As there is another thread about this currently on the front page, I've merged all the comments that are not just bickering about the title into that one:
Hey all, I’m the lead author on this work. Thanks to the OP for posting, and for all the interest and feedback. To clarify a few points: the goal of this work is to investigate the extent to which an LLM can manage memory and different memory hierarchies, applying lessons from operating systems to extend effective context lengths. All of our code is open sourced at https://github.com/cpacker/MemGPT so you can try it out yourselves!
This is a fascinating approach! I posted this as a separate comment but would be curious about your response directly.
What do you think about the tradeoff of having an agent manage its own memory as opposed to having a separate agent whose job it is to manage the memory and the other agent just focuses on the conversation?
It’s an interesting paper by credible academics but the name is just downright terrible. I suppose I’m extra annoyed because I don’t really have an interest in LLMs to begin with but I do in operating systems.
Sorry if this is a bit meta but why is it so popular and accepted to Just Say Stuff in AI land? Sometimes it seems to me that AI people have begun hallucinating as much as the models they’re training.
This article seems pretty cool but in no way has it anything to do with its title. “Operating system” has a clearly defined meaning and it’s not “a thing that has memory”. To me these kinds of grandiose claims with no substance undermine the credibility of the authors.
I mean what’s wrong with “Tiered memory layers to provide extended AI context windows”?
IMO, as an NLP researcher, it's a product of the "fast science" culture of AI where there are more papers than anyone could possibly read even in a given subfield at a particular conference, and an "old" paper is anything that was published over two years ago. (Who knew Bollywood movies and AI publications would have so much in common?) Maybe there are good reasons why publications should move so much more quickly in AI as opposed to, say, scholastic philosophy, but it's undeniable that it has become a struggle to get others to even read an abstract of your work, and as such we should not be surprised when sensationalization verges on becoming a requirement for your work to get engagement.
So to answer your question--perhaps cynically--“Tiered memory layers to provide extended AI context windows” is not the best title you could have for an AI paper because it has all the rizz of a shipping manifest. If you want to maximize citations, you need to market more.
I also considered whether to blame the field's very young reviewer population for not having the proper disdain for sensationalization that I'd expect from an older researcher who would surely have more restraint than to speak so much more of the sizzle than of the steak, but then I remembered that the paper which introduced the Transformer model was titled "Attention Is All You Need".
I love this response because it's at least honest about the reality of "scientific publishing" today - who cares about the accuracy, verifiability or ability to replicate - you just need to maximize your citations via hype!!"
Thanks for your feedback @skrebbel. We called it MemGPT - in reference to teaching LLMs how to manage their own memory hierarchy. The idea and title evolved when we realized we were teaching the LLM to read/write, and this enabled perpetual chatbots, doc QA, etc. Various capabilities by a few simple abstractions, inspired by the original unix paper's 4 key abstractions - read/write/open/close. Further, in the title we say "Towards", indicating a qualified step in that direction.
Perpetual chatbots are a super cool achievement but I fail to see how that resembles an OS, a thing that handles files and lets you run programs, in any way. If MemGPT is as spectacular as you suggest then I think an honest, down to earth title would have gotten you a lot more attention. It seems to me that limited context windows are one of the key roadblocks to wider real-world LLM applicability, beyond demos that do well on Twitter I mean.
I think a title like "MemGPT: Operating System inspired memory management for LLMs" would be okay. The current title makes it seem as if you are building an OS based on the LLM technology, which is quite misleading.
still this is not an "operating system" nor does it in any way related to them;
please use the right words when you have to communicate, otherwise you just sound like scammers who want to sell pots with holes by passing them off as colanders
In defense of the authors, I find it helpful to explore analogies between the nascent paradigm of LLM programming and other much older conceptual frames. This paper isn't just tiered memory, but also the analogy between an LLM and a CPU, interrupts, event processing etc.
This is why we need to get AI (or DL) out of the realm of CS, quickly. Similar to how psychology is a different field than physics even if the former is built on the latter.
I hate this take, no matter how often it's repeated. Yes there is hype fever pitch around AI, yes it attracts grifters (as does every hype), but in comparison to crypto, at the very least gen AI provides tractable benefits in a wide array of usecases. What has nft/web3/x-coins ever given us apart from speculative investments (aka gambling)?
>> What has nft/web3/x-coins ever given us apart from speculative investments (aka gambling)?
Cryptocurrencies modernized the malware industry.
Before cryptocurrencies, ransomware used to require victims to pay using gift cards or other ridiculous means.
After cryptocurrencies, victims could now pay their "captors" directly over the Internet in cryptocurrency.
Before cryptocurrencies, malware entities and actors acted largely independently unless they were part of a state-sponsored group or larger organized crime "business".
After cryptocurrencies, "ransomware as a service" emerged where different groups specialize in particular components and services of ransomware: tools to establish footholds, tools to take over and encrypt victim systems, services to demand and take payments, and services to wash and launder payments. Payments between these groups are made in cryptocurrency.
The last bull saw the rise of all kinds of new technologies which will enable the use cases people were talking about (Avalanche is my favorite example, but theres some new stuff like SUI coming). The thing web3 did absolutely right is create a true interchangeable format (ie tokens) the downside is this technology has been developing sooooooo slowly. Whenever people start to get hope the technology might be ready for use, the gambling use cases rise to the top and dominate. I think crypto is also dominated by a certain hardline techy. Or at least it's critics are. Crypto should be working side by side with traditional finance. Instead they view it as their enemy. FTX US is an example of how regulation can be good (the funds on FTX-US were more or less unmolested) yet most crypto people seem to argue against any regulation at all. It will be their demise.
We always thought dangerous AIs will be funded by the military, or out of internal R&D budgets of evil multinational megacorps-that-do-everything. Turns out instead it's all been funded by grifters and scammers, both large and small. Literally. It's not DARPA or some Umbrella Corporation, but the advertising companies that did all the AI R&D, while shitcoiners spent a decade unintentionally funding the hardware AI needs.
If you told me a decade ago that this is where we would land, I would tell you to at least make the lie a bit more believable. And yet here we are! Strange times indeed.
Well, check out the insane profits (not just revenue) NVIDIA has made this year. That profit margin can fit quite a few mansions as well as the biggest AI companies.
People are joking about the crypto bros funding it, but the leap in profit suggests it's not crypto fueled at all.
The bugs in programs aren't actual bugs either, windows aren't actual windows, and many would argue that computer science is still more of an art than a science.
This comment section is wild right now btw, no one is responding to the content, just the, i guess, misleading title.
The OS in the title is making a comparison to how page caches work. Relevant data is moved into the context the llm needs to answer chat questions. When it runs out of memory it moves a condensed, searchable version of the content to another data store. Its trying to solve the problem of relatively low max token limits.
I wonder if this makes it easier to run on small devices?
Thanks for checking out our work! Yeah, great point, this is even more critical in those scenarios when you have very limited memory. You can play around with different context sizes using the code on GitHub: https://github.com/cpacker/MemGPT/blob/main/memgpt/constants...
I think the problem with stuff like this is that in many cases the authors have an idea like "LLMS + ???? = Operating System !"
which is not that hard of an idea to come up with. Generally in each instance of LLM + ???? = "thing", the first papers that come out to fill the ???? with an answer do that in a rush to get on arxiv and so naturally the work is lackluster (since they have barely had time to think about an actual good solution for the ????).
Hey OP, thank you for the post. In this work, we are looking at one particular functionality of an OS - memory management. The project started off motivated by trying to extend context length for a project we were working on. The analogy to OSes came naturally as the project progressed and there was growing similarity to the caching and memory hierarchy relation.
My initial reaction when reading the title was the same. Their abstract does a better job at explaining what they actually do, and where the connection with an OS comes in:
> To enable using context beyond limited context windows, we propose virtual context management, a technique drawing inspiration from hierarchical memory systems in traditional operating systems that provide the appearance of large memory resources through data movement between fast and slow memory.
People have to stop treating LLMs and transformer models like they're the solution to everything. Never before had I seen so much jump on a hype by academics (I'm an academic too). Honestly it feels like mostly clout seeking and a way to improve citation numbers.
Transformers are a primitive we have not fully figured out yet. So it’s perfectly fine to see where this primitive fits.
For example, this week several papers on time series forecasts indicate they may have use there.
For example they seem to do better job on translation than previous approaches.
For example they seem to do a better job at transcription than previous approaches.
Probably will do better on OCR than previous approaches.
Probably has flaws that limits scenarios requiring high precision we may or may not overcome
Possibly will better on autonomous decision making (Does x include a privacy leak should be investigated?) than previous approaches (keyword scanning).
We are in a technological wave of discovery and experimentation, calls for restraint of curiosity and research betray fears
It goes way beyond that. Transformers are the first practical, scalable, general-purpose differentiable (i.e. trainable with gradient descent) algorithm. We haven't come close to seeing the limits of what they can do, because everything so far points to the fact that their only limit is our current hardware. And hardware is improving at a much faster and steadier rate than algorithms in computer science these days.
everything? they've solved reinforcement learning? they can handle continuous domains, like robot motion? that's funny, I thought they could only handle sequences of tokens.*
yes, they're exciting, and they are the most general architecture we've found so far, but there are important problems in AI (like anything continuous), that they're really not suited for.
I think there's better architectures out there for many tasks, and I'm a little dismayed that everyone seems to be cargo-culting the GPT architecture rather than taking the lessons for transformers and experimenting with more specialized algorithms.
*btw they don't need quantized tokens, there's no reason they can't just work on continuous vectors directly, and they don't have to be causal or limited to one sequence, but "transformer" seems to mean GPT in everyone's mind, and even though the original transformer was an encoder-decoder model we rarely seem to see those these days for some reason.
There's two ways you can solve currently intractable problems: Find better algorithms or improve the hardware. It's actually insanely hard to come up with new algorithms, that is why machine learning and AI were lagging behind most of computer science for decades.
> calls for restraint of curiosity and research betray fears
That kind of rhetoric is dangerously close to the cryptocurrency shills saying all critics are people who are annoyed because they “missed the boat” and didn’t get rich. It’s the kind of generic comment which can be used to discredit anything the interlocutor wants.
"There is more where this came from" reads differently based on how interesting/valuable "this" is. Cryptocurrency wasn't producing anything interesting beyond new flavors of old scams, and renewed appreciation for financial regulation. Transformer models and LLMs in particular seem to be blazing through any problem we throw at them, including previously unsolved problems, and every marginal improvement is instantly useful (both in terms of capabilities and in terms of $$$ produced). We hit the motherlode here - and it shows no sign of running out just yet.
That's the difference.
And even when we hit the limit of improvement on those models, when scaling up won't make a qualitative difference, we can expect many years of further breakthroughs in applications, as R&D focuses on less obvious applications and making more efficient use of the capabilities available.
That’s not the part of the comment I had an issue with, which is why it’s not the part I quoted. What I commented on is the end, which implies negative intentions on the part of the original commenter.
> Research and curiosity a miles away from “get instantly rich” and “solution in search of a problem” we’ve seen in crypto.
And what the original comment boils down to is “let’s not make LLMs and transformers solutions in search of problems, let’s not try to fit them to solve everything”. It seems you might agree.
My issue with the response has nothing to do with specific technologies, but that it painted another view as having an agenda (fear). That is the generic defence. You take something you believe and then say those who disagree do so due to <negative connotation>.
By the way, this is not the point but there are plenty of “solution in search of a problem” and “get instantly rich” (including full-on scams) cases in the current wave of AI.
I'm struggling to find where these authors are treating LLMs and transformers like they're the solution to everything, or where they are just following hype around? These are long-time AI researchers at Berkeley who've developed a framework for dealing with limited context windows, drawing inspiration from operating systems.
In the paper, they're talking about managing LLM context windows the way operating systems manage memory and files. They're not saying LLMs should be used as operating systems.
Always cringeworthy when the average bum coder on HN with zero attention span or intellectual curiosity criticizes papers way out of their intellectual depth without reading further than the title. They are describing a new concept using "operating system" almost as a metaphor. This is a way to describe novel concepts. The article abstract makes the point clear and only takes a few seconds to read.
The branding on this is a bit much, it’s not an operating system. However LLMs are the real deal, some papers claim they are achieving SOTA or significant breakthroughs in various domains. Surely they’re computationally intensive, but if you have read the most of the papers out of Berkeley, Microsoft, Meta, Google/DeepMind/Waymo… I think you’d have an change in opinion.
Weird thing is it was designed to model language. It’s surprising that it returns sound answers as often as it does. But that’s also kind of the problem, it’s “surprising”, i.e. we don’t really know what happened.
You wouldn’t fly on a jetliner that’s “surprising it flies without disintegrating midair”.
Science being iterative, they definitely weren't the first to fly, and not even in a heavier than aircraft, what they did acheive was the first time the pilot had 3-axis control, and was the first powered heavier than air manned flight.
> Weird thing is it was designed to model language. It’s surprising that it returns sound answers as often as it does.
Is this surprising? Can you point to researchers in the field being “surprised” by LLMs returning sound answers?
> “surprising”, i.e. we don’t really know what happened.
This ie reads like a sort of popsci conclusion.
We know exactly what happened. We programmed it to perform these calculations. It’s actually rather straightforward elementary mathematics.
But, what happens is so many interdependent calculations grow the complexity of the problem until we are unable to hold it in it our minds, and to analyze its decisions computationally necessitates similar levels of computation for each decision being made as what was used to compute the weights.
As for its effectiveness, familiarity with the field of computational complexity points to high dimensional polynomial optimization problems being broadly universal solvers.
> Is this surprising? Can you point to researchers in the field being “surprised” by LLMs returning sound answers?
It's surprising because it wasn't the intent of LLMs. LLMs are just predictive models that guess the most likely next word. Having the results make sense was never a priority. Early version, GPT1/2, all return mostly complete nonsense. It was only with GPT3 when the model got large enough that it started returning results that are convincing and might even make sense often enough.
Even more mind boggling is the fact that randomness is part of its algorithm, i.e. temperature, and that without it the output is kind of meh.
> It's surprising because it wasn't the intent of LLMs. LLMs are just predictive models that guess the most likely next word. Having the results make sense was never a priority.
If you took the same amount of data for the GPT3+ but scrambled it's tokenization before training THEN I would agree with you that its current behaviour is surprising, but the model was fed data that has large swaths that are literal question and answer constructions. It's over fitting behavior is largely why it's parent company is facing so much legal backlash.
> Even more mind boggling is the fact that randomness is part of its algorithm
The randomness is for token choice rather than any training time tunable so fails to support the "i.e. we don’t really know what happened" sentiment. We do know, we told it to flip a coin, and it did.
> i.e. temperature, and that without it the output is kind of meh.
Both without it and with it. You can turn up the temperature and get bad results as well as you can turn it down and get bad results.
If adding a single additional dimension to the polynomial of the solution space turned a nondeterministic problem into a deterministic one, then yes, I would agree with you, that would be surprising.
> so fails to support the "i.e. we don’t really know what happened" sentiment
It's less that we don't know what's happening on a micro-level but more that it's surprising that it's producing anything coherent at all on a macro-level - especially with a (necessary) element of randomness in the process.
For most part we don't seem particularly knowledgeable about what happens on a macro-level. Hallucinations remain an unsolved problem. AI companies can't even make their "guardrails" bulletproof.
I think this is an uncharitable reading of this thread.
I’m arguing against the breathless use of “surprising”.
My gp explains what I think you overlooked in this dismissive response.
> to analyze its decisions computationally necessitates similar levels of computation for each decision being made as what was used to compute the weights.
Explainable but intractable is still far from surprising for me.
> It's surprising because it wasn't the intent of LLMs. LLMs are just predictive models that guess the most likely next word. Having the results make sense was never a priority.
If you read through what Hinton or any of his famous students have said, it genuinely was and is surprising. Everything from AlexNet to the jump between GPT-2 to GPT-3 was surprising. We can't actually explain that jump in a formal way, just reasonable guesses. If something is unexplainable, it's unpredictable. Prediction without understanding is a vague guess and the results will come as a surprise.
>Is this surprising? Can you point to researchers in the field being “surprised” by LLMs returning sound answers?
Lol researchers were surprised by the mostly incoherent nonsense pre-transformer RNNs were spouting years go, nevermind the near perfect coherency of later GPT models. To argue otherwise is just plain revisionism.
> What made this result so shocking at the time was that the common wisdom was that RNNs were supposed to be difficult to train (with more experience I’ve in fact reached the opposite conclusion). Fast forward about a year: I’m training RNNs all the time and I’ve witnessed their power and robustness many times, and yet their magical outputs still find ways of amusing me.
This reads more like humanizing the language of the post then any legitimate surprise from the author.
The rest of the post then goes into great detail showing that “we DO really know what happened” to paraphrase the definition the op provides for their use of “surprise”.
> Conclusion We’ve learned about RNNs, how they work, why they have become a big deal, we’ve trained an RNN character-level language model on several fun datasets, and we’ve seen where RNNs are going.
I am pushing back on people conflating the innate complexity of a high dimensional polynomial with a misplaced reverence of incomprehensibility.
> In fact, it is known that RNNs are Turing-Complete in the sense that they can to simulate arbitrary programs (with proper weights).
Mathematically proven to be able to do something is about as far from surprise as one can get.
>This reads more like humanizing the language of the post then any legitimate surprise from the author.
Lol Sure
>I am pushing back on people conflating the innate complexity of a high dimensional polynomial with a misplaced reverence of incomprehensibility.
We don't know what the models learn and what they employ to aid in predictions. That is fact. Going on a grad descent rant is funny but ultimately meaninglessness. It doesn't tell you anything about the meaning of the computations.
There is no misplaced incomprehensibility because the internals and how they meaningfully shape predictions is incomprehensible.
>Mathematically proven to be able to do something is about as far from surprise as one can get.
Magic the gathering is turing complete. I'm sorry but "therotically turing complete" is about as meaningless as it gets. Transformers aren't even turing complete.
Depends what "operating system" means. I'd say things like democracy, marketing/journalism/propaganda, etc are operating systems of some sort, in that they perform orchestration of humans, modify reality, etc. Lack of memory is a big handicap for LLM's if they want to play in that league.
This wave of AI hype really brought out the curmudgeons, didn't it? Here we are, with this exciting architecture that is surprisingly applicable to a wide range of problem domains, yet we get "this is like crypto all over again!!1" and "oh everything looks like a nail now, huh?". I mean dang.
If the object of hype adds useful novelty, the interest could be justified. If, as it's often the case, it is not quite known - it's a question to figure out.
Granted, intuition is worth something, but it's still not a certainty, so somebody having a different opinion still could see something useful here.
I would agree, but on the other hand, Transformer (or attention really) based models seem to be the first time that computers are generating ad hoc human text on a variety of topics, so I do believe the hype is justified. I mean... people have spent entire careers in pursuit of this goal, and it's here... as long as what you want to talk about is less than 4096 / some K tokens.
Given how little progress (relatively) was made until transformers, it seems totally reasonable to pursue att ention models.
Interesting: I do wonder about slightly more complex languages which have declensions and verb "gender" (eg. in Serbian "pevala" means "(a female) sang", whereas "pevao" means that a male did. Or nouns and adjectives can be in 7 declensions: "plavom olovkom" means "with a blue pen", whereas "a blue pen" is just "plava olovka".
ChatGPT always mixes these up, hallucinates a bunch of words (inappropriate prefixes, declensions etc and is very happy to explain the meaning of these imaginary words), and I can imagine smaller, more complex languages like Serbian needing even larger corpuses than English, yet that's exactly the hard part: there is simply less content to go off of.
I agree they are probably being over-applied but I for one am fascinated and excited for the future and all the possibilities. It’s amazing how much effort is being applied across the board and failures to me just progress other research.
they are very clearly the most powerful technological innovation i've witnessed in my adult life. we don't even know what they're capable of yet, we're like one year into discovering practical usecases and they are improving at a constant dizzying rate. this is the time for exploration and experimentation.
I'm a software developer, I live in the concrete world.
To put it in layman's terms, LLMs are to software developers what power tools were for tradesmen.
Sure you can still use your old screw driver, and for some work, it's not worth getting your electric drill out; but it's a game changer.
Sometimes, the hype is justified. I believe at the OS layers, LLM support would make sense. It's full of old mental constructs and "I have to remember how to do this".
> Every single problem does not need to be solved using an LLM
How do you know if a problem is suitable for transformers/LLMs unless you try? We have this great generalized architecture, I would hope people throw everything at it and see what sticks, because what's the downside? Less focus on bespoke predictive models? Oh no.
https://news.ycombinator.com/item?id=37901902