He did not say what kind of research strategies or techniques might take its place. In the paper describing GPT-4, OpenAI says its estimates suggest diminishing returns on scaling up model size. Altman said there are also physical limits to how many data centers the company can build and how quickly it can build them.
> In the paper describing GPT-4, OpenAI says its estimates suggest diminishing returns on scaling up model size.
I read the two papers (gpt 4 tech report, and sparks of agi) and in my opinion they don't support this conclusion. They don't even say how big GPT-4 is, because "Given both the competitive landscape and the safety implications of large-scale models like GPT-4, this report contains no further details about the architecture (including model size), hardware, training compute, dataset construction, training method, or similar."
> Altman said there are also physical limits to how many data centers the company can build and how quickly it can build them.
OK so his argument is like "the giant robots won't be powerful, but we won't show how big our robots are, and besides, there are physical limits to how giant of a robot we can build and how quickly we can build it." I feel like this argument is sus.
OpenAI has likely run into a wall (or is about to) for model size given it's funding amount/structure[1] - unlike its competition who actually own data centers and have lower marginsl costs. It's just like when peak-iPad Apple claimed that a "post-PC" age was upon us.
1. What terms could Microsoft wring out of OpenAI for another funding round?
If you are worried about AI, this shouldn't make you feel a ton better. GPT4 is just trained to predict the next word, a very simple but crude approach and look what it can do!
Imagine when a dozen models are wired together and giving each other feedback with more clever training and algorithms on future faster hardware.
FWIW we had thin clients in computer labs in middle school / high school 15 years ago (and still today these are common in enterprise environments, e.g. Citrix).
Biggest issue is network latency which is limited by the speed of light, so I imagine if computers in 10 years require resources not available locally it would likely be a local/cloud hybrid model.
Machine learning is actually premised on being “simple” to implement. The more priors you hardcode with clever algorithms, the closer you get to what we already have. The point is to automate the process of learning. We do this now with relatively simple loss functions and models containing relatively simple parameters. The main stipulation is that they are all defined to be continuous so that you can use the chain rule from calculus to calculate the error with respect to every parameter without taking so long that it would never finish.
I agree that your suggested approach of applying cleverness to what we have now will probably produce better results. But that’s not going to stop better architectures, hardware and even entire regimes from being developed until we approach AGI.
My suspicion is that there’s still a few breakthroughs waiting to be made. I also suspect that sufficiently advanced models will make such breakthroughs easier to discover.
Personally, I'm less worried about AI than I am about what people using these models can do to others. Misinformation/disinformation, more believable scams, stuff like that.
I have repeatedly argued against this notion of „just predicting the next word“. No. It‘s completing a conversation. It‘s true that it is doing this word by word, but it‘s kind of like saying a CNN is just predicting a label. Sure, but how? It‘s not doing it directly. It‘s doing it by recovering a lot of structure and in the end boiling that down to a label. Likewise a network trained to predict the next word may very well have worked out the whole sentence (implicitly, not as a text) in order to generate the next word.
I actually have high hopes for the hybrid architecture Ben Goertzel has been working on at OpenCog. I think the LLMs are soon going to hit a S curve w/o introduction of additional scientific knowledge like physics and notion of energy (wrt AGI development, they'll still be good for tonnes of other jobs displacing things).
I agree, probably for a different reason. IMO the fact that Sam is saying this likely means that the LLMs are close to the upper knee of the S-Curve and after GPT5 they likely dont have many new fundamental ideas (additional SCurves) to throw at it. since they are ahead of the rest, it likely means we might be headed for an AI improvement pause for couple of years after GPT5.
Still good enough to upset the balance in search/ad market. Interesting times.
> it likely means we might be headed for an AI improvement pause for couple of years after GPT5.
I suspect that a pause in base LLM performance won’t be an AI improvement pause; there’s a whole lot of space to improve the parts of AI systems around the core “brain in a jar” model.
I agree, there will be other things to be improved in AI system, but IMHO (tea leaves reading really) it would only lead to incremental improvements in overall systems. Also there is a lot of 'interfacing' work that needs to happen & i suspect that would end up filling the pause, which really is LLM productization loosely speaking.
far as AGI is concerned I dont believe LLMs are really the right architecture for it, AGI likely needs some symbolic logic and a notion of physicality (ie.. physical laws & energy/power).
> but IMHO (tea leaves reading really) it would only lead to incremental improvements in overall systems.
It will reach a point where that is the case, sure; it is not there now, and if we are within one model generation of exhausting (for now) major core model improvements, I don’t think we’ll have reached the point of gradual incremental improvement from rest-of-system improvements yet.
No we haven't. the title is misleading. there's plenty of scale room left. part of it might just not be economical (parameter sie) but there's data. If you take this to mean, "we're at a dead end" you'd be very wrong
Before I get nuked from orbit for daring to entertain humor, if someone is running ahead of me in a marathon, and running so far ahead, yet still broadcasting things to the back for the slow people (like myself), then eventually we catch up to them, and they suddenly say, you know what guys, we should stop running in this direction, there's nothing to see here right before anyone else is able to verify the veracity of their statement, perhaps it would still be in the public interest for at least one person to verify what they are saying. Given how skeptical the internet at large has been of Musk's acquisition of a company, it's interesting that the skepticism is suddenly put on hold when looking at this part of his work...
Saying “hey don’t go down the path we are on, where we are making money and considered the best in the world.. it’s a dead end” rings pretty hollow.. like “don’t take our lunch please?” Might be a similar statement it feels..
It's a pretty sus argument for sure when they're scared to release even parameter size.
although the title is a bit misleading on what he was actually saying. still, there's a lot left to go in terms of scale. Even if it isn't parameter size(and there's still lots of room here too, it just won't be economical), contrary to popular belief, there's lots of data left to mine
Nah - GPT-4 is crazy expensive, paying 20$/mo only get's you 25messages/3hours and it's crazy slow. The api is rather expensive too.
I'm pretty sure that GPT-4 is ~1T-2T parameters, and they're struggling to run it(at reasonable performance and profit). So far their strategy has been to 10x the parameter count every GPT generation, and the problem is that there's diminishing returns everytime they do that. AFAIK they've now resorted to chunking GPT through the GPUs because of the 2 to 4 terabytes of VRAM required (at 16bit).
So now they've reached the edge of what they can reasonably run, and even if they do 10x it the expected gains are less. On top of this, models like LLaMa have shown that it's possible to cut the parameter count substantially and still get decent results (albiet the opensource stuff still hasn't caught up).
On top of all of this, keep in mind that at 8bit resolution 175B parameters (GBPT3.5) requires over 175GB of VRAM. This is crazy expensive and would never fit on consumer devices. Even if you use quantization and use 4bit, you still need over 80GB of VRAM.
This definitely is not a "throw them off the trail" tactic - in order for this to actually scale the way everyone envisions both in performance and running on consumer devices - research HAS to be on improving the parameter count. And again there's lots of research showing its very possible to do.
> Reasonable to assume that in 1-2 years it will also come down in cost.
Definitely. I'm guessing they used something like quantization to optimize the vram usage to 4bit. The thing is that if you can't fit the weights in memory then you have to chunk it and that's slow = more gpu time = more cost. And even if you can fit it in GPU memory, less memory = less gpus needed.
But we know you _can_ use less parameters, and that the training data + RLHF makes a massive difference in quality. And the model size linearly relates to the VRAM requirements/cost.
So if you can get a 60B model to run at 175B's quality, then you've almost 1/3rd your memory requirements, and can now run (with 4bit quantization) on a single A100 80GB which is 1/8th the previously known 8x A100's that GPT-3.5 ran on (and still half GPT-3.5+4bit).
Also while openai likely doesn't want this - we really want these models to run on our devices, and LLaMa+finetuning has shown promising improvements (not their just yet) at 7B size which can run on consumer devices.
Yeah I am noticing this as well. GPT enables you to do difficult things really easily, but then it is so expensive you would need to replace it with custom code for any long term solution.
For example: you could use GPT to parse a resume file, pull out work experience and return it as JSON. That would take minutes to setup using the GPT API and it would take weeks to build your own system, but GPT is so expensive that building your own system is totally worth it.
Unless they can seriously reduce how expensive it is I don't see it replacing many existing solutions. Using GPT to parse text for a repetitive task is like using a backhoe to plant flowers.
> For example: you could use GPT to parse a resume file, pull out work experience and return it as JSON. That would take minutes to setup using the GPT API and it would take weeks to build your own system, but GPT is so expensive that building your own system is totally worth it.
True, but an HR SaaS vendor could use that to put on a compelling demo to a potential customer, stopping them from going to a competitor or otherwise benefiting.
And anyway, without churning the numbers, for volumes of say 1M resumes (at which point you've achieved a lot of success) I can't quite believe it would be cheaper to build something when there is such a powerful solution available. Maybe once you are at 1G resumes... My bet is still no though.
I work for a company with the web development team. We have ~6 software developers.
I'd love to be able to just have people submit their resume's and extract the data from there, but instead I'm going to build a form and make applicants fill it out because chatGPT is going to be at least $0.05USD depending on the length of the resume.
I'd also love to have mini summeries of order returns summerized in human form, but that also would cost 0.05USD per form.
the tl;dr here is that there's a TON of usecases for a LLM outside of your core product (we sell clothes) - but we can't currently justify that cost. Compare that to the rapidly improving self-hosted solutions which don't cost 0.05USD for literally any query (and likely more for anything useful).
The problem is that it would take us the same amount of time to just add a form with django. Plus you have to handle failure cases, etc.
And yeah I agree this would be a great use-case, and isn't that expensive.
I'd like to do this in lots of places, and the problem is I have to convince my boss to pay for something that otherwise would have been free.
The conversation would be "We have to add these fields to our model, and we either tell django to add a form for them, which will have 0 ongoing cost and no reliance on a third party,
or we send the resume to openai, pay for them to process it, make some mechanism to sanity check what GPT is responding with, alert us if there's issues, and then put it into that model, and pay 5 cents per resume."
> 1-3 hours of a fully loaded engineers salary per year.
That's assuming 0 time to implement, and because of our framework it would take more hours to implement the openai solution (that's also more like 12 hours where we are).
> $500 per 10k.
I can't stress this enough - the alternative is 0$ per 10k. My boss wants to know why we would pay any money for a less reliable solution (GPT serialization is not nearly as reliable as a standard django form).
I think within the next few years we'll be able to run the model locally and throw dozens of tasks just like this at the LLM, just not yet.
I have tried GPT3.5 and GPT4 for this type of task - the "near perfect results" is really problematic because you need to verify that it's likely correct, notify you if there's issues, and even then you aren't 100% sure that it selected the correct first/last name.
This is compared to a standard html form. Which is.... very reliable and (for us) automatically has error handling built in, including alerts to us if there's a 504.
For a big company that is nothing but if you are bootstrapping and trying to acquire customers with an MVP racking up a $500 bill is frightening. What if you offer a free trial and blow up and end up with 5k+ bill.
It's never been in OpenAIs interest to make their model affordable or fast, they're actually incentivized to do the opposite as an excuse to keep the tech locked up.
This is why Dall-e 2 ran in a data centre and Stable Diffusion runs on a gamer GPU
I think you're mixing the two. They do have an incentive to make it affordable and fast because that increases the use cases for it, and the faster it is the cheaper it is for them, because the expense is compute time (half the time ~= half the cost).
> This is why Dall-e 2 ran in a data centre and Stable Diffusion runs on a gamer GPU
This is absolutely why they're keeping it locked up. By simply not releasing the weights, you can't run Dalle2 locally, and yeah they don't want to do this because they want you to be locked to their platform, not running it for free locally.
On the other hand though, Chinchilla and multimodal approaches already showed how later AIs can be improved beyond throwing petabytes of data at them.
It is all about variety and quality from now on I think. You can teach a person all about the color zyra but without actually ever seeing it, they will never fully understand that color.
It does seem, though, that using chinchilla like techniques does not create a copy with the same quality as the original. It's pretty good for some definition of the phrase, but it isn't equivalent, it's a lossy technique.
I agree on the lossy. There is a tradeoff between efficiency and comprehensiveness, kind of. It would be pretty funny if in the end, the most optimal method turns out to be the brain we already have. Extremely efficient, hardware optimized, but slow as hell and misunderstand stuff all the time unless prompted with specific phrases.
eh, I haven't personally found a usecase for LLMs yet given the fact that you can't trust the output and it needs to be verified by a human (which might as well be just as time consuming/expensive as actually doing the task yourself)
I've enjoyed using it for very small automation tasks. For instance, it helped me write scripts to take all my audiobooks with poor recording quality, split them into 59-minute chunks, and upload them to Adobe's free audio enhancement site to vastly improve the listening experience.
I’d reconsider the “might as well just be as time consuming” thing. I see this argument about Copilot a lot, and it’s really wrong there, so it might be wrong here too.
Like, for most of the time I’m using it, Copilot saves me 30 seconds here and there and it takes me about a second to look at the line or two of code and go “yeah, that’s right”. It adds up, especially when I’m working with an unfamiliar language and forget which Collection type I’m going to need or something.
> Like, for most of the time I’m using it, Copilot saves me 30 seconds here and there and it takes me about a second to look at the line or two of code and go “yeah, that’s right”.
I've never used Copilot but I've tried to replace StackOverflow with ChatGPT. The difference is, the StackOverflow responses compile/are right. The ChatGPT responses will make up an API that doesn't exist. Major setback.
They're good for tasks where generation is hard but verification is easy. Things like "here I gesture at a vague concept that I don't know the name of, please tell me what the industry-standard term for this thing is" where figuring out the term is hard but looking up a term to see what it means is easy. "Create an accurate summary of this article" is another example - reading the article and the summary and verifying that they match may be easier than writing the summary yourself.
Thing is, you can't trust what you find on stack overflow or other sources either. And searching, reading documentation and so on takes a lot of time too.
I've personally been using it to explore using different libraries to produce charts. I managed to try out about 5 different libraries in a day with fairly advanced options for each using chatGPT.
I might have spent a day in the past just trying one and not to the same level of functionality.
So while it still took me a day, my final code was much better fitted to my problem with increased functionality. Not a time saver then for me but a quality enhancer and I learned a lot more too.
Maybe, maybe not. I get useful results from it, but it doesnt always work. And it's usually not quite what I'm looking for, so then I have to go digging around to find out how to tweak it. It all takes time and you do not get a working solution out of the box most of the time.
No? I use it all the time to help me, for example, read ML threads when I run into a term I don't immediately understand. I can do things like 'explain this at the level of a high school student'
> YouTubers upload about 720,000 hours of fresh video content per day. Over 500 hours of video were uploaded to YouTube per minute in 2020, which equals 30,000 new video uploads per hour. Between 2014 and 2020, the number of video hours uploaded grew by about 40%.
But what are you mostly "teaching" the LLM then? Mundane everyday stuff? I guess that would make them better at "being average human" but is that what we want? It already seems that prompting the LLM to be above-average ("pretend to be an expert") improves performance.
This whole conversation about training set size is bizarre. No one ever asks what’s in the training set. Why would a trillion tokens of mundane gossip improve a LLMs ability to do anything valuable at all?
If a scrape of the general internet, scientific papers and books isn’t enough, a trillion trillion trillion text messages to mom aren’t going to change matters.
Right. They've already sucked in most of the good general sources of information. Adding vast amounts of low-quality content probably won't help much and might degrade the quality of the trained model.
Ilya Sutskever (OpenAI Chief Scientist): "Yeah, I would say the data situation is still quite good. There's still lots to go" - https://youtu.be/Yf1o0TQzry8?t=685
There was a rumor that they were going to use Whisper to transcribe YouTube videos and use that for training. Since it's multimodal, incorporating video frames alongside the transcriptions could significantly enhance its performance.
One way would be to get people to let AI watch as they interact with computer (watch YouTube or perform other tasks). You might even be able to outsource some of the computing to the local system.
Yeah, but it's not like the videos are private. Surely Amazon has the real advantage, given they have a ton of high quality tokens in the form of their kindle library and can make it difficult for OpenAI to read them all.
You can transcribe all spoken words everywhere and keep the model up to date? Keep indexing new data from chat messages, news articles, new academic work etc.
What about all the siloed content kept inside corporate servers? You won't get normal GPT to train on it, of course, but IBM could build a "IBM-bot" that has all the GPT-4 dataset + all of IBM's internal data.
That model might be very well tuned to solve IBM's internal problems.
I don't think you can just feed it data. You've got to curate it, feed it to the LLM, and then manually check/further train the output.
I also question that most companies have the volume and quality of data worth training on. It's littered with cancelled projects, old products, and otherwise obsolete data. That's going to make your LLM hallucinate/give wrong answers. Especially for regulated and otherwise legally encumbered industries. Like can you deploy a chat bot that's wrong 1% or 0.1% of the time?
Well, IBM has 350k employees. If training a LLM on curated data costs tens of millions of dollars but ends up reducing headcount by 50k, it would be a massive win for any CEO.
You have to understand that all the incentives are perfectly aligned for corporations to put this to work, even spending tens of millions in getting it right.
The first corporate CEO who announces that his company used AI to reduce employee costs while increasing profits is going to get such a fat bonus that everyone will follow along.
Since Chat-GPT-4 is being integrated into the MS Office suite, this is an "in" to corporate silos. The MS cloud apps can see inside a great many of those silos.
If you were devious enough, you could be listening in on billions of phone conversations and messages and adding that to your data set.
This also makes me doubt that NSA hasn't already cracked this problem. Or that China won't eventually beat current western models since it will likely have way more data collected from its citizenry.
I wonder what percentage of phone calls would add anything meaningful to models, I imagine that the nature of most phone calls are both highly personal and fairly boring.
You can generate textual examples that teach logic, multi-dimensional understanding and so on. Similar to the ones that are in math books, but in a massive scale.
I doubt they have trained on 0.1% of the tokens that are 'easily' available (that is, available with licencing deals that are affordable to OpenAI/MSFT).
They might have trained on a lot of the 'high quality' tokens, however.
> Once you've trained on the internet and most published books (and more...) what else is there to do? You can't scale up massively anymore.
Dataset size is not relevant to predicting the loss threshold of LLMs. You can keep pushing loss down by using the same sized dataset, but increasingly larger models.
Or augment the dataset using RLHF, which provides an "infinite" dataset to train LLMs on. Limited by the capabilities of the scoring model which, of course, you can scale the scoring model infinitely so again the limit isn't dataset size but training compute.
> Dataset size is not relevant to predicting the loss threshold of LLMs. You can keep pushing loss down by using the same sized dataset, but increasingly larger models.
Deepmind and others would disagree with you! No-one really knows in actual fact.
I don't recall the Chinchilla paper disputing my point. They establish "training-compute optimal" scaling laws, but none of their findings suggest that loss hits any kind of asymptote.
Perhaps we're talking past each other, is "loss threshold" a specific term in LLM literature?
Merely pointing out that the debate as to whether we are compute or data limited (OP) has not concluded at all; There are lots of compelling theories on relationship between the two.
You could have it start talking to itself in the way that AlphaGO learns to get better at Go. All that needs to be done is find some fitness function that indicates that useful knowledge has been produced. In Go and Chess this is easy.
It can start posting synthesized ideas on social media and see how many likes it gets. Coupled with a metric containing dissimilarity to current information, this could be a useful way to progress to superhuman insights.
Videos - all of youtube, all the movies, everything that's ever been captured on film. Transcribe the audio, automatically describe the images and try to predict the next one.
people seem to have forgotten about the multi-modal GPT-4
There's a ton of potential left on the table. The question is if transformers have hit their limit with GPT-4 or not.
It's a pretty simple equation when you think about it this way and why Sam would say they have hit their limit. Sam is basically Microsoft and they want to retain their lead.
Once Google learns to put their data to use correctly, it's almost guaranteed game over for OpenAI if they want it to be.
In short it seems like virtually all of the improvement in future AI models will come from better algorithms, with bigger and better data a distant second, and more parameters a distant third.
Of course, this claim is itself internally inconsistent in that it assumes that new algorithms won't alter the returns to scale from more data or parameters. Maybe a more precise set of claims would be (1) we're relatively close to the fundamental limits of transformers, i.e., we won't see another GPT-2-to-GPT-4-level jump with current algorithms; (2) almost all of the incremental improvements to transformers will require bigger or better-quality data (but won't necessarily require more parameters); and (3) all of this is specific to current models and goes out the window as soon as a non-transformer-based generative model approaches GPT-4 performance using a similar or lesser amount of compute.
I'd bet on a 2030 model trained on the same dataset as GPT-4 over GPT-4 trained with perfect-quality data, hands down. If data quality were that critical, practitioners could ignore the Internet and just train on books and scientific papers and only sacrifice <1 order of magnitude of data volume. Granted, that's not a negligible amount of training data to give up, but it places a relatively tight upper bound on the potential gain from improving data quality.
It's possible that this effect washes out as data increases, but researchers have shown that for smaller data set sizes average quality has a large impact on model output.
So true. There are still plenty of areas where we lack sufficient data to even approach applying this sort of model. How are we going to make similar advances in something like medical informatics where we not only have less data readily available but its much more difficult to acquire more data
All the LC grinding may come in handy after all! /s
What algorithms specifically show the most results upon improvement? Going into this I thought the jump of improvements were really related more advanced automated tuning and result correction, in which it could be done at scale as it were allowing a small team of data scientists to tweak the models until desired results were being achieved.
Are you saying instead, that concrete predictive algorithms need improvement or are we lumping the tuning into this?
We need more data efficient neural network architectures. Transformers work exceptionally well because they allow us to just dump more data into it, but ultimately we want to learn advanced behavior without having to feed it Shakespeare
I think it's unlikely that the first model to be widely considered AGI will be a transformer. Recent improvements to computational efficiency for attention mechanisms [0] seem to improve results a lot, as does RLHF, but neither is a paradigm shift like the introduction of transformers was. That's not to downplay their significance - that class of incremental improvements has driven a massive acceleration in AI capabilities in the last year - but I don't think it's ultimately how we'll get to AGI.
I'm using AGI here as arbitrary major improvement over the current state of the art. But given that OpenAI has the stated goal of creating AGI, I don't think it's a non-sequitur to respond to the parent comment's question
> Are you saying instead, that concrete predictive algorithms need improvement or are we lumping the tuning into this?
in the context of what's needed to get to AGI - just as if NASA built an engine we'd talk about its effectiveness in the context of space flight.
Traditional CS may have something to do with slightly improving the performance by allowing more training for the same compute, but it won't be an order of magnitude or more. The improvements to be gained will be found more in statistics than CS per se.
I'm not sure. Methods like Chinchilla and Quantization have been able to reduce compute by more than an order of magnitude. There might very well be a few more levels of optimizations within the same statistical paradigm.
I don't think LLMs are over [0]. I think we're relatively close to a local optimum in terms of what can be achieved with current algorithms. But I think OpenAI is at least as likely as any other player to create the next paradigm, and that it's at least as likely as likely as any other player to develop the leading models within the next paradigm regardless of who actually publishes the research.
Separately, I think OpenAI's current investors have a >10% chance to hit the 100x cap on their returns. Their current models are already good enough to address lots of real-world problems that people will pay money to solve. So far they've been much more model-focused than product-focused, and by turning that dial toward the product side (as they did with ChatGPT) I think they could generate a lot of revenue relatively quickly.
[0] Except maybe in the sense that future models will be predominantly multimodal and therefore not strictly LLMs. I don't think that's what you're suggesting though.
It already is relatively trivial to fine-tune generative models for various use cases. Which implies huge gains to be had with targeted applications not just for niche players but also OpenAI and others to either build that fine-tuning into the base system, build ecosystems around it, or just purpose build applications on top.
I think it's more exciting if compute stops being the core differentiation, as purpose trained models is exactly where I suspect real value lies.
Especially as a differentiation for a company. If everyone is using ChatGPT, then they're all offering the same thing and I can just as well go to the source and cut out the middleman.
The other fun development to come is well performing self hosted models, and the idea of light weight domain specific interface models that curate responses from bigger generalist models.
ChatGPT is fun but it is very general, it doesn't know about my business nor keep track of it or interface with it. I fully expect to see "Expert Systems" of old come back, but trained on our specific businesses.
Improvements will not come from collecting more and more samples for current large models, but will come from improvements to algorithms, that also may focus on improving the quality and use of input data.
I don't think there is such a clear separation between algorithms and data as your comment suggests.
An amusing thought I've had recently is whether LLMs are in the same league as the millions of monkeys at the keyboard, struggling to reproduce one of the complete works of William Shakespeare.
But I think not, since monkeys probably don't "improve" noticeably with time or input.
>"the company’s CEO, Sam Altman, says further progress will not come from making models bigger. “I think we're at the end of the era where it's going to be these, like, giant, giant models,” he told an audience at an event held at MIT late last week. “We'll make them better in other ways.”
So to reiterate, he is not saying that the age of giant AI models is over. Current top-of-the-line AI models are giant and likely will continue to be. However, there's not point in training models you can't actually run economically. Inference costs need to stay grounded which means practical model sizes have a limit. More effort is going to go into making models efficient to run even if it comes at the expense of making them less efficient to train.
Yes, but it also tells us that if Altman is honest here, then he doesn’t believe GPT-like models can scale to near level human performances (because even if the cost of compute was 10x or even 100x it would still be economically sound).
For one thing they're already at human performance.
For another, i don't think you realize how expensive inference can get. Microsoft with no scant amount of available compute is struggling to run gpt-4 such that they're rationing it between subsidiaries while they try to jack up compute.
So saying, it would be economically sound if it cost x10 or x100 what it costs now is a joke.
This tells me you haven't really stress tested the model. GPT is currently at the stage of "person who is at the meeting, but not really paying attention so you have to call them out". Once GPT is pushed, it scrambles and falls over for most applications. The failure modes range from contradicting itself, making up things for applications that shouldn't allow it, to ignoring prompts, to simply being unable to perform tasks at all.
Still waiting to see those plugins rolled out and actual vector DB integration with GPT 4, then we'll see what it can really do. Seems like the more context you give it the better it does, but the current UI really makes it hard to provide that.
Plus the recursive self prompting to improve accuracy.
We have given it extensions, and really the extensions do a lot of the work. The tool that judges the style and correctness of the text based on the embedding is doing much of the heavy lifting. GPT essentially handles generating text and dense representations of the text.
How are they at human performance? Almost everything GPT has read on the internet didn‘t even exist 200 years ago and was invented by humans. Heck, even most of the programming it does wasn‘t there 20 years ago.
Not every programmer starting from scratch would be brilliant, but many were self taught with very limited resources in the 80s form example and discovered new things from there.
GPT cannot do this and is very far from being able to.
Because it performs at least average human level (mostly well above average) on basically every task it's given.
"Invest something new" is a nonsensical benchmark for human level intelligence. The vast majority of people have never and will never invent anything new.
If your general intelligence test can't be passed by a good chunk of humanity then it's not a general intelligence test unless you want to say most people aren't generally intelligent.
I would argue some programmers do in fact invent something new. Not all of them, but some. Perhaps 10%.
Second the point is not whether everyone is by profession an inventor but whether most people can be inventors. And to a degree they can be. I think you underestimate that by a large margin.
You can lock people in a room and give them a problem to solve and they will invent a lot if they have the time to do it. GPT will invent nothing right now. It‘s not there yet.
Quality over quantity. Just building a model with a gazillion parameters isn't indicative of quality, you could easily have garbage parameters with tons of overfitting. It's like megapixel counts in cameras: you might have 2000 gigapixels in your sensor, but that doesn't mean you're going to get great photos out of it if there are other shortcomings in the system.
What overfitting? If anything, LLMs suffer from underfitting, not overfitting. Normally, overfitting is characterized by increasing validation loss while training loss is decreasing, and solved by early stopping (stopping before that happens). Effectively, all LLMs are stopped early, so they don't suffer from overfitting at all.
I don't disagree with you, these models may be underfitted, but overfitting is not explicitly defined by val vs. training loss, but rather how closely its output matches training data.
If you trained a MLP model where the number of parameters exceeded the data, it would be able to memorize the data and return a zero loss on training data. The larger the models are, the greater chance it memorizes the data, rather than the latent variables or distribution of the data.
Early LLMs, GPT2 (circa 2019) for example was definitely overfitting. I would frequently copy and paste output and find a reddit comment with the exact words.
Intelligence is the single most expensive resource on the planet. Hundreds of individuals have to be born, nurtured, and educated before you might get an exceptional 135+ IQ individual. Every intelligent person is produced at a great societal cost.
If you can reduce the cost of replicating a 135 IQ, or heck, even a 115 IQ person to a few thousand dollars, you're beating biology by a massive margin.
But we're still nowhere near that, or even near surpassing the skill of an average person at a moderately complex information task, and GPT-4 supposedly took hundreds of millions to train. It also costs a decent amount more to run inference on it vs. 3.5. It probably makes sense to prove the concept that generative AI can be used for lots of real work before scaling that up by another order of magnitude for potentially marginal improvements.
Also, just in terms of where to put your effort, if you think another direction (for example, fine-tuning the model to use digital tools, or researching how to predict confidence intervals) is going to have a better chance of success, why focus on scaling more?
There are a lot of employees at large tech consultancies that don't really do anything that can't be automated away by even current models.
Sprinkle in some more specific training and I can totally see entire divisions at IBM and Accenture and TCS being made redundant.
The incentive structures are perversely aligned for this future - the CEO who manages to reduce headcount while increasing revenue is going to be very handsomely rewarded by Wall Street.
Are intelligent people that valuable? There's lots of them at every university working for peanuts. They don't seem to be that valued by society, honestly.
If you ask any Fortune 500 CEO if he could magically take all the 135 IQ artists and academics and vagabonds, erase all their past traumas, put them through business or tech school, and put them to work in their company, they would all say 100% yes.
An equivalent AI won't have any agency and will be happy doing the boring work other 135 IQ humans won't.
IQ isn't all that. Mine is 140+ and I'm just a somewhat well paid software engineer. It's TOO abstract a metric in my view - for sure it doesn't always translate into real world success.
Right were very much in the same boat. I'm good at pattern recognition I guess. I learn things quickly. What else? I don't have magic powers really. I still get headaches and eat junk food.
Mine is 150-160 (varies by how much I’m sleep deprived during the IQ test) and I’m told that I’m exceptionally intelligent by teachers, friends, colleagues, most everyone I met since early childhood. I guess the more the difference to average, the more it stands out. From my experience, I believe higher IQ is nothing but better pattern recognition and being smart or genius means merely higher IQ + very good memory capability. I believe those two are interlinked [0]. By memory capability I mean not forgetting anything you’ve ever seen, not in the sense of being able to recall every minute of your entire life, but in the sense of reliably and always being able to recall all info regarding X if you’ve seen X only once in your life. Higher intelligence doesn’t mean you’re automatically better off than everyone also. It just means that you can be far ahead of everyone in any cognitive task with a far smaller amount of effort put in. Note that this doesn’t conflict with IQ being a reliable predictor of financial success, rather I believe this is the reason for it.
The reason we put everyone through school is we believe that it’s in society’s best interest to educate everyone to the peak of their abilities. It’s good for many different reasons.
It would be much easier to identify gifted kids and only educate them, but I happen to agree that universal education is better.
There’s downsides and tradeoffs but yes, if we wanted to we could stop trying to teach kids with below average IQs calculus, unless they specifically wanted to.
This only makes sense if you use "IQ" ignoring the actual definition of "IQ", in which case it's silly to use numbers in your post to make it look technical.
IQ 1. can't be compared against generations of IQ tests 2. supposedly doesn't test education (of course, it actually does) 3. isn't real.
I've been training large 65b models on "rent for N hours" systems for less than 1k per customized model. Then fine tuning those to be whatever I want for even cheaper.
2 months since gpt 4.
This ride has only just started, fasten your whatevers.
Finetuning cost are nowhere near representative of the cost to pre-train those models.
Trying to replicate the quality of GPT-3 from scratch, using all the tricks and training optimizations in the books that are available now but weren't used during GPT-3 actual training, will still cost you north of $500K, and that's being extremly optimistic.
GPT-4 level model would be at least 10x this using the same optimism (meaning you are managing to train it for much cheaper than OpenAI).
And That's just pure hardware cost, the team you need to actually makes this happen is going to be very expensive as well.
edit: To quantify how "extremely optimistic" that is, the very model you are finetuning, which I assume is Llama 65B, would cost around ~$18M to train on google cloud assuming you get a 50% discount on their listed GPU prices (2048 A100 GPUs for 5 months). And that's not even GPT-4 level.
As I stated in my comment, $5M is assuming you can do a much much better job than OpenAI at optimizing your training, only need to make a single training run, your employees salaries are $0, and you get a clean dataset for essentially free.
Real cost is 10-20x that.
That's still a good investment though. But the issue is you could very well sink $50M into this endeavour and end up with a model that actually is not really good and gets rendered useless by an open-source model that gets released 1 month later.
OpenAI truly has unique expertise in this field that is very, very hard to replicate.
No I'm not, it's the full model on 8 gpus for a couple hundred.
After training I fine tune for chats but mostly command and control tools, and then you fine-tune for application.
Gates has refuted saying this. Are you implying by analogy that Altman hasn't said/will disclaim saying that "the age of giant AI models is almost over"?
Just that there is tremendous hubris in the statement—at least when the statement stands alone. Vastly larger LLMs will probably become one or more relatively small components or layers of much larger systems that run on whatever we use as telephones in 20 years.
I suspect what he means is that OpenAI is finding diminishing returns from throwing money and hardware at larger models right now and that they are investigating other and/or composite AI techniques that make more optimal use of their hardware investment.
"...for the current cycle, in our specific public-facing market."
As most here well know "over" is one of those words like "never" which particularly in this space should pretty much always be understood as implicitly accompanied by a footnote backtracking to include near-term scope.
I'd bet that what he, and the competition, is realizing is that the bigger models are too expensive to run.
Pretty sure Microsoft swapped out Bing for something a lot smaller in the last couple of weeks; Google hasn't even tried to implement a publicly available large model. And OpenAI still has usage caps on their GPT-4.
I'd bet that they can still see improvement in performance with GPT-5, but that when they look at the usage ratio of GPT3.5 turbo, gpt3.5 legacy, and GPT4, they realized that there is a decreasing rate of return for increasingly smart models - most people don't need a brilliantly intelligent assistant, they just need a not-dumb assistant.
Obviously some practitioners of some niche disciplines (like ours here) would like a hyperintelligent AI to do all our work for us. But even a lot of us are on the free tier of ChatGPT 3.5; I'm one of the few paying $20/mo for GPT4; and idk if even I'd pay e.g. $200/mo for GPT5.
> I'd bet that what he, and the competition, is realizing is that the bigger models are too expensive to run.
I think it's likely that they're out of training data to collect. So adding more parameters is no longer effective.
> most people don't need a brilliantly intelligent assistant, they just need a not-dumb assistant.
I tend to agree, and I think their pathway toward this will all come from continuing advances in fine tuning. Instruction tuning, RLHF, etc seem to be paying off much more than scaling. I bet that's where their investment is going to be turning.
OpenAI has gone from open-sourcing its work, to publishing papers only, to publishing papers that omit important information, to GPT-4 being straight-up closed. And Sam Altman doesn't exactly have a track record of being overly concerned about the truth of his statements.
I had a fun conversation (more like argument) with ChatGPT about the hypocrisy of OpenAI. It would explicitly contradict itself and then began starting every reply with “I can see why someone might think…” and then just regurgitating fluff about democratizing AI. I finally was able to have it define democratization of technology and then recognize the absurdity of using that label to describe a pivot to gating models and being for-profit. Then it basically told me “well it’s for safety and protecting society”.
An AI, when presented with facts counter to what it thought it should say, agreed and basically went: “Won’t someone PLEASE think of the children!”
Why are you discussing OpenAI with ChatGPT? I’m honestly interested.
I would imagine that any answer of ChatGPT on that topic is either (a) „hallucinated“ and not based on any verifiable fact or (b) scripted in by OpenAI.
The same question pops up for me whenever someone asks ChatGPT about the internals and workings of ChatGPT. Am I missing something?
Simple curiosity. I wanted to see if it could explain the shift in OpenAIs operating in a way that might give some interesting or perhaps novel insight (even if hallucinated) other than what their corpo-speak public facing reasoning is.
For the most part it just regurgitated the corpo-speak with an odd sense of confidence. I know that’s the point of the model, but it can also be surprisingly honest when it incorporates what it knows about human motivation and business.
It’s pretty easy to have chatGPT contradict itself, point it out and have the LLM respond « well, I’m just generating text, nobody said it had to be correct »
It was trained on corpus full of mainstream media lies, why would you have expected otherwise? It's by far the most common deflection in its training set.
It's easy to recognize and laugh at the AI replying with the preprogrammed narrative, I'm still waiting for the majority of people realizing they are given the same training materials, non-stop, with the same toxic narratives, and becoming programmed in the same way, and that is what results in their current worldview.
And no, it's not enough to be "skeptic" of mainstream media. It's not even enough to "validate" them. Or to go to other sources. You need to be reflective enough to realize that they a pushing a flawed reasoning methods, and then abusing them again and again, to get you used to their brand of reasoning.
Their brand of reasoning is just basically reasoning with brands. You're given negative sounding words for things they want you to think are bad, and positive sounding words for things they want you to think are good, and continuously reinforce these connections. They brand true democracy (literally rule of the people) as populism and tell you it's a bad thing. They brand freedom of speech as "misinformation". They brand freedom as "choice" so that you will not think of what you want to do, but which of the things they allow you to do will you do. Disagree with the scientific narrative? You're "science denier". Even as a professional scientist. Conspiracy theory isn't a defined word - it is a brand.
You're trained to judge goodness or badness instinctively by their frequency and peer pressure, and produce the explanation after your instinctive decision, instead of the other way around.
"Then it basically told me “well it’s for safety and protecting society”."
That was pretty much OpenAI's argument when they first published that GPT-3 paper. "Oh no so scary people might use it for wrong stuff, only we should have control of it."
This trend has happened in the small for their APIs as well. They've been dropping options - the embeddings aren't the internal embeddings any more, and you don't have access to log probabilities. It's all closing up at every level.
I don't think these comments are driven from financial incentives. It's a distraction and only a fool would believe Altman here. What this likely means is they are prioritizing adding more features to their current models while they train the next version. Their competitors scramble to build an LLM with some sort of intelligence parity, when that happens no one will care because ChatGPT has the ecosystem and plugins and all the advanced features....and by the time their competitors reach feature parity in that area, OpenAI pulls its Ace card and drops GPT5. Rinse and repeat.
That's my theory and if I was a tech CEO in any of the companies competing in this space, that is what I would plan for.
Training an LLM will be the easy part going forward. It's building an ecosystem around it and hooking it up to everything that will matter. OpenAI will focus on this, while not-so-secretly training their next iterations.
text-davinci-003 but cheaper and runs on your own hardware is already a massive selling point. If you you release a foundational model at parity with GPT4 you'll win overnight because OpenAI's chat completions are awful even with the super advanced model.
In this case I think it's Wired that's lying. Altman didn't say large models have no value, or that there will be no more large models, or that people shouldn't invest in large models.
He said that we are at the end of the era where capability improvements come primarily from making models bigger. Which stands to reason... I don't think anyone expect us to hit 100T parameters or anything.
Like Altman said, it's comparable to the GHz race in the 1990's. If 4GHz is good, 5GHz is better, why not 10GHz?
Turns out there are diminishing returns and advances come from other dimensions. I've got no opinion on whether he's right or not, but he's certainly in a better position to opine that current scale has hit diminishing returns.
In any event, there's nothing special about 1T parameters. It's just a round base-10 number. It is no more magic than 900B or 1.3T.
But just look at what all Lincoln accomplished with 640KB of memory. In the grand examination of time, one might even say that Lincoln is a more important figure than ChatGPT itself.
Anyone with the expertise to have insightful takes in AI also has a financial incentive to steer the conversation in particular directions. This is also the case for many, many other fields! You do not become an expert by quarantining your livelihood away from your expertise!
The correct response is not to dismiss every statement from someone with a conflict of interest as "basically worthless", but to talk to lots of people and to be reasonably skeptical.
First of all, if Altman continually makes misleading statements about AI he will quickly lose credibility, and that short term gain in whatever 'financial incentive' that birthed the lie would be eroded in short order by a lack of trust of the head of one of the most visible AI companies in the world.
Secondly, all the competitors of OpenAI can plainly assess the truth or validity of Altman's statements. There are many companies working in tandem on things at the OpenAI scale of models, and they can independently assess the usefulness of continually growing models. They aren't going to take this statement at face value and change their strategy based on a single statement by OpenAI's CEO.
Thirdly, I think people aren't really reading what Altman actually said very closely. He doesn't say that larger models aren't useful at all, but that the next sea change in AI won't be models which are orders of magnitude bigger, but rather a different approach to existing problem sets. Which is an entirely reasonable prediction to make, even if it doesn't turn out to be true.
All in all, "his word is basically worthless" seems much to harsh an assessment here.
It is possible that GP meant that Altman’s word is basically worthless to them, in which case that’s not something that can be argued about. It’s a factually true statement that that is their opinion of that man.
I personally can see why someone could arrive at that position. As you’ve pointed out, taking Sam Altman at face value can involve suppositions about how much he values his credibility, how much stock OpenAI competitors put in his public statements, and the mindsets people in general have when reading what he writes.
I've seen Altman say in an interview that training GPT-4 took "hundreds of little things".
I don't find this implausible, but it folds slightly to Ockham's razor when you consider that this is the exact type of statement that would be employed to obfuscate a major breakthrough.
It just makes me crook my eyebrow and look to more credible sources.
Yeah, I also had a hunch he wasn't an AI. (I assume you meant "AI researcher" there :))
All joking aside, I wonder how that's affecting company morale or their ability to attract top researchers. I know if I was a top AI researcher, I'd probably rather work at a company where the CEO was an expert in the field (all else being equal).
It might be true in general; however, AI research laboratories are typically an exception, as they are often led by experienced AI researchers or scientists with extensive expertise in the field.
And that's why they have a hard time getting their stuff out there and getting the money they need. I mean, trying to run a business like a research lab is kind of flawed, you know? And you don't always want some Musk-like character messing around with the basics of the company
Honestly I'm not sure it matters that much. CEOs who are not experts or researches in a domain can still build great companies and empower their employees to do incredible work. Lots of tech people absolutely love to point out that Steve Jobs was not an engineer, but under his leadership the company invented three products that totally revolutionized different industries. Now, I'm not going to sit here and say Altman is Jobs, but running a company, knowing how to raise money, knowing how to productize technologies, etc are all very important skills that industry researchers aren't always good at.
Ilya gives numerous talks and interviews, and he's well worth listening to about technical matters. I listened to many of his talks recently, and the main theme is that scaling up compute works, and will continue to do so. His optimism about the potential of scaling to support deep learning has clearly guided his entire career, starting with his early success on AlexNet.
IIRC Altman has no financial stake in the success or failure of OpenAI to prevent these sorts of conflicts of interests between OpenAI and society as a whole
> OpenAI’s ChatGPT unleashed an arms race among Silicon Valley companies and investors, sparking an A.I. investment craze that proved to be a boon for OpenAI’s investors and shareholding employees.
> But CEO and co-founder Sam Altman may not notch the kind of outsize payday that Silicon Valley founders have enjoyed in years past. Altman didn’t take an equity stake in the company when it added the for-profit OpenAI LP entity in 2019, Semafor reported Friday.
Do you think GPT-4 was trained and then immediately released to the public? Training finished Aug 2022. They spent the next 6 months improving it in other ways (eg human feedback). What he is saying is already evident therefore.
Right. All the evidence points to more potential being left on the table for emergent abilities. It would make no sense that the model would develop all of these complex skills for better predicting the next token, then just stop.
It's a massive bet for a company to push compute into the billion dollar range - if saying something like this has the potential to help ward off those decisions, I don't see what's stopping them from saying it.
It's incredible that people are so eager to eat up these unsupported claims.
This is the second [1] OpenAI claim in the span of a few days that conveys a sense of "GPT-4 represents a plateau of accomplishment. Competitors, you've got time to catch up!".
And it's not just a financial incentive, it's a survival incentive as well. Given a sufficiently sized (unknowable ahead of time) lead, the first actor that achieves AGI and plays their cards right, can permanently suppress all other ongoing research efforts should they wish to.
Even if OpenAI's intentions are completely good, failure to be first could result in never being able to reach the finish line. It's absolutely in OpenAI's interest to conceal critical information, and mislead competing actors into thinking they don't have to move as quickly as they can.
Something kind of funny (but mostly annoying), about this announcement is the people arguing that OpenAI is, in fact, working on GPT-5 in secret.
To my knowledge, NFT/crypto hype never got so bad that conspiracy theories began to circulate (though I’m sure there were some if you looked hard enough).
Isn’t it obvious? Q is definitely an LLM, trained on trillions of words exfiltrated from our nation’s secure systems. This explains why it’s always wrong in its predictions: it’s hallucinating!
Hm, all right, I'm guessing that huge models as a business maybe are over until economics are figured out, but huge models as experts for knowledge distillation seems reasonable. And if you pay a super premium can you use huge model.
524 comments
[ 3.6 ms ] story [ 123 ms ] threadHe did not say what kind of research strategies or techniques might take its place. In the paper describing GPT-4, OpenAI says its estimates suggest diminishing returns on scaling up model size. Altman said there are also physical limits to how many data centers the company can build and how quickly it can build them.
I read the two papers (gpt 4 tech report, and sparks of agi) and in my opinion they don't support this conclusion. They don't even say how big GPT-4 is, because "Given both the competitive landscape and the safety implications of large-scale models like GPT-4, this report contains no further details about the architecture (including model size), hardware, training compute, dataset construction, training method, or similar."
> Altman said there are also physical limits to how many data centers the company can build and how quickly it can build them.
OK so his argument is like "the giant robots won't be powerful, but we won't show how big our robots are, and besides, there are physical limits to how giant of a robot we can build and how quickly we can build it." I feel like this argument is sus.
1. What terms could Microsoft wring out of OpenAI for another funding round?
Imagine when a dozen models are wired together and giving each other feedback with more clever training and algorithms on future faster hardware.
It is still going to get wild
FWIW we had thin clients in computer labs in middle school / high school 15 years ago (and still today these are common in enterprise environments, e.g. Citrix).
Biggest issue is network latency which is limited by the speed of light, so I imagine if computers in 10 years require resources not available locally it would likely be a local/cloud hybrid model.
Wouldn't these models hallucinate more than normal, then?
I agree that your suggested approach of applying cleverness to what we have now will probably produce better results. But that’s not going to stop better architectures, hardware and even entire regimes from being developed until we approach AGI.
My suspicion is that there’s still a few breakthroughs waiting to be made. I also suspect that sufficiently advanced models will make such breakthroughs easier to discover.
Really you’re still just predicting the next word, but with extra steps.
Really you're just switching switches on and off, but with extra steps.
Still good enough to upset the balance in search/ad market. Interesting times.
I suspect that a pause in base LLM performance won’t be an AI improvement pause; there’s a whole lot of space to improve the parts of AI systems around the core “brain in a jar” model.
far as AGI is concerned I dont believe LLMs are really the right architecture for it, AGI likely needs some symbolic logic and a notion of physicality (ie.. physical laws & energy/power).
It will reach a point where that is the case, sure; it is not there now, and if we are within one model generation of exhausting (for now) major core model improvements, I don’t think we’ll have reached the point of gradual incremental improvement from rest-of-system improvements yet.
but sir, that means the same thing
Throw this heretic into the pit of terror.
Fine, to the outhouse of madness then.
Before I get nuked from orbit for daring to entertain humor, if someone is running ahead of me in a marathon, and running so far ahead, yet still broadcasting things to the back for the slow people (like myself), then eventually we catch up to them, and they suddenly say, you know what guys, we should stop running in this direction, there's nothing to see here right before anyone else is able to verify the veracity of their statement, perhaps it would still be in the public interest for at least one person to verify what they are saying. Given how skeptical the internet at large has been of Musk's acquisition of a company, it's interesting that the skepticism is suddenly put on hold when looking at this part of his work...
although the title is a bit misleading on what he was actually saying. still, there's a lot left to go in terms of scale. Even if it isn't parameter size(and there's still lots of room here too, it just won't be economical), contrary to popular belief, there's lots of data left to mine
I'm pretty sure that GPT-4 is ~1T-2T parameters, and they're struggling to run it(at reasonable performance and profit). So far their strategy has been to 10x the parameter count every GPT generation, and the problem is that there's diminishing returns everytime they do that. AFAIK they've now resorted to chunking GPT through the GPUs because of the 2 to 4 terabytes of VRAM required (at 16bit).
So now they've reached the edge of what they can reasonably run, and even if they do 10x it the expected gains are less. On top of this, models like LLaMa have shown that it's possible to cut the parameter count substantially and still get decent results (albiet the opensource stuff still hasn't caught up).
On top of all of this, keep in mind that at 8bit resolution 175B parameters (GBPT3.5) requires over 175GB of VRAM. This is crazy expensive and would never fit on consumer devices. Even if you use quantization and use 4bit, you still need over 80GB of VRAM.
This definitely is not a "throw them off the trail" tactic - in order for this to actually scale the way everyone envisions both in performance and running on consumer devices - research HAS to be on improving the parameter count. And again there's lots of research showing its very possible to do.
tl;dr: smaller = cheaper+faster+more accessible+same performance
GPT3 on release was more expensive ($0.06/1000 tokens vs $0.03 input and $0.06 output for GPT4).
Reasonable to assume that in 1-2 years it will also come down in cost.
Definitely. I'm guessing they used something like quantization to optimize the vram usage to 4bit. The thing is that if you can't fit the weights in memory then you have to chunk it and that's slow = more gpu time = more cost. And even if you can fit it in GPU memory, less memory = less gpus needed.
But we know you _can_ use less parameters, and that the training data + RLHF makes a massive difference in quality. And the model size linearly relates to the VRAM requirements/cost.
So if you can get a 60B model to run at 175B's quality, then you've almost 1/3rd your memory requirements, and can now run (with 4bit quantization) on a single A100 80GB which is 1/8th the previously known 8x A100's that GPT-3.5 ran on (and still half GPT-3.5+4bit).
Also while openai likely doesn't want this - we really want these models to run on our devices, and LLaMa+finetuning has shown promising improvements (not their just yet) at 7B size which can run on consumer devices.
For example: you could use GPT to parse a resume file, pull out work experience and return it as JSON. That would take minutes to setup using the GPT API and it would take weeks to build your own system, but GPT is so expensive that building your own system is totally worth it.
Unless they can seriously reduce how expensive it is I don't see it replacing many existing solutions. Using GPT to parse text for a repetitive task is like using a backhoe to plant flowers.
True, but an HR SaaS vendor could use that to put on a compelling demo to a potential customer, stopping them from going to a competitor or otherwise benefiting.
And anyway, without churning the numbers, for volumes of say 1M resumes (at which point you've achieved a lot of success) I can't quite believe it would be cheaper to build something when there is such a powerful solution available. Maybe once you are at 1G resumes... My bet is still no though.
I'd love to be able to just have people submit their resume's and extract the data from there, but instead I'm going to build a form and make applicants fill it out because chatGPT is going to be at least $0.05USD depending on the length of the resume.
I'd also love to have mini summeries of order returns summerized in human form, but that also would cost 0.05USD per form.
the tl;dr here is that there's a TON of usecases for a LLM outside of your core product (we sell clothes) - but we can't currently justify that cost. Compare that to the rapidly improving self-hosted solutions which don't cost 0.05USD for literally any query (and likely more for anything useful).
And yeah I agree this would be a great use-case, and isn't that expensive.
I'd like to do this in lots of places, and the problem is I have to convince my boss to pay for something that otherwise would have been free.
The conversation would be "We have to add these fields to our model, and we either tell django to add a form for them, which will have 0 ongoing cost and no reliance on a third party,
or we send the resume to openai, pay for them to process it, make some mechanism to sanity check what GPT is responding with, alert us if there's issues, and then put it into that model, and pay 5 cents per resume."
> 1-3 hours of a fully loaded engineers salary per year.
That's assuming 0 time to implement, and because of our framework it would take more hours to implement the openai solution (that's also more like 12 hours where we are).
> $500 per 10k.
I can't stress this enough - the alternative is 0$ per 10k. My boss wants to know why we would pay any money for a less reliable solution (GPT serialization is not nearly as reliable as a standard django form).
I think within the next few years we'll be able to run the model locally and throw dozens of tasks just like this at the LLM, just not yet.
I have tried GPT3.5 and GPT4 for this type of task - the "near perfect results" is really problematic because you need to verify that it's likely correct, notify you if there's issues, and even then you aren't 100% sure that it selected the correct first/last name.
This is compared to a standard html form. Which is.... very reliable and (for us) automatically has error handling built in, including alerts to us if there's a 504.
To show an MVP to a customer you only need 10 resumes (or 1 in most demos I've been in).
So 50c.
This is why Dall-e 2 ran in a data centre and Stable Diffusion runs on a gamer GPU
> This is why Dall-e 2 ran in a data centre and Stable Diffusion runs on a gamer GPU
This is absolutely why they're keeping it locked up. By simply not releasing the weights, you can't run Dalle2 locally, and yeah they don't want to do this because they want you to be locked to their platform, not running it for free locally.
On the other hand though, Chinchilla and multimodal approaches already showed how later AIs can be improved beyond throwing petabytes of data at them.
It is all about variety and quality from now on I think. You can teach a person all about the color zyra but without actually ever seeing it, they will never fully understand that color.
Like, for most of the time I’m using it, Copilot saves me 30 seconds here and there and it takes me about a second to look at the line or two of code and go “yeah, that’s right”. It adds up, especially when I’m working with an unfamiliar language and forget which Collection type I’m going to need or something.
I've never used Copilot but I've tried to replace StackOverflow with ChatGPT. The difference is, the StackOverflow responses compile/are right. The ChatGPT responses will make up an API that doesn't exist. Major setback.
I've personally been using it to explore using different libraries to produce charts. I managed to try out about 5 different libraries in a day with fairly advanced options for each using chatGPT.
I might have spent a day in the past just trying one and not to the same level of functionality.
So while it still took me a day, my final code was much better fitted to my problem with increased functionality. Not a time saver then for me but a quality enhancer and I learned a lot more too.
Eh. An outdated answer will be called out in the comments/downvoted/updated/edited more often than not, no?
> YouTubers upload about 720,000 hours of fresh video content per day. Over 500 hours of video were uploaded to YouTube per minute in 2020, which equals 30,000 new video uploads per hour. Between 2014 and 2020, the number of video hours uploaded grew by about 40%.
If a scrape of the general internet, scientific papers and books isn’t enough, a trillion trillion trillion text messages to mom aren’t going to change matters.
There was a rumor that they were going to use Whisper to transcribe YouTube videos and use that for training. Since it's multimodal, incorporating video frames alongside the transcriptions could significantly enhance its performance.
But I am quite sure that if you start doing it at scale, google will notice.
You could be sneaky, but people in this business talk (since they know another good paying job is just around the corner) so It would likely come out.
Google is also owns a lot of it's own backbone, so It would be a lot easier for them to play network games.
And they could even try to be sneaky and try poisoning the data if it comes to that.
And since OpenAI konw that, since they probably have people that used to work at google at some point, they are unlikely to try.
Even less likely if Microsoft would know. MS is probably the only company that has even more layers than Oracle and they would not approve.
The data is not finite.
That model might be very well tuned to solve IBM's internal problems.
I also question that most companies have the volume and quality of data worth training on. It's littered with cancelled projects, old products, and otherwise obsolete data. That's going to make your LLM hallucinate/give wrong answers. Especially for regulated and otherwise legally encumbered industries. Like can you deploy a chat bot that's wrong 1% or 0.1% of the time?
You have to understand that all the incentives are perfectly aligned for corporations to put this to work, even spending tens of millions in getting it right.
The first corporate CEO who announces that his company used AI to reduce employee costs while increasing profits is going to get such a fat bonus that everyone will follow along.
This also makes me doubt that NSA hasn't already cracked this problem. Or that China won't eventually beat current western models since it will likely have way more data collected from its citizenry.
Then again it would give you data on every accent in the country, so the holy grail for modelling human speech.
They might have trained on a lot of the 'high quality' tokens, however.
Dataset size is not relevant to predicting the loss threshold of LLMs. You can keep pushing loss down by using the same sized dataset, but increasingly larger models.
Or augment the dataset using RLHF, which provides an "infinite" dataset to train LLMs on. Limited by the capabilities of the scoring model which, of course, you can scale the scoring model infinitely so again the limit isn't dataset size but training compute.
Deepmind and others would disagree with you! No-one really knows in actual fact.
[1] https://www.deepmind.com/publications/an-empirical-analysis-...
Merely pointing out that the debate as to whether we are compute or data limited (OP) has not concluded at all; There are lots of compelling theories on relationship between the two.
It can start posting synthesized ideas on social media and see how many likes it gets. Coupled with a metric containing dissimilarity to current information, this could be a useful way to progress to superhuman insights.
Maybe if you trained it on movies before CGI existed ?
There's a ton of potential left on the table. The question is if transformers have hit their limit with GPT-4 or not.
It's a pretty simple equation when you think about it this way and why Sam would say they have hit their limit. Sam is basically Microsoft and they want to retain their lead. Once Google learns to put their data to use correctly, it's almost guaranteed game over for OpenAI if they want it to be.
In short it seems like virtually all of the improvement in future AI models will come from better algorithms, with bigger and better data a distant second, and more parameters a distant third.
Of course, this claim is itself internally inconsistent in that it assumes that new algorithms won't alter the returns to scale from more data or parameters. Maybe a more precise set of claims would be (1) we're relatively close to the fundamental limits of transformers, i.e., we won't see another GPT-2-to-GPT-4-level jump with current algorithms; (2) almost all of the incremental improvements to transformers will require bigger or better-quality data (but won't necessarily require more parameters); and (3) all of this is specific to current models and goes out the window as soon as a non-transformer-based generative model approaches GPT-4 performance using a similar or lesser amount of compute.
What algorithms specifically show the most results upon improvement? Going into this I thought the jump of improvements were really related more advanced automated tuning and result correction, in which it could be done at scale as it were allowing a small team of data scientists to tweak the models until desired results were being achieved.
Are you saying instead, that concrete predictive algorithms need improvement or are we lumping the tuning into this?
[0] https://hazyresearch.stanford.edu/blog/2023-03-27-long-learn...
"Sammy A thinks we've made the best engine with the tools at hand" -> "this will never get us out of the solar system"
Sorry to unload on you. It is frustrating to constantly see AGI get brought up needlessly on HN
> Are you saying instead, that concrete predictive algorithms need improvement or are we lumping the tuning into this?
in the context of what's needed to get to AGI - just as if NASA built an engine we'd talk about its effectiveness in the context of space flight.
Separately, I think OpenAI's current investors have a >10% chance to hit the 100x cap on their returns. Their current models are already good enough to address lots of real-world problems that people will pay money to solve. So far they've been much more model-focused than product-focused, and by turning that dial toward the product side (as they did with ChatGPT) I think they could generate a lot of revenue relatively quickly.
[0] Except maybe in the sense that future models will be predominantly multimodal and therefore not strictly LLMs. I don't think that's what you're suggesting though.
Especially as a differentiation for a company. If everyone is using ChatGPT, then they're all offering the same thing and I can just as well go to the source and cut out the middleman.
The other fun development to come is well performing self hosted models, and the idea of light weight domain specific interface models that curate responses from bigger generalist models.
ChatGPT is fun but it is very general, it doesn't know about my business nor keep track of it or interface with it. I fully expect to see "Expert Systems" of old come back, but trained on our specific businesses.
I don't think there is such a clear separation between algorithms and data as your comment suggests.
But I think not, since monkeys probably don't "improve" noticeably with time or input.
Maybe once tons of bananas are introduced...
So to reiterate, he is not saying that the age of giant AI models is over. Current top-of-the-line AI models are giant and likely will continue to be. However, there's not point in training models you can't actually run economically. Inference costs need to stay grounded which means practical model sizes have a limit. More effort is going to go into making models efficient to run even if it comes at the expense of making them less efficient to train.
For one thing they're already at human performance.
For another, i don't think you realize how expensive inference can get. Microsoft with no scant amount of available compute is struggling to run gpt-4 such that they're rationing it between subsidiaries while they try to jack up compute.
So saying, it would be economically sound if it cost x10 or x100 what it costs now is a joke.
Because, yeah, “brain in a jar” GPT isn’t enough for most tasks beyond parlor-trick chat, but being used as a brain in a jar isn’t the point.
Plus the recursive self prompting to improve accuracy.
Not every programmer starting from scratch would be brilliant, but many were self taught with very limited resources in the 80s form example and discovered new things from there.
GPT cannot do this and is very far from being able to.
Because it performs at least average human level (mostly well above average) on basically every task it's given.
"Invest something new" is a nonsensical benchmark for human level intelligence. The vast majority of people have never and will never invent anything new.
If your general intelligence test can't be passed by a good chunk of humanity then it's not a general intelligence test unless you want to say most people aren't generally intelligent.
I would argue some programmers do in fact invent something new. Not all of them, but some. Perhaps 10%.
Second the point is not whether everyone is by profession an inventor but whether most people can be inventors. And to a degree they can be. I think you underestimate that by a large margin.
You can lock people in a room and give them a problem to solve and they will invent a lot if they have the time to do it. GPT will invent nothing right now. It‘s not there yet.
Lol Okay
>And to a degree they can be. I think you underestimate that by a large margin.
Do i? Because i'm not the one making unverifiable claims here.
>You can lock people in a room and give them a problem to solve and they will invent a lot if they have the time to do it.
If you say so
Just listen to what you're saying:
- GPT isn't at human level because GPT isn't able to invent something new
- Not all programmers invent something new, but some. Perhaps 10%
I'm pretty sure this implies literally that 90% programmers aren't human level.
The lengths to which people are willing to go to dismiss GPT's abilities is mind boggling to me.
No, GPT4 fails at some very basic tasks. It can't count letters passed 15.
If you trained a MLP model where the number of parameters exceeded the data, it would be able to memorize the data and return a zero loss on training data. The larger the models are, the greater chance it memorizes the data, rather than the latent variables or distribution of the data.
Early LLMs, GPT2 (circa 2019) for example was definitely overfitting. I would frequently copy and paste output and find a reddit comment with the exact words.
Intelligence is the single most expensive resource on the planet. Hundreds of individuals have to be born, nurtured, and educated before you might get an exceptional 135+ IQ individual. Every intelligent person is produced at a great societal cost.
If you can reduce the cost of replicating a 135 IQ, or heck, even a 115 IQ person to a few thousand dollars, you're beating biology by a massive margin.
Also, just in terms of where to put your effort, if you think another direction (for example, fine-tuning the model to use digital tools, or researching how to predict confidence intervals) is going to have a better chance of success, why focus on scaling more?
Sprinkle in some more specific training and I can totally see entire divisions at IBM and Accenture and TCS being made redundant.
The incentive structures are perversely aligned for this future - the CEO who manages to reduce headcount while increasing revenue is going to be very handsomely rewarded by Wall Street.
Edit: I don’t understand the downvotes. I don’t mean this in any disparaging way, just that an AGI is probably going to be a lot higher than that.
Even if we don't expend any cost on education the number of people with IQ 135 stays the same.
An equivalent AI won't have any agency and will be happy doing the boring work other 135 IQ humans won't.
[0]: https://saveall.ai/blog/learning-is-remembering
It would be much easier to identify gifted kids and only educate them, but I happen to agree that universal education is better.
Is it so easy?
IQ 1. can't be compared against generations of IQ tests 2. supposedly doesn't test education (of course, it actually does) 3. isn't real.
2 months since gpt 4.
This ride has only just started, fasten your whatevers.
Trying to replicate the quality of GPT-3 from scratch, using all the tricks and training optimizations in the books that are available now but weren't used during GPT-3 actual training, will still cost you north of $500K, and that's being extremly optimistic.
GPT-4 level model would be at least 10x this using the same optimism (meaning you are managing to train it for much cheaper than OpenAI). And That's just pure hardware cost, the team you need to actually makes this happen is going to be very expensive as well.
edit: To quantify how "extremely optimistic" that is, the very model you are finetuning, which I assume is Llama 65B, would cost around ~$18M to train on google cloud assuming you get a 50% discount on their listed GPU prices (2048 A100 GPUs for 5 months). And that's not even GPT-4 level.
Real cost is 10-20x that.
That's still a good investment though. But the issue is you could very well sink $50M into this endeavour and end up with a model that actually is not really good and gets rendered useless by an open-source model that gets released 1 month later.
OpenAI truly has unique expertise in this field that is very, very hard to replicate.
ahem Bard ahem
[1] https://hazyresearch.stanford.edu/blog/2023-03-27-long-learn...
Bill Gates
I suspect what he means is that OpenAI is finding diminishing returns from throwing money and hardware at larger models right now and that they are investigating other and/or composite AI techniques that make more optimal use of their hardware investment.
It costs the same to train as these giant models. You merely spend they money on training it for longer instead of larger.
As most here well know "over" is one of those words like "never" which particularly in this space should pretty much always be understood as implicitly accompanied by a footnote backtracking to include near-term scope.
Pretty sure Microsoft swapped out Bing for something a lot smaller in the last couple of weeks; Google hasn't even tried to implement a publicly available large model. And OpenAI still has usage caps on their GPT-4.
I'd bet that they can still see improvement in performance with GPT-5, but that when they look at the usage ratio of GPT3.5 turbo, gpt3.5 legacy, and GPT4, they realized that there is a decreasing rate of return for increasingly smart models - most people don't need a brilliantly intelligent assistant, they just need a not-dumb assistant.
Obviously some practitioners of some niche disciplines (like ours here) would like a hyperintelligent AI to do all our work for us. But even a lot of us are on the free tier of ChatGPT 3.5; I'm one of the few paying $20/mo for GPT4; and idk if even I'd pay e.g. $200/mo for GPT5.
I think it's likely that they're out of training data to collect. So adding more parameters is no longer effective.
> most people don't need a brilliantly intelligent assistant, they just need a not-dumb assistant.
I tend to agree, and I think their pathway toward this will all come from continuing advances in fine tuning. Instruction tuning, RLHF, etc seem to be paying off much more than scaling. I bet that's where their investment is going to be turning.
Altman has a financial incentive to lie and obfuscate about what it takes to train a model like GPT-4 and beyond, so his word is basically worthless.
An AI, when presented with facts counter to what it thought it should say, agreed and basically went: “Won’t someone PLEASE think of the children!”
Love it.
I would imagine that any answer of ChatGPT on that topic is either (a) „hallucinated“ and not based on any verifiable fact or (b) scripted in by OpenAI.
The same question pops up for me whenever someone asks ChatGPT about the internals and workings of ChatGPT. Am I missing something?
I was curious about state persistence between prompt, or how to get my prompt better, or having a idea of the training data.
Only got crap and won’t spend time doing that again
For the most part it just regurgitated the corpo-speak with an odd sense of confidence. I know that’s the point of the model, but it can also be surprisingly honest when it incorporates what it knows about human motivation and business.
It’s pretty easy to have chatGPT contradict itself, point it out and have the LLM respond « well, I’m just generating text, nobody said it had to be correct »
It's easy to recognize and laugh at the AI replying with the preprogrammed narrative, I'm still waiting for the majority of people realizing they are given the same training materials, non-stop, with the same toxic narratives, and becoming programmed in the same way, and that is what results in their current worldview.
And no, it's not enough to be "skeptic" of mainstream media. It's not even enough to "validate" them. Or to go to other sources. You need to be reflective enough to realize that they a pushing a flawed reasoning methods, and then abusing them again and again, to get you used to their brand of reasoning.
Their brand of reasoning is just basically reasoning with brands. You're given negative sounding words for things they want you to think are bad, and positive sounding words for things they want you to think are good, and continuously reinforce these connections. They brand true democracy (literally rule of the people) as populism and tell you it's a bad thing. They brand freedom of speech as "misinformation". They brand freedom as "choice" so that you will not think of what you want to do, but which of the things they allow you to do will you do. Disagree with the scientific narrative? You're "science denier". Even as a professional scientist. Conspiracy theory isn't a defined word - it is a brand.
You're trained to judge goodness or badness instinctively by their frequency and peer pressure, and produce the explanation after your instinctive decision, instead of the other way around.
Excellent post, especially this part. Sums up the problem perfectly.
That was pretty much OpenAI's argument when they first published that GPT-3 paper. "Oh no so scary people might use it for wrong stuff, only we should have control of it."
In reality they just saw $$$ blink.
That's my theory and if I was a tech CEO in any of the companies competing in this space, that is what I would plan for.
Training an LLM will be the easy part going forward. It's building an ecosystem around it and hooking it up to everything that will matter. OpenAI will focus on this, while not-so-secretly training their next iterations.
He said that we are at the end of the era where capability improvements come primarily from making models bigger. Which stands to reason... I don't think anyone expect us to hit 100T parameters or anything.
Turns out there are diminishing returns and advances come from other dimensions. I've got no opinion on whether he's right or not, but he's certainly in a better position to opine that current scale has hit diminishing returns.
In any event, there's nothing special about 1T parameters. It's just a round base-10 number. It is no more magic than 900B or 1.3T.
- Abraham Lincoln
The correct response is not to dismiss every statement from someone with a conflict of interest as "basically worthless", but to talk to lots of people and to be reasonably skeptical.
Secondly, all the competitors of OpenAI can plainly assess the truth or validity of Altman's statements. There are many companies working in tandem on things at the OpenAI scale of models, and they can independently assess the usefulness of continually growing models. They aren't going to take this statement at face value and change their strategy based on a single statement by OpenAI's CEO.
Thirdly, I think people aren't really reading what Altman actually said very closely. He doesn't say that larger models aren't useful at all, but that the next sea change in AI won't be models which are orders of magnitude bigger, but rather a different approach to existing problem sets. Which is an entirely reasonable prediction to make, even if it doesn't turn out to be true.
All in all, "his word is basically worthless" seems much to harsh an assessment here.
I personally can see why someone could arrive at that position. As you’ve pointed out, taking Sam Altman at face value can involve suppositions about how much he values his credibility, how much stock OpenAI competitors put in his public statements, and the mindsets people in general have when reading what he writes.
I don't find this implausible, but it folds slightly to Ockham's razor when you consider that this is the exact type of statement that would be employed to obfuscate a major breakthrough.
It just makes me crook my eyebrow and look to more credible sources.
Yeah, I also had a hunch he wasn't an AI. (I assume you meant "AI researcher" there :))
All joking aside, I wonder how that's affecting company morale or their ability to attract top researchers. I know if I was a top AI researcher, I'd probably rather work at a company where the CEO was an expert in the field (all else being equal).
> OpenAI’s ChatGPT unleashed an arms race among Silicon Valley companies and investors, sparking an A.I. investment craze that proved to be a boon for OpenAI’s investors and shareholding employees.
> But CEO and co-founder Sam Altman may not notch the kind of outsize payday that Silicon Valley founders have enjoyed in years past. Altman didn’t take an equity stake in the company when it added the for-profit OpenAI LP entity in 2019, Semafor reported Friday.
It's a massive bet for a company to push compute into the billion dollar range - if saying something like this has the potential to help ward off those decisions, I don't see what's stopping them from saying it.
I basically see Microsoft talking when Sam talks.
This is the second [1] OpenAI claim in the span of a few days that conveys a sense of "GPT-4 represents a plateau of accomplishment. Competitors, you've got time to catch up!".
And it's not just a financial incentive, it's a survival incentive as well. Given a sufficiently sized (unknowable ahead of time) lead, the first actor that achieves AGI and plays their cards right, can permanently suppress all other ongoing research efforts should they wish to.
Even if OpenAI's intentions are completely good, failure to be first could result in never being able to reach the finish line. It's absolutely in OpenAI's interest to conceal critical information, and mislead competing actors into thinking they don't have to move as quickly as they can.
[1] https://news.ycombinator.com/item?id=35570690
To my knowledge, NFT/crypto hype never got so bad that conspiracy theories began to circulate (though I’m sure there were some if you looked hard enough).
Can’t wait for an AIAnon community to emerge.