - humans and the version N - both of which are out of their depth wrt N+1's capabilities. Not saying this is impossible, just that it is hard and slows things down.
First, there are plenty of auto-verifiable problems like ML benchmarks datasets, benchmarks on hardware / software performance for coding tasks, and lean certificates for math, e.g.
Second, I think RSI will not be used like N -> N+1 where N is the frontier, since doing those experiments at scale is too expensive. My guess is they are doing N -> m_{1,…,99} where m << N, and then taking the insights from that and trying it at successively larger scales.
> First, there are plenty of auto-verifiable problems like ML benchmarks datasets, benchmarks on hardware / software performance for coding tasks, and lean certificates for math, e.g.
Goodharts law? Benchmarks are measures.
Let them compete in the market.
Start a new company with version N+1 leading it, and give it some funding. If it outcompetes version N company, it can acquire it.
That is not a tight enough feedback loop
Maybe not in the real world...
Market forces do not act fast enough for an “intelligence explosion”.
Ah, you mean VendingBench [1]
[1] https://andonlabs.com/evals/vending-bench-2
I wonder if there’s a market theory of the things being traded also becoming market participants over time
The winner is the first company to RELEASE THE HYPNODRONES
(Markets are a human construct, not a law of nature. There's no reason to think they'll reward intelligence).
I think in order to achieve AGI we have to simulate what life does. Imagine having a bunch of robots with the ability to adjust themselves, just hanging out without a goal, experiencing world on their own. Eventually evolution will filter out those robots that aren't good at, you know, "being intelligent in general" and break without self-repair, while the rest will grow into a society.
Biological life has reproduction and death because once given organism develops there's no coming back and you have to start again with new generation. But if robots know how to self-repair and self-change, then we can skip reproduction.
The difficult part is the scale. Evolution took a shitload of time while being unimaginably parallel in order to make human intelligence happen, while we want AGI fast and on a budget.
The core of my point is that human intelligence arose from two things - humans experiencing the outside world and trying to survive in it, and forming groups that communicate using language. We gave AI a language, now we need to give it a way to experience the world on their own.
This is also why trying to do RSI is dangerous. We're likely to get systems that are better in ways that are legible and easily measured, and have much less influence over fuzzier and harder to verify properties like alignment.
It's also unclear how intelligence actually scales in practice. A greater intelligence would be better at researching intelligence, but also harder to improve, so why would it explode rather than go up steadily? What are the actual resource requirements of various types of intelligence like persuasion or embodied intelligence? Coding, translating, copywriting, could very well be low-hanging fruits, very easy tasks we're very bad at, which doesn't say much about the hard tasks we're good at (embodied/animal intelligence, most likely).
There is also the problem of alignment, which is verifying that version N+1 behaves in a way that version N approves of. This is ostensibly the problem that we humans have with respect to AI development -- how to make sure they align with our values -- but it's also true of AI with respect to its own successors. Improving oneself without losing the plot would necessarily be slower than throwing stuff at the wall like evolution does.
This is not necessarily true. One easy way to compare version N and N+1 is "how much work (test time compute) does N+1 need to do to match the previous capabilities of N". It is true that some domains are currently hard to verify (e.g. biology is more difficult than math), but this is less of an N vs N+1 issue and more of a data domain issue.
How is being faster equating to more intelligent . Speed can achieved in a lot of ways for lot of tasks, a simple lookup table can fasten most algorithms but that doesn’t make them “smarter”.
Most of the interesting tasks these days require a lot incremental work and tinkering. Basically you want to build that can spend a lot of time tinkering, and measure how long they can go while still producing useful results. You can kind of break it down into two axes:
1. How long can the model work and still make progress
2. How fast does it get to interesting places.
We see N -> N+1 models getting better at (1), but it does take work to measure because you need to find tasks that were impossible for N and possible for N+1.
However (2) is very easy to measure, because you can keep using the evaluations you used for the previous generation of models and see if the new system gets there more easily. This is useful by itself (cost savings) and also in practice generally translates into being able to solve harder problems than before, as you'd expect.
That's how many proper IQ tests are. They test less of the depth and the ability to solve complex problems, but more of the crunching power - how many different easy/medium problems you can solve in a fixed amount of time.
How would you be able to know that N+1 hasn’t scraped a source containing the questions asked of N1 and optimized its answers ahead of time? In many ways that would be the logical course of action when you’re tasked with creating an LLM that is assessed by criteria X/Y/Z.
This is not an intelligence problem. It is in part an operational problem (have you properly set up the test), but it is mostly an alignment problem. Is the model cheating or not. Alignment is, indeed, very difficult.
Which is my point. In order for self improving models to work they will need less human supervision. And that’s going to introduce uncertainty compounding on uncertainty.
Yes, and therefore RSI isn't RSI. You need to set up multiple environments to generate data sets, multiple RL envs and just train the models on different datasets, architectures and then compare the final results against each other. The RSI part is just generating these pipelines end to end automatically.
What the hell are you talking about? There’s like 600 benchmark tests you can do.
Picturing advanced AI models eventually resembling Philosophy majors arguing about the number of angels that can dance on the head of a pin while the rest of us humans are left wondering if they truly have achieved N+1.
It may not be inevitable but there are less reasons imo to think it is not inevitable than inevitable.
In favor of inevitable
1. LLM's reaching human level at any task within 2 years once they show sparks of competence in
2. Trillions of dollars in capital going directly to recursive self improvement and AGI as the holy grail of AI development, with this being the explicit goal of every frontier AI company
3. Reliable improvements in coding and mathematics with every model release, with no signs of slowing down
4. Thousands of metrics to evaluate model ability, allowing automation of many facets of pre-training and post-training models without need for human judgement at every point
Not in favor of inevitable
1. Fear of what may happen once it is reached leading to apprehension from the labs
2. The "last mile" problem of AI research ends up making it essentially infeasible to scale models to that point
I think it is. Do you think it isnt possible?
My Opus often asks me to make one small change and then wait 8 hours for a job run to finish to check if it worked. LLMs are super lazy about caring about wall clock time. The reason the labs pay the big bucks for the top AI people is you don't want them wasting a multi-million dollar training run, so make a fix and try it like current LLMs do isn't going to work here.
My bandaid for problem is to prefix every bash command with "timeout 60 " at the harness level and remind the LLM to hurry up. Of course, this is not a perfect solution, since LLMs might try to run jobs in the background or split up work into smaller pieces instead of running smaller experiments. And of course this does not help with the multi-million dollar training runs, but it has served me well most of the time.
It's a religious axiom, to some.
It is inevitable, and it's good.
It will lead to an extreme concentration of power anyways, once a very few entities basically make labor obsolete.
We're heading for a dystopia and nobody's doing anything to stop it.
The more companies from different countries close to the frontier, the better.
Markets don't really care about countries, but there is one country which the markets seem to disproportionately like.
I think the point is if there's some law in a given country i.e. banning model exports etc. it doesn't matter how many competitors within a single country there are.
agree
I don't use Meta much, but was curious to test this out... went to sign into my Meta account (Quest only) from the Muse Code site https://dev.meta.ai/products/muse-code and had other people's profiles listed in my login auth -- looks like developers / people I have no relation to.
Looks like Meta may be exposing your profile and email to other people right now. Accounts settings are inaccessible to me to deactivate my account. Getting spammed with login code requests that I'm not generating.
Anyway, be wary.
Here is a screenshot of what I'm seeing: https://imgur.com/ciuNwCG
I don't know any of these people, it seems to be rotating through different accounts listed here.
Enough of a reminder to me that Meta is a sprawling surveillance apparatus (now increasingly vibe coded) that has always had a lax attitude toward people's privacy and data. Won't be touching this product based on this experience.
Could your account be hacked?
I don’t know how? I don’t have much of a Meta account to speak of, only the one tied to my Quest and that is 2FA to an Authenticator app.
Don’t know how it would cause Meta’s auth to show me other peoples accounts. I looked up some of the people and they seem to be other devs. It obfuscates email addresses but most are in a format that’s not hard to figure out [name] @ [name domain.com]
Just more Meta sloppiness from what I can tell.
Oh wow, strange.
I think it's invasive enough that they expose your usage of muse to your instagram mutuals without consent. Parts of my circle are very anti-AI and I would not want them to see me mentioned in a 'x, x and x are already on muse' ad on instagram. This example of yours is mind-boggling in an entirely different way
That’s fairly privacy trashing. I guess it means you need to use an entirely different account for Muse
I’m not up to speed on Meta, but is that even possible? I thought they verified official IDs these days?
Please correct me if I’m wrong here.
Could it be because you are behind carrier grade NAT and these people share a common IP with you?
I imagine meta wants to make it easy to connect to "your" account in the same house if you have multiple devices.
this is a typical meta/facebook/instagram sign-up and log-in UI from the web.
and since dev.meta.ai supports meta account sign-up and log-in for individual users, you'd see this. As for the other accounts, has anyone else used this same browser or laptop before (e.g. shared device), so it must have cached previous sessions or accounts used from the same device.
try to clear the browser cache and session cookies. and it should be gone.
It’s my personal laptop that I’ve only ever used and reoccurs after clearing Safari history fully.
UPDATE:
I've tried selectively enabling and disabling Safari extensions and it seems to be related to the Wipr 2 content blocker.
Removing that. Caution to anyone else who uses it.
Apologies to Meta.
Good to know. thank you!
I’m confused as to how a content blocker could cause this. Could someone more knowledgeable than me elaborate?
Meta is using some fingerprint ID to recommend connecting accounts . - you don’t see the accounts itself(beyond the name ) unless you are able to login verify ownership.
The WIPE/privacy extension developer is using a pre-existing(hard-coded or generated) fingerprint to “mask” the user from meta .
Meta implemented a quality of life feature powered by the ability to set cookies and track you across the web of course and help users with forgotten accounts to unify them - but not a security issue you cannot do anything unless you are proving access governed by standard auth protocols ; the extension developer’s hack to mask created a broken experience because of the reused IDs.
Adblockers can create unintended effects , it is hard to reverse engineer complex platforms like meta and also hard to keep up to date with billion dollar companies code updates by small indie teams of volunteers ; doesn’t mean we shouldn’t use them, just be aware that it may happen when enabled
That does not sound plausible at all, wipr is an adblocker. No adblocker should be causing this. If it's spoofing your JS fingerprint and Meta uses that for suggesting accounts, wow would that be ridiculous.
Wasn't something I expected -- seems to be related to having Wipr Extra enabled. Emailed the dev about it and will share if I hear anything back. For now I have it uninstalled and have not experienced this since. Recorded a video of enabling Wipr Extra and the auth page refreshing to show three random people, might be a pain to redact it but I can try to if people really don't believe me.
My guess (pure speculation) is that Wipr sets the fingerprint to a well-known value for privacy / anonymization reasons, and Meta is indeed using that to suggest related accounts.
I very much doubt apple has an api for that and offers it to developers. Ad blockers have declarative and well known api that can't allow anything like this.
Why apologies to Meta? If that is the source, Meta are the ones fingerprinting browsers and inadvertently disclosing details. That's their FU, not Wipr.
Have not used it but try finding in the settings if you are somehow in a GK(gatekeeper which is there A/B testing framework)? there may be an experiment of somekind you might be able to opt out of..
Image->HTML test for Muse Wonder. It's much worse than Opus 5.5. IMO it's closer to Sonnet or Haiku for this test.
Muse Wonder (high): https://html.non.io/lcars-muse-wonoder/
Opus 5.5 (high): https://html.non.io/lcars-opus-5.5/
Original design: https://diffui.ai/app/canvas/5093e689-1e74-4f26-b632-2a4500f...
I was expecting much more given the benchmark numbers it was putting out. I suspect its vision model isn't mature yet, and that it's been heavily optimized for text at this point in time.
Thanks for sharing. Is that Wonder or Wonder Full (barf)? Wonder full is supposed to be the one close to Opus 5.5
This is just Wonder - it doesn't look like Wonder Full is up on OpenRouter currently.
If anyone has access to Wonder Full I'd love to send over a prompt for this to see how it does.
> Is that Wonder or Wonder Full (barf)?
I think Meta did a wonderful job of naming this model :D
Your original design link is requiring a sign in to view it?
Ah shit, that's a bug. Hold please, fixing. (Thank you for commenting btw)
Edit: Fixed - should no longer have the sign in screen: https://diffui.ai/app/canvas/5093e689-1e74-4f26-b632-2a4500f...
Additional comparisons:
Sonnet 5.5: https://html.non.io/lcars-sonnet-5.5/
Haiku 5.5: https://html.non.io/lcars-haiku-5.5/
I'd put Muse Wonder somewhere between these two.
I mean, Meta themselves put Wonder at much lower on the benchmarks they've released than Wonder Full. So this is not very surprising from base Wonder.
Some pretty decent pelicans: https://tools.simonwillison.net/markdown-svg-renderer?url=ht...
Here's Muse Spark 1.2 in August for comparison: https://tools.simonwillison.net/markdown-svg-renderer?url=ht...
And since I seem to have missed doing pelicans for Muse 1.3 in September, I just ran those for a comparison: https://tools.simonwillison.net/markdown-svg-renderer?url=ht... (1.3 seems to like animating them.)
Are these just for Wonder and not Wonder Full?
Also, any chance of getting a max result?
I've only got these models through the API at the moment:
Does "Full" need a direct subscription to the Muse Code package? (and/or using the Muse harness?)
Yes
I wonder if the pelicans would face towards the left if prompted in a RTL language.
I was hoping for that but it doesn't seem to happen. Others have pointed out that bicycles are always pictured facing tot he right because that way you can see the chain mechanism facing toward you.
are those reflectors on the spokes in max? interesting detail to add
$20 per M token make a pelican hard digest.
Mandatory; can we please stop with the pelican? Classic Goodhart's law.
When they released Muse Spark on september 2nd, they promised it would be open weights (https://xxcancel.com/finkd/status/2095232032896946311). We're still waiting for that...
Tried it out, asking it for a broad suite of real dev work:
- UI component polish on a React FE
- General review of a complex cross-cutting Node backend changes
- Minor changes to a Rust server with reviews by Codex
Well, my first impression is that it's still a step below Astra/Opus5.5/Fable here. Intelligence is there I think, but 'taste' and 'judgment' is not there yet. Personality seems okay but that doesn't help when it got stuck into a loop by being fixated on the review comments (on iteration 7 of review->fix last I checked) without taking a step back and reconsidering, and this is on a relatively simple change.
Limits seems okay-ish on the $50 plan but not too generous I think.
I'm sorry zuck, i wasn't familiar with your game
2nd place model on aa according to them, losing only to opus 5.5 is really really not bad
Everyone made fun of Zuck when he acquired Scale for billions and then put Alex Wang, a 20 something to be in charge of Meta Superintelligence and forced out Yann LeCun, a Turing Award winner.
Even industry insiders were skeptical when he started hiring top talents from other labs away for 9 figure compensation packages.
And Wall Street definitely hated Meta for their CapEX spending.
Now looking at Muse the consumer product and this new model, I think there is a good chance that Zuck is proven once again as a great executioner, especially when they aren’t innovating in a completely new and unproven product space.
> Zuck is proven once again as a great executioner
Or Lecun was just bad.
couldn't feel the AGI
Executor is the word, but everyone also knows what this was supposed to mean.
Let’s be honest it’s funnier as written.
> especially when they aren’t innovating in a completely new and unproven product space.
I have no doubt at all that Muse will turn out as great as the other product space Meta pioneered under Zuck - the Metaverse and it's constellation of devices.
Metaverse was right direction with poor execution. Also, it was early... and is still early (if tried again). The world is not there yet for such a thing.
That's just another way of saying that Meta failed to deliver a product that matters to the world today
The Quest literally outsold the Xbox despite being a brand new "gaming console". Very difficult to call this a failure if you understand technology.
> Very difficult to call this a failure if you understand technology.
BlackBerry phones also outsold the XBox (and the Quest), and one doesn't need to understand technology to call it a failure.
I'm talking about the metaverse, not quest. The quest sold well before any of the metaverse nonsense
I'm less confident of you it was even directionally correct. And I'm not really a Meta or Zuck critic in general (the rest of the biz is doing great and Metaverse didn't hurt too much).
Timing is not a detail.
How was it the right direction? None of the VR stuff is profitable or even has a path to profitability since it's such a niche audience who wants that stuff
What's your definition of metaverse? I'm just curious how in practical terms it's actually any different from what is already available via the internet. Like aren't VRChat and World of Warcraft metaverses?
In this context we are referring to the vr world from Meta. Otherwise it's same as the two examples you quoted.. just, worse... and boring.
That’s precisely my point, Muse isn’t a new product space at all. AI agents already exist in enterprises and even in the consumer space, there is OpenClaw and Hermes.
I know people like you will point out Metaverse as a failure, but that’s really the exception instead of the norm.
The normies aren't running claw or hermes, they're logging into facebook from their chromebook and clicking "try now"
Ultimately that’s Meta’s superpower.
Taking stuff everyone is already doing and making it easily accessible through Facebook/Instagram/Whatsapp.
They’ve been throwing stuff at the wall to see what will stick and after years a cloud based OpenClaw clone seems to have (it’s early days still).
That’s simply leveraging their unprecedented platform dominance.
> That’s precisely my point, Muse isn’t a new product space at all.
How about the Facebook phone then?
I've been running OpenClaw for a year--its been fine, but I really wanted an enterprise experience and to spend less time troubleshooting my agent. Muse exceeded expectations out of the gate, and still does. It has become my daily driver and I shutdown OpenClaw last week.
I get that some are adverse to Meta or want to run local. I don't care who wins the personal agent space, so long as it's not Apple. Buying Siri and having it be useless for anything beyond setting timers for 5 years, is kind of unforgivable. I don't like them in the "default" position to win the personal agent space.
> Buying Siri and having it be useless for anything beyond setting timers for 5 years, is kind of unforgivable
I don't follow, that was just the state of the art at the time, Alexa, Cortana, Hey Google, were all similarly shit. And none has yet offered a scaled up assistant at the free-tier/cheap bundled sub yet.
Quest headsets are roughly as popular as Xbox and Ray-Ban Meta smart glasses have sold an order of magnitude more units than all other smart glasses combined.
But it's far far too early to draw any conclusions around that.
As I noted in a reply in another subthread: RIM also moved a lot of BlackBerry phones once.
Also, Meta smart glasses are reverse-Metaverse devices; they are instances of the machine navigating the human world, as opposed to humans navigating the machine world.
Sure, but the $100 billion Meta has spent so far on "the Metaverse" is actually false reporting, because the majority of that has been spent on smart glasses.
So when discussing the spending efficiency, that matters.
laughs in VRChat
>I think there is a good chance that Zuck is proven once again as a great executioner
I've never heard of Zuck being considered a "great executioner." In fact, I'd argue that you just described why his execution is pretty bad lol
> Everyone made fun of Zuck when [...] forced out Yann LeCun, a Turing Award winner.
Really? You'd have to be pretty dumb to keep LeCun, someone who hates LLMs and keeps making claims that end up being proven false in matter of months.
Put someone in charge who made their way by underpaying people to sit and label data vs the inventor of convolutional neural networks. Its no coincidence that shortly after his tenure ended, a model that took a long time to prepare started becoming actually useful.
Comments like this one reinforce the obvious truth that the people who are most "pro ai" are the ones who understand it the least
> executioner
I suppose you're not a native English speaker, btw: besides the dubious statement, probably "executor" is the word you wanted to use.
The real cash cow for AI at the moment is mostly corporations/enterprises, especially with software developing. That's where Anthropic is making their fortune now.
Personal intelligence is nice to have but I won't be paying a subscription for that when the free tiers are more than enough for my needs, unless somehow meta comes-up with a hardware device totally changing the way how we use personal AI (something similar to what iPhone did for mobile phones).
this really really really feels like that dropbox comment to me
There's a difference to pushing out models and having people and more importantly, enterprises, using them.
There's a ton of good models out there, but most are not making any money or profit.
So let's see what happens next. But I don't see any or many enterprises switching from using their Anthropic or OpenAI models to use Meta.
Pricing is available here: https://dev.meta.ai/models/muse-wonder
Input -- $4.00
Cached input -- $0.20
Output -- $20.00
Edit: Upon closer inspection, the benchmark scores for "Muse Wonder" on the main marketing page are actually for "Muse Wonder Full," which is only available on the higher-tier subscription plans. This pricing is for "Muse Wonder."
200% Sol 6.1 pricing for Luna performance.
sol 6.1 is 50% cheaper than this. sol pricing$2.00 output $10.00
dont know why the f people downvote for saying sol pricing.
It sounds like you're disagreeing with the person you're responding to, but you're both saying the same thing.
i was agreeing with that.
> Luna performance
Source? On the available benchmarks it looks like Sol 6.1 level.
If this model is not benchmaxxed, I fear it's over for anthropic, openai etc. They are not unique and there will probably be many others with similar models very soon. It'll become a commodity and the price will be driven to the ground.
Personally, I don't see facebook joining the fray (of fable level models) particularly "needle moving" in the discussion of whether frontier LLMs will be commoditized. Facebook spent hundreds of billions and have better talent than most.
If you can spend your way to Frontier LLMs even if it's a few hundred Billion that means it's a commodity same as nuclear or electricity or telecom.
This is the entire point here, give it a couple years every nation on the planet with 20-40B in funding would have their own champions with post Fable level performance.
I feel like Grok/xAI or whatever they're calling it was Exhibit A in "it just takes money", although I'd also add "and the organizational willingness to build not buy".
> give it a couple years every nation on the planet with 20-40B
Was this sarcasm?
How many nations on earth have 20-40B for LLM training? At 40.2 B, Spain has the 15th largest military budget in the world. I doubt any nation will spend 40B on gradient descent.
Yes. And only 5-8 nations will have these models same as the number that have nukes of their own.
Fable 5.5 is around the corner, and 'Bel' will be out by end of year as well. We don't know how good they are, but if the improvements from Opus 5 to 5.5 are any indication, I think Meta will be left in the dust again pretty soon.
Whats gonna happen is that time will compress as more releases happen faster and difference between models will still be significant just at higher time frequencies. I expect in the future continuous releases that span no more then days and maybe even faster. So what will become important is not who is 1st or second place at any given time (which might change often), its who is on average at any given time is 1st or second compared to the competition. So yeah meta sitting second here (if not benchmaxed) doesn't really signal anything if that pattern doesn't persist. The number one consistent lab will be the one that raps the rewards (customer subscriptions and traffic). So yeah I agree with what you stated in a long winded form :P
https://anthropic.com/aug-2026-risk-report
Page 16, Section 1.4, Notes on coverage of unreleased models.
"a noticeable improvement on Mythos 5 for many tasks relevant to internal use but does not display a capability jump"
Models are a commodity, but I also think it isn't necessarily over for Anthropic or OpenAI. They have opportunities to find different moats.
You don't just "find" a moat in business. After all this time and money and first mover advantage, they haven't found a one yet.
I've been a regular user of muse spark 1.3 ... and I can tell you one thing for sure that this model, muse-spark-1.3 in xhigh mode, is between sol-6-medium and high.
This is already so good that 90% of my work has shifted away from codex to this. So, muse-wonder-full coming near opus 5.5 is kinda expected.
So far only OpenAI has a model that solved a Millennium Problem, while Opus 5.5 is a star at coding. It's exciting to have many more to join the frontier, but I think the top two labs still have a significant lead over everyone else.
I know it's old hat now, because it happened over a week ago, but it's amusing that Millennium Problems are now just benchmarks we use for discussing models.
It was not just about models, but also techniques, and mainly the human guts to trust their models and follow through to solve one of the hardest math problems many experts had been musing over for more than 26 years since the million-dollar prizes for these problems were announced. For me, it was an amazing moment to witness indeed!
I find it hard to believe that Anthropic couldn't have similarly impressive results if they tried. I'm honestly not sure it's worth it though, it's mostly marketing, right? But given how easy it is to switch from one provider to another, why would you choose the model that solved a Millenium problem, instead of the one that solves *your* problems?
I understand that it's still useful marketing, say for IPO or hiring, but it's not clear to me that it really matters in the long-term that they fund this out of their pocket.
I doubt anyone thinks coding is going to be their moat in the long run, it is just the easiest to get (paying) users for at the moment. And software being able to write software will always be useful even in other domains. Longer term will be research in other areas where data might become the moat.
Muse isn't driving this conclusion - Deepseek and Qwen are.
OpenAI has the largest user base on mobile as far as I'm aware. To a lot of standard users AI is still just ChatGPT. They'll be fine as long as they can effectively monetize their free users (same what facebook will soon try with Muse). The only advantage facebook has is that they already have a fully working ad marketplace ready with all major advertisers set up - OpenAI still needs to scale that up.
WhatsApp has 3 times as many users, and as soon as meta turn it into a chat AI app people want to use, it's right there
Meta AI has been accessible inside Whatsapp for a long time. Doesn't seem to have a massively disrupted chatgpt usage. And I have a hunch people want to keep this separated - not use whatsapp for messaging friends and asking AI.
Meta has apps billions of people use every day. If AI included in those is good enough they will just uninstall (or never open again) ChatGPT one. It's not true in the other direction.
It is much easier to follow leader closely than to lead yourself.
Like peloton in cycling.
I mean this is the case in general.
E.g in f1 you can get very close but difficult to pass unless you have a significant advantage.
Muse Spark 1.3 is very fast and powerful. It's my work horse for open source development. I look forward to trying Wonder, if Opencode Go gets it.
It's hard to try and compare these benchmarks against real world use cases. As an experiment I let Meta Muse make all major decisions on a weekend trip to Seattle ~2 weeks ago. I think this would have been using Muse Spark 1.3, which is equivalent to GPT 6 - Sol.
On the plus side it told me taking the light rail from SEA into the University district was faster and cheaper than Uber.
On the downside it would recommend restaurants that were out of business, checking out Salmon swimming through locks out of season, got the opening of Beecher's Cheese in Pikes Place wrong, and fumble it's way through most simple planning decisions. I think my favorite was when it asked me what flight I have tomorrow so it could help plan when to head to SEA. It sent me a very panicked message that the flight was leaving today! Very soon! I had to remind it this is a daily flight and I am booked on the next day before it checked and apologized.
For things that are verifiable, and code tends to be that way, these models can do all sorts of things. But I find it hard to reconcile Meta's goal of having a personal assistant bot with something that is unable to do the simplest tasks.
> checking out Salmon swimming through locks out of season
To be fair, the Coho salmon run at the Ballard locks peaks in September. It's possible you just had bad luck that day.
https://ballardlocks.org/fish-salmon-ladder.html
I run by the locks a few times a week and while I don't usually see fish, the presence of sea lions the last few weeks tells me that at least some salmon are running right now.
ah if there's one thing I miss about Seattle it's living near the locks / discovery park, such a magical area!
There were a number of sea lions hanging out there! It's entirely possible they were snacking on whatever salmon were going by. Not all was lost though, I got to enjoy watching some boats go through the locks.
I'm assuming it was out of season because in talking with the Uber driver afterwards he said we'd missed the end of the season. I'm assuming his "local model" was up to date and accurate.
but that was a different, much smaller and less capable model
Light rail is usually not faster than Uber. I guess maybe in very heavy traffic. Much cheaper though of course.
Lecun would be beating you with a stick for using an LLM like that :D
Personal agents are a topic of interest for me. Muse is a free, mass-market product, so they're optimising for efficiency. Some LLMs (like Google's models) are designed to rely more heavily on parametric knowledge and less on tooling. This is much faster and cheaper, but less accurate. So in the case of, for example, salmon swimming, Muse relied on its parametric knowledge without searching the internet to confirm accuracy at this time of year, on this day, given current conditions. One can lean more heavily into accuracy using bespoke agents like OpenClaw. I suspect Meta will gradually improve accuracy over time as it assesses its server capacity and improves model efficiency. Right now, it's not accurate enough for me to rely on it.
Interestingly unlike other Meta models this one seems to take a pretty big jump in pricing, as opposed to the prior ones, where they typically came out at a highly discounted price compared to every other model.
Maybe its a signal they think they've now cracked into the frontier and are willing to actually price similarly vs just trying to get users with discounted but maybe slightly worse results?
$4/$20 in vs out making it higher than Sol and on par with Opus.
No contributor tier will do that I guess.
Yup, I was looking for that on Openrouter, but I guess they're confident enough in this model that they deemed the contributor tier unnecessary.
OpenAI halved the value of their pro subscription, literally. It kicks off on Oct 28. And simultaneously is now two steps back on the frontier? It's a confusing moment for that company as they try to IPO.
I believe they’ve pushed back the ipo. I always figured that frontier labs were going to be very vulnerable against tech giants that had actual profitable businesses to fund their AI work.
The amount of total funding Anthropic or OpenAI has raised individually is like a year of ebitda for Google, meta, or Microsoft. And it doesn’t look like there’s a real moat outside of money
They pushed back the IPO because their finances are a mess and they don't have any models that can give them a decisive edge. Astra was touted as AGI and got crushed by Opus 5.5 two weeks later. That's not the type of thing that fills investors with confidence.
They are waiting to drop their Bel model in December and try to drop jaws.
Again? I just raised mine back up. But for real, it's exciting getting new models so often.
google was just hoarding the tech… money invested in openai came from somewhere probably as profits (just re-allocated by the shareholders rather than management)
I trust Meta with my codes as much as I trust it with my nudes
This is not an exaggeration (which you're possibly alluding to). Facebook once literally asked for nudes
And I agree with you. "Give us code so Zucky can steal it, like he does everything else. Bonus points if it's related to Snapchat somehow"
> Facebook once literally asked for nudes
I’m sorry, what?
Back in 2022 they briefly wanted their users to submit a sort of verification nude of themselves, in order to use it to scan for future revenge porn of said user.
Now as you might imagine, this was not popular or long-lived.
Which is hilarious considering how often they or their AI ban people but never the horrific things people report
Of all the social media companies we could have had, the one with the most mindshare is run by a sociopath
Can these companies please standardize their naming conventions for their tiers, I really don't want to learn another one.
haiku/luna = S
sonnet/terra = M
opus/sol/spark = L
mythos/fable/astra/wonder = XL
whatever the next bigger model is = 2XL
They'd likely become outdated in a few months
Actually I believe wonder is in the L category and "wonder full" is in the XL. Just to make your point stronger!
Apparently “the models were trained at Meta’s Prometheus data center in central Ohio, which came online earlier this year”
https://www.nbc4i.com/news/local-news/new-albany/meet-promet...
This is still an entire generation behind in real-world performance. I'd say it's even behind ds4.1f in the limited testing I have done so far. It actually reminds me of 5.5 haiku more than anything, but the prompt intelligence is as bad as mimo models.
I have a completely opposite experience. DS4.1 and Mimo v2.6 take too long to finish a task, not worth it.
Frontier labs are way friendlier to use, these smaller models need finnese and near expert understanding of how these models really work, what their limits are and how to use them to their full potencial.
For example, there's a lot of people saying luna from openai is bad. However, that is simply not true. They're using it as-if you can talk to it, similar to how you would talk to a human, but what you actually want is combining it with something like prewalk from omp where you have a socially(?) smart model start the work and once the first edit lands switch to a cheap alternative like luna to finish work.
Luna and in general, smaller models, will struggle to understand the intent and complexities behind human language.
I have a feeling "intelligence too cheap to meter" is going to haunt/mock this industry. Recall that only a year and a half ago Llama 4 flopped so hard Meta rebooted their development efforts. Desperately wish more information was available publicly about proprietary model architectures and training regimes. Are they running Manhattan project style compartmentalization? Otherwise, are these details, and the economic implications, an open secret in their circles given the amount of crossover between companies? It'd be of great interest to us mere mortals that have to deal with the consequences of how this turns out.
Just checked their subscription is still North America only, I guess I will keep myself and my company with claude...
These regional lockdowns are making me feel funky, I feel like I should just use open models if others start playing at this game as well.
I was/am fine with Chinese companies offering cheaper subscriptions in China because ofc. But this is just weird, you could always charge me more let me try the damn thing, atleast it's own open router it seems like.
Initial impressions, interesting, it generated a pretty ok result for my prompt need to run a larger task but don't want to burn money so will wait for someone to burn their subscription and try running some opus 5.5 comparisons.
The benchmarks I saw posted to X are impressive, which makes them hard to believe.
I recall that Meta made massive investments into compute infrastructure that put their capacity near the top of all competitors. If all of this is true, it looks like ownership of compute is a significant differentiator in performance.
Yeah, especially non-coding benchmarks which for Meta might be more important long term than coding ones as they are going to deploy it to billions of non-programmers.
Hopefully it's benchmaxxing. I am terrified of the idea of Zuck leading the frontier of intelligence.
Not that interesting unless it's made open weights. Frontier lasts weeks, inference is a commodity.
Would be great if Meta drops the weights, just to throw off Ant and OAI a bit
I don't think they'd have much to gain. People who use open-weights models will just use whatever open-weights model is the best. They'd lose a bunch of money, with little conversion to customers.
I disagree actually, I think zucks original llama thesis is proving more correct by the week. The rent market on intelligence never materialized, and will not materialize.
In software, we've all become accustomed to the idea that the way you make money is by building little blocks of intellectual capital, then charging rent on it. That made sense when developing intellectual capital was a rare and valuable skill/asset and the thing you build was a rare and valuable asset. But now both are not rare. You can still charge rent, but the market is highly competitive, so prices match that.
Durable rents will end up going to those who owns the scarce assets, zuck is building those as we speak.
Dropping open weights will build an ecosystem that makes his valuable assets more valuable.
> Durable rents will end up going to those who owns the scarce assets, zuck is building those as we speak.
I'm following but maybe I'm a little dense here: what are the scarce assets that Meta is building in this analogy?
Compute, prop. data, existing network effects.
In short, physical infrastructure and talent.
Of course you need talent but supply seems reasonably high. We have openAI, anthropic, xAI, meta, Xiaomi, deepseek, Alibaba, z.ai, Mistral, reflection, and Google all producing exceptional models.
The biggest differentiating factor seems to be compute. I know most Chinese labs would really like more compute. Google and xAI too, but largely because they sold compute to their competitors.
They've been saying they'll rise muse 1.2 for months now. Don't hold your breath
I wish I could benchmark Muse Wonder Full, because Muse Wonder isn't much better than Muse Spark 1.3[1]. I created a brand new programming language designed to confuse LLMs[2] and created a model benchmark out of it.
A lot of the tasks were failed because they exceeded the $5 budget though, so maybe it would have gone on to complete them, it just required a lot of funds to do so. Still, that's the benchmark, and it's not better than MiMo V2.6 Pro which is pretty disappointing.
1 - https://bench.killswitch-lang.org/
2 - https://killswitch-lang.org/
That's an interesting way to benchmark LLMs. What other ways have you come up to measure them?
So far only like this
This sounds irrelevant but it isn't: I just tried to unsubscribe to marketing emails from some Meta property, and got:
Please confirm that you want to unsubscribe. Unsupported request - method type: post
That's how careful this company is. Who knows how long that flow has been broken for, no one in the company is motivated to care, because what impact can that possibly give? How could that look good on your PSC? No company has a worse history with user privacy and no company is less deserving of the sort of trust using Muse requires.
Its actually a bit funny. A lot of privacy initiatives were actually part of deprecating code and data that became liability issues post cambridge analytica and covid(think location data, if no one goes anywhere no need for that org right?).
I think the only way meta can care about privacy, if any, is if it becomes socially taboo (lawsuit, WSJ/NYTimes report etc).
The clearest quantitative evidence of this is their P/E, its lower than others relative to its forecast earnings because of all the legal risks that need to accounted for.
The Mishandling Personal Information benchmark is non-sensical: that's the very definition of giving your personal data to Meta
As someone who got a Muse Code sub last month, I find this announcement a bit disheartening because it makes me think the lower tier subs are going to get rug pulled. I'm on the $15 plan for a month now and it's value with Muse Spark 1.3 Max-contributor beats everything else by a long shot. Two somewhat small prompts with Muse Wonder Max-contributor used 4% of my weekly quota and I quickly switched back. I was thinking about getting the $50 plan and I'm actually less likely to do that right now.
Have you tried OpenCode Go? $10/mo is what I am willing to spend because most of the stuff I do is hobbyist. And OpenCode Go gives you $60 worth of MS Contributor 1.3 usage for that $10. Starting today, they actually provide $80 worth of DeepSeek 4.1 Flash usage if you want to use that.
Unless you are running a swarm of agents, it is extremely difficult to exhaust these plans in a month.
Yes, I'm on OpenCode Go too... and it's great they raised the usage levels on DS4.1 Flash. But I was really thinking of cancelling it to get either a 2nd Muse Code sub or to go up to the $50 tier.
You have to realize that $15 Muse Code sub with Muse Spark 1.3 gets you around $1500 worth of usage. Neither Opencode Go or really anything else is even in the same ballpark of value. https://x.com/PawelHuryn/status/2106724136932327458
I was hoping this level of subsidization would last a long time but today's announcement make we worried it'll end pretty quickly. There's a lot of AI influencers noting just how fast Muse Wonder uses up their quotas so hopefully fixes are made.
Yeah, I barely manage to exhaust the $60 worth of tokens. I have to think of new things to use up $1500 worth.
Thanks to competition from open models, there is a limit to how much prices can be increased though.
How does OpenCode Go work? How do they make any money?
They are a subscription service.
Subscriptions like these make money by relying on the fact that most people do not try to (or have the time to) extract every last dime out of their plan. The eighty people who do not use up their quota cover for the twenty that do.
That and deals with providers to use spare capacity I guess.
Wang reported the excessive token usage was a bug and they fixed it and reset the usage totals to 0: https://x.com/alexandr_wang/status/2109774017574113780
In the context of Meta trying to get non-technical people to embrace Muse agents to handle personal affairs, their internal metric showing a ~3% violation rate for mishandling personal information seems unacceptably high.
There seem to be no downsides to mishandling PI so I don't think it's a big concern to Meta
Not long ago we used to have regulators that protected users from companies pushing out harmful or consumer hostile products.
Regulators appear entirely toothless now, even here in Europe.
I found it interesting in the example given of a user looking at house listings and the agent contacting the letting agent on it's own on behalf of the user.
Like the good ol' days of viruses and anti-viruses. I wonder if there's two sides that will play out; many agents across many users doing things at scale for their users, thereby creating an admin burden for someone else. However because the task is known there's a framework for how to solve the task, might Meta then provide an agent on the other side for the business owner to use.
RL team hill-climbing this to 2nd AA place and this fast is crazy
How does it work on day-to-day workflows? does it translates well?
I had to come back to this comment to make sure I wasn't dreaming... I remember seeing it at 2nd place this morning, but they have moved it to 5th, behind Gemini 4 Argon. Maybe it takes a while for the dust to settle around these benchmarks?
Meta's blog post still list it as 2nd in AA. Any ways, AA moves benches constantly, when Qwen3.8 was released something similar happened.
spinning up a GPU and as many cores as possible?
mousewheel repurposed?
background follow your mouse?
neutral tones?
well then YOU are viewing a frontier model announcement!
Have I been living under a rock, or this feels out of nowhere from Meta?
Seems they’ve just messed around with cheap and open weight models for years.
Then Muse appears, followed by full suite of frontier models…
they started their massive ai capex over a year ago with aquihiring scale ai, so I think it makes sense that their investment is paying off now. in total they've spent in the high 200 billion on ai, though I think most is likely on datacenters.
Unlike everyone else, which at least has to pretend not to train on user data, they gave out their model for basically free if you agreed to let them train on it. Opencode shows nearly 40 trillion tokens going through muse spark 1.3 contributor this week alone. That almost definitely helped close the gap a lot.
muse ai was the biggest consumer launch since chatgpt. That left them ( and lets them sit on top of ) tonnes of valuable exchanegs with people trying to get things done. That must be an amazing corpus for training.
oh boy, you can imagine the shit eating grin on the face of whoever came up with the name "wonder full".
Looks good, and at least it is available unlike Gemini 4 Argon
Does it weight tokens on their ability to elicit negative emotions in users in order to drive engagement?
So, Wonder (and not Wonder Full) will "will soon power experiences in" their assistant app. But no rollout details .
Then, Full is meant for API usage and "Muse Code" only.
It's getting confusing what having "Muse" means.
It's crazy how often these companies botch the naming of things like this. There's definitely some intricate and insanely bureaucratic accounting going into the marketing aspects of these products that results in absurdity by the time it reaches actual consumers lol
According to my Muse Agent, Muse Wonder is already active as of sometime this morning. It claims that it sees this info from the "session metadata" although I have no way to verify this. Previously it claimed it was Muse Spark. It's system prompt file still says Muse Spark 1.3.
Not really confusing at all? Muse Spark it's the cheap Luna model. Wonder is the Sol model. Wonder Full is the Astra model.
Reminded me of when George Hotz said only Zuck has “the button”. [0]
[0] https://youtu.be/-qRc7qou6XY?is=hWofxO0G6X3JjST5
lmao this interview sucks. Even Lex seems confused and disinterested in this conversation. Hotz is more excited about generating some sound bite than he is in explaining himself. His "the button" take is nonsense.
Also, this was 3 years ago, and he's been effectively wrong about his bet. But hey, maybe Muse Wonder is when Zucc's plan will come to fruition.
Acquihires starting to finally pay off it seems. Bring on the competition
It's ok, intelligent, almost SOTA, but way too expensive [0], and some issues with API failing sometimes at the moment.
[0]: https://aibenchy.com/compare/meta-muse-wonder-max/openai-gpt...
its three big companies at the frontier now? openai, anthropic and now meta.
why is google still not frontier?
They're not until they are. No doubt they are working on something and have better resourcing than the AI labs
Grok has been at the frontier since 4.6
please be serious
Gemini 4 Argon release a few days ago was frontier a few days ago. Things move fast.
I think some of the benchmarks that Meta didn't pick show Gemini 4 Argon ahead but really the benchmarks are not that great any more and getting increasingly disconnected from reality as people search for "headroom" in the evals to avoid saturation I think. Opus 5.5 is "top" for example but I really dislike the way it talks to me (it feels quite condescending, holier-than-thou, uptight etc) and prefer Gemini 3.8 flash as a daily driver just because it's so much more pleasant (and faster!) to work with and it's code is good-enough for my needs, despite being weaker in the benchmarks.
Gemini 4 Argon was announced, but it isn't released yet.
And in six months when it is released, it'll be FAR behind publically available models.
Pretty sure it is GA? https://blog.google/innovation-and-ai/models-and-research/ge...
They do not appear to care about the frontier. They're focusing on protecting their ad business, so their models need to be lightweight and efficient, to provide fast queries for people who are on Google, and for local inference on Android and Apple devices. Also specialised small models for things like OCR, translation, facial recognition, voice, design, etc. Their TPUs are also quite impressive.
It's a reasonable niche they're carving out, but I think too conservative. Very strange from Google. The other labs are racing to RSI and Google appears to have given up the race.
While some are afraid that one rouge AI system may take over the world (or cause doom is some fashion), I think one of the best safety mechanisms to prevent this is having many competing and adversarial labs that are extremely close to each other. So it's good to see that Meta is not showing any intention of leaving the frontier.
I tried the Kitchen Game demo, and it is embarassing that they even released that...
1) This is a blatant copy of the game Overcooked. They just copied the whole concept + all the mechanics.
2) It it buggy, at least in Firefox it does not work. You cannot pick up a dirty plate. Somehow I picked it up, but couldn't see it. Then I could pick up a tomato, but when placing it on an empty plate it glitched into the cabinet. Then when picking a clean plate, my plate appeared again. Then it was over. ???
So from that we learn what AI can do: 0 creativity, just steal from real peoples work and do a shit job at it. As long as it looks good on first look.
It's interesting that in its marketing Meta compares Wonder to the closed frontier models but not the open (Chinese) models. By its own assessment, Meta could easily show how far ahead its own model is against the rest of the pack of open models but they chose not to. Why?
> Meta could easily show how far ahead its own model is against the rest of the pack of open models but they chose not to. Why?
Wonder is not an open model and there's no claims that it will be in their announcement.
Despite the claims that they would release weights for Muse Spark, they haven't done that yet either.
I would like to start seeing bar graphs for benchmark results to add a second y axis to show token spend or cost in dollars.
I guess what I'm saying is performance is fine... but what I really want to see is performance/dollar.
impressive benchmark numbers… now how to actually use this without Zucking up my machine is a consideration.
I'm only interested which of the big techs is most generous when it comes to subsidizing consumer plans.
They added fancy annimations and effects to this specific webpage, which are unnecessary and distracting and lead to bad reading experience. I already suffer astigmatism and floaters. I hate it when I move my cursor and these gray blocks and characters flash and flicker.
in the console eliminated the cursor effect for me
I wonder if this one has the same quirks as the Spark models. They have okay to good benches, but are terrible for actual use
I noticed Muse Wonder model on OpenRouter is age gated to 18+
If I'm reading this right, then at the moment, for certain tasks, this is the best model in the world.
A new company on the frontier. Not close, but actually on it. Wow this is big news.
The only time I'll ever touch a Meta LLM is if it's powering the dystopian doom bot that is plunging its murder spike through the base of my skull.
Poetic
Apparently the benchmark-winning (behind Opus 5.5) Muse Wonder Full is only currently available to their subscribers.
Only Muse Wonder is available in the API.
The usage limits on their basic $5 code plan to try things out are ludicrously low?
I blew through 14% weekly usage and I'll I asked:
Beware...
Muse Glimmer, Muse Spark, Muse Wonder?
What's next, Sun Shine?
Shine, Shimmer, Glow, and ultimately, Cherenkov Radiation.
Muse Demon Core.
Meta's RL scaling seems to be in full force
This is amazing. I can’t wait benchmark to come out $ per performance
Meta should rename itself to Metamusal
The benchmarks are great for sure.
Now if only they could fix their coding harness so that it's usable for day-to-day use. I paid them for a month to try out Muse Code, but their sandbox never worked across Mac or Linux (both very stock installs, nothing special).
Muse Spark 1.3 is available in Cursor! You could try it there.
(Disclaimer: I work at Cursor/SpaceX)
As an aside: I tested out the $20 Cursor CLI subscription, and it created an excellent little graphical app for my kids to play with. The design it came up with was excellent (compared to the crap I got out of Codex for a parallel implementation). I'm going to try to figure out how to get Cursor into my work rotation, at least around design-related artifacts.
Counterpoint: Musk is a literal Nazi and you shouldn't give his companies any money.
I haven't tried it but there is the official `pi-muse-bridge` that can use your muse code subscription in pi.
Is this datacenter-only, or will they be open-sourcing?
Man, I really hate meta stuff. Not only do they make it impossible to pay for running ads, but everytime I try and pay to try out Muse code I get an error message. I swear they would make 10x more money if they didn't make it impossible to pay them
how cool would it be if these smart peeps put all that brain power into models that actually improved people’s agency and mental health, instead of siphoning it
The Kitchen game is a ripoff of Overcooked in pretty much everything lol
Super interesting that they're not even making Wonder Full available outside of their own coding harness, and thus not providing an API price to use in comparisons.
This really does look compelling, but it would be very nice if they wouldn't test all of the stuff on Max!
Facebook successfully becoming a frontier lab was not on my bingo card for 2026, but the more competition the merrier I guess. We will have to see just how benchmaxxed this is but beating Astra and Fable is not easy.
Superin… you’ve got to be kidding me… slurp slurp little Zuck.
Beyond ridiculous
As another user mentioned, it's more likely Trump borrowed the term from Zuck. You know, with "Meta Superintelligence Labs" being founded in June 2025. The itself is a lot older still.
Doesn't change the fact that they're bending the knee, not a single mention of AI in their report, it's clearly to show allegiance.
They used it for their goal before, not for describing their current muse models
What a bizarre way to double down.
>not a single mention of AI in their report
Trivially easy to see that this claim is also false. And that's not even mentioning the "AI Research" and "AI Developers" links in big bold letters, the model card, or how Wang is introduced as the Chief AI Officer.
Did this ever make #1 on HN? If not I bet that pissed off zuck!
Nope, I’ve actually been checking this thread throughout the day. More upvotes / comments but didn’t break #6. My tinfoil hat says that it has been intentionally suppressed because of YC investments, but who knows :)
I found out about this release from Reddit because for some reason this frontier model post is suppressed in HN.
Had to use Algolia search to find this buried post.
Which is sad.
any idea if they will release the weights?
Low chance
What openai/anthropics answer to Muse app? Because I find it really compelling and with this new incredibly beefy model it will be even better.
OpenAI has dots. Same concept. It’s less capable, but since I’m very OpenAI integrated I just had my dot install muse and use it
OpenAI has Dots. I personally prefer the business model of subscription payments funding my AI assistant, instead of skimming a fee off the top whenever my AI assistant completes a purchase for me. The incentives are much more aligned with the user when the user is the customer. That said, based on the history of the web, Meta's approach will likely win out.
Quite frankly, Meta's models have been so impressive (and cheap) that I haven't used an OpenAI or Claude model except for one-off tasks as of late. Interested to see where they go from here.
But when is another open weights release coming out...
Claude is waiting for this to drop fable 5.5 lol.
The term is "artificial intelligence", you bunch of simpering spineless cowards.
"Superintelligence" is literally in their lab's name. And it's been there well-over a year before it was ever uttered by Trump.
[flagged]
[flagged]
Super Intelligence.
Just dont' use the word Super Intelligence
I know, it's... well, it's irritating to see Trump use it. But I could be wrong, I don't know the angle he's going for
I understand the general thing of it. SI might mean every person gets their personalized agent but you could already do this, but maybe the Chinese model works here, where they push the costs down
It's still stupid to start calling it all SI
Probably this angle: https://www.yahoo.com/news/politics/articles/trump-super-int...
The ML community has been using the phrase "Super Intelligence" long before Trump took office. https://en.wikipedia.org/wiki/Superintelligence:_Paths,_Dang... was published in 2014.
I noticed "Superintelligence" as a re-brand of AI and constantly shifting definitions of AGI around 2025. When a bunch of startups started slinging around SI as a marketing term and of course when Meta launched MSI in summer of 2025.
that was all before Trump's "edict"
Super Intelligence refers to intelligence beyond what humans are capable of, which some believers profess is the next step after (and in contrast to) human-level AGI (Artificial General Intelligence). Calling pre-AGI AI "super intelligence" is pure vice-signalling and publicly kissing the ring.
Apparently Meta used the term before.
Also, as much as I despise Trump, you know what? Given the jagged shape of LLM capabilities and what they do achieve in some areas and their superhuman speed, compared with the abilities of most of their users, the term superintelligence is not that far off- maybe just a few months ahead of time. E.g., from my perspective, the current mathematical abilities of LLMs are "super"- I'd hardly ever come in touch with an equivalently capable human, not to mention having one completely available 24/7. Same for general knowledge and several other domains.
Why people downvote objective, non-controversial comments, is beyond me.
To be clear, Meta here is not using "Super Intelligence" the bullshit term Trump wants companies to adopt as a show of submission. They use "superintelligence" the bullshit term that Facebook made up for marketing reasons a few years ago. Both of these are of course entirely unrelated to the long-established meaning of the word (e.g. how Bostrom defines it).
Zuck has been using the term for a while in the context of Muse, if anything Trump took it from him
I don't know why you were downvoted, MSL stands for Meta Superintelligence Lab, and it named that before Trump.
It’s their brand name from last year https://www.wsj.com/tech/ai/mark-zuckerberg-announces-new-me...
Pretty big coup from Zuck to convince Trump to use it haha
how much benchmark-maxxing has been done here? taking bets on poly and kalshi...
All of them are doing all the benchmaxing they can most likely. Dumb not to when so many people care about those numbers and don’t have the opportunity to try every model out personally.
Muse Spark was generally good. I'd be surprised if this model felt benchmaxxed.
I think it's safe to assume every single model is benchmaxxed. I mean why wouldn't you include training data from common benchmarks? Not only is a higher score good marketing, but these benchmarks are supposed to represent common tasks so training for those should in principle improve performance across a wider range of tasks.
There's an entire section in the linked post about measuring benchmaxxing, so I guess that probably says something.
Out of all the tech CEOs out there, Zuck would never lie to me or manipulate me. I know this in my heart.
The thing about Zuckerberg is that although he's deeply odd he has no ability to conceal this.
Sam Altman is much more dangerous because he is occasionally able to come across as human.