LLMs are not just for generating text anymore. Over the past several months I've realized that we are moving towards a world where anything that can be done on a computer can be done by LLMs.

Recently I've been hearing messages like "Our education system is outdated because of AI and kids should be learning AI". I think AI skills are important but I'm skeptical of this message. I think instead it's a better strategy to invest in timeless skills like reading and writing that have been around for thousands of years and will therefore probably be useful a thousand years from now (i.e. the Lindy Effect). Think about what *won't* change as well as what will change.

AI may be the first true "silver bullet" in software engineering: en.wikipedia.org/wiki/No_Sil…

Interesting and underrate fact: the DeepSeek R1 and Kimi 1.5 RL papers were published on the same day on 22 January 2025 on Arxiv though DeepSeek R1 gets most of the credit for RL scaling. arxiv.org/abs/2501.12599 arxiv.org/abs/2501.12948

Researchers at Anthropic have discovered the J-space which is like the equivalent of working memory for LLMs: lesswrong.com/posts/3PaLrzxa…

There’s a lot of focus on using AI to increase productivity and throughput but I haven’t seen many people focusing on techniques for increasing quality while keeping throughput constant.

This Substack post has some good career advice in the context of AI: jasmi.news/p/2026-advice

A good analysis on the economic and social effects of future AI: longwalls.substack.com/p/the…

Interesting paper showing how there is an "AI assistant" steering vector: steering towards the AI assistant direction results in more stereotypical AI assistant behavior. Steering away causes the AI to adopt a wide variety of dramatic and entertaining personalities. The Assistant Axis: Situating and Stabilizing the Default Persona of Language Models: arxiv.org/pdf/2601.10387

Software engineering in the 2020s seems to be going through the kind of industrial revolution the textile industry went through in the 1700s: chrisloy.dev/post/2025/12/30…

Interesting how AI models see the world in a different way to humans: "Interestingly, we also found many cases where humans strongly agree on an answer, but the AI models get it wrong. For the third example below, most people agree the starfish is the odd one out. But most vision models focus more on superficial features like background color and texture, and choose the cat instead." deepmind.google/blog/teachin…

I wrote a new blog post on LLM interpretability: lesswrong.com/posts/XGHf7EY3…

"Consciousness in Artificial Intelligence: Insights from the Science of Consciousness" is a great paper on applying different theories of consciousness to understand whether AI models are conscious: arxiv.org/pdf/2308.08708 The conclusion is that current AI models don't seem to be conscious but that could change in the future.

Interesting article on some benefits of AI: openai.com/index/ai-as-the-g…

Good article on the role of people in an advanced AI economy: ninapanickssery.substack.com… > "Technical education will become more critical than ever—it will be needed to train the human auditors capable of overseeing these powerful systems. We will still need technical specialists, but their primary role will shift from generation (doing the task) to discrimination (auditing the AI's work)."

Good TED talk by Eric Schmidt about the future of AI: youtube.com/watch?v=id4YRO7G…

Dario Amodei (CEO of Anthropic) recently wrote a blog post on the importance of interpretability in advanced AI models: darioamodei.com/post/the-urg…