Posted by srean 3 hours ago
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.
Almost all statements Yann LeCun made back then (LLMs won't be able to do this or that) turned out to be wrong. They were right to fire him.
RCA developed the first good color TV cameras. They were monsters. Originally, electronic TV (as opposed to mechanical scanning) used Zworykin's iconoscope. This needed huge amounts of light. It didn't have a photo-multiplier stage to boost sensitivity, but it did accumulate charge from light over the whole frame time. There was an alternative approach, Farnsworth's image dissector, which also needed huge amounts of light. It had a photo-multiplier stage but only the light from the dot currently being scanned by the scanning beam mattered.
So RCA threw money at the problem. They developed the image orthicon camera tube. It combined the advantages of both approaches, integration over a full frame and a photo-multiplier stage. It was insanely expensive, costing thousands of dollars in the 1950s. It needed about twenty different voltages and multiple independent power supplies. It worked reasonably well.
Then came color. That took three image orthicons (RGB) inside a box the size of a small refrigerator, along with optics to split the incoming image into three colors. Plus a rack of analog electronics per camera, connected through a large cable. Plus an alignment problem from hell, with a huge number of adjustment knobs.[1]
Scaling worked! From the 1950s to the 1970s, color TV broadcasting was powered by those monsters.
Now, independent of this, other people were working on other approaches. Vidicons, silicon detectors inside camera tubes, and finally fully solid state image sensors. Eventually, in 1983, after thirty years of image orthicon monsters, RCA produced a usable solid state camera. Today, a cell phone camera costs about US$10 in bulk, and it's far better than the monsters.
It took thirty years of low-priority development to get past the giant image orthicon cameras.
Electronic fluorescent lamp ballasts had a different problem. That was a suppressed invention. The magnetic ballast industry was dominated by Magnetek. Someone invented a solid state replacement, which was smaller and cheaper, and licensed their patent to Magnetek for royalties. Magnetek just sat on it, not wanting to gut their core market. The inventor hired Townsend, Townsend and Crew in Palo Alto and sued. (Read this in the law firm's brag book in their office.) Eventually, the inventor got a hundred million or so, solid state ballasts went into production, and Magnetek seems to be gone. (The company called Magnetek today, which makes crane controls, is an offshoot of a Litton unit, and seems to be unrelated.)
There are also parallels from the steam locomotive industry, the Baldwin Locomotive Works being the best known. Baldwin actually built the first Diesel locomotive. They made a few half-hearted tries at Diesels over several decades, then gave up and went back to steam. Baldwin ceased all locomotive production in 1956. The corporation lingered on for a while, not doing much.
[1] https://www.tvcameramuseum.org/rca/tk40a/pix/rcatk40aart.pdf
But even in computing, when there is a promising way forward that the basic reaearch has been done on, its hard to justify side quests.
...And because they have monopoly money. Basically, research labs are usually vanity spending for corporations with too much money. Xerox parc was funded by a monopoly. Bell labs was funded by a monopoly. Google funds its moonshots with its monopoly.
Seems weird when it's become apparent that LLMs aren't really special sauce outside of just having the compute.
He was against LLMs, reinforcement learning, etc. Claiming that those are a dead-end. He was wrong. His own "JEPA" ideas have basically gone nowhere. Meta was right in moving on without him.
10% chance to be market leader doesn’t sound as good as 60% second place to them for example
Once you get close to the top, the risks don't look worth it.
It's his call for sure, but rather funny as an argument.
The author states "I consider the scaling-only focus to be both misdirected and not fun.". I get the point but also totally get the focus on scale out for now.
LLMs work well for search problems with the help of a strong verifier.
As it turns out, every problem is ultimately a search problem. But not every domain has an easily available verifier.
Which is superior, the big f*king monstrosity or the elegant on-demand, small batches, reduce waste?
Meta will go in the shitter and Deepseek will make the small batches.
It does (Lean at least) incorporate controlled experimentation, but tightly controlled with a both a laser focus on what the goal is and tight resource constraints. And Lean (at least several writers on Lean) favors taking multiple paths with a a quick and low risk and a longer-payoff, higher-risk, higher reward path for some efforts—one source used the US Navy’s two-path approach to SLBMs with Polaris and Poseidon as a model for this—but that's different than “maybe there is something down this path” basic research, which is outside of the concern a production model like TPS/Lean and often completely outside the model of a product-oriented firm; basic research is something firms consume and incorporate more than generate, because until the basics have been proven out, the rationally perceived chance of any particular avenue paying off is near zero for any firm.
Toyota is also a good example of what not to do. Their lack of earlier research investment in EVs has led to being severely behind the market, in addition to their lack of focus on quality causing poorer sales (their quality used to be one of the only reasons to buy them). KIA basically stole their playbook by creating loads of good-quality cheap hybrids and EVs. Toyota could've had that market if the invested more ahead of time in research.
And that's what's happening with AI. While there are plenty of players now, it's the ones that aren't looking ahead that will be left behind. Nothing scales forever; when you hit the ceiling, you're trapped in a box of your own making.
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.
(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”.
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.
Going close to as fast as possible while saving a stream for the other approaches makes sense as an approach, but let's not ignore that there's a compromise/tradeoff here. Might not matter in orgs with near infinite resources
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.
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.
LLMs need an awful lot of handholding all the time, meanwhile humans get pretty good at it on their own.
I mean we literally do this every few seconds already with natural birth processes.
Trying to estimate equivalent "parameter size" of the brain when we only superficially understand behavior of individual neurons let alone how intelligence emerges from the connectome seems too speculative to be a particularly useful metric.
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?
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 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).
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
Where's the money coming for those 10X and 100X ventures?
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.
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 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.
Perhaps "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?
> a clever idea pursued by one person it was likely to arrive late and unconvincing because one person could not move fast enough and the baselines which one person could assemble just were not strong enough to really show that an idea was worth taking flight
may no longer be true. The limiting factors are: frontier model token and compute budgets, researcher skill, drive, foresight and imagination, not the ability to work in teams. Of course, working in teams still helps, and benefits from internal use, peer review and promotion of the research line are certainly there, but compared to some time prior to 2026, it's pretty much not a factor.
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.
https://en.wikipedia.org/wiki/Digg#2025%E2%80%93present:_The...
Seems like it quickly pivoted quickly to cash in on "AI" hype.
The more interesting idea is that MS has abandoned AI research entirely, presumably because their zillions invested in other companies are a better bet in that area.
I think the real point of the post is that office politics and poor management got in the way of innovation due to infighting and poor execution in transfer of technology from research to product.
It was my introduction to feature hashing.
https://microsoft.ai/news/frontier-pre-training-at-microsoft...