Category: Research Topics

OpenAI’s o1-preview has been all the buzz lately. While this model is based on the GPT-4o general architecture, it boasts much improved reasoning capabilities: it can ponder the question for about a minute, reason through multiple possibilities, and arrive at solutions that could not be generated from a single try of GPT-4o. In this post, I discuss the o1-preview model but mainly present the most striking advantage of o1-preview over all previous LLMs: it can meaningfully answer questions from a quiz game called “What? Where? When?”. At this point, it probably does not sound all that exciting compared to winning math competitions and answering PhD level questions on science, but let me elaborate.

We continue our series on LLMs and various ways to make them better. We have already discussed ways to increase the context size, world models that arise in LLMs and other generative models, and LLM fine-tuning including RLHF, LoRA, and more. Today we consider another key idea that can make LLMs far more effective and useful in practice: retrieval-augmented generation, or RAG. We discuss the basic idea of RAG, its recursive agentic extensions, the R[e]ALM approach that integrates retrieval into LM training, some key problems of modern RAG approaches, discuss in detail knowledge graphs and how they are being used in RAG, and conclude with a reminder that even simple approaches can work well and a list of directions for future work.

We continue our series on generative AI. We have discussed Transformers, large language models, and some specific aspects of Transformers – but are modern LLMs still running on the exact same Transformer decoders as the original GPT? Yes and no; while the basics remain the same, there has been a lot of progress in recent years. Today, we briefly review some of the most important ideas in fine-tuning LLMs: RLHF, LoRA, instruction tuning, and recursive self-improvement. These ideas are key in turning a token prediction machine into a useful tool for practical applications.

The announcement of Gemini 1.5 by Google was all but eclipsed by OpenAI’s video generation model Sora. Still, there was one very important thing there: the promise of processing a context window of up to 1 million tokens. A very recent announcement of new Claude models by Antropic also boasts context windows of up to 1M tokens, with 200K tokens available at launch. Today, we discuss what context windows are, why they are a constraint for Transformer-based models, how researchers have been trying to extend the context windows of modern LLMs, and how we can understand if a large context window usefully works. By virtue of nominative determinism, this is a a very long post even by the standards of this blog, so brace yourself and let’s go!

One of the most interesting AI-related news for me recently was a paper by DeepMind researchers that presented a new mathematical result found by large language models: new constructions for the cap set problem. In this post, we take a step back and discuss the general relation between math and AI. A mathematical proof is easy to verify but may be very hard to find. But there are AI-shaped holes in looking for a proof: math involves multi-step reasoning and planning, hard theorems need to be decomposed into lemmas, there are search strategies involved… However, mathematics has turned out to be unexpectedly difficult for AI. In this post we discuss what people have been doing with AI in math and how LLMs can help mathematicians right now.

Here at Synthesis AI, we have decided to release the “Generative AI” series in an e-book form; expect a full-fledged pdf with all the images soon. But when I started collecting the posts into a single coherent whole, I couldn’t help but feel the huge, glaring omission of the most important topic in modern AI, the secret sauce that drives the entire field of ML nowadays: self-attention layers introduced in the original Transformer architecture. I haven’t planned to cover them before since there are plenty of other excellent sources, but in a larger format Transformers have become an inevitability. So today, I post the chapter on Transformers, which seems to be by far the longest post ever on this blog. We will discuss how the Transformer works, introduce the two main families of models based on self-attention, BERT and GPT, and discuss how Transformers can handle images as well.