This has nothing to do with superintelligence, it's just the people that were working on the paper prior to the re-org happened to publish after the name change.
Though it is notable that contrary to many (on HN and Twitter) that Meta would stop publishing papers and be like other AI labs (e.g. OpenAI). They're continued their rapid pace of releasing papers AND open source models.
What model(s) have Meta released since the Lab re-org?
Also, that wasn't based on purely hearsay, Zuck explicitly said:
> We believe the benefits of superintelligence should be shared with the world as broadly as possible. That said, superintelligence will raise novel safety concerns. We'll need to be rigorous about mitigating these risks and careful about what we choose to open source. Still, we believe that building a free society requires that we aim to empower people as much as possible. [0]
This is the right terminology. Model weights are literally compiled binary data; they are the output of an algorithm run on a bunch of source data. That training dataset is the "source" of the model. Training data (or the scripts used to generate it) is human-readable and modifiable, like source code. Binary weights are not.
It's kinda funny, Meta has long had some of the best in the field, but left them untapped. I really think if they just took a step back and stop being so metric focused and let their people freely explore then they'd be winning the AI race. But with this new team, I feel like meta mostly hired the people who are really good at gaming the system. The people that care more about the money than the research.
A bit of this is true at every major lab. There's tons of untapped potential. But these organizations are very risk adverse. I mean why not continue with the strategy that got us to the point we're at in the first place. Labs used to hire researchers and give them a lot of free reign. But those times ended and AI progress also slowed down. Maybe if you want to get ahead you gotta stop thinking like everyone else
Well meta... you can "hold me hostage" for a lot cheaper than those guys. I'm sure this is true for hundreds of passionate ML researchers. I'd take a huge pay cut to have autonomy and resources. I know for a fact there's many working at Mets right now that would do the same. Do maybe if you're going to throw money at the problem, diversify a bit and look back at what made SV what it is today and what made AI take leaps forward
My theory is that as more people compete, the top candidates become those who are best at gaming the system rather than actually being the best. Someone has probably studied this. My only evidence is job applications for GAFAM and Tinder tho.
I've spent most of my career working, chatting and hanging out with what might be best described as "passionate weirdos" in various quantitative areas of research. I say "weirdos" because they're people driven by an obsession with a topic, but don't always fit the mold by having the ideal combination of background, credentials and personality to land them on a big tech company research team.
The other day I was spending some time with a researcher from Deep Mind and I was surprised to find that while they were sharp and curious to an extent, nearly every ounce of energy they expended on research was strategic. They didn't write about research they were fascinated by, they wrote and researched on topics they strategically felt had the highest probability getting into a major conference in a short period of time to earn them a promotion. While I was a bit disappointed, I certainly didn't judge them because they are just playing the game. This person probably earns more than many rooms of smart, passionate people I've been in, and that money isn't for smarts alone; it's for appealing to the interests of people with the money.
You can see this very clearly by comparing the work being done in the LLM space to that being done in the Image/Video diffusion model space. There's much more money in LLMs right now, and the field is flooded with papers on any random topic. If you dive in, most of them are not reproducible or make very questionable conclusions based on the data they present, but that's not of very much concern so long as the paper can be added to a CV.
In the stable diffusion world it's mostly people driven by personal interest (usually very non-commericial personal interests) and you see tons of innovation in that field but almost no papers. In fact, if you really want to understand a lot of the most novel work coming out of the image generation world you often need to dig into PRs made by an anonymous users with anime themed profile pic.
The bummer of course is that there are very hard limits on what any researcher can do with a home GPU training setup. It does lead to creative solutions to problems, but I can't help but wonder what the world would look like if more of these people had even a fraction of the resources available exclusively to people playing the game.
This is such a nuanced problem. Like any creative endeavour, the most powerful and significant research is driven by an innate joy of learning, creating, and sharing ideas with others. How far the research can be taken is then shaped by resource constraints. The more money you throw at the researchers, the more results they can get. But there seems to be a diminishing returns kind of effect as individual contributors become less able to produce results independently. The research narrative also gets distorted by who has the most money and influence, and not always for the better (as recent events in Alzheimer's research has shown).
The problem is once people's livelihoods depend on their research output rather than the research process, the whole research process becomes steadily distorted to optimise for being able to reliably produce outputs.
Anyone who has invested a great deal of time and effort into solving a hard problem knows that the 'eureka' moment is not really something that you can force. So people end up spending less time working on problems that would contribute to 'breakthroughs' and more time working on problems that will publish.
> I certainly didn't judge them because they are just playing the game.
Please do judge them for being parasitical. They might seem successful by certain measures, like the amount of money they make, but I for one simply dislike it when people only think about themselves.
As a society, we should be more cautious about narcissism and similar behaviors. Also, in the long run, this kind of behaviour makes them an annoying person at parties.
You dislike them because they don’t benefit you indirectly by benefiting society at large.
The incentive structure is wrong, incentivizing things that benefit society would be the solution not judging those that exist in the current system by pretending altruism is somehow not part of the same game.
I agree that the system itself is dysfunctional, and I understand the argument that individuals are shaped or even constrained by it. However, in this case, we are talking about people who are both exceptionally intelligent and materially secure. I think it's reasonable to expect such individuals to feel some moral responsibility to use their abilities for broader good.
As for whether that expectation is "selfish" on my part, I think that question has been debated for centuries in ethics, and I'm quite comfortable landing on the side that says not all disapproval is self-interest. In my own case, I'm not benefiting much either :)
There is a difference between being selfish in the sense that you want others to contribute back to the society that we are all part of, and being selfish in the sense that you want to compete for exclusive rewards.
You can call this difference whatever you want, don't pretend that they are morally or effectively equivalent.
Please don't read too much into this single word. The comment above mentioned "nearly every ounce of energy they expended on research was strategic", and I was keeping that in mind while writing my remark.
Please read my sibling comment where I expand a bit on what I meant to say.
Thanks for sharing. I did not know this law existed and had a name. I know nothing about nothing but it appears to be the case that the interpretation of metrics for policies assume implicitly the "shape" of the domain. E.g. in RL for games we see a bunch of outlier behavior for policies just gaming the signal.
There seems to be 2 types
- Specification failure: signal is bad-ish, a completely broken behavior --> local optimal points achieved for policies that phenomenologically do not represent what was expected/desired to cover --> signaling an improvable reward signal definition
- Domain constraint failure: signal is still good and optimization is "legitimate", but you are prompted with the question "do I need to constraint my domain of solutions?"
- finding a bug that reduces time to completion of a game in a speedrun setting would be a new acceptable baseline, because there are no rules to finishing the game earlier
- shooting amphetamines on a 100m run would probably minimize time, but other factors will make people consider disallowing such practices.
I view Goodhart's law more as a lesson for why we can never achieve a goal by offering specific incentives if we are measuring success by the outcome of the incentives and not by the achievement of the goal.
This is of course inevitable if the goal cannot be directly measured but is composed of many constantly moving variables such as education or public health.
This doesn't mean we shouldn't bother having such goals, it just means we have to be diligent at pivoting the incentives when it becomes evident that secondary effects are being produced at the expense of the desired effect.
If I hadn't seen it in action countless times, I would belive you. Changelists, line counts, documents made, collaborator counts, teams lead, reference counts in peer reviewed journals...the list goes on.
You are welcome to prove me wrong though. You might even restore some faith in humanity, too!
The Zoological Survey of India would like to know but hasn't figured out a good way to do a full census. If you have any ideas they would love to hear them.
Naja naja has Least Concern conservation status, so there isn't much funding in doing a full count, but there are concerns as encroachment both reduces their livable habitat and puts them into more frequent contact with humans and livestock.
If it were a good metric there wouldn't be a few phone books worth of regulations on what you can do before and during running 100 meters. From banning rocket shoes, to steroids, to robot legs the 100 meter run is a perfect example of a terrible metric both intrinsically as a measure of running speed and extrinsically as a measure of fitness.
What do you mean? People start doping or showing up with creatively designed shoes and you need to layer on a complicated system to decide if that's cheating, but some of the methods are harder to detect and then some people cheat anyway, or you ban steroids or stimulants but allow them if they're by prescription to treat an unrelated medical condition and then people start getting prescriptions under false pretexts in order to get better times. Or worse, someone notices that the competition can't set a good time with a broken leg.
So what is your argument, that it doesn't apply everywhere therefore it applies nowhere?
You're misunderstanding the root cause. Your example works as the the metric is well aligned. I'm sure you can also think of many examples where the metric is not well aligned and maximizing it becomes harmful. How do you think we ended up with clickbait titles? Why was everyone so focused on clicks? Let's think about engagement metrics. Is that what we really want to measure? Do we have no preference over users being happy vs users being angry or sad? Or are those things much harder to measure, if not impossible to, and thus we focus on our proxies instead? So what happens when someone doesn't realize it is a proxy and becomes hyper fixated on it? What happens if someone does realize it is a proxy but is rewarded via the metric so they don't really care?
Your example works in the simple case, but a lot of things look trivial when you only approach them from a first order approximation. You left out all the hard stuff. It's kinda like...
Edit: Looks like some people are bringing up metric limits that I couldn't come up with. Thanks!
> So what is your argument, that it doesn't apply everywhere therefore it applies nowhere?
I never said that. Someone said the law collapses, someone asked for a link, I gave an example to prove it does break down in some cases at least, but many cases once you think more about it. I never said all cases.
If it works sometimes and not others, it's not a law. It's just an observation of something that can happen or not.
You're right. My bad. I inferred that through the context of the conversation.
> If it works sometimes and not others, it's not a law.
I think you are misreading and that is likely what lead to the aforementioned misunderstanding. You're right that it isn't a scientific law, but the term "law" gets thrown around a lot in a more colloquial manner. Unfortunately words are overloaded and have multiple meanings. We do the same thing to "hypothesis", "paradox", and lots of other things. I hope this clarifies the context. (even many of the physics laws aren't as strong as you might think)
But there are many "laws" used in the same form. They're eponymous laws[0], not scientific ones. Read "adage". You'll also find that word used in the opening sentence on the Wiki article I linked as well as most (if not all) of them in [0]
I disagree with all of those examples, they are misunderstanding what it means for the metric to break down in the context of the law, but alas. "If you run a different race" lol.
could you explain what you think the difference is?
a metric is chosen, people start to game the system by doing things that make the metric improve but the original intent is lost. increasingly specific rules/laws have to be made up to make the metric appear to work, but it becomes a lost cause as more and more creative ways are found to work around the rules.
Exactly, that's the definition. It doesn't apply to timing a 100m race. There's many such situations that are simple enough and with perfect information available where this doesn’t break down and a metric is just a metric and it works great.
Which is not to the detriment of the observation being true in other contexts, all I did was provide a counter example. But the example requires the metric AND the context.
Do you have an example that doesn't involve an objective metric? Of course objective metrics won't turn bad. They're more measurements than metrics, really.
I'd like to push back on this a little, because I think it's important to understanding why Goodhart's Law shows up so frequently.
*There are no /objective/ metrics*, only proxies.
You can't measure a meter directly, you have to use a proxy like a tape measure. Similarly you can't measure time directly, you have to use a stop watch. In a normal conversation I wouldn't be nitpicking like this because those proxies are so well aligned with our intended measures and the lack of precision is generally inconsequential. But once you start measuring anything with precision you cannot ignore the fact that you're limited to proxies.
The difference of when we get more abstract in our goals is not too dissimilar. Our measuring tools are just really imprecise. So we have to take great care to understand the meaning of our metrics and their limits, just like we would if we were doing high precision measurements with something more "mundane" like distance.
I think this is something most people don't have to contend with because frankly, very few people do high precision work. And unfortunately we often use algorithms as black boxes. But the more complex a subject is the more important an expert is. It looks like they are just throwing data into a black box and reading the answer, but that's just a naive interpretation.
Sure, if you get a ruler from the store it might be off by a fraction of a percent in a way that usually doesn't matter and occasionally does, but even if you could measure distance exactly that doesn't get you out of it.
Because what Goodhart's law is really about is bureaucratic cleavage. People care about lots of diverging and overlapping things, but bureaucratic rules don't. As soon as you make something a target, you've created the incentive to make that number go up at the expense of all the other things you're not targeting but still care about.
You can take something which is clearly what you actually want. Suppose you're commissioning a spaceship to take you to Alpha Centauri and then it's important that it go fast because otherwise it'll take too long. We don't even need to get into exactly how fast it needs to go or how to measure a meter or anything like that, we can just say that going fast is a target. And it's a valid target; it actually needs to do that.
Which leaves you already in trouble. If your organization solicits bids for the spaceship and that's the only target, you better not accept one before you notice that you also need things like "has the ability to carry occupants" and "doesn't kill the occupants" and "doesn't cost 999 trillion dollars" or else those are all on the chopping block in the interest of going fast.
So you add those things as targets too and then people come up with new and fascinating ways to meet them by sacrificing other things you wanted but didn't require.
What's really happening here is that if you set targets and then require someone else to meet them, they will meet the targets in ways that you will not like. It's the principal-agent problem. The only real way out of it is for principals to be their own agents, which is exactly the thing a bureaucracy isn't.
But there is no way to know who is truly the 'best'. The people who position and market themselves to be viewed as the best are the only ones who even have a chance to be viewed as such. So if you're a great researcher but don't project yourself that way, no one will ever know you're a great researcher (except for the other great researchers who aren't really invested in communicating how great you are). The system seems to incentivize people to not only optimize for their output but also their image. This isn't a bad thing per se, but is sort of antithetical to the whole shoulder of giants ethos of science.
The problem is that the best research is not a competitive process but a collaborative one. Positioning research output as a race or a competition is already problematic.
Yeah I think this is a general principle. Just look at the quality of US presidents over time, or generations of top physicists. I guess it’s just a numbers game: the number of genuinely interested people is relatively constant while the number of gamers grows with the compensation and perceived status of the activity. So when compensation and perceived status skyrockets the ratio between those numbers changes drastically.
I think the number of generally interested people goes up. Maybe the percent stays the same? But honestly, I think we kill passion for a lot of people. To be cliche, how many people lose the curiosity of a child? I think the cliche exists for a reason. It seems the capacity is in all of us and even once existed.
I have seen absolutely incredible, best in the world type engineers, much smarter than myself, get fired from my FAANG because of the performance games.
I persist because I'm fantastic at politics while being good enough to do my job. Feels weird man.
> Labs used to hire researchers and give them a lot of free reign.
I can't think of it ever really paying off. Bell Labs is the best example. Amazing research that was unrelated to the core business off the parent company. Microsoft Research is another great one. Lots of interesting research that .. got MS some nerd points? But has materialized into very very few actual products and revenue streams. Moving AI research doesn't help Meta build any motes or revenue streams. It just progresses our collective knowledge.
On the "human progress" scale it's fantastic to put lots of smart people in a room and let them do their thing. But from a business perspective it seems to almost never pay off. Waiting on the irrational charity of businesses executive is probably not the best way to structure thing.
I'd tell them to go become academics.. but all the academics I know are just busy herding their students and attending meetings
Perhaps these companies just end up with so much money that they can't possibly find ways to spend all of it rationally for purely product driven work and just end up funding projects with no clear business case.
Indeed. And it feels like there is this untold in-between where if you belong to an unknown applied AI team, you don’t have to deal with academia’s yak shaving, you don’t have to deal with Meta’s politics and you end up single handedly inventing TRMs.
Also it is what big tech was doing until LLMs hit the scene
So I'm not sure what you mean by it never paying off. We were doing it right up till one of those things seemed to pay off and then hyper focused on it. I actually think this is a terrible thing we frequently do in tech. We find promise in a piece of tech, hyper focus on it. Specifically, hyper focus on how to monetizing it which ends up stunting the technology because it hasn't had time to mature and we're trying to monetize the alpha product instead of trying to get that thing to beta.
> But from a business perspective it seems to almost never pay off.
So this is actually what I'm trying to argue. It actually does pay off. It has paid off. Seriously, look again at Silicon Valley and how we got to where we are today. And look at how things changed in the last decade...
Why is it that we like off the wall thinkers? That programmers used to be known as a bunch of nerds and weirdos. How many companies were started out of garages (Apple)? How many started as open source projects (Android)? Why did Google start giving work lifestyle perks and 20% time?
So I don't know what you're talking about. It has frequently paid off. Does it always pay off? Of course not! It frequently fails! But that is pretty true for everything. Maybe the company stocks are doing great[0], but let's be honest, the products are not. Look at the last 20 years and compare it to the 20 years before that. The last 20 years has been much slower. Now maybe it is a coincidence, but the biggest innovation in the last 20 years has been in AI and from 2012 to 2021 there were a lot of nice free reign AI research jobs at these big tech companies where researchers got paid well, had a lot of autonomy in research, and had a lot of resources at their disposal. It really might be a coincidence, but a number of times things like this have happened in history and they tend to be fairly productive. So idk, you be the judge. Hard to conclude that this is definitely what creates success, but I find it hard to rule this out.
> I'd tell them to go become academics.. but all the academics I know are just busy herding their students and attending meetings
For just 10% of those money a $100M mathematician can hire 10 $1M mathematicians or a whole math dept in some European university to do the work and the thinking for them and thus beat any pressure while resting and vesting on the remaining 90%.
I thought Alex Wang was a very curious choice. There are so many foundational AI labs with interesting CEOs... I get that Wang is remarkable in his own right, but he basically just built MTurk and timed the bubble.
A lot of people also don't know that many of the well known papers are just variations on small time papers with a fuck ton more compute thrown at the problem. Probably the strongest feature that correlates to successful researcher is compute. Many have taken this to claim that the GPU poor can't contribute but that ignores so many other valid explanations... and we wonder why innovation has slowed... It's also weird because if compute was all you need then there's a much cheaper option than Zuck paid. But he's paying for fame.
> A lot of people also don't know that many of the well known papers are just variations on small time papers with a fuck ton more compute thrown at the problem.
I worked for a small research heavy AI startup for a bit and it was heart breaking how many people I would interact with in that general space with research they worked hard and passionately on only to have been beaten to the punch by a famous lab that could rush the paper out quicker and at a larger scale.
There were also more than a few instances of high-probability plagiarism. My team had a paper that had been existing for years basically re-written without citation by a major lab. After some complaining they added a footnote. But it doesn't really matter because no big lab is going to have to defend themselves publicly against some small startup, and their job at the big labs is to churn out papers.
> only to have been beaten to the punch by a famous lab that could rush the paper out quicker and at a larger scale.
This added at least a year to my PhD... Reviewers kept rejecting my works saying "add more datasets" and such comments. That's nice and all, but on the few datasets I did use I beat out top labs and used a tenth of the compute. I'd love to add more datasets but even though I only used a tenth of the compute I blew my entire compute budget. Guess state of the art results, a smaller model, higher throughput, and 3rd party validation were not enough (use an unpopular model architecture).
I always felt like my works were being evaluated as engineering products, not as research.
> a few instances of high-probability plagiarism
I was reviewing a work once and I actually couldn't tell if the researchers knew that they ripped me off or not. They compared to my method, citing, and showing figures using it. But then dropped the performance metrics from the table. So I asked. I got them in return and saw that there was no difference... So I dove in and worked out that they were just doing 99% my method with additional complexity (computational overhead). I was pretty upset.
I was also upset because otherwise the paper was good. The results were nice and they even tested our work in a domain we hadn't. Were they just upfront I would have gladly accepted the work. Though I'm pretty confident the other reviewers wouldn't have due to "lack of novelty."
It's a really weird system that we've constructed. We're our own worst enemies.
> their job at the big labs is to churn out papers.
I'd modify this slightly. Their job is to get citations. Churning out papers really helps with that, but so does all the tweeting and evangelizing of their works. It's an unfortunate truth that as researchers we have to sell our works, and not just by the scientific merit that they hold. People have to read them after all. But we should also note that it is easier for some groups to get noticed more than others. Prestige doesn't make a paper good, but it sure acts as a multiplying factor for all the metrics we use for determining if it is good.
I learnt the hard way that communications/image/signal processing research basically doesn’t care about Computer Architecture at the nuts and bolts level of compiler optimization and implementation.
When they encounter a problem whose normal solution requires excessive amounts of computation, they reduce complexity algorithmically using mathematical techniques, and quantify the effects.
They don’t quibble about a 10x speed up, they reduce the “big O()” complexity. They could care less whether it was implemented in interpreted Python or hand-optimized assembly code.
On one hand, I know there’s a lot of talent in AI today. But throwing hardware at the problem is the dumbest way forward.
WiFI adapters would be wheeled luggage if we had the same mentality during their development.
At some point it becomes difficult to improve the O() complexity. How do you do better that the O(n-squared) of the Transformer, with acceptable tradeoffs? Many big brains in all the big labs are very aware of the importance of algorithmic advances. There is no low hanging fruit, but they're doing their best.
Then in parallel to that looking at compiler optimizations, and other higher-level algorithmic innovations such as Flash Attention (a classic at this point) which had a drastic impact on performance due to cache awareness, without changing the O() complexity.
> They don’t quibble about a 10x speed up, they reduce the “big O()” complexity. They could care less whether it was implemented in interpreted Python or hand-optimized assembly code.
I can at least say that's not all of us. But you're probably right that this is dominating. I find it so weird since everyone stresses empirics yet also seems to not care about them. It took me my entire PhD to figure out what was really going on. I've written too many long winded rants on this site though
You make it sound like reducing the big O complexity is a dumb thing to do in research, but this is really the only way to make lasting progress in computer science. Computer architectures become obsolete as hardware changes, but any theoretical advances in the problem space will remain true forever.
The reportings at the time said that he was Mark’s 5th choice or similar. It is fairly clear he would prefer Ilya, Murati, Mark Chen, and perhaps others, but they said no, and Alex Wang was the first one to say yes.
One thing I don't get about the ever-reoccuring RAG discussions and hype men proclaiming "Rag is dead", is that people seem to be talking about wholly different things?
My mental model is that what is called RAG can either be:
- a predefined document store / document chunk store where every chunk gets a a vector embedding, and a lookup decides what gets pulled into context as to not have to pull whole classes of document, filling it up
- the web search like features in LLM chat interfaces, where they do keyword search, and pull relevant documents into context, but somehow only ephemerally, with the full documents not taking up context in the future of the thread (unsure about this, did I understand it right?) .
with the new models with million + tokens of context windows, some where arguing that we can just throw whole books into the context non-ephemerally, but doesnt that significantly reduce the diversity of possible sources we can include at once if we hard commit to everything staying in context forever? I guess it might help with consistency? But is the mechanism with which we decide what to keep in context not still some kind of RAG, just with larger chunks of whole documents instead of only parts?
I'd be extatic if someone who really knows their stuff could clear this up for me
Technically, RAG is anything that augments generation with external search. However, it often has a narrower meaning: "uses a vector DB."
Throwing everything into one large context window is often impractical - it takes much more time to process, and many models struggle to find information accurately if too much is going on in the context window ("lost in the middle").
The "classic" RAG still has its place when you want low latency (or you're limited by VRAM) and the results are already good enough.
In both cases for "Question Answering" it's about similarity search but there are two main orthogonal differences between RAG and Non-RAG :
-Knowing the question at the time of index building
-Higher order features : the ability to compare fetched documents with one another and refine the question
Non-RAG, aka multi-layer (non-causal) transformer with infinite context, is the more generic version, fully differentiable meaning you can use machine learning to learn how to Non-RAG better. Each layer of the transformer can use the previous layer to reason and refine the similarity search. (A causal transformer know the question at the time when it is feed the question, and can choose to focus it's attention on different part of the previously computed features of the provided documents but may benefit from having some reflection token, or better : be given the question before being presented the documents (provided you've trained it to answer it like that).)
RAG is an approximation of the generic case to make it faster and cheaper. Usually it breaks end-to-end differentiability by using external tools, so this mean that if you want to use machine learning to learn how to RAG better you will need to use some variant of Reinforcement Learning which is slower to learn things. RAG usually don't know the question at the time of index building, and documents are treated independently of each other, so no (automatic) higher order features (embeddings are fixed).
A third usual approximation, is to feed the output of RAG into Non-RAG, to hopefully get the best of both world. You can learn the Non-RAG given RAG with machine learning (if you train it with some conversations where it used RAG), but the RAG part won't improve by itself.
Non-RAG need to learn so it needs a big training dataset, but fortunately it can pick-up question answer pair in an unsupervised fashion when you feed it the whole web, and you only need a small instruction training and preference optimization dataset to shape it to your need. If performance isn't what you expect in a specific case, you can provide more specific examples and retrain the model until it gets it and you get better performance for the case you were interested in. You can improve the best case but it's hard to improve the worst case.
RAG has more control on what you feed it but content should be in a more structured way. You can prevent worst cases more easily but it's hard to improve good case.
We can't throw in infinite things in the context though.
My impression is that GPT-5 gets confused, not quite right away, but after a couple of pages it has no idea. It doesn't take pages on pages before it forgets things.
I’m currently experimenting with prompts of ~300k tokens for a certain classification task and I think I might be able to make it work. GPT5 chokes but Gemini 2.5 Pro is showing promise. Jury’s still out and I might change my tune in a couple of weeks.
It should also be said, that what I say here is focused on things where these models have problems.
For example, I consider the model confused when it starts outputting stereotyped or cliche responses, and I intentionally go at problems that I know that the models have problems with (I already know they can program and do some maths, but I want to see what they can't do). But if you're using them for things they're made for, and which aren't confusing, such as people arguing with each other, then you are probably likely to succeed.
Prompts with lots of examples are reasonable and I know they can get very long.
> But RAG is a very real world, practical topic for something as significant as a new lab’s first paper.
I would expect exactly the opposite - that a new lab would put out a few random papers that happen to be in areas their researchers were interested in and already working on, and once people had been working together a while and developed some synergy they would maybe come out with something really groundbreaking.
do people really view a "first paper" as something deeply significant and weighty? because that just seems like a good way to get bogged down in trying to second guess whether any given paper was good enough to be your all-important debut!
As an academic I would expect the same as you, and no, to my knowledge "first paper" is meaningless, at least in academia. Most people's first paper is some small contribution to what their PhD supervisor is doing at the time, where the student tries their best at writing but it ends up so heavily edited that probably 90% of the final text comes from the supervisor :) So typically first papers don't define or represent a researcher. When you start you just don't have the experience to have a great idea and carry it through to a good paper.
Of course here we are talking about a lab, not an individual person, but still I haven't heard of first papers being considered special in any way, even for labs.
I'm curious whether this is work that was specifically begun under the "superintelligence" umbrella, or if it's just that the people who were working on it had been shifted to the Superintelligence team by the time they wrote the paper. I would guess the former?
I am not surprised because the culture at meta is not at all, even in the slightest, to focus on science for the sake of it. It’s actively actively purged out of you. The focus is on metrics and how the bottom line is impacted. So this is in line with that
"People are using our service more!" turns out to be a horrible metric when they outright lie to you (x has sent you a message! - when no message exists)
I couldn't immediately see in their graphs/tables any comparison against simple lexical/statistical based context compression, such as candidate selection of chunks using TF-IDF, word overlap etc. For most of us in the industry we need to find these quick wins that give us equivalent performance to sending huge amount of information to the LLM, while compressing by 10x.
If you can collapse "retrieve this complex chunk when it is needed" into a single token, what else can you put into a token?
"Send this through the math coprocessor." "Validate against the checklist." "Call out to an agent for X." "Recheck against input stream Y." And so on.
Retrieval augmentation is only one of many uses for this. If this winds up with better integration with agents, it is very possible that the whole is more than the sum of its parts.
It’s unlikely that the existing LLM architecture will evolve into anything that resembles superintelligence any more than it does already.
Which means that modifications to the architecture, and combining it with other components and approaches, are the next likely step. This paper fits that.
I can start brushing my teeth 30 times faster but it won't change my life. This is nice for RAG but it's a very localized improvement. And 30× sounds big but is just an order of magnitude improvement also.
So this looks essentially like continuous prompting (see prefix tuning) with RL-driven selection of what to present as tokens and what as continuous inputs (embeddings).
This was inevitable. You can't keep training LLMs and expect that's the answer to the evolution of AI. Yes it'll happen and we'll keep creating new more refined and bigger models but it's like DNA or something like the cortex of the brain. After that you need these systems that essentially "live" for years digesting information and develop a more refined way to process, store and retrieve the information. Compression of RAG was also inevitable. It's like the btree index of a database. The thing is, we're probably one or two iterations away from being good enough on the RAG pipeline and then we'll need to focus more on the other pieces of sensory input that need to be connected and processed at higher throughput. Right now it's not fast or efficient enough. This is where the likes of Google will shine. They are probably two decades ahead of everyone on internal technology and there is some team with the breakthrough but it hasn't seen the light of day yet. What's coming out of DeepMind is really a forced effort in productization and publication of work in a consumable format but internally they are likely way ahead. I don't have as much faith in Meta's efforts despite seeing things like this. Quite frankly those people, the ones doing the work should move to more honourable companies. Not feed crack addiction in the form of Meta's universe.
Did a "superintelligence" lab publish a superintelligence related paper with no results for intelligence? What measured improvements did this proposal make in their LLM's intelligence?
Interesting. All developers I know who tinkered around with embeddings and vector similarity scoring were instantly hooked. The efficiency of computing the embeddings once and then reusing as many times as needed, comparing the vectors with a cheap <30-line function is extremely appealing. Not to mention the indexing capabilities to make it work at scale.
IMO vector embedding is the most important innovation in computing of the last decade. There's something magical about it. These people deserve some kind of prize. The idea that you can reduce almost any intricate concept including whole paragraphs to a fixed-size vector which encapsulates its meaning and proximity to other concepts across a large number of dimensions is pure genius.
Vector embedding is not an invention of the last decade. Featurization in ML goes back to the 60s - even deep learning-based featurization is decades old at a minimum. Like everything else in ML this became much more useful with data and compute scale
Yes but it doesn't generalize very well. Even on simple features like gender. If you go look at embeddings you'll find that man and woman are neighbors, just as king and queen are[0]. This is a better explanation for the result as you're just taking very small steps in the latent space.
Here, play around[1]
mother - parent + man = woman
father - parent + woman = man
father - parent + man = woman
mother - parent + woman = man
woman - human + man = girl
Or some that should be trivial
woman - man + man = girl
man - man + man = woman
woman - woman + woman = man
Working in very high dimensions is funky stuff. Embedding high dimensions into low dimensions results in even funkier stuff
Vector embeddings are slightly interesting because they come pre-trained with large amounts of data.
But similar ways to reduce huge numbers of dimensions to a much smaller set of "interesting" dimensions have been known for a long time.
Examples include principal component analysis/single value decomposition, which was the first big breakthrough in face recognition (in the early 90s), and also used in latent semantic indexing, the Netflix prize, and a large pile of other things. And the underlying technique was invented in 1901.
Dimensionality reduction is cool, and vector embedding is definitely an interesting way to do it (at significant computational cost).
Vector embeddings are so overhyped. They're decent as a secondary signal, but they're expensive to compute and fragile. BM25 based solutions are more robust and WAY lower latency, at the cost of some accuracy loss vs hybrid solutions. You can get the majority of the lift from hybrid solutions with ingest time semantic expansion/reverse hyde type input annotation with a sparse embedding BM25 at a fraction of the computational cost.
> the core insight here is actually: if embeddings are generated by layers within the LLM, it makes no sense to convert them back to natural language, just for another LLM to compress those tokens back to embeddings.
Doesn't this tie the two layers together in a way that they can't evolve separately?
Working in big tech it's pretty wild to see how integral AI has become to our work internally, vs the public perception of it. People are NOT prepared.
Personal experience here in a FAANG, there has been a considerable increase in:
1. Teams exploring how to leverage LLMs for coding.
2. Teams/orgs that already standardized some of the processes to work with LLMs (MCP servers, standardized the creation of the agents.md files, etc)
3. Teams actively using it for coding new features, documenting code, increasing test coverage, using it for code reviews etc.
Again, personal, experience, but in my team ~40-50% of the PRs are generated by Codex.
“Teams exploring how to leverage [AI]s for [anything]” is true for about a decade now in every large multinational companies at every level. It’s not new at all. AI is the driving buzzword for a while now, even well before ChatGPT. I’ve encountered many people who just wanted the stamp that they use AI, no matter how, because my team was one of the main entry point to achieve this at that specific company. But before ChatGPT and co, you had to work for it a lot, so most of them failed miserably, or immediately backtracked when they realized this.
There are places that offer Copilot to any team that wants it, and then behind the scenes they informed their managers that if the team (1+ persons) adopts it they will have to shed 10%+ human capacity (lose a person, move a person, fire a person) in the upcoming quarters next year.
It is true sometimes, but other times it saves hours. We're all still in the learning stage of how best to use these new tools, and their capabilities are growing constantly.
AI may slow coding a bit but dramatically reduces cognitive load.
The real value of AI isn't in helping coding. It's in having a human-like intelligence to automate processes. I can't get into details but my team is doing things that I couldn't dream of three years ago.
It does dramatically reduce cognitive load. I think that part is understated and lost to the headline of how it writes two thousand lines of code in 30 seconds.
Sure but I'm the real world do you think businesses are going to deploy piles of code into production generated this way? No, non technical people will continue to whip up MS PowerApps. AI generated code has no value to many businesses.
The value of AI is not in generating code. That's just a "nice-to-have."
The value of AI is in having a scalable, human-like decision maker that you can plug into anything, anywhere. This has unlocked countless use cases for my team, that we could scarcely imagine a few years ago.
"Human-like decision maker" except it's just as if not more unpredictable than a human, has no understanding of what it's actually outputting or the impact of it, and it isn't concerned with losing their job or facing legal repercussions for their actions.
There are plenty of ways to manage those drawbacks, and a mind-boggling number of use cases where it's "good enough" already.
But it's not my job to convince you, my lived experience working with the tech is enough to convince me, and that's all I care about, to be honest. Everyone else will get there sooner or later.
I find it absurd that, compared to the past, large companies now have more abundant stock prices and cash than ever before, yet nearly every AI Lab in these companies is facing greater pressure than ever, being asked to generate short-term profits. In the midst of AI's unprecedented boom, the research environment and atmosphere in the industry seem to have worsened compared to the past.
• MSI’s first paper, REFRAG, is about a new way to do RAG.
• This slightly modified LLM converts most retrieved document chunks into compact, LLM-aligned chunk embeddings that the LLM can consume directly.
• A lightweight policy (trained with RL) decides which chunk embeddings should be expanded back into full tokens under a budget; the LLM runs normally on this mixed input.
• The net effect is far less KV cache and attention cost, much faster first-byte latency and higher throughput, while preserving perplexity and task accuracy in benchmarks.
I wish more long posts followed this model of a scientific paper.
Though it is notable that contrary to many (on HN and Twitter) that Meta would stop publishing papers and be like other AI labs (e.g. OpenAI). They're continued their rapid pace of releasing papers AND open source models.
Also, that wasn't based on purely hearsay, Zuck explicitly said:
> We believe the benefits of superintelligence should be shared with the world as broadly as possible. That said, superintelligence will raise novel safety concerns. We'll need to be rigorous about mitigating these risks and careful about what we choose to open source. Still, we believe that building a free society requires that we aim to empower people as much as possible. [0]
[0]: https://www.meta.com/superintelligence/
[1] https://ethz.ch/en/news-and-events/eth-news/news/2025/07/a-l...
[2] https://news.ycombinator.com/item?id=44535637
MSL is not only those few high profile hires.
A bit of this is true at every major lab. There's tons of untapped potential. But these organizations are very risk adverse. I mean why not continue with the strategy that got us to the point we're at in the first place. Labs used to hire researchers and give them a lot of free reign. But those times ended and AI progress also slowed down. Maybe if you want to get ahead you gotta stop thinking like everyone else
Well meta... you can "hold me hostage" for a lot cheaper than those guys. I'm sure this is true for hundreds of passionate ML researchers. I'd take a huge pay cut to have autonomy and resources. I know for a fact there's many working at Mets right now that would do the same. Do maybe if you're going to throw money at the problem, diversify a bit and look back at what made SV what it is today and what made AI take leaps forward
The other day I was spending some time with a researcher from Deep Mind and I was surprised to find that while they were sharp and curious to an extent, nearly every ounce of energy they expended on research was strategic. They didn't write about research they were fascinated by, they wrote and researched on topics they strategically felt had the highest probability getting into a major conference in a short period of time to earn them a promotion. While I was a bit disappointed, I certainly didn't judge them because they are just playing the game. This person probably earns more than many rooms of smart, passionate people I've been in, and that money isn't for smarts alone; it's for appealing to the interests of people with the money.
You can see this very clearly by comparing the work being done in the LLM space to that being done in the Image/Video diffusion model space. There's much more money in LLMs right now, and the field is flooded with papers on any random topic. If you dive in, most of them are not reproducible or make very questionable conclusions based on the data they present, but that's not of very much concern so long as the paper can be added to a CV.
In the stable diffusion world it's mostly people driven by personal interest (usually very non-commericial personal interests) and you see tons of innovation in that field but almost no papers. In fact, if you really want to understand a lot of the most novel work coming out of the image generation world you often need to dig into PRs made by an anonymous users with anime themed profile pic.
The bummer of course is that there are very hard limits on what any researcher can do with a home GPU training setup. It does lead to creative solutions to problems, but I can't help but wonder what the world would look like if more of these people had even a fraction of the resources available exclusively to people playing the game.
The problem is once people's livelihoods depend on their research output rather than the research process, the whole research process becomes steadily distorted to optimise for being able to reliably produce outputs.
Anyone who has invested a great deal of time and effort into solving a hard problem knows that the 'eureka' moment is not really something that you can force. So people end up spending less time working on problems that would contribute to 'breakthroughs' and more time working on problems that will publish.
Please do judge them for being parasitical. They might seem successful by certain measures, like the amount of money they make, but I for one simply dislike it when people only think about themselves.
As a society, we should be more cautious about narcissism and similar behaviors. Also, in the long run, this kind of behaviour makes them an annoying person at parties.
You dislike them because they don’t benefit you indirectly by benefiting society at large.
The incentive structure is wrong, incentivizing things that benefit society would be the solution not judging those that exist in the current system by pretending altruism is somehow not part of the same game.
As for whether that expectation is "selfish" on my part, I think that question has been debated for centuries in ethics, and I'm quite comfortable landing on the side that says not all disapproval is self-interest. In my own case, I'm not benefiting much either :)
You can call this difference whatever you want, don't pretend that they are morally or effectively equivalent.
The key word there is only. Nothing in the post you suggested only. You have one vignette about one facet of this guy’s life.
I really dislike the resurgence in Puritanism.
Please read my sibling comment where I expand a bit on what I meant to say.
https://en.wikipedia.org/wiki/Goodhart%27s_law
There seems to be 2 types
- Specification failure: signal is bad-ish, a completely broken behavior --> local optimal points achieved for policies that phenomenologically do not represent what was expected/desired to cover --> signaling an improvable reward signal definition
- Domain constraint failure: signal is still good and optimization is "legitimate", but you are prompted with the question "do I need to constraint my domain of solutions?"
This is of course inevitable if the goal cannot be directly measured but is composed of many constantly moving variables such as education or public health.
This doesn't mean we shouldn't bother having such goals, it just means we have to be diligent at pivoting the incentives when it becomes evident that secondary effects are being produced at the expense of the desired effect.
You are welcome to prove me wrong though. You might even restore some faith in humanity, too!
Naja naja has Least Concern conservation status, so there isn't much funding in doing a full count, but there are concerns as encroachment both reduces their livable habitat and puts them into more frequent contact with humans and livestock.
https://en.wikipedia.org/wiki/Perverse_incentive
Context: track athlete
Does it cease to be a good metric? No. After this you can likely come up with many examples of target metrics which never turn bad.
> Context: track athlete
> Does it cease to be a good metric? No.
What do you mean? People start doping or showing up with creatively designed shoes and you need to layer on a complicated system to decide if that's cheating, but some of the methods are harder to detect and then some people cheat anyway, or you ban steroids or stimulants but allow them if they're by prescription to treat an unrelated medical condition and then people start getting prescriptions under false pretexts in order to get better times. Or worse, someone notices that the competition can't set a good time with a broken leg.
You're misunderstanding the root cause. Your example works as the the metric is well aligned. I'm sure you can also think of many examples where the metric is not well aligned and maximizing it becomes harmful. How do you think we ended up with clickbait titles? Why was everyone so focused on clicks? Let's think about engagement metrics. Is that what we really want to measure? Do we have no preference over users being happy vs users being angry or sad? Or are those things much harder to measure, if not impossible to, and thus we focus on our proxies instead? So what happens when someone doesn't realize it is a proxy and becomes hyper fixated on it? What happens if someone does realize it is a proxy but is rewarded via the metric so they don't really care?
Your example works in the simple case, but a lot of things look trivial when you only approach them from a first order approximation. You left out all the hard stuff. It's kinda like...
Edit: Looks like some people are bringing up metric limits that I couldn't come up with. Thanks!
I never said that. Someone said the law collapses, someone asked for a link, I gave an example to prove it does break down in some cases at least, but many cases once you think more about it. I never said all cases.
If it works sometimes and not others, it's not a law. It's just an observation of something that can happen or not.
But there are many "laws" used in the same form. They're eponymous laws[0], not scientific ones. Read "adage". You'll also find that word used in the opening sentence on the Wiki article I linked as well as most (if not all) of them in [0]
[0] https://en.wikipedia.org/wiki/List_of_eponymous_laws
a metric is chosen, people start to game the system by doing things that make the metric improve but the original intent is lost. increasingly specific rules/laws have to be made up to make the metric appear to work, but it becomes a lost cause as more and more creative ways are found to work around the rules.
Which is not to the detriment of the observation being true in other contexts, all I did was provide a counter example. But the example requires the metric AND the context.
Yes if you run anything other than the 100m
*There are no /objective/ metrics*, only proxies.
You can't measure a meter directly, you have to use a proxy like a tape measure. Similarly you can't measure time directly, you have to use a stop watch. In a normal conversation I wouldn't be nitpicking like this because those proxies are so well aligned with our intended measures and the lack of precision is generally inconsequential. But once you start measuring anything with precision you cannot ignore the fact that you're limited to proxies.
The difference of when we get more abstract in our goals is not too dissimilar. Our measuring tools are just really imprecise. So we have to take great care to understand the meaning of our metrics and their limits, just like we would if we were doing high precision measurements with something more "mundane" like distance.
I think this is something most people don't have to contend with because frankly, very few people do high precision work. And unfortunately we often use algorithms as black boxes. But the more complex a subject is the more important an expert is. It looks like they are just throwing data into a black box and reading the answer, but that's just a naive interpretation.
Sure, if you get a ruler from the store it might be off by a fraction of a percent in a way that usually doesn't matter and occasionally does, but even if you could measure distance exactly that doesn't get you out of it.
Because what Goodhart's law is really about is bureaucratic cleavage. People care about lots of diverging and overlapping things, but bureaucratic rules don't. As soon as you make something a target, you've created the incentive to make that number go up at the expense of all the other things you're not targeting but still care about.
You can take something which is clearly what you actually want. Suppose you're commissioning a spaceship to take you to Alpha Centauri and then it's important that it go fast because otherwise it'll take too long. We don't even need to get into exactly how fast it needs to go or how to measure a meter or anything like that, we can just say that going fast is a target. And it's a valid target; it actually needs to do that.
Which leaves you already in trouble. If your organization solicits bids for the spaceship and that's the only target, you better not accept one before you notice that you also need things like "has the ability to carry occupants" and "doesn't kill the occupants" and "doesn't cost 999 trillion dollars" or else those are all on the chopping block in the interest of going fast.
So you add those things as targets too and then people come up with new and fascinating ways to meet them by sacrificing other things you wanted but didn't require.
What's really happening here is that if you set targets and then require someone else to meet them, they will meet the targets in ways that you will not like. It's the principal-agent problem. The only real way out of it is for principals to be their own agents, which is exactly the thing a bureaucracy isn't.
I persist because I'm fantastic at politics while being good enough to do my job. Feels weird man.
I can't think of it ever really paying off. Bell Labs is the best example. Amazing research that was unrelated to the core business off the parent company. Microsoft Research is another great one. Lots of interesting research that .. got MS some nerd points? But has materialized into very very few actual products and revenue streams. Moving AI research doesn't help Meta build any motes or revenue streams. It just progresses our collective knowledge.
On the "human progress" scale it's fantastic to put lots of smart people in a room and let them do their thing. But from a business perspective it seems to almost never pay off. Waiting on the irrational charity of businesses executive is probably not the best way to structure thing.
I'd tell them to go become academics.. but all the academics I know are just busy herding their students and attending meetings
Also it is what big tech was doing until LLMs hit the scene
So I'm not sure what you mean by it never paying off. We were doing it right up till one of those things seemed to pay off and then hyper focused on it. I actually think this is a terrible thing we frequently do in tech. We find promise in a piece of tech, hyper focus on it. Specifically, hyper focus on how to monetizing it which ends up stunting the technology because it hasn't had time to mature and we're trying to monetize the alpha product instead of trying to get that thing to beta.
So this is actually what I'm trying to argue. It actually does pay off. It has paid off. Seriously, look again at Silicon Valley and how we got to where we are today. And look at how things changed in the last decade...Why is it that we like off the wall thinkers? That programmers used to be known as a bunch of nerds and weirdos. How many companies were started out of garages (Apple)? How many started as open source projects (Android)? Why did Google start giving work lifestyle perks and 20% time?
So I don't know what you're talking about. It has frequently paid off. Does it always pay off? Of course not! It frequently fails! But that is pretty true for everything. Maybe the company stocks are doing great[0], but let's be honest, the products are not. Look at the last 20 years and compare it to the 20 years before that. The last 20 years has been much slower. Now maybe it is a coincidence, but the biggest innovation in the last 20 years has been in AI and from 2012 to 2021 there were a lot of nice free reign AI research jobs at these big tech companies where researchers got paid well, had a lot of autonomy in research, and had a lot of resources at their disposal. It really might be a coincidence, but a number of times things like this have happened in history and they tend to be fairly productive. So idk, you be the judge. Hard to conclude that this is definitely what creates success, but I find it hard to rule this out.
Same problem, different step of the ladder[0] https://news.ycombinator.com/item?id=45555175
Quite the statement for anybody who follows developments (without excluding xAI).
Doesn't really scream CEO of AGI to me.
I worked for a small research heavy AI startup for a bit and it was heart breaking how many people I would interact with in that general space with research they worked hard and passionately on only to have been beaten to the punch by a famous lab that could rush the paper out quicker and at a larger scale.
There were also more than a few instances of high-probability plagiarism. My team had a paper that had been existing for years basically re-written without citation by a major lab. After some complaining they added a footnote. But it doesn't really matter because no big lab is going to have to defend themselves publicly against some small startup, and their job at the big labs is to churn out papers.
I always felt like my works were being evaluated as engineering products, not as research.
I was reviewing a work once and I actually couldn't tell if the researchers knew that they ripped me off or not. They compared to my method, citing, and showing figures using it. But then dropped the performance metrics from the table. So I asked. I got them in return and saw that there was no difference... So I dove in and worked out that they were just doing 99% my method with additional complexity (computational overhead). I was pretty upset.I was also upset because otherwise the paper was good. The results were nice and they even tested our work in a domain we hadn't. Were they just upfront I would have gladly accepted the work. Though I'm pretty confident the other reviewers wouldn't have due to "lack of novelty."
It's a really weird system that we've constructed. We're our own worst enemies.
I'd modify this slightly. Their job is to get citations. Churning out papers really helps with that, but so does all the tweeting and evangelizing of their works. It's an unfortunate truth that as researchers we have to sell our works, and not just by the scientific merit that they hold. People have to read them after all. But we should also note that it is easier for some groups to get noticed more than others. Prestige doesn't make a paper good, but it sure acts as a multiplying factor for all the metrics we use for determining if it is good.I learnt the hard way that communications/image/signal processing research basically doesn’t care about Computer Architecture at the nuts and bolts level of compiler optimization and implementation.
When they encounter a problem whose normal solution requires excessive amounts of computation, they reduce complexity algorithmically using mathematical techniques, and quantify the effects.
They don’t quibble about a 10x speed up, they reduce the “big O()” complexity. They could care less whether it was implemented in interpreted Python or hand-optimized assembly code.
On one hand, I know there’s a lot of talent in AI today. But throwing hardware at the problem is the dumbest way forward.
WiFI adapters would be wheeled luggage if we had the same mentality during their development.
Then in parallel to that looking at compiler optimizations, and other higher-level algorithmic innovations such as Flash Attention (a classic at this point) which had a drastic impact on performance due to cache awareness, without changing the O() complexity.
The right people to deliver immense progress dont exist right now.
The people are always there, you just need to find them and enable them.
Why do you say that?
Shareholders should be livid if they knew a single thing about what was going on.
This is very true, and more than just in ai.
I think if they weren’t so metric focused they probably wouldn’t have hit so much bad publicity and scandal too.
Well for starters you need a leader who can rally the troops who "think(s) different" - something like a S Jobs.
That person doesnt seem to exist in the industry right now.
- a predefined document store / document chunk store where every chunk gets a a vector embedding, and a lookup decides what gets pulled into context as to not have to pull whole classes of document, filling it up
- the web search like features in LLM chat interfaces, where they do keyword search, and pull relevant documents into context, but somehow only ephemerally, with the full documents not taking up context in the future of the thread (unsure about this, did I understand it right?) .
with the new models with million + tokens of context windows, some where arguing that we can just throw whole books into the context non-ephemerally, but doesnt that significantly reduce the diversity of possible sources we can include at once if we hard commit to everything staying in context forever? I guess it might help with consistency? But is the mechanism with which we decide what to keep in context not still some kind of RAG, just with larger chunks of whole documents instead of only parts?
I'd be extatic if someone who really knows their stuff could clear this up for me
Throwing everything into one large context window is often impractical - it takes much more time to process, and many models struggle to find information accurately if too much is going on in the context window ("lost in the middle").
The "classic" RAG still has its place when you want low latency (or you're limited by VRAM) and the results are already good enough.
In both cases for "Question Answering" it's about similarity search but there are two main orthogonal differences between RAG and Non-RAG :
-Knowing the question at the time of index building
-Higher order features : the ability to compare fetched documents with one another and refine the question
Non-RAG, aka multi-layer (non-causal) transformer with infinite context, is the more generic version, fully differentiable meaning you can use machine learning to learn how to Non-RAG better. Each layer of the transformer can use the previous layer to reason and refine the similarity search. (A causal transformer know the question at the time when it is feed the question, and can choose to focus it's attention on different part of the previously computed features of the provided documents but may benefit from having some reflection token, or better : be given the question before being presented the documents (provided you've trained it to answer it like that).)
RAG is an approximation of the generic case to make it faster and cheaper. Usually it breaks end-to-end differentiability by using external tools, so this mean that if you want to use machine learning to learn how to RAG better you will need to use some variant of Reinforcement Learning which is slower to learn things. RAG usually don't know the question at the time of index building, and documents are treated independently of each other, so no (automatic) higher order features (embeddings are fixed).
A third usual approximation, is to feed the output of RAG into Non-RAG, to hopefully get the best of both world. You can learn the Non-RAG given RAG with machine learning (if you train it with some conversations where it used RAG), but the RAG part won't improve by itself.
Non-RAG need to learn so it needs a big training dataset, but fortunately it can pick-up question answer pair in an unsupervised fashion when you feed it the whole web, and you only need a small instruction training and preference optimization dataset to shape it to your need. If performance isn't what you expect in a specific case, you can provide more specific examples and retrain the model until it gets it and you get better performance for the case you were interested in. You can improve the best case but it's hard to improve the worst case.
RAG has more control on what you feed it but content should be in a more structured way. You can prevent worst cases more easily but it's hard to improve good case.
My impression is that GPT-5 gets confused, not quite right away, but after a couple of pages it has no idea. It doesn't take pages on pages before it forgets things.
For example, I consider the model confused when it starts outputting stereotyped or cliche responses, and I intentionally go at problems that I know that the models have problems with (I already know they can program and do some maths, but I want to see what they can't do). But if you're using them for things they're made for, and which aren't confusing, such as people arguing with each other, then you are probably likely to succeed.
Prompts with lots of examples are reasonable and I know they can get very long.
> But RAG is a very real world, practical topic for something as significant as a new lab’s first paper.
I would expect exactly the opposite - that a new lab would put out a few random papers that happen to be in areas their researchers were interested in and already working on, and once people had been working together a while and developed some synergy they would maybe come out with something really groundbreaking.
do people really view a "first paper" as something deeply significant and weighty? because that just seems like a good way to get bogged down in trying to second guess whether any given paper was good enough to be your all-important debut!
Of course here we are talking about a lab, not an individual person, but still I haven't heard of first papers being considered special in any way, even for labs.
(preferably using representative language from the article)
https://www.youtube.com/watch?v=Ek0tZootK00
In general we need to make it simpler for LLMs to take in different forms of embeddings. At least frameworks that simplify it.
"Send this through the math coprocessor." "Validate against the checklist." "Call out to an agent for X." "Recheck against input stream Y." And so on.
Retrieval augmentation is only one of many uses for this. If this winds up with better integration with agents, it is very possible that the whole is more than the sum of its parts.
Which means that modifications to the architecture, and combining it with other components and approaches, are the next likely step. This paper fits that.
IMO vector embedding is the most important innovation in computing of the last decade. There's something magical about it. These people deserve some kind of prize. The idea that you can reduce almost any intricate concept including whole paragraphs to a fixed-size vector which encapsulates its meaning and proximity to other concepts across a large number of dimensions is pure genius.
The fact that dot product addition can encode the concept of royalty and gender (among all other sorts) is kind of magic to me.
Here, play around[1]
Or some that should be trivial Working in very high dimensions is funky stuff. Embedding high dimensions into low dimensions results in even funkier stuff[0] https://projector.tensorflow.org/
[1] https://www.cs.cmu.edu/~dst/WordEmbeddingDemo/
Having set of "king - male + female = queen" like relations, including more complex phrases to align embeddings.
It seems like terse, lightweight, information dense way to address essence of knowldge.
But similar ways to reduce huge numbers of dimensions to a much smaller set of "interesting" dimensions have been known for a long time.
Examples include principal component analysis/single value decomposition, which was the first big breakthrough in face recognition (in the early 90s), and also used in latent semantic indexing, the Netflix prize, and a large pile of other things. And the underlying technique was invented in 1901.
Dimensionality reduction is cool, and vector embedding is definitely an interesting way to do it (at significant computational cost).
Doesn't this tie the two layers together in a way that they can't evolve separately?
2. Wild claim that the companies that sell LLMs are actually downplaying their capabilities instead of hyping them
Again, personal, experience, but in my team ~40-50% of the PRs are generated by Codex.
Ready for the impending lay off fella?
https://www.infoworld.com/article/4061078/the-productivity-p...
The real value of AI isn't in helping coding. It's in having a human-like intelligence to automate processes. I can't get into details but my team is doing things that I couldn't dream of three years ago.
Non-software devs are actually making functional programs for themselves for the first time ever. The value is crazy.
The value of AI is in having a scalable, human-like decision maker that you can plug into anything, anywhere. This has unlocked countless use cases for my team, that we could scarcely imagine a few years ago.
But it's not my job to convince you, my lived experience working with the tech is enough to convince me, and that's all I care about, to be honest. Everyone else will get there sooner or later.
It means you're reading into it too much and need to be let down, gently, from the hype train.
Which other under pressure labs are you talking about?
TL;DR
• MSI’s first paper, REFRAG, is about a new way to do RAG.
• This slightly modified LLM converts most retrieved document chunks into compact, LLM-aligned chunk embeddings that the LLM can consume directly.
• A lightweight policy (trained with RL) decides which chunk embeddings should be expanded back into full tokens under a budget; the LLM runs normally on this mixed input.
• The net effect is far less KV cache and attention cost, much faster first-byte latency and higher throughput, while preserving perplexity and task accuracy in benchmarks.
I wish more long posts followed this model of a scientific paper.