Our internal data shows Claude is accelerating AI development—a possible path to recursive self-improvement, or AI autonomously building a more capable successor. It’s happening faster than we thought, and the implications deserve greater attention. anthropic.com/institute/recu…

For those of you trying to intuitively understand test time compute and why it's a big deal. LLMs are next token predictors. They take a bunch of words and predicts what comes next. This is a "greedy" algorithm that predicts one token at a time - which turns out to be powerful, but limited in terms of its ability to do long term planning. Test-time compute allows the model to "roll out" many tokens forward at once and compare different completion options, which creates the "thinking" behavior. Concretely, the model can predict many tokens, evaluate if the end product is good, and then do that many times over with slightly different options and pick the best full scale completion (side note: picking the "best" completion is a separate hard problem). The benefits of this are allowing the model to simulate a process of weighing and then determining the ultimate implications of its predictions before it spits out the next word. It's easy to think about this in the context of a game like Go. A next token predictor would say, "if I look at a history of all prior Go games, what is the most likely move given this board state?" Test-time compute, which was the innovation behind AlphaGo, allows it to intelligently simulate multiple different moves forward and understand the end state of the board. It's the model asking itself "if I do this move, what will the next 10 moves be, and what will the end state be?" and then comparing that to many other options. In the context of a reasoning problem, this is a huge, huge upgrade. Imagine asking a model to solve a complex math problem. You can definitely learn patterns of tweaking a formula and luck your way into the answer, but if you've ever tried to solve one of these yourself, for complex problems, you usually need to try a couple of different ways to solve the problem. Another good example would be writing out a long, persuasive essay. Yes, you can spit out one token at a time, but you're much better if you can try a few attempts at how the argument flows together cohesively and pick the best one. LLMs needed some effective mechanism for "planning" and "thinking through options" that didn't exist prior to o1. I can't understate how much of a big deal this is simply because it gives us a mechanism for actual problem solving rather than hoping reasoning and logic are emergent from the inherent properties of historical language data. There are still questions about these models such as: 1. How well this will scale and how much will performance improve? How will we get enough of a broad set of "chain-of-thought" data labelled? 2. Will they be limited by the underlying distribution of the data an LLM is exposed to? 3. Do we need other techniques to build true AGI? Personally, I think it almost doesn't matter. These models are likely to have sufficient performance and and generalized enough underlying distribution to solve most of the relevant tasks that humans do today. The key will still be how they get deployed into products in a way that works. Long latency models that have to simulate thinking will introduce lots of design and deployment challenges. This will get solved over time but will need short term design solutions to get these things into products that need fast response times. Distilled smaller models also hold promise here for applications that need fast reasoning in more limited domains. Ultimately though, it feels very likely this is going to unlock a huge step forward. o3 feels likely to surprise and o4/o5 feel even more likely to be massive steps forward. Strap in. Things are going to get weird.

I had lots of fun at OpenAI Dev Day. I left very inspired by what's possible ahead. Here is a summary of my 5 key takeaways based on conversations and presentations: - Prompt Engineering, RAG, and Finetuning can all be leveraged but which one you use and when you use it depends on the task at hand. There is a flow that works great but it's important to start with a simple baseline using prompt engineering (e.g., few-shot learning). Then you can use RAG if context needs knowledge enrichment. Finetuning helps when you want the model to perform a specific task such as converting an email to a specific tone or a blog of text to a specific structure. - Evaluation is hard! Regardless of how you use and apply LLMs, it won't matter much if you are not consistently evaluating. Hence it's important to start with a simple baseline that allows you to put together a good evaluation pipeline for the future. You should always be evaluating, even after you launch your AI application. - AI assistants are here and they are the next wave of AI innovation. How to build the most powerful and robust personalized assistant will be challenging but they can help transform our personal and professional lives. It's also a great opportunity and time for businesses to think about what can be done with personalized AI assistants in their business and how to leverage state-of-the-art LLMs and large-scale data. This is not easy but it's easier to get started today with the introduction of new features like GPTs. - LLMOps (LLM operations) is going to a be challenge and something we need to start to think about now. Building with LLMs is becoming more complex as it's no longer just a simple LLM call and getting back a text completion. These LLM APIs we are using will now involve complex components like retrievers, threads, prompt chains, access to tools, etc. that will need to be logged and monitored. Access to LLM APIs is becoming cheaper - For AI businesses, it's important to think about ways to quickly adapt and embrace AI strategically. You should do so responsibly, always trying to innovate on product experience, and always putting your users first. Not everything will require AI but it's important you experiment, explore, and measure the success of AI applications. It's still a good time to get involved but it has to happen now. No more sitting back. Thanks to the OpenAI team for the invite. Hoping to be in the next one.