LLMs aren't "trained on the internet" anymore
allenpike.com
allenpike.com
And Phi-3 is something else, even from relatively limited time playing with it, so that’s useful signal for anyone who hadn’t looked at it yet. Wildly cool stuff.
It seems weird to not mention Anthropic or Mistral or FAIR pretty much at all: they’re all pretty clearly on more modern architectures at least as concerns capability per weight and Instruct-style stuff. I’m part of a now nontrivial group who regards Opus as basically shattering GPT-4-{0125, 1106}-Preview (which is basically the same as 4o for pure language modalities) on basically everything I care about, and LLaMA3 is just about there as well, maybe not quite Opus, comparable if you ignore trivially gamed metrics like MMLU.
And I have no idea why we’re talking about GPT-5 when there’s little if any verifiable evidence it even exists as a training run tracking to completion. Maybe it is, maybe not, but let’s get a look at it rather than just assume that it’s going to lap the labs that are currently pushing the pace now?
Opus historically had issues with minor typographical errors, though recently that seems to not happen often, lots of very sharp people at Anthropic.
So a month ago if I wanted something from Opus I’d run it through a cleanup pass courtesy of one of the other ones, but even my old standby dolphin-8x7 can clean up typos. 1106 can as well, but all else equal I don’t want to be sending my stuff to any black box data warehouse and I’m always surprised so many other sophisticated people don’t share the preference.
My personal eyeball capability check is to posit a gauge symmetry and ask what it thinks the implied conserved quantity is, and I’ve yet to see Opus not crush that relative to anything else, including real footnotes.
On coding I usually hand it a Prisma schema and ask for a proto3/gRPC definition that is a good way to interact with it, Opus in my personal experience also dominates there.
If you have an example of a task that represents a counter example I’d be grateful for another little integration test for my personal ad-hoc model card. I want to know the best tool for every job.
I really don't like how condescending your root comment is, I don't have any idea how all these extra things you're disappointed in are relevant at all to the actual topic.
I’d likewise be grateful if you’d lend me an example or two that should be on my little ad-hoc task set.
Assuming the former, here's an example I had yesterday where Opus was never able to formulate a correct response even with multiple follow-ups but GPT-4o got it with only two follow ups:
> I have a postgres database with a table called projects. On that table is a column called designs that is of json type and is nullable as well as a column called id that is of type text. The data in the designs column will be an array of objects (assuming it's not null). Each object in the array will have a field called price that will be a string. Most of these prices will be in the format $#,### where # is a number. However some will be in the format # ### US$. This second format where the numbers are separated by spaces instead of commas and the dollar sign appears at the end of the string instead of the beginning with an extraneous US is incorrect. What SQL can I run to migrate all prices that are incorrectly formatted to the correct format?
Alternatively, this is a non-tech related one that GPT-4 gets correct that no other model does:
> In 388 BC, the boxer Eupolus of Thessaly defeated three opponents at the ancient Olympics. Funded by the men involved, several bronze statues of Zeus were erected, with inscriptions detailing the events. Why did the four men prefer not to have the statues around?
I like that one because Opus will tell you that you're wrong and mistaken, which is pretty funny while GPT-4 answer correctly with the actual context.
However, if I start with: "What is the earliest record of cheating in the olympic games?" Then all models get the question right. It's surprising that GPT-4 gets it right on the first go.
But it’s not by a ton, and it could be just less hyper-scale network infrastructure.
I don’t recall what Anthropic has raised but IIRC it wasn’t “peer with everyone” levels.
[1] https://www.bloomberg.com/news/articles/2023-12-21/anthropic...
[2] https://www.anthropic.com/news/anthropic-series-c
[3] https://www.aboutamazon.com/news/company-news/amazon-anthrop...
It's great that we get to keep saying this. I wonder if that's because we have no objective statistics to measure these projects by.
Make it easy to try everything, let people decide for themselves what works best for them.
Ya know, like a market.
An actual market requires informed consumers and standardized metrics between competitors. They also typically include publications about the market positioned towards the consumer, almost like ya know, Consumer Reports.
https://openai.com/index/openai-board-forms-safety-and-secur...
But this is not the reality of modern LLMs by a long shot, they are trained in increasingly large parts from custom built datasets that are created by countless paid individuals, hidden behind stringent NDAs.
The author here seems to see that as a strength, an opportunity for unbounded growth and potential, I think this is the opposite, this approach is close to a gigantic whack a mole game, effectively unbounded, but in the wrong way.
Something that is new this time around, AFAIK, is that we haven’t previously had general ML systems that businesses and consumers are paying billions of dollars a year to use. So if, say, 10% of revenue goes back in to making better data sets every year, I can imagine continued improvement on certain economically valuable use cases – though likely with diminishing returns.
If this is so prevalent, wouldn't there be a proportional amount of data leaks? If there any particular evidence, even of doubtful authenticity, of this being the case?
Are you talking about leaks of the actual training data? The secret sauce of modern LLMs? That is like leaking the google3 source code or the recipe for Coca Cola - a ludicrously risky move. And for what gain?
Do they ask PhD to explain root of negative 1 and why is it complex? Is it like Quora but private only and high quality answers by top researchers in their respective field?
> Do they ask PhD to explain root of negative 1 and why is it complex?
This isn't necessarily impossible, but I would consider it to be infeasible with existing labelling workforces. Of course if you really needed a dataset like this and you were sufficiently resourced and willing to spend, you could maybe make it work (I would question whether you really needed PhDs though, that might be hard to swing at any price point).
But the core idea behind your question is correct - this is what a dataset might look like and hiring/contracting appropriately-skilled people and asking them to do repetitive tasks with some guidance is how you would go about getting it. Depending on the need, it can be quite a bit more complex too - if you needed self-driving car driving behavior data maybe you build a simulator and hire people to drive in the simulator and use that as training data (made up and probably crap example, but it illustrates the possibilities).
Some people think that labelling workforces are all low skill and there is a lot of good things low skill workforces can do well (visual stuff, basic language and emotion tasks), but you might be surprised at the ability to get skilled labelers. There are lots of smart/educated people around the world and there is ridiculous amounts of money flowing into this space.
Or, as the case may be... humans are biased? Also "generate some better examples" sounds like fudging data to fit the expected outcome. It smells of clutching at straws hoping to come up with something before the world looses interests and investor money runs out.
If you want to see how LLMs fail at coming up with original responses ask your favourite hallucinating bot to come up with fifty different ways of encouraging people to "Click the Subscribe button" in a YT video. Not only it will not come up with anything original, but it will simply start repeating itself (well, not itself, it will start repeating phrases found in YT video transcripts).
LLMs are tools. As a tool author, you have certain desired outcomes for certain use cases. If the current data you’re training on isn’t giving you those outcomes, it is absolutely reasonable to "fudge" the data. This might mean reducing bias, or adding bias, or any number of nudges. Training an LLM is not a scientific study, it’s a product development effort.
The article states that this human data is PhDs, poets, and other experts but my recollection from some info about programming LLM training is that there was a small army of low paid Indian programmers feeding it with data.
Even if it's actually experts now I have to wonder when that will switch to 3rd worlders making $1/hour.
The problem with hiring that group of people is presumably that they are not poor enough to lack ambition in their career, which every dummy can spot from miles away is an utter dead end feeding some LLM.
The difference between trusting 5 random people to tell you how they think quantum mechanics works versus asking 5 presently publishing physicists.
Does no one remember why expert systems fell apart? Because you have to keep paying experts to feed the beast. Because they are bound to the whims and limitations of experts. Making up data isnt going to get us there, we already failed with this method ONCE.
Open AI's bet with MS and the resignation of all the safety people says everything you need to know. MS gets everthin up to AGI... IF you thought you were close, if you thought that you were going to get there with a bigger model and more data then you MIGHT want MS's money. And MS had its own ML folks publish papers with "hints of AGI", The google engineer saying "it's AGI" before getting laughed at...
I suspect that everyone at OpenAI was high on their own supply. That they thought AGI would emerge, or sapience, or sentience if they shoved enough data at it. I think the safety minding folks leaving points to the fact that they found the practical limitations.
Show me the paper that has progress on hallucination. Show me the paper that doubles effectiveness and halves the size. These are where we need progress for this to become more than grift, than NFT's.
LLM's have pretty clearly been the most rapidly advancing technology in the history of humankind. Are you not entertained?!
Many of the current generation of AI experts mostly either did not pay great attention to the history of AI or they believe this time is completely different. They would do well to spend more time learning about history.
However, your view doesn't strike me as correct either. Expert system fell apart because the world was more complex than researchers realized and enumeration was essentially discovered to be infeasible (more or less as you say). But the impossibility of enumerating the world isn't news, everyone knows "the bitter lesson". And this isn't the past - now everyone on earth carries around a computer, a video camera and a microphone. They talk to each other through the internet. Remote workers screens' are recorded. Billions of vehicles with absurd numbers of sensors are roaming around the world. More of the arenas that matter to humanity are digital and thus effective domains for automated exploration and data generation.
The information about how the world operates exists or can be generated, the only real question is how to get your hands on it.
I'm sure I could read all the information for an astrophysics course in a relatively short time. Understanding it is a different matter.
The other use case is to make them hallucinate tropey stories and concepts for brainstorming, this isn't nearly as useful though due to it adding so much low quality stuff here when it hallucinates.
The hubris of mathematics. At what scale does whether prediction become 100 percent accurate? How large of a model do you need, and how big of a computer to run it?
Do we thing that reducing the world to a model and feeding it through (what isnt even close to a model) of "thought" or "interaction" or ... what ever you want to bill and LLM as is going to be any more accurate than weather prediction?
It remains to be seen, but the days where I scoffed at that idea are firmly in the past where they belong. Today we are building machines with intelligence high enough that it is forcing us to reconsider and redefine what intelligence is. And there is a huge amount of progress just sitting in front of us, waiting to be fed into models.
There were many reasons. One of them was the "Knowledge Acquisition Bottleneck", but that was not about the cost of paying experts, rather the cost of creating and maintaining a potentially very large knowledge base (i.e. one big mother of a database of production rules). Also, the fact that many experts' knowledge is tacit and not easily formalisable.
Modern machine learning began in the 1980's as an effort to overcome the Knowledge Acquisition Bottleneck. Accordingly many early machine learning approaches were designed to learn production rules for expert systems. Decision trees, one of the staple classifiers in data science, come from that era; you can tell, because decision trees are a symbolic, logic-based "model".
There were other problems with expert systems, e.g. their infamous "brittleness". But modern, statistical machine learning systems, are also criticised for "brittleness" too.
There were also purely political reasons and nothing to do with science or technology considerations. Then there was the 5th Generation Computer Project, and the AI winter, and then there were no more expert systems.
The journal of Expert Systems with Applications is still alive and well, on the other hand, although it mostly publishes on machine learning and neural nets these days. With an occasionally cool article, like one about Wolf Colony Optimisation I spotted recently. Too tired to look for links now, sorry.
And I was like, y'saying, it's a zipped list of edge cases...
Those PHDs better up their negotiating skills then.
So the headline is bullshit then.
The thrust of this article is essentially “LLMs are trained on the internet, _but wait_, they are also trained on other stuff in these very rarified and specific cases”. So even you concede that, LLMs are still, for the most part, trained on the internet.
(Apologies to Mitch Hedberg. https://www.youtube.com/watch?v=VqHA5CIL0fg )
My core point is that the “Trained On the Internet” mental model is becoming less true over time, which makes it a poor model for predicting the long term performance of models. These titles would be better:
1. LLMs Aren’t Just “Trained On the Internet” Anymore
2. LLMs Aren’t Simply Being “Trained On the Internet”
3. Future LLMs Won’t Just Be “Trained On the Internet”
I’ve swapped in the first one. Thanks for the feedback.
> Common Crawl (filtered) 410 billion tokens 60% of GPT-3 training data but only 44% of it was used i.e. 0.44 epochs (from the paper published May 28, 2020)
From Aligning language models to follow instructions January 27, 2022 [0]
> ...these models can also generate outputs that are untruthful, toxic, or reflect harmful sentiments. This is in part because GPT-3 is trained to predict the next word on a large dataset of Internet text, rather than to safely perform the language task that the user wants.
> To make our models safer, more helpful, and more aligned, we use an existing technique called reinforcement learning from human feedback (RLHF). On prompts submitted by our customers to the API, our labelers provide demonstrations of the desired model behavior, and rank several outputs from our models. We then use this data to fine-tune GPT-3.
> The resulting InstructGPT models are much better at following instructions than GPT-3. They also make up facts less often, and show small decreases in toxic output generation. Our labelers prefer outputs from our 1.3B InstructGPT model over outputs from a 175B GPT-3 model, despite having more than 100x fewer parameters. At the same time, we show that we don’t have to compromise on GPT-3’s capabilities, as measured by our model’s performance on academic NLP evaluations.
> One way of thinking about this process is that it “unlocks” capabilities that GPT-3 already had, but were difficult to elicit through prompt engineering alone
Note about the mention of a 1.3B InstructGPT: They trained InstructGPT in a few different sizes including 1.3B, 6B and 175B [1]
From Training language models to follow instructions with human feedback March 4, 2022 [1]
> We start with a pretrained language model, a distribution of prompts on which we want our model to produce aligned outputs, and a team of trained human labelers. We then apply the following three steps:
> Step 1: Collect demonstration data, and train a supervised policy. Our labelers provide demonstrations of the desired behavior on the input prompt distribution. We then fine-tune a pretrained GPT-3 model on this data using supervised learning.
> Step 2: Collect comparison data, and train a reward model. We collect a dataset of comparisons between model outputs, where labelers indicate which output they prefer for a given input. We then train a reward model to predict the human-preferred output.
> Step 3: Optimize a policy against the reward model using PPO. We use the output of the RM as a scalar reward. We fine-tune the supervised policy to optimize this reward using the PPO algorithm
> Steps 2 and 3 can be iterated continuously; more comparison data is collected on the current best policy, which is used to train a new RM and then a new policy
> The cost of increasing model alignment is modest relative to pretraining. The cost of collecting our data and the compute for training runs, including experimental runs is a fraction of what was spent to train GPT-3: training our 175B SFT model requires 4.9 petaflops/s-days and training our 175B PPO-ptx model requires 60 petaflops/s-days, compared to 3,640 petaflops/s-days for GPT-3 (Brown et al., 2020). At the same time, our results show that RLHF is very effective at making language models more helpful to users, more so than a 100x model size increase.
From Introducing ChatGPT November 30, 2022 [2]
> We trained this model using Reinforcement Learning from Human Feedback (RLHF), using the same methods as InstructGPT, but with slight differences in the data collection setup. We trained an initial model using supervised fine-tuning: human AI trainers provided conversations in which they played both sides—the user and an AI assistant. We gave the trainers access to model-written suggestions to help them compose their responses. We mixed this new dialogue dataset with the InstructGPT dataset, which we transformed into a dialogue format.
> To create a reward model for reinforcement learning, we needed to collect comparison data, which consisted of two or more model responses ranked by quality. To collect this data, we took conversations that AI trainers had with the chatbot. We randomly selected a model-written message, sampled several alternative completions, and had AI trainers rank them. Using these reward models, we can fine-tune the model using Proximal Policy Optimization. We performed several iterations of this process.
> ChatGPT is fine-tuned from a model in the GPT-3.5 series, which finished training in early 2022.
In all likelihood, the usage data collected from ChatGPT is being used to continuously train and update a reward model which is being used to continue fine-tuning and improve performance. As for sources of new data, the article lays those out pretty clearly. I agree with the article that models like Phi3 show that higher quality data is more important than simple volume of data and that paying experts to produce, edit and/or grade data will get you much further ahead than only focusing on increasing token count especially when RLHF can 100x the effectiveness of that data. The main reason for finding more sources of training data would be to increase the breadth of knowledge and fill in larger gaps that can't be tackled by a small number of experts. More internet slop won't do that so the contribution of scraped web data can be expected to steadily decrease.
[0] https://openai.com/index/instruction-following/
OK. And then what?
Which dream is collapsing? I’m not disputing it, legitimately curious which of several collapsing dreams you mean.
I agree that LLMs are not 100% trained on Internet posts, but that they are mostly trained on good Internet posts.
When I ask an LLM a question I expect a simulation if a good answer from an Internet discussion board specializing in that topic.
What if those expert just published the stuff freely online instead. Surely that would be more productive and trustworthy. Reality is truly stupid.
That kind of work also sounds incredibly boring. They probably have to pay their experts a lot more than it would normally cost to hire the same caliber of experts. Which would mean they are not generating very much private data.
We are not. Open AI is
Because nobody else is willing to pay them to do that. Research grants, peer review publications, etc, are not set up to reward stuff like “produce one thousand random examples of expressing thoughtful uncertainty with respect to questions relevant to your discipline for which there is currently no consensus answer”
> What if those expert just published the stuff freely online instead. Surely that would be more productive and trustworthy.
I’m sure many of them would be happy to do so if someone is paying them
A "realistic-sounding, 80% correct version of an expert opinion" only costs US$20/month–and if you are willing to compromise a bit on correctness and/or privacy, you can even get it for free. No way you are getting a real expert at that price.
Whereas, pool $20/month across millions of subscribers, and suddenly you have enough money – and, more importantly, future revenue growth prospects with which to convince investors to give you even more money – to afford roomfuls of real experts to try to get that "80% correct" percentage higher.
???
Um, that's kind of obvious.
Because they want to make money.
I think the point of the article is that it may get to the point where these internal datasets are so much more valuable than the sea of fake nonsense on the open internet that the practice of keeping the internal datasets private will become a no-brainer.
Who is willing to give up billions of dollars just 'cuz? Even if you can't run a company, one of the big companies will happily pay you for access to that data. There's no way you make the data public if that data is able to provide an edge in training.
In fact, let's be realistic, the more of an edge it provides, the more money it will be worth.
All these models still regularly fail to provide me with correct code to use even the most common open source libraries (stuff like numpy and matplotlib), even though there were certainly trained on that code.
Actually,