This refers specifically to "AI Frontiers", a research group within Microsoft Research that focuses on fundamental advances in AI. There is no end in sight to real AI frontiers.
Perhaps "The End of Microsoft Research's AI Frontiers Group"
would be a less ambiguous title.
Yeah, I went down the rabbit hole scratching my head the whole time. This is about the group at Microsoft Research that was disbanded or reorged into what is now Microsoft Research.
To the author: 100% need this kind of thinking. It’s a balancing act. Historically though, dedicated labs haven’t had the best track record at commercializing their tech. The skillset differs vastly from running a product-focused business.
That said, even within this narrow scope - there are areas ripe for research.
Some of my particular interests include:
(1) Latent thinking/reasoning, which could arguably be more efficient but also opaque is one
(2) Separation of “intelligence” from knowledge - ie. keep the ability to reason like a frontier model and make tool calls (calling the ‘oracle’ or searching the web) but remove all parameters associated with niche knowledge/facts.
These are a couple off the top of my head that are very much aligned to scaling.
That statement is broad enough to be unfalsifiable, but if you narrow it down just a bit to scaling LLMs, maybe there is an end in sight. At least financial markets are skeptical of there being no end in sight.
If scaling LLMs really was the road to God in a box, then compute demand would be infinite, and nobody building to meet that demand would have a junk debt rating. Did those money people just not get it?
The real news is "The End of Microsoft Research". Langford's group is just one of many that have been parted out over the last year or so. The division that was T&R under Kevin Scott was quietly dissolved about a year ago and MSR was reorged under Azure Security with the AzSec leader as hatchetman for the department. MSR itself doesn't have a future beyond a few years at most I reckon. I don't suspect Kevin Scott will be around much longer either since his department keeps getting reduced and is fully eligible for VEO.
Langford has been at MSR a very long time and has done amazing research doing that time. He's bounced around a bit at MSR during his tenure and it's a shame to lose him. AI Frontiers wasn't around that long.
Much of this is Microsoft performing a very large and uncommunicated strategy update. All updates of MSR reorgs are just that "MSR is becoming more lean and focused" when it's obvious this isn't the case. This gets into the conversation that many Microsoft employees tend to have these days which is "Why does culture no longer matter at Microsoft and why did Satya turn into another heartless CEO". I won't speculate there, but the place isn't nearly as fun as it used to be.
simply owning most of OpenAI via a convertible note on very cheap inference sold to it could turn out to be a bigger deal than the rest of what microsoft does.
To me, all signs point to US over investment in hardware capacity to support the "frontier" when in reality the next major breakthroughs are likely radical optimizations that end in increased consumer training and inference efficiency.
I'm not sure how that, combined with open models inching up on frontier intelligence, is compatible with these massive AI IPO valuations?
It seems like closed model AI demand has a lot of contingencies that are ultimately fragile and possibly short sighted.
Everyone wants radically more efficient AI. That doesn’t mean it’s coming any time soon. And when models are twice as efficient, guess what happens next. We try making one twice as big for the same cost.
If they make more efficient models won’t people just use the big data centres to make and run even bigger more powerful models than they can currently support? Or massive swarms of agents to look at all different avenues of a problem?
They've offered voluntary redundancy to a quite large proportion of the staff (a quarter?) so they'll also soon be losing anyone with employment prospects elsewhere.
I see a lot of news about waves of layoffs in large tech companies. Is MS the only shitty place now? I think all of them are, maybe except OpenAI & Anthropic, since they're building something new and are kind of on the "good" side right now.
Yep, from what I hear and see, it's like this everywhere. It actually started with the Twitter layoffs and has spread across the industry. At first, it seemed like a reaction to the ZIRP-era overhiring, but it's kept going long after correcting for it, and with AI has been cemented as the modus operandi for many companies.
From the outside, Microsoft seems to actually be one of the better places, and is making at least token efforts to reduce headcount in more amicable ways, as exemplified the buyout offer.
But in general, the strategy seems to be overload people until they burn out and quit, because hey, AI will replace them anyway.
The transformer based approach is seen as a highway to race down in terms of training and building better models with minimal organizational distractions allowed while every company puts its foot down on this highway. Exploratory research into other AI methods is highly discouraged for now and organizations are encouraging transformer based scale out and iteration as a top priority.
As an example Meta literally fired it's previously head of AI who held views that transformers won't scale. He wanted to veer off the highway and heavily fund exploratory new types of models. They got someone willing to scale the current method out. Essentially the messaging it "Foot down, there's a straight highway ahead don't divert".
I have the view this makes sense too. Current top models are mere (mere in big datacenter terms) trillions of parameters right now. We definitely haven't hit the ceiling and everyone's noticing "Hey this Opus 5.5 with the increased parameters seems much better than the last one". Well that's a trillion parameter model. We'll soon start hitting 10's of trillions and 100's of trillions of parameter models. Unless there's a ceiling which nobody expects they should keep scaling just as they have been. I really do agree that it'd be insane to veer off this highway with exploratory research at this particular point in time. Foot down, get to 10 trillion parameters, see if the scaling holds, if so off to 100 trillion parameters. If the scaling continues to hold we'll be asking the AI to do the exploratory research for us soon enough.
It's worth it to keep exploiting transformers while doing just a bit of exploration of alternatives because they work so well. Heck, just exploiting transformers more will accelerate exploration of alternatives...
This is a dumb question since I know very little about LLMs and transformers but I keep hearing similar things such as "more GPUs/parameters/computer" leads to a predictable scaling in inteligence.
Is that observed scaling part of some equation or law? Is there a limit to how far that scaling goes before a ceiling or diminishing returns hit? Or is it simply based on more GPUs/parameters/compute plus how much human knowledge is in existence and will scale predictably forever?
It should be noted that these laws aren't, well, laws; "a neural scaling law is an empirical scaling law that describes how neural network performance changes as key factors are scaled up or down.". Notice the word empirical here.
These empirical observations on quantities which aren't "intelligence" since no one has yet defined intelligence.
Data isn't "smart", it doesn't contain intelligence. I would argue that LLMs are functionally intelligent, they have learned to what functionally amounts to thinking by training on text (and images, in some cases). They are capable of performing experiments autonomously and make decisions based on out-of-distribution data. Besides (current) lack of online learning, they are limited by their biases and working memory - but so are humans, and I'd like to see a convincing argument on why these mathematical laws should not also apply to humans, which are arguably intelligent.
Are you smart if you see that cognitive scientists had something figured out, connected it to something in your field of chemistry and made a breakthrough treatment for some disease?
Does any of our existing intelligence build on completely unknown information?
You could argue that any new solution/emergent intelligence is an association of existing information in new ways.
It also brings us to what is brute forcing? Humans are brute forcing stuff most of the time. You will have 10 stupid thoughts for every good thought. 10 stupid ideas for every good idea, you will discard most of them so fast that you won't think of that you did that. It's very rare, if not near impossible, that a non-mathematician with no prior knowledge, goes into a room on day 1 of deciding to think about a conjecture they don't know anything about, and solve it on the blackboard. They need to use their collect associations, formalize their axioms, think of first principles (known information), and associate it.
So to answer your question you need to answer:
1) What does "smarter" refer to? Is it accumulation of information about the universe?(if yes, a database is smart) Is it an efficient, abstracted representation of accumulated? Is it prediction of the future state of the universe given a learned abstraction? Is it compression?
Or going further, is it solving tasks useful for the human in the world? Is it acting as an agent with it's own goals and interests in the universe? What if the goals are hard for humans to understand?
2) What does "genuinely new approaches" refer to? Is it fine to use math axioms to prove things? Does interpolation between two existing solutions count as new? At what granularity of concept does novelty happen (Does using two higher abstraction concepts, like combining some new, modern neural network architecture with a database count as a new thing), even if the underlying lower abstraction of information as storage, latency and compute is not novel?
---
As you can imagine, this is all not really defined or agreed upon. Here are some examples of projects that are set up to drive these ideas:
The Hutter Prize poses that intelligence is defined by compression, so if you manage to compress the English Wikipedia efficiently, you are closer to better intelligence (http://prize.hutter1.net/)
Compute: Intel CPU
The ARC-AGI foundation poses that being able to be goal-oriented and understand higher level abstractions is intelligence: https://arcprize.org/leaderboard
Yes, the Chinchilla scaling laws from 2022: https://arxiv.org/abs/2203.15556 More parameters directly and mechanically means less prediction error for each token.
(This moves the argument to "is being really good at predicting the next token the same as being intelligent?" which has happily diverted many minds for years now)
The Chinchilla paper is probably the poster child of “paper that aged like milk” in LLM field though. Everyone over the past 4 years have trained with several orders of magnitude more data than the Chinchilla optimal number.
If you don’t know which data has the highest value, though more at it. Chinchilla is probably still true, but with that caveat. And it could be like advertising… you know a lot of it is worthless, but you don’t know which part.
Your phrasing (_observed_ scaling _part of_ some equation) might indicate you assume / think there’s some theoretical equation that relate intelligence to those parameters and the we observe the data to confirm it. As far as I know, there isn’t - it’s purely empirical curve fitting. Train a model with different amount of compute / data / parameters, measure performance, fit curve.
'Scaling law' is the prediction that increasing the training data and model size will cause the result to be better. It's not a law, but it was the prediction that OpenAI staked their fledgling startup on to get to the ChatGPT moment.
> Foot down, get to 10 trillion parameters, see if the scaling holds, if so off to 100 trillion parameters. If the scaling continues to hold we'll be asking the AI to do the exploratory research for us soon enough.
Where's the money coming for those 10X and 100X ventures?
100 trillion parameters doesn't make any sense if your dataset is ~30 trillion tokens. My uneducated guess is 1/10 compression ratio is good and that point to ~300 to ~1000 trillion tokens, which again, doesn't make any practical sense.
> Hey this Opus 5.5 with the increased parameters seems much better than the last one"
Is this true?
I barely notice any improvements ever since... 4.6? At least for software development.
Opus makes the same sort of mistakes and has the same sort of strengths ever since that point in my perception. My impressions is that increasing the number of parameters clearly has diminishing returns.
In fact, the best improvement I got from Claude was Haiku 5.5. Good cheap workhorse to implement plans.
As in your statement to me indicates that you clearly see the scaling working over longer timespans. You chose a relatively recent 2026 model as your baseline after all, not a 2025 model or earlier. You clearly see that the 2026 model is better than ones in the past.
Might i suggest that you perhaps just don't see the improvement as clearly when comparing iterations that are fairly close together? If you're not anecdotally seeing improvement between an early 2026 model and one from ~6 months later perhaps just give it another iteration? Personally i do see 5.5 as much better but i can understand not seeing a lot of difference to 4.6 since they literally aren't that far apart in releases.
> Might i suggest that you perhaps just don't see the improvement as clearly when comparing iterations that are fairly close together?
No, you can't. Because you invalidate your argument by adding later:
> Personally i do see 5.5 as much better
It has to be one or the other.
Either each minor release Anthropic farts is a major improvement from what came before, or I don't see much difference because it has been very little time in between releases. Holding both arguments as true would require me to be excessively apologetic for a company that desires a 2T IPO.
And to be clear, I do think that there is an argument that not much time has passed since Opus 4.6 and Chat GPT 5.3 Codex days.
Part of this conundrum may be because the models reached a good enough point for software development. Maybe walking 80% of the way was easier than the final 20%. I am nowadays more excited about how cheap models can be useful for it (hence my positive remark about Haiku 5.5 and Chinese models in general).
> Current top models are mere (mere in big datacenter terms) trillions of parameters right now. ... We'll soon start hitting 10's of trillions and 100's of trillions of parameter models.
Chinchilla Law says no, we wont. We're likely well past the point of diminishing returns but MOE hides a fair bit of it.
We already know that power laws apply to scaling so 10x the size is going to result in minor gains - where costs become a real factor (as if they weren't already).
FTA "The core philosophy there is scaling, and aggressively not research."
This is where the real gains for AI come from - good enough and CHEAP is going to win the day here (brute force, mixture of models). It's the engineers displacing the academics - and that will mean a bloodbath for revenue (see DeepInfer for what the end game looks like).
I'm not an AI booster (in a "should" not "can" sense), but in terms of "can" we are going to see next generation hardware (i.e. Vera Rubin GPU architecture) and improvements in model density (i.e. loosely improvements in intelligence per millions of weights), so the models will get bigger, the models will get more dense, and we may see more "expert models" i.e. in the sense that for a while there were human generalists that knew all of science and mathematics until the 1800s, after a certain point we had to specialize since there was too much to know. The models will need to partially specialize after hitting the hardware ceiling.
Bioinspired algorithms may eventually lead to vastly improved intelligence per a weight too.
I don't think we should do this, but I feel doomed in that I suspect that we "can".
> The models will need to partially specialize after hitting the hardware ceiling.
I'm incredibly skeptical of anything that goes "oh, we need to split AI into specialist systems".
I suspect that a lot of the delta between "our best models" and "base models they're trained on" lies in relatively low amounts of parameter changes driven by SFT and RL.
Model capacity is not the limit - the limit is our ability to drive those changes efficiently. Even things like "good research taste" could be represented as tiny in relation to total model capacity, if they could be trained into a model efficiently.
At least for now, those keys haven't been completely unlocked, so parameter scaling is the primary way corporations are achieving increases in ability. If they hit the hardware ceiling, maybe the research direction changes then.
No, parameter scaling still works because it broadly increases sample efficiency. We don't have the data to make a Chinchilla optimal 100T, but that doesn't mean there are no gains in building up to 100T.
I don't think "biology" and "elegantly" belongs in the same sentence. Let alone the one you wrote.
I suspect that the "elegant" workings of the human brain are a pile of hacks. Evolution crammed the brain full of patterns that are a good fit for the specific problems it typically solves - the way curve detectors and scale invariance are a good fit for CNNs.
We don't have millions of years of evolution, but what we have is enormous datasets and compute at scale.
I'm seeing the same pattern at other multinationals with huge R&D departments. Mostly colleagues with academic pedigree in traditional ai research who have been for years in denial that transformers are actually - in combination with recent advanced patterns in software development - the key to ai. it's really simple as that! except they weren't part of the revolution and still don't want to accept the reality that their research has become obsolete.
Is it the key to AI? It probably depends on the squishy definition of AI and what different people want from it. If your definition includes self awareness and sentience I’m not sure it is the key (I’m also not entirely convinced it isn’t, either), but it’s not clear either way in my opinion.
The one thing that is clear is if it can become more cost effective it’s world changing. Even at its subsidized standards now it’s world changing. And it’s reasonably classified as AI IMO, but it’s not self aware.
I think a lot of your general traditional researchers are fundamentally looking to understand ourselves through machines as proxies in a sense. Something self aware that senses, learns, adapts, so on. The current path isn’t there or necessarily clearly the key to achieving that. That may be the disconnect. I think LLMs may give us some hints at that but not the full picture.
Still incredibly powerful for things we once considered nearly impossible.
Except doing both..is..so..hard when the future payoff of one is more unclear and you are faced with quarterly earnings reports.
Look at what happened to Meta where they squandered a lead, had some mixed results, and reorged to more product certainty and away from fundamental research.
Theoretically one need doesn't preclude the other. The OP gives a detailed argument and his own experience of one need precluding the other in practice.
Both of these approaches require quite a bit of institutional support notably.
It seems extremely likely to me that scaling produces benefits, it will get better at particular things things continue but scaling by itself isn't heading to general intelligence.
I should also say that I use ChatGpt daily and enjoy it for researching algorithms. I tried to set up Claude last week and it/he was so overwhelming as to be unusable. So I think frontier models can be "unwieldy" or something for average use.
This doesn't engage with the article in the precise way that transformer-based improvements and fundamental research that way are not being incorporated into models, at Microsoft AI, which is concerning.
is scaling working? It is also my understanding that Opus 5.5 is probably a well under 10T parameter model. Probably a similar size class to Fable 5 earlier. And supposedly even GPT-4 was a 1T model. Clearly most of the gains are coming from something other than scaling up the model size. Yes, scaling up RL counts as scaling too, but the fact that labs have pivoted to a different axis of scaling is evidence that it's not just a "straight highway" to improve LLMs.
From the outside it seemed like GPT 4.5 was the apogee of parameter size scaling and since then a huge investment in RLVR became the new scaling axis.
So yeah, LLMs keep getting more powerful/useful but because of an industry wide pivot not a "straight highway" as imagined in the GPT-4 era, when simply going bigger and bigger seemed like a reasonable roadmap for the foreseeable future.
What if scaling fails to hold at a certain point? Whether that is a steep or gradual dropoff in returns. What happens to the empires of accumulated compute? How much value do they immediately lose once the returns to scaling ceiling is reached?
> Foot down, get to 10 trillion parameters, see if the scaling holds, if so off to 100 trillion parameters
Moore's law: Computing power is multiplied by 10 every 6 months? Can't wait for the optimization and rationalization phase, where we'll recover a trillionth of that power to run powerful databases.
> Essentially the messaging it "Foot down, there's a straight highway ahead don't divert".
Feels like this is leaving major breakthroughs on the table for the competition. Sure, transformers might get us much farther. But if someone figures out an approach that unlocks a performance boost or better resolution then they are going to kill the competition.
Feels like the American industry is high on all the capitol being shoveled into the furnace without caring how efficient the furnace is. At some point someone is going to make a breakthrough and burst that bubble. This could crater the value as we are currently betting on the low power efficiency of the models as the barrier of entry and point of profit. Of course that would lead to cheaper AI but is that profitable vs local models once (or if) it becomes affordable to do so?
These corporations have an obscene amount of money
This is a false dichotomy if I ever saw one
Why not hedge and just keep an ai skunkworks team working? These companies are big enough that they can have these people far away from the main work without distraction or without starving those teams of bandwidth
Ignoramus question: does the 100s of trillions of parameters world still exist with a constrained domestic data center build out? Democrats are signaling they are about to go psycho mode on data center contracts.
Like the sister comment, if not in US then in other countries.
US rightfully placed a bunch of labor/environmental regulations around manufacturing which moved to China but stupidly didn't place restrictions on goods manufactured with severe labor and environmental violations.
> US rightfully placed a bunch of labor/environmental regulations around manufacturing which moved to China but stupidly didn't place restrictions on goods manufactured with severe labor and environmental violations.
The 'US' isn't a monolithic thing. US market leaders convinced US politicians to open US markets to China, and the lack of labor and environmental regulations was a feature, not a bug in the plan.
Im at a company with its own set of machine learning engineers, a few separate organizations internally have them so probably a few hundred MLEs in total.
More and more, the talk of using generative models (e.g. transformers) is coming up to replace existing models in all concepts. Anything from understanding data, image recognition, labeling, ranking, etc.
As someone who is not an MLE and really doesnt understand these models well, it seems like the transformer approach is a "one size fits all" solution that may not be optimal for all problem categories but is good enough that its not worth investing in specialized models to get incremental improvements.
Isn't the thing with transformers is that they work really well, but only if you have a lot of training data? If data is sparse, you need more specialized models
Which many credible, AI researchers believe is already starting up an increasingly steep scaling wall of diminishing returns that eventually ends at a brick wall.
That depends on your definition of scale. Yes! We were to literally convert the whole world and its resources into data centers.
But the whole problem is that these things aren't acting like other technological revolutions like the internet or Google where lots of profit comes from relatively little spend. They scale linearly maybe if you are super smart and good about your design.
Meanwhile, citizens all over the world are fighting data centers tooth and nail because we are hitting a scaling limit in terms of the resources people will allow to be put into this highway project and its equally unwise and untethered from reality to pretend this doesn't matter in a strategic sense.
"citizens all over the world" will not lift a finger for CAFOs or trawl fishing. i really don't think anyone "fighting data centers tooth and nail" is doing so because of a belief in the limits of scaling or any understanding of scarcity.
I'm still convinced the model advancements are largely US firms jingling a set of keys in front of investors/markets/the public to distract while they scramble to build serious patent moats around the harness/orchestration/infrastructure technologies.
I think that we're well into diminishing returns already. The models have improved, but to me the far bigger advances have been in feeding them contexts, building them into loops, etc.
It's like we invented the wheel in 2022, the crowds are fighting over capital to build wheel factories, and savvier tech firms are quietly patenting automobile designs while everyone's focused on making slightly rounder wheels.
I would say I used to believe how the models would scale, now I'm a bit more skeptical. I would be incredibly surprised if oai, anthropic, etc hadn't at least TRIED this yet. They have the compute to do it. They're either sitting on it and saving it for a rainy day, or there is something wrong with it.
We know nothing about their internal models that, for instance, produced revolutionary results in mathematics. The public no longer gets to see what’s at the frontier, and we may never see much of it. Also, the cost to train these is immense. China literally does not have enough chips in circulation to compete. I’m not sure where US lab capabilities tap out but I suspect they are at least partly bound by GPU production as well at this point.
We know that if this immense cost had in fact been spent, and the internal models really were that revolutionary, these companies would be selling access to it for cash instead of taking on mountains of debt.
I agree they would not shut up about it if they had something to reveal, so we probably have a semi-accurate vision of the capabilities if not the implementation details.
You’re underestimating the cost of training and serving a 10T model on current hardware. The fair market pricing to make this available to external customers napkin maths out on the order of $10k/hr. It really just doesn’t make sense.
The hardware that’s taping out ~now is going to bring this down considerably.
This is why I’m personally convinced we’re in a massive bubble, not because AI is going to stop scaling but because they spent too much money too quickly without anticipating rapid hardware improvements. H100s are still being sold for tens of thousands of dollars a piece, when we’re less than 500 days away from hardware which is going to make them completely obsolete. No amount of distillation or quantization is going to save them.
GPUs were designed first and foremost to be useful as a toy for adult children gamers, second to be useful as a general solution to embarrassingly parallel math problems. We need neither here, theres a very specific set of operations that are actually needed. Specialized hardware for this problem was always going to win in the end, but it took a few years for it to be designed and validated and produced at scale. Too many companies failed to keep powder dry in anticipation of that playing out.
> "Hey this Opus 5.5 with the increased parameters seems much better than the last one". Well that's a trillion parameter model.
Opus is 5T, not 1T.
> We'll soon start hitting 10's of trillions
Fable is already 10T.
> "Foot down, there's a straight highway ahead don't divert". I have the view this makes sense too.
The roadblock they ignored is politics.
Three days ago the states where most of the new datacenters are being built (Texas, Ohio, Georgia and others) elected governors who campaigned on halting datacenter construction. The highway metaphor implies money in => intelligence out. Turns out there's at least one other ingredient, and we just ran out of it.
More parameters without significantly more data fails; it loses the ability to generalize and simply memorizes.
And we’ve consumed pretty much all the data, and that part is getting harder as data is now expensive, Balkanized by owners, litigated and charged for. Couple that with model collapse as models eat their own output (not solved for any models, despite more than a decade of the problem), and add that lots of previously human curated datasets (stack overflow style), and it very well may be the models are not getting much better.
Very large models trained with sgd variants don't tend to over fit, especially when using few epochs. See double descent.
The coding/math breakthroughs were caused by training models on the subset of their own outputs which pass a verifier, see reinforcement learning from verified rewards.
That's the historical wisdom, but there are reasons to believe the opposite: that past a certain point more parameters are always better. Basically, that overfitting don't matter, because a large enough model will memorize and build the general solution simultaneously. Then weight decay and SGD noise encourage forgetting of the memorized solution so the general solution takes over.
Do they really? If your experience is with the free version of ChatGPT, I wouldn't be surprised if you're running a MoE with like 8-12B active params and <100B total.
(Just speculation on params, but I find a lot of "LLMs still hallucinate on non-coding problems" come from people using very-cost-optimised free models and drawing the wrong conclusions).
> If your experience is with the free version of ChatGPT,
Even the free version of ChatGPT shouldn't hallucinate, after all it uses a good enough model. And the answer is no, it's a combination of paid and free models coming from OpenAI, Claude and Google's Gemini (with this latter one actually badly miss-identifying a WW2 military dictator a couple of weeks ago, wrote about it here [1] )
Why wouldn't the explore/exploit balance leave at least a little to the explore side? Huge amounts of money are going into AI, why not save a little bit to keep an eye on the things that don't currently look like they're going to scale up?
Like maybe a couple dozen people on an explore team?
>> We definitely haven't hit the ceiling and everyone's noticing "Hey this Opus 5.5 with the increased parameters seems much better than the last one".
Opus is better (not necessarily "very much") than Sonnet in coding. That's because of a gigantic engineering effort to make it better at coding. That is a regression: we've gone from "sparks of AGI" and systems that were said to be "genearl-purpose" to systems that do one thing well and everything else... well, there's no real measure of how better Opus is from Sonnet in replying to your emails so it's hard to say.
Same goes for OpenAI in maths. Industry wide, there is a regression from general-purpose to special-purpose systems, like in Good, Old-Fashioned AI. LeCun is absolutely right to complain that this is not the path to general intelligence. I don't think his preferred approach, JEPA, is going anywhere either, but at least he's trying. Other luminaries of the field have expressed similar doubts, e.g. Richard Sutton's tweet was in my news feed yesterday saying that brute force is ngmi and algorithmic advances are needed [1]. Schmidhuber I think has said multiple times something along the lines of "there is no intelligence without physical intelligence".
We are very far from that. The recent trend for Vision-Language-Action Transformers (VLA-Transformers) in robotics has conclusively demonstrated that you can't solve the real world by overfitting to a set of discrete tasks, and there is no source of data that we can harness to repeat the feat of LLMs. Algorithmic advances will be necessary, but as the article suggests, these have now be put on hold.
There is a dumbing down, slowing down of research as all resources are spent to try and scale what we have now. What we have now is not enough for intelligence, but that doesn't matter. The industry's goal is not intelligence, it's making more money than anyone has ever made before; in fact more money than all humans put together have ever made before. That's the current race: who will win the money game. Obviously that's not likely to do anything for intelligence.
____________
[1] Brute force computation is not enough for intelligence. Clever new methods are needed. The science of AI is far from finished. IMHO
This sort of tallies with the claim that every significant technology in the iPhone was initially developed through state funding. Basic research and fundamental inventions are expensive, long term, high risk projects that are generally incompatible with companies driven by quarterly profits and stock price (with some notable exceptions such as pharmaceutical companies).
Same in deepmind too. Bit of a shuffle as well to the frontier AI group, jeff dean and pratik jain headed it now jeff is gone, so pratik is in MTV. And most of the different research groups are all moved to LLMs + no publication mode now.
I wouldn't interpret it as some sort of philosophical position. After a few years once the current buildout settles down they will go back to the usual setup.
Right now it is stupid to not dedicate all organisational resources towards LLMs, when you are lighting your free cash flow on fire and issuing bonds like there's no tomorrow, you might as well have all your smart people working on it. JEPA can wait a few years.
It's mentioned in the article as "post-inflection" which means when Mustafa's startup was acquihired, and Mustafa became CEO of msft AI. (Microsoft likes to call the heads of various orgs "CEO".)
Mustafa has since been demoted from head of MAI (which includes applications) to just head of the superintelligence team. (Of course not externally broadcasted as a demotion.)
If you value the freedom to pursue your own research direction and share your work freely, then choose academia. All the perks of working in an industry research role are provided in exchange for this freedom.
If your research aligns with the incentives of industry, you can get the best of both worlds in industry. This was the case for machine learning researchers in the post-AlexNet, pre-ChatGPT era. The good times can't last forever, though.
>
If you value the freedom to pursue your own research direction and share your work freely, then choose academia.
The working environment in academia is often precarious :-( , so such people often have no alternatives to going into industry, even though they still deeply appreciate and are influenced by the values of academia.
>If you value the freedom to pursue your own research direction and share your work freely, then choose academia
Perhaps it is more free, but my understanding has always been that professional academia still restricts freedom of research. You have to apply for funding, convince the board that your direction will be fruitul.
An archaeologist can't just dig wherever. They compete with their peers for a limitted funding pool. If the school's board is more interested in the bronze age than the paleolithic, you're probably not going to get your dino dig approved.
The same is true for CS, I assume. You can't just "i would like spend university resources developing brainfuck 2.0" because the other teams are more likely to get their funding over you getting yours.
Both corporations and university's have finite budgets, both with bottom lines and margins. The uni wants their research done in a popular field that will attract students because students generate income that funds research that attracts students that generate income that funds research that...
Academic research is free in the same sense as running your own business. You can try whatever you want if you can find the funding, but you still have to deal with people and regulations.
Universities are mostly just administrative organizations. They have little control over your research, they rarely fund it, and they are usually not even interested in it.
That’s not the issue here, it’s not like AIF was doing purely academic work.
They just think it makes sense to explore things, rather than just rinse and repeat with more cash.
The good news for John is that when 99.99% of the competitors are doing all the same, it’s when the rewards for looking around can be the biggest. Those are the real moments for disruption, and companies disregard them at their peril.
In such cases I think it's only fair that HN points get shared between the posters. I am a little embarrassed to get them all, especially for just pointing towards a news item.
121 comments
[ 9.1 ms ] story [ 2699 ms ] threadPerhaps "The End of Microsoft Research's AI Frontiers Group" would be a less ambiguous title.
To the author: 100% need this kind of thinking. It’s a balancing act. Historically though, dedicated labs haven’t had the best track record at commercializing their tech. The skillset differs vastly from running a product-focused business.
That said, even within this narrow scope - there are areas ripe for research.
Some of my particular interests include:
(1) Latent thinking/reasoning, which could arguably be more efficient but also opaque is one
(2) Separation of “intelligence” from knowledge - ie. keep the ability to reason like a frontier model and make tool calls (calling the ‘oracle’ or searching the web) but remove all parameters associated with niche knowledge/facts.
These are a couple off the top of my head that are very much aligned to scaling.
That statement is broad enough to be unfalsifiable, but if you narrow it down just a bit to scaling LLMs, maybe there is an end in sight. At least financial markets are skeptical of there being no end in sight.
If scaling LLMs really was the road to God in a box, then compute demand would be infinite, and nobody building to meet that demand would have a junk debt rating. Did those money people just not get it?
Langford has been at MSR a very long time and has done amazing research doing that time. He's bounced around a bit at MSR during his tenure and it's a shame to lose him. AI Frontiers wasn't around that long.
Much of this is Microsoft performing a very large and uncommunicated strategy update. All updates of MSR reorgs are just that "MSR is becoming more lean and focused" when it's obvious this isn't the case. This gets into the conversation that many Microsoft employees tend to have these days which is "Why does culture no longer matter at Microsoft and why did Satya turn into another heartless CEO". I won't speculate there, but the place isn't nearly as fun as it used to be.
I'm not sure how that, combined with open models inching up on frontier intelligence, is compatible with these massive AI IPO valuations?
It seems like closed model AI demand has a lot of contingencies that are ultimately fragile and possibly short sighted.
If so, then yeah, increasing efficiency 100x means we hit that point much faster.
If not, then it just means we make improvements 100x faster.
From the outside, Microsoft seems to actually be one of the better places, and is making at least token efforts to reduce headcount in more amicable ways, as exemplified the buyout offer.
But in general, the strategy seems to be overload people until they burn out and quit, because hey, AI will replace them anyway.
The transformer based approach is seen as a highway to race down in terms of training and building better models with minimal organizational distractions allowed while every company puts its foot down on this highway. Exploratory research into other AI methods is highly discouraged for now and organizations are encouraging transformer based scale out and iteration as a top priority.
As an example Meta literally fired it's previously head of AI who held views that transformers won't scale. He wanted to veer off the highway and heavily fund exploratory new types of models. They got someone willing to scale the current method out. Essentially the messaging it "Foot down, there's a straight highway ahead don't divert".
I have the view this makes sense too. Current top models are mere (mere in big datacenter terms) trillions of parameters right now. We definitely haven't hit the ceiling and everyone's noticing "Hey this Opus 5.5 with the increased parameters seems much better than the last one". Well that's a trillion parameter model. We'll soon start hitting 10's of trillions and 100's of trillions of parameter models. Unless there's a ceiling which nobody expects they should keep scaling just as they have been. I really do agree that it'd be insane to veer off this highway with exploratory research at this particular point in time. Foot down, get to 10 trillion parameters, see if the scaling holds, if so off to 100 trillion parameters. If the scaling continues to hold we'll be asking the AI to do the exploratory research for us soon enough.
Already happening: https://news.ycombinator.com/item?id=50397994
Obviously we'll see how it pans out, but other approaches feel a bit like trying to ignore the bitter lesson.
Is that observed scaling part of some equation or law? Is there a limit to how far that scaling goes before a ceiling or diminishing returns hit? Or is it simply based on more GPUs/parameters/compute plus how much human knowledge is in existence and will scale predictably forever?
And you can in practice often see a large network follow an expected training trajectory for loss/entropy.
In practice data is getting more sparse than compute, so you are usually overtraining using more tokens and being suboptimal.
These empirical observations on quantities which aren't "intelligence" since no one has yet defined intelligence.
I wonder about this a lot.
There are all these publications from OpenAI and Anthropic that are solving complex math & science problems.
Are they just "brute forcing" it? Able to try a lot more "known" or reasonable approaches given existing knowledge?
Or are they showing "emergent intelligence" and able to try genuinely new approaches that can't be expected from the knowledge in their dataset?
There's a mathematical sense in which it cannot be: the data-processing inequality (https://en.wikipedia.org/wiki/Data_processing_inequality). The extent to which that sense is totally counterintuitive to how things normally work in ML nowadays has led to investigating different foundations for information theory: https://arxiv.org/abs/2601.03220.
Are you smart if you see that cognitive scientists had something figured out, connected it to something in your field of chemistry and made a breakthrough treatment for some disease?
Does any of our existing intelligence build on completely unknown information? You could argue that any new solution/emergent intelligence is an association of existing information in new ways.
It also brings us to what is brute forcing? Humans are brute forcing stuff most of the time. You will have 10 stupid thoughts for every good thought. 10 stupid ideas for every good idea, you will discard most of them so fast that you won't think of that you did that. It's very rare, if not near impossible, that a non-mathematician with no prior knowledge, goes into a room on day 1 of deciding to think about a conjecture they don't know anything about, and solve it on the blackboard. They need to use their collect associations, formalize their axioms, think of first principles (known information), and associate it.
So to answer your question you need to answer:
1) What does "smarter" refer to? Is it accumulation of information about the universe?(if yes, a database is smart) Is it an efficient, abstracted representation of accumulated? Is it prediction of the future state of the universe given a learned abstraction? Is it compression?
Or going further, is it solving tasks useful for the human in the world? Is it acting as an agent with it's own goals and interests in the universe? What if the goals are hard for humans to understand?
2) What does "genuinely new approaches" refer to? Is it fine to use math axioms to prove things? Does interpolation between two existing solutions count as new? At what granularity of concept does novelty happen (Does using two higher abstraction concepts, like combining some new, modern neural network architecture with a database count as a new thing), even if the underlying lower abstraction of information as storage, latency and compute is not novel?
---
As you can imagine, this is all not really defined or agreed upon. Here are some examples of projects that are set up to drive these ideas:
The Hutter Prize poses that intelligence is defined by compression, so if you manage to compress the English Wikipedia efficiently, you are closer to better intelligence (http://prize.hutter1.net/)
Compute: Intel CPU
The ARC-AGI foundation poses that being able to be goal-oriented and understand higher level abstractions is intelligence: https://arcprize.org/leaderboard
Compute: NVIDIA T40s
Modded NanoGPT competition poses that intelligence is minimizing entropy/loss over a dataset in the least amount of steps: https://github.com/kellerjordan/modded-nanogpt
Compute: NVIDIA H100
---
(This moves the argument to "is being really good at predicting the next token the same as being intelligent?" which has happily diverted many minds for years now)
> for every doubling of model size the number of training tokens should also be doubled
Well, I don't think many claim that increasing the model size while also increasing the training tokens doesn't make sense.
The problem is how to get the training tokens...
And the answer to your question is basically just “synthetic data”.
Where's the money coming for those 10X and 100X ventures?
Is this true?
I barely notice any improvements ever since... 4.6? At least for software development.
Opus makes the same sort of mistakes and has the same sort of strengths ever since that point in my perception. My impressions is that increasing the number of parameters clearly has diminishing returns.
In fact, the best improvement I got from Claude was Haiku 5.5. Good cheap workhorse to implement plans.
As in your statement to me indicates that you clearly see the scaling working over longer timespans. You chose a relatively recent 2026 model as your baseline after all, not a 2025 model or earlier. You clearly see that the 2026 model is better than ones in the past.
Might i suggest that you perhaps just don't see the improvement as clearly when comparing iterations that are fairly close together? If you're not anecdotally seeing improvement between an early 2026 model and one from ~6 months later perhaps just give it another iteration? Personally i do see 5.5 as much better but i can understand not seeing a lot of difference to 4.6 since they literally aren't that far apart in releases.
No, you can't. Because you invalidate your argument by adding later:
> Personally i do see 5.5 as much better
It has to be one or the other.
Either each minor release Anthropic farts is a major improvement from what came before, or I don't see much difference because it has been very little time in between releases. Holding both arguments as true would require me to be excessively apologetic for a company that desires a 2T IPO.
And to be clear, I do think that there is an argument that not much time has passed since Opus 4.6 and Chat GPT 5.3 Codex days.
Part of this conundrum may be because the models reached a good enough point for software development. Maybe walking 80% of the way was easier than the final 20%. I am nowadays more excited about how cheap models can be useful for it (hence my positive remark about Haiku 5.5 and Chinese models in general).
Chinchilla Law says no, we wont. We're likely well past the point of diminishing returns but MOE hides a fair bit of it.
We already know that power laws apply to scaling so 10x the size is going to result in minor gains - where costs become a real factor (as if they weren't already).
FTA "The core philosophy there is scaling, and aggressively not research."
This is where the real gains for AI come from - good enough and CHEAP is going to win the day here (brute force, mixture of models). It's the engineers displacing the academics - and that will mean a bloodbath for revenue (see DeepInfer for what the end game looks like).
Bioinspired algorithms may eventually lead to vastly improved intelligence per a weight too.
I don't think we should do this, but I feel doomed in that I suspect that we "can".
I'm incredibly skeptical of anything that goes "oh, we need to split AI into specialist systems".
I suspect that a lot of the delta between "our best models" and "base models they're trained on" lies in relatively low amounts of parameter changes driven by SFT and RL.
Model capacity is not the limit - the limit is our ability to drive those changes efficiently. Even things like "good research taste" could be represented as tiny in relation to total model capacity, if they could be trained into a model efficiently.
Thinking tokens have nothing to do with parameter count and where some of the biggest gains have come from.
We're really past the point of scaling parameters being viable. We simply lack the content to make "bigger" better at this point.
I suspect that the "elegant" workings of the human brain are a pile of hacks. Evolution crammed the brain full of patterns that are a good fit for the specific problems it typically solves - the way curve detectors and scale invariance are a good fit for CNNs.
We don't have millions of years of evolution, but what we have is enormous datasets and compute at scale.
Sure, frontier LLMs require timeslots on massive server arrays. But they also get things done a lot faster than humans do. The net favors the LLM.
The one thing that is clear is if it can become more cost effective it’s world changing. Even at its subsidized standards now it’s world changing. And it’s reasonably classified as AI IMO, but it’s not self aware.
I think a lot of your general traditional researchers are fundamentally looking to understand ourselves through machines as proxies in a sense. Something self aware that senses, learns, adapts, so on. The current path isn’t there or necessarily clearly the key to achieving that. That may be the disconnect. I think LLMs may give us some hints at that but not the full picture.
Still incredibly powerful for things we once considered nearly impossible.
I see it that most of legacy research is not obsolete but can be significantly amplified by AI.
Race down the transformer highway as fast as possible and explore other AI methods.
One need not preclude the other.
Look at what happened to Meta where they squandered a lead, had some mixed results, and reorged to more product certainty and away from fundamental research.
Both of these approaches require quite a bit of institutional support notably.
I should also say that I use ChatGpt daily and enjoy it for researching algorithms. I tried to set up Claude last week and it/he was so overwhelming as to be unusable. So I think frontier models can be "unwieldy" or something for average use.
It
So yeah, LLMs keep getting more powerful/useful but because of an industry wide pivot not a "straight highway" as imagined in the GPT-4 era, when simply going bigger and bigger seemed like a reasonable roadmap for the foreseeable future.
Moore's law: Computing power is multiplied by 10 every 6 months? Can't wait for the optimization and rationalization phase, where we'll recover a trillionth of that power to run powerful databases.
Feels like this is leaving major breakthroughs on the table for the competition. Sure, transformers might get us much farther. But if someone figures out an approach that unlocks a performance boost or better resolution then they are going to kill the competition.
Feels like the American industry is high on all the capitol being shoveled into the furnace without caring how efficient the furnace is. At some point someone is going to make a breakthrough and burst that bubble. This could crater the value as we are currently betting on the low power efficiency of the models as the barrier of entry and point of profit. Of course that would lead to cheaper AI but is that profitable vs local models once (or if) it becomes affordable to do so?
This is a false dichotomy if I ever saw one
Why not hedge and just keep an ai skunkworks team working? These companies are big enough that they can have these people far away from the main work without distraction or without starving those teams of bandwidth
US rightfully placed a bunch of labor/environmental regulations around manufacturing which moved to China but stupidly didn't place restrictions on goods manufactured with severe labor and environmental violations.
We are repeating the mistake again.
The 'US' isn't a monolithic thing. US market leaders convinced US politicians to open US markets to China, and the lack of labor and environmental regulations was a feature, not a bug in the plan.
Yeah, we all know that.
Who voted for the politicians? Who clamored for cheaper goods and fell for the lies?
Anthropic, which is more open about this has shown the numbers, it works.
More and more, the talk of using generative models (e.g. transformers) is coming up to replace existing models in all concepts. Anything from understanding data, image recognition, labeling, ranking, etc.
As someone who is not an MLE and really doesnt understand these models well, it seems like the transformer approach is a "one size fits all" solution that may not be optimal for all problem categories but is good enough that its not worth investing in specialized models to get incremental improvements.
Which many credible, AI researchers believe is already starting up an increasingly steep scaling wall of diminishing returns that eventually ends at a brick wall.
But the whole problem is that these things aren't acting like other technological revolutions like the internet or Google where lots of profit comes from relatively little spend. They scale linearly maybe if you are super smart and good about your design.
Meanwhile, citizens all over the world are fighting data centers tooth and nail because we are hitting a scaling limit in terms of the resources people will allow to be put into this highway project and its equally unwise and untethered from reality to pretend this doesn't matter in a strategic sense.
I think that we're well into diminishing returns already. The models have improved, but to me the far bigger advances have been in feeding them contexts, building them into loops, etc.
It's like we invented the wheel in 2022, the crowds are fighting over capital to build wheel factories, and savvier tech firms are quietly patenting automobile designs while everyone's focused on making slightly rounder wheels.
We definitely know that.
The hardware that’s taping out ~now is going to bring this down considerably.
This is why I’m personally convinced we’re in a massive bubble, not because AI is going to stop scaling but because they spent too much money too quickly without anticipating rapid hardware improvements. H100s are still being sold for tens of thousands of dollars a piece, when we’re less than 500 days away from hardware which is going to make them completely obsolete. No amount of distillation or quantization is going to save them.
GPUs were designed first and foremost to be useful as a toy for adult children gamers, second to be useful as a general solution to embarrassingly parallel math problems. We need neither here, theres a very specific set of operations that are actually needed. Specialized hardware for this problem was always going to win in the end, but it took a few years for it to be designed and validated and produced at scale. Too many companies failed to keep powder dry in anticipation of that playing out.
Opus is 5T, not 1T.
> We'll soon start hitting 10's of trillions
Fable is already 10T.
> "Foot down, there's a straight highway ahead don't divert". I have the view this makes sense too.
The roadblock they ignored is politics.
Three days ago the states where most of the new datacenters are being built (Texas, Ohio, Georgia and others) elected governors who campaigned on halting datacenter construction. The highway metaphor implies money in => intelligence out. Turns out there's at least one other ingredient, and we just ran out of it.
And we’ve consumed pretty much all the data, and that part is getting harder as data is now expensive, Balkanized by owners, litigated and charged for. Couple that with model collapse as models eat their own output (not solved for any models, despite more than a decade of the problem), and add that lots of previously human curated datasets (stack overflow style), and it very well may be the models are not getting much better.
The coding/math breakthroughs were caused by training models on the subset of their own outputs which pass a verifier, see reinforcement learning from verified rewards.
https://arxiv.org/pdf/1912.02292
There's plenty more. Read Google scholar
https://youtu.be/lKoJJxjUfdw is a really excellent lecture that changed how I think about it.
The LLMs still hallucinate way too much when there's no computer code involved, but the computer programmers in here won't be aware of that.
(Just speculation on params, but I find a lot of "LLMs still hallucinate on non-coding problems" come from people using very-cost-optimised free models and drawing the wrong conclusions).
Even the free version of ChatGPT shouldn't hallucinate, after all it uses a good enough model. And the answer is no, it's a combination of paid and free models coming from OpenAI, Claude and Google's Gemini (with this latter one actually badly miss-identifying a WW2 military dictator a couple of weeks ago, wrote about it here [1] )
[1] https://news.ycombinator.com/reply?id=50256321&goto=threads%...
Like maybe a couple dozen people on an explore team?
Opus is better (not necessarily "very much") than Sonnet in coding. That's because of a gigantic engineering effort to make it better at coding. That is a regression: we've gone from "sparks of AGI" and systems that were said to be "genearl-purpose" to systems that do one thing well and everything else... well, there's no real measure of how better Opus is from Sonnet in replying to your emails so it's hard to say.
Same goes for OpenAI in maths. Industry wide, there is a regression from general-purpose to special-purpose systems, like in Good, Old-Fashioned AI. LeCun is absolutely right to complain that this is not the path to general intelligence. I don't think his preferred approach, JEPA, is going anywhere either, but at least he's trying. Other luminaries of the field have expressed similar doubts, e.g. Richard Sutton's tweet was in my news feed yesterday saying that brute force is ngmi and algorithmic advances are needed [1]. Schmidhuber I think has said multiple times something along the lines of "there is no intelligence without physical intelligence".
We are very far from that. The recent trend for Vision-Language-Action Transformers (VLA-Transformers) in robotics has conclusively demonstrated that you can't solve the real world by overfitting to a set of discrete tasks, and there is no source of data that we can harness to repeat the feat of LLMs. Algorithmic advances will be necessary, but as the article suggests, these have now be put on hold.
There is a dumbing down, slowing down of research as all resources are spent to try and scale what we have now. What we have now is not enough for intelligence, but that doesn't matter. The industry's goal is not intelligence, it's making more money than anyone has ever made before; in fact more money than all humans put together have ever made before. That's the current race: who will win the money game. Obviously that's not likely to do anything for intelligence.
____________
[1] Brute force computation is not enough for intelligence. Clever new methods are needed. The science of AI is far from finished. IMHO
https://x.com/RichardSSutton/status/2119142310894334262?a=ry...
I wouldn't interpret it as some sort of philosophical position. After a few years once the current buildout settles down they will go back to the usual setup.
Right now it is stupid to not dedicate all organisational resources towards LLMs, when you are lighting your free cash flow on fire and issuing bonds like there's no tomorrow, you might as well have all your smart people working on it. JEPA can wait a few years.
Mustafa has since been demoted from head of MAI (which includes applications) to just head of the superintelligence team. (Of course not externally broadcasted as a demotion.)
Hah this is as funny as banks calling their junior employees VP of something something
If your research aligns with the incentives of industry, you can get the best of both worlds in industry. This was the case for machine learning researchers in the post-AlexNet, pre-ChatGPT era. The good times can't last forever, though.
The working environment in academia is often precarious :-( , so such people often have no alternatives to going into industry, even though they still deeply appreciate and are influenced by the values of academia.
Perhaps it is more free, but my understanding has always been that professional academia still restricts freedom of research. You have to apply for funding, convince the board that your direction will be fruitul.
An archaeologist can't just dig wherever. They compete with their peers for a limitted funding pool. If the school's board is more interested in the bronze age than the paleolithic, you're probably not going to get your dino dig approved.
The same is true for CS, I assume. You can't just "i would like spend university resources developing brainfuck 2.0" because the other teams are more likely to get their funding over you getting yours.
Both corporations and university's have finite budgets, both with bottom lines and margins. The uni wants their research done in a popular field that will attract students because students generate income that funds research that attracts students that generate income that funds research that...
Universities are mostly just administrative organizations. They have little control over your research, they rarely fund it, and they are usually not even interested in it.
They just think it makes sense to explore things, rather than just rinse and repeat with more cash.
The good news for John is that when 99.99% of the competitors are doing all the same, it’s when the rewards for looking around can be the biggest. Those are the real moments for disruption, and companies disregard them at their peril.