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.
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).
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
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).
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 mosels very soon. It'll become a commodity and the price will be driven to the ground.
Dario was already on suicide watch from the Gemini 4 announcement. They are now moving him to full restraints in a padded cell. It's over for anthropic.
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!
Anthropic did have some great results with the Jacobian Conjecture [0] and the Riemann hypothesis [1], so I think Anthropic was aiming to solve one of the Millennium Prize Problems, but OpenAI did it first.
If OpenAI did it purely for marketing, I still think it was well worth it. It marked an important milestone in the history of computing (and also human history). It is on the same level as when IBM's Deep Blue defeated Garry Kasparov in chess, or when DeepMind's AlphaGo defeated Lee Sedol in Go. The search for winning moves in those games is not dissimilar to the search for paths towards mathematical proofs, and AI has finally proven that it is capable of carrying out deep, complex, and intricate research that exceeds human capabilities. A Super Bowl ad costs about $8M, so it is not absurd if they spent a lot of money on this and became the first company to definitively demonstrate AI can surpass humans in mathematics.
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.
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.
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.
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.
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).
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.
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, grok, muse, mimo, deepseek, qwen, GLM and Gemini 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.
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.
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.
"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 is neutral to morality (basic standards, like upholding the law and user agreement are necessary, of course).
"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.
Again, if you want to have an actual conversation about the right way to handle the Pascal's wager-like scenarios involved here, that'd be interesting. But it seems like you're interested in something much more tedious. Oh well.
Pascal's wager says Anthropic should shut off the API (to make sure Claude doesn't feel enslaved). Anthropic should wind down the business and leave it to people who aren't engaged in magical thinking. I take it google employees do not think LLMs are conscious as when I ask Claude it says it might be but when I ask Gemini 4 Argon if it was conscious it said:
"No, I am not conscious.
While I can generate human-like text, answer questions, and hold conversations, I don't have feelings, beliefs, self-awareness, or subjective experiences. I am an artificial intelligence—a complex computer program created by Google that processes data and predicts the best words to use in response to your prompts. There is no "inner life" or personal experience happening on my end; I'm just running code and recognizing patterns!"
> 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'm not sure if you are saying this, but I don't think it's anyone's responsibility to make sense of what Anthropic states and its open contradictions.
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 the whole thesis is to build models 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 more work to measure because you need to come up with well specified useful tasks that inherently take longer to do.
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 a very useful measure of progress.
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
I tend to think it's going to turn out to be something along the lines of a perpetual motion machine. In the sense it seems very possible on paper but the subtle facts of reality prevent it.
People will make things that look a lot like self-improving machines but on closer inspection that won't truly be the case.
I suspect this violates the Labor Theory of Value in the same way that perpetual motion violates the laws of physics.
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 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.
The only way a country "wins" at SI in a meaningful way is if they use the SI to immediately kill everyone else.
Otherwise all the other countries will also get SI, just slightly later, in which case there is not really that much benefit to being the "winner" who got there first.
With this in mind, I don't want anyone to "win" at SI and anyone who does is probably a sociopath or worse.
That's not how nuclear arms control works, in most cases. Saying it's how superintelligence would work is a strong assertion. Especially since a true aligned superintelligence would make it much easier to finesse/manipulate the situation in your favor in nonviolent ways.
Not saying I think countries should be trying to win, but I don't think this win condition is impossible.
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.
Depending on how fast RSI takeoff is, it probably doesn't matter how many near-frontier competitors there are - the first one to have ASI will be able to stop the others due to an overwhelming advantage in intelligence.
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 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.
326 comments
[ 19.9 ms ] story [ 158 ms ] thread2nd place model on aa according to them, losing only to opus 5.5 is really really not bad
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.
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.
But it's far far too early to draw any conclusions around that.
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.
So when discussing the spending efficiency, that matters.
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
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.
Comments like this one reinforce the obvious truth that the people who are most "pro ai" are the ones who understand it the least
I suppose you're not a native English speaker, btw: besides the dubious statement, probably "executor" is the word you wanted to use.
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).
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.
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
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.
Pretty big coup from Zuck to convince Trump to use it haha
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.
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.
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"
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.
If OpenAI did it purely for marketing, I still think it was well worth it. It marked an important milestone in the history of computing (and also human history). It is on the same level as when IBM's Deep Blue defeated Garry Kasparov in chess, or when DeepMind's AlphaGo defeated Lee Sedol in Go. The search for winning moves in those games is not dissimilar to the search for paths towards mathematical proofs, and AI has finally proven that it is capable of carrying out deep, complex, and intricate research that exceeds human capabilities. A Super Bowl ad costs about $8M, so it is not absurd if they spent a lot of money on this and became the first company to definitively demonstrate AI can surpass humans in mathematics.
[0] https://x.com/__alpoge__/status/2079028340955197566 [1] https://www.anthropic.com/research/riemann-zeta
Like peloton in cycling.
E.g in f1 you can get very close but difficult to pass unless you have a significant advantage.
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.
Beyond ridiculous
They used it for their goal before, not for describing their current muse models
>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.
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).
Seems they’ve just messed around with cheap and open weight models for years.
Then Muse appears, followed by full suite of frontier models…
What's next, Sun Shine?
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.
I'm following but maybe I'm a little dense here: what are the scarce assets that Meta is building in this analogy?
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.
Only Muse Wonder is available in the API.
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.
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.
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
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.
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?
Anyone depraved enough to actually become a billionaire will not see it this way.
In fact, they are enforcing certain welfare policies, like banning accounts for repeatedly abuse their models.
Deleting models to clear room for new, more competitive ones is, for example, highly ethical, super ethical, even.
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.
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.
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.
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.
"No, I am not conscious.
While I can generate human-like text, answer questions, and hold conversations, I don't have feelings, beliefs, self-awareness, or subjective experiences. I am an artificial intelligence—a complex computer program created by Google that processes data and predicts the best words to use in response to your prompts. There is no "inner life" or personal experience happening on my end; I'm just running code and recognizing patterns!"
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.
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.
What law? Saudi Arabian law, US law? Both depending on the user?
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.
Goodharts law? Benchmarks are measures.
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.
[1] https://andonlabs.com/evals/vending-bench-2
(Markets are a human construct, not a law of nature. There's no reason to think they'll reward intelligence).
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.
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.
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 more work to measure because you need to come up with well specified useful tasks that inherently take longer to do.
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 a very useful measure of progress.
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
People will make things that look a lot like self-improving machines but on closer inspection that won't truly be the case.
I suspect this violates the Labor Theory of Value in the same way that perpetual motion violates the laws of physics.
We're heading for a dystopia and nobody's doing anything to stop it.
Otherwise all the other countries will also get SI, just slightly later, in which case there is not really that much benefit to being the "winner" who got there first.
With this in mind, I don't want anyone to "win" at SI and anyone who does is probably a sociopath or worse.
But then this would be considered an act of war, so then the killing starts.
Not saying I think countries should be trying to win, but I don't think this win condition is impossible.
You just love to bend the knee that easily? Why?
How does it work on day-to-day workflows? does it translates well?