Shortly after OpenAI’s surprise release of its highly anticipated GPT-4 model yesterday, there was a slew of online criticism about what accompanied the announcement: a 98-page technical report about the “development of GPT-4.”
Many said the report was particularly notable for what it did not Involving. A section called Scope and Limitations of this technical report states: “Given both the competitive landscape and the security implications of large-scale models such as GPT-4, this report does not provide further details on the architecture (including model size), hardware, training computer, dataset construction, training method or the like.
“I think we can call it closed on ‘Open’ AI: The 98-page paper introducing GPT-4 proudly declares that they reveal *nothing* about the contents of their training set,” tweeted Ben Schmidt, VP of information design at Nomic AI.
And David Picard, an AI researcher at the Ecole des Ponts ParisTech, tweeted: “Please @OpenAI change your name as soon as possible. It is an insult to our intelligence community to call yourself ‘open’ and release such a ‘technical report’ that does not contain any technical information.”
A notable critic of the report is William Falcon, CEO of Lightning AI and creator of PyTorch Lightning, an open-source Python library that provides a high-level interface to the popular deep learning framework PyTorch. After he posted the following meme, I reached out to Falcon for comment. This interview has been edited and shortened for clarity.
VentureBeat: There is currently a lot of criticism of the recently released GPT-4 research paper. What are the biggest problems?
Willem Valk: I think what’s bothering everyone is that OpenAI made a whole paper about 90 pages long. That makes it feel like it’s open source and academic, but it’s not. They literally do not describe anything there. If an academic paper says benchmarks, it says “Hey, we’ve done better than this and here’s a way to validate that.” There is no way to validate that here.
That’s not a problem if you’re a company and you say, “My thing is 10x faster than this.” Let’s take that with a grain of salt. But if you’re posing as research, that’s the problem.
When I publish, or someone in the community publishes an article, I compare it to things that people already have, and they’re public and I put the code there and I tell them exactly what the data is. Usually there is code on GitHub that you can run to reproduce this.
VB: Is this different from when ChatGPT came out? Or DALL-E? Did they disguise themselves as research in the same way?
Falcon: No they were not. Remember that GPT-4 is based on the Transformer architecture that was open sourced from Google for many years. So we all know that’s exactly what they use. They usually had a code to verify. It wasn’t fully reproducible, but you could make it happen if you knew what you’re doing. This is not possible with GPT-4.
My company is not competitive with OpenAI. So we don’t really care. Many of the other people tweeting are competitors. So their problem is mainly that they can’t replicate the results. That’s totally fair – OpenAI doesn’t want you to keep copying their models, that makes sense. As a company, you have every right to do so. But you pretend to do research. That is the problem.
From GPT to ChatGPT, what made it work really well is RLHF, aka reinforcing learning from human feedback. OpenAI showed that it worked. They didn’t have to write a paper on how it works, because that’s a well-known research technique. When we’re cooking, it’s like we all know how to bake, so let’s try this. That’s why there are many companies like Anthropic that have replicated many of OpenAI’s results because they knew what the recipe was. So I think what OpenAI is trying to do now, to avoid recopiing GPT-4, is not letting you know how to do it.
But there’s something else they’re doing, a version of RLHF that’s not open, so no one knows what that is. It is most likely a slightly different technique that makes it work. Honestly, I don’t even know if it works better. It sounds like it is. I’m hearing mixed results about GPT-4. But the thing is, there’s a secret ingredient in it that they don’t tell anyone what it is. That is confusing for everyone.
VB: So in the past, even though it wasn’t exactly replicable, at least you knew what the basic ingredients of the recipe were. But now here’s a new ingredient that no one can identify, like the secret KFC recipe?
Falcon: Yeah, that’s exactly what it is. It could even be their data. Maybe there is no change. But think about it when I give you a fried chicken recipe – we all know how to make fried chicken. But suddenly I do something slightly different and you think, why is this different? And you can’t even identify the ingredient. Or maybe it’s not even baked. Who knows?
It’s like from 2015-2019, as a field of research, we were trying to figure out what foods people wanted to eat. We found burgers to be a hit. From 2020-2022 we learned to cook them properly. And in 2023, we’re apparently now adding secret sauces to the burgers.
VB: Is the fear that this is where we’re going – that the secret ingredients aren’t even shared let alone the model itself?
Falcon: Yeah, it’s going to set a bad precedent. I’m a little sad about this. We all came from academia. I am an AI researcher. So our values are rooted in open source and academia. I came from Yann LeCun’s lab at Facebook where everything they do is open source and he continues to do that and he does that a lot on FAIR. I think LLaMa recently introduced one that is a really good example of that thinking. Most of the AI world has done that. My company is open source, everything we’ve done is open source, other companies are open source, we drive a lot of those AI tools. So we’ve all given a lot to the community to help AI be where it is today.
And OpenAI has generally supported that. They played well. Now, because they have this pressure to make money, I think today is literally the day they really went closed source. They just separated themselves from the community. They say, we don’t care about academia, we sell out to Silicon Valley. We all have VC funding, but we all still maintain academic integrity.
VB: So would you say this move is beyond anything from Google, Microsoft or Meta?
Falcon: Yeah, Meta is the most open – I’m not biased, I came from there, but they’re still the most open. Google still has private models, but they always write papers for you to replicate. Now it might be really hard, like the chef or some crazy restaurant writing a recipe where four people in the world can replicate that recipe, but it’s there if you want to try it. Google has always done that. All of these companies have. I think the first time I see this is not possible, based on this article.
VB: What are the dangers of this in terms of ethics or responsible AI?
Falcon: First, there’s a whole bunch of companies that are starting to come out that aren’t from the academic community. It’s Silicon Valley startup types who start companies, and they don’t really bring these ethical AI research values. I think OpenAI sets a bad precedent for them. They’re basically saying, it’s cool, just do your thing, we don’t care. So you get all those companies that are not going to be incentivized anymore to make things open source, to tell people what they’re doing.
Second, if this model goes wrong, and it will, you’ve already seen it with hallucinations and giving false information, how should the community respond? How should ethical researchers go and actually propose solutions and say, this way doesn’t work, maybe modify it to do this other thing? The community loses all of this, so these models can become super dangerous very quickly, without people even noticing them. And it’s just really hard to control. It’s kind of like a bank that doesn’t belong to FINRA, like how are you supposed to regulate it?
VB: Why do you think OpenAI is doing this? Is there another way they could have both protected and opened GPT-4 from replication?
Falcon: There could be other reasons, I know Sam a little bit, but I can’t read his mind. I think they are more concerned with making the product work. They are definitely concerned about ethics and making sure things don’t harm people. I think they’ve thought that through.
In this case I think it’s just about people not replicating because if you find them launching something new every time [it gets replicated]. Let’s start with stable diffusion. Stable Diffusion came out many years ago by OpenAI. It took a few years to replicate, but it was done in open source by Stability AI. Then ChatGPT came out and it’s only a few months old and we already have a pretty good version that’s open source. So time is getting shorter and shorter.
At the end of the day, what matters is what data you have, not the specific model or techniques you use. So what they can do is protect the data, which they already do. They don’t really tell you what they train on. So that’s about the most important thing people can do. I just think companies in general need to stop worrying so much about the models themselves being closed source, and start worrying more about the data and the quality you’re championing.
VentureBeat’s mission is to become a digital city plaza where tech decision makers can learn about transformative business technology and execute transactions. Discover our Briefings.
