This is the first part of a 3-post series. See end of this post for a peek into the topics we will be covering later.

Note. Most of the models we are discussing in this post are LLMs (Large Language Models). We will say 'LLM' when speaking about them specifically, and 'AI models' when we are talking more broadly and including other (current or future) architectures as well. If you are not familiar with the term 'LLM', don't worry - there will soon be a follow-up post where we explain what LLMs are. For now, just think of well-known examples of families of LLMs, like Chat-GPT, Claude, DeepSeek, Gemini, etc.

Introduction

Lately, AI seems to dominate the news. But how to discern truth from exaggerations? The news coverage is confusing and extremely polarized about the AI topic. For example, you might see videos with titles "AI will replace all jobs in next 2 years" next to a video titled "AI progress has hit a wall" or "AI bubble is about to pop". Of course, clickbait sells, and so do extreme titles - in either direction.

And even from a perspective of an expert the progress can look very jagged - periods of rapid improvement alternate with periods of relative stagnation. In order to get a clearer view, one must zoom out and look at the progress trends over a longer timeframe. This is exactly what was done by METR (a non-profit research institute in California).

Read more…

AI is a rapidly advancing field. The progress in the last decade has been enormous and unexpected. Just think what would people have answered in 2010 if asked when will an AI system be created that can both:

• find security flaws in Linux kernel (see e.g. here); and

• solve famous mathematical theorems that have remained unsolved for 50+ years (example, comprehensive list).

At that time, the likely answers from specialists would have been in the range of 2040 to 2050 or later. For example, this study from 2016 estimated that we were 30-50 years away from AI systems winning math tournaments and solving publishable theorems. However, both in the fields of computer security and mathematics, the future has indeed arrived.

Read more…