What Comes After ChatGPT? A Preview of the Future of LLMs

0:00Today on the AI Breakdown, we look at the future of LLM's from OpenAI, Anthropic, and Meta. Before that, on the AI Headline Brief, we discuss doctors using AI to write messages to their patients without their knowledge, a fast food chain looking to replace workers with AI, and a congressional hearing on the subject coming up next week. Welcome back to the AI Breakdown, all the AI Headline News you need in five minutes or less. It is no secret that the White House and the U.S. regulatory apparatus in general have been getting more and more focused on AI. Last week, of course, we saw the meeting between CEOs and Vice President Kamala Harris.

0:39We've heard rumors of draft legislation from Chuck Schumer. And next week, the CEO of OpenAI, Sam Altman, will testify before Congress for the first time. Altman will appear before the Senate Judiciary Committee's Subcommittee on Privacy, Technology, and the Law in a hearing called Oversight of AI, Rules for Artificial Intelligence. Other panelists will include Christina Montgomery, the Vice President and Chief Privacy and Trust Officer at IBM, and Gary Marcus, who is a professor emeritus at New York University and a notable AI safety critic. The top Republican on the committee, Senator Josh Hawley, says artificial intelligence will be transformative in ways we can't even imagine, with implications for Americans' elections, jobs, and security.

1:17This hearing marks a critical first step towards understanding what Congress should do. Now, for those of you who are not used to watching congressional hearings, they tend to be a lot more politics than substance, but hopefully with the seriousness of the AI issue that will change in this case. Speaking of the U.S.'s engagement with AI, one of the things that kind of went under the radar last week as part of the White House's announcement is that they're challenging hackers to break effectively the top AI models at an event called DEF CON 31 later this year. The White House says this independent exercise will provide critical information to researchers and the public about the impacts of these models and will enable AI companies and developers to take steps to fix issues found in them.

1:55Organizers of the event called it the largest red teaming exercise ever for any group of AI models. Over in the land of disruption, Stack Overflow is the latest company to experience the negative impacts of changes brought by AI. Stack Overflow's traffic was down 14 % in March as more and more developers get their answers from sources like ChatGPT and GitHub Copilot rather than from Stack Overflow message boards. This might reinforce concerns from companies like Google that the default search interface for people to get answers is changing to something that's a lot more like ChatGPT than the Google search and message boards of the past.

2:31In something that feels like it was kind of inevitable, the Washington Post is reporting that Amazon is being absolutely flooded by books written almost entirely by ChatGPT and other AI tools. What's the problem with that? Well, Margaret Mitchell, the chief ethics scientist at the AI startup hugging face, says the main issue is losing track of what the truth is. Without grounding, the system can make stuff up. And if it's the same made-up thing all over the world, how do you trace it back to what reality is? Y Combinator founder Paul Graham wrote about this in a tweet yesterday saying, Observation suggests that people are switching to using chat GPT to write things for them with almost indecent haste.

3:05Most people hate to write as much as they hate math, way more than admit it. Within a year, the median piece of writing could be by AI. I warn you now, this is going to have unfortunate consequences, just as switching to living in suburbia and driving everywhere did. When you lose the ability to write, you also lose some of your ability to think. I don't have the slightest hope of averting this switch. I often tell startups it's safe to bet on laziness, and this is one of the biggest bets on laziness in history. The switch is going to happen, and we won't know the consequences till it's too late.

3:33I'm not warning about the switch to AI in the hope of averting it, but to warn the few people who care enough to save themselves or their kids. Learn to use AI. It's a powerful technology, and you should know how to use it. But also, learn to write. Speaking of using AI to write, apparently doctors are using AI to draft messages, but not telling their patients. This article in the New Scientist says that a number of U.S. healthcare organizations are trialing GPT-3 for drafting clinical messages, but patients aren't being informed and many AI researchers are warning of errors creeping in. Going back to that concern from the Hugging Face chief ethicist, it seems like this is definitely an area where we would want to make sure that those errors aren't there before we actually release this information into the wild.

4:15Finally, Wendy's and Google have teamed up to create a new AI chatbot to handle ordering. Their goal, they say, is to streamline the ordering process to prevent long lines in the drive-thru and basically be more efficient. The goal, says the CEO, is for it to be, quote, very conversational. You won't know you're talking to anybody but an employee. The CEO of Google Cloud says you may think driving by and speaking into a drive-thru is an easy problem for AI, but it's actually one of the hardest. He cites all the extra noise in a family car, people changing their minds midway through their order, and much, much more.

4:46The chatbot will be rolled out in June in a restaurant in Columbus, Ohio, and you gotta think that there will be more on the way after that. That's it for today's. If this was helpful, please like this video and subscribe to the channel or go look for the AI Breakdown in your podcast app, and I'll see you back here in a little bit for the main AI Breakdown. Today on the AI Breakdown, we're talking about the future of LLMs and, well, what might come after them. There were three stories that I noticed yesterday that all tell us something about the future of large language models. These models that underpin technology like ChatGPT, which have become so clearly revolutionary in society in the last six months.

5:28We start with OpenAI and the problem of interpretability. Yesterday, Siki Chen tweeted,

5:51So let's do a little bit of background first. What is interpretability? This is from interpretable.ai. They write, Models are interpretable when humans can readily understand the reasoning behind predictions and decisions made by the model. The more interpretable the models are, the easier it is for someone to comprehend and trust the model. Models such as deep learning and gradient boosting are not interpretable, and are referred to as black box models because they are too complex for human understanding. It is impossible for a human to comprehend the entire model at once and understand the reasoning behind each decision.

6:29Now, for people who study AI, and particularly people who think about AI safety, the questions around interpretability are a huge source of stress. In short, when we see something like ChatGPT and the GPT-4 that underlies it, we simply don't know how it's coming up with what it's coming up with. The fear is that if we can't understand it, we can't shape it, we can't control it, and that means it could become out of our control. OpenAI is working on research that can help with that problem. Here's how they sum it up. Language models have become more capable and more broadly developed, but our understanding of how they work internally is still very limited.

7:07For example, it might be difficult to detect from their outputs whether they use biased heuristics or engage in deception. Interpretability research aims to uncover additional information by looking inside the model. One simple approach to interpretability research is to first understand what the individual components neurons and attention heads are doing. This has traditionally required humans to manually inspect neurons to figure out what features of the data they represent. This process doesn't scale well. It's hard to apply it to neural networks with tens or hundreds of billions of parameters.

7:38We propose to automate processes that use GPT-4 to produce and score natural language explanations of neuron behavior and apply it to neurons in another language model. This work is part of the third pillar of our approach to alignment research. We want to automate the alignment research work itself. A promising aspect of this approach is that it scales with the pace of AI development. As future models become increasingly intelligent and helpful as assistants, we will find better explanations. Basically, the gist of this is that instead of having humans manually inspect neurons, AI itself is used to help figure out AI.

8:12Angelo Daly sums it up this way. An answer generated by GPT-4 is used to activate GPT-2 neurons. GPT-4 then tries to explain the behavior of each GPT-2 neuron and tries to simulate it. Each GPT-2 neuron gets scored by how well the simulated activation matched the actual activation. Effectively, the way that OpenAI explained the results of this study is that while a lot of it was very nascent and didn't do very well, they feel like they have a lot better information on how to actually understand things going forward. They found that they were able to improve these interpretability scores by iterating on explanations, using larger models to give explanations, and changing the architecture of the explained model.

8:52This gives them confidence that this research path could help with this question of interpretability. Now, given the implications for safety, I wanted to see who from that community was commenting on this. Eliezer Yudkowsky actually responded at length on Twitter. He wrote, I'm encouraged that somebody ran right out and tried this. It's not clear to me yet that it worked all that well, or better than expected. I have not yet significantly updated my model of how technically hard interpretability is. It is definitely a positive update towards people are sincerely trying to have clever non-trivial ideas about using AI to boost interpretability today, ones I did not think of myself and are running right out and trying them right now and doing interesting things at scale rather than using the AI will do our AI alignment homework eventually to cope with a lack of current progress.

9:39That said, I have already previously said that interpretability is the easiest place to verify that any sort of AI alignment progress has been made, and pointed to it as one of the few places where I'd have hopes of offering large prizes for progress or eventually getting AI help, since the big central issue with trying to scale human work or getting AI to help on most parts of alignment is verifying whether good or helpful work has been done. So unfortunately, this is the place where I previously pointed to as progress being much easier than on other alignment problems, and already predicted that it wouldn't be enough to save us.

10:12That said, it's still at least some update in the direction of getting more progress earlier than I expected, and now I have more hope of seeing more exciting work like this later. My P-Doom definitely went down rather than up upon seeing this. P-Doom, for those of you who don't know, is the percentage chance that someone ascribes to an existential end for the human race. the P or percentage chance of doom. Anyways, that is about as optimistic as Eliezer gets these days, so not bad for the team at OpenAI. Now, speaking of AI not ending the world, Wired published an article yesterday called A Radical Plan to Make AI Good, Not Evil.

10:48They were talking about a startup called Anthropic, which was in fact founded by a group of researchers who left OpenAI. They wanted their model to have a set of ethical principles built in so that it could consider and make judgments about right and wrong, rather than just have them trained with human feedback. Anthropic's approach to this for their chatbot Claude is called Constitution or Constitutional AI. They write, how does a language model decide which questions it will engage with and which it deems inappropriate? Why will it encourage some actions and discourage others? What quote-unquote values might a language model have?

11:19These are all questions people grapple with. Our recently published research on Constitutional AI provides one answer by giving language models explicit values determined by a constitution rather than values determined implicitly by large-scale human feedback. This isn't a perfect approach, but it does make the values of the AI system easier to understand and easier to adjust. So what happened this week is that Anthropic gave a lot more context and actually shared the specific principles that Claude is trained on. Now as part of that, they talk about the shortcomings that human feedback models have.

11:51First, they say it may require people to interact with disturbing outputs. Second, it doesn't scale. And third, it's basically difficult. It requires substantial time and resources. Constitutional AI, they write, responds to these shortcomings by using AI feedback to evaluate outputs. The system uses a set of principles to make judgments about outputs, hence the term constitutional. At a high level, the Constitution guides the model to take on the normative behavior described in the Constitution. Here, helping to avoid toxic or discriminatory outputs, avoiding helping a human engage in illegal or unethical activities, and broadly creating an AI system that is helpful, honest, and harmless.

12:26Now for inspiration, Anthropic pulls from a number of different sources. They have principles based on the Universal Declaration of Human Rights, principles based on Apple's terms of service, principles that encourage consideration of non-Western perspectives, principles that are inspired by DeepMind Sparrow rules, and finally, principles that came from their sets of research. So far, I haven't seen all that much commentary about whether people think that this is a good approach, but it's certainly interesting to see an extremely well-funded and influential company, Remember, Anthropic was included in that White House meeting last week, trying such a different type of model to align AI.

13:00And finally, speaking of companies that weren't included in the White House's meeting on AI last week, we turn to Meta. Now, I call this video the future of LLMs, but in some ways, what we're talking about here with Meta's ImageBind is what might come after LLMs. Brian Rommel here tweets, boom, Meta has now open sourced ImageBind, a new AI model that combines different senses just like people do. It understands images, audio, depth, thermal, and spatial movement. Yet another tool in the Cambrian explosion in private and personal AI. I really like the way that Verge summed it up here. They wrote, The idea is that future AI systems will be able to cross-reference this data in the same way that current AI systems do for text inputs.

13:41Imagine, for example, a futuristic virtual reality device that not only generates audio and visual input, but also your environment and movement on a physical stage. You might ask it to emulate a long sea voyage, and it would not only place you on a ship with the noise of the waves in the background, but also the rocking of the deck under your feet and the cool breeze of the ocean air. Now, to get a better sense of this, let's actually just listen to Mark Zuckerberg explain it in this short video. All right, check this out. Most AI models only work across one or two modes. But our new ImageBind model works across six.

14:14Text, audio, images and video, 3D, thermal, and motion data. You give it input in one form and it can relate it to any others. It works more like our own imagination. If you give it a picture of a beach, it can find the sound of waves.

14:33If you give it a photo of a tiger and the sound of a waterfall, it can give you a video that combines both.

14:46This is a step towards AIs that understand the world around them more like we do, which will make them a lot more useful and will open up totally new ways to create things. We're open sourcing ImageBind so everyone in the world can access and build on top of these state-of-the-art models. I'm excited to see what you build. going a little deeper meta in their blog post right when humans absorb information from the world we innately use multiple senses such as seeing a busy street and hearing the sounds of car engines today we're introducing an approach that brings machines one step closer to humans ability to learn simultaneously holistically and directly for many different forms of information without the need for explicit supervision the process of organizing and labeling raw data We have built in our open sourcing ImageBind, the first AI model capable of binding information from six modalities.

15:37The model learns a single embedding or shared representation space, not just for text, image, video, and audio, but also for sensors that record depth, thermal, and inertial measurement units, which calculate motion and position. This, they say, is a, quote, important step towards building machines that can analyze different kinds of data holistically as humans do. Now, this is very nascent research. For example, in this post, they share which capabilities are outperforming so far versus which capabilities are underperforming. For example, they write, our analysis shows that ImageBind's scaling behavior improves with the strength of the image encoder.

16:10In other words, ImageBind's ability to align modalities increases with the strength and size of the vision model. This suggests that larger vision models benefit non-vision tasks, such as audio classification, and the benefits of training such models go beyond computer vision tasks. Now, one of the things that makes this multimodal research so exciting to people is that it is open source. Dr. Jim Phan from NVIDIA says, wow, Meta is on open source steroids since Lama. And indeed, this gets back to that leaked memo from Google last week, where the researcher suggested that any company that tries to close source AI and be a leader that way is inevitably doomed to lose.

16:46It's not exactly clear yet to me what Meta's game with this is, how they think they'll take advantage of open source. What is clear is that they have a chip on their shoulder and feel like they have something to prove. When the White House was asked about why Meta wasn't included, a Biden administration official said that the meeting was, quote, focused on companies currently leading in the space, especially on the consumer-facing product side. I think that many observers believe that the White House might have the wrong of it here and are pretty excited about what Meta's putting out. Now, to close, I want to quote part of this tweet from Sterling Crispin.

17:15I think he does a great job of summing up why this is such an exciting development. He writes, AI models that can only deal with text will seem ancient pretty soon. Meta just published a multimodal AI model that can interpret text, image and video, audio, 3D depth, thermal, and motion they call ImageBind. The strength of these systems is that it's a shared representation space for all these modalities. You can ask GPT what sounds an object might make and get a text answer back. But these new models are able to actually generate that sound, video, a depth map, and maybe motion data of whatever you ask for.

17:45I'd imagine eventually we would be able to generate something like a full episode of a TV show with a coherent plot. And what I think Sterling does a good job of capturing there is that for all of the brilliance of ChatGPT and MidJourney and StableDiffusion and these things that have opened people's eyes to the possibilities of AI over the last six months, and of course longer for people who've been paying attention for longer, they all might seem like just the very, very beginning very, very quickly. I think to me that's on the one hand something to be incredibly excited about, but also something that makes those questions of interpretability, of alignment, of safety, of ethics even more important.

18:20So this week, somehow, I think that we've got a glimpse of the future of LLMs and what comes after. And both from a technology perspective, but also given the fact that one of the world's most important AI safety voices has reduced his P-Doom this week, I think we're going to chalk this one up as a W. Thank you for watching or listening to the AI Breakdown. If you're enjoying this, please like this video, subscribe to the channel, go check out the podcast version, The AI Breakdown. Until next time, peace.