I’m really excited for the new year. Looking back on the past year the rate of change of AI and the amount of compute that’s exchanged hands is mindboggling. I expect these trends to continue into 2026 but even more so. I wanted to lay out my predictions for what AI progress and societal shifts in 2026.

AI:

  1. There is no wall. Models will get more intelligent, lower latency, think for longer if needed, hallucinate less, and have longer memory.

  2. Context windows will become a dated idea as frontier models can handle this new sort of almost infinite memory via techniques like compaction. This makes frontier models very sticky and keeps open source at bay due to their superior product experience.

  3. OpenAI launches their “super assistant” as a software tool and then bundled into a hardware product experience by EOY. This super assistant tracnends the current “chat” model and can be proactive and complete almost any (internet based) everyday task for a user with basic guidance.

  4. The “Claude-Code/Codex/Cursor” experience is felt for all industries. This has already been said by employees at Big Labs like Sholto Douglass from Anthropic but I expect the reaction to be much more profound. Namely due to the fact that we as devs have been slowly conditioned to the incremental progress of models on our work for the past 2 years inside Cursor. Most of society really haven’t had this experience and it’s safe to say that the majority of knowledge work is simpler than programming. This jump in capabilities from basically being able to do nothing useful to doing 90% of their work will be shocking.

  5. Programming as work will no longer be a valid option to do manually outside of education or hobbies. The progress in software development, testing, debugging, has completely changed in the past year. At the end of 2024 all we had was o1, which was an amazing model at the time but mostly incapable of writing good, maintainable code. Now Opus 4.5 or GPT-5.2-Codex can do this reliably. The key concept here is that incremental improvements in every dimension on benchmarks haven’t been substantial however the experience as a dev has been qualitatively improved. Super-senior engineers have gone from 0-90% of their code written by AI within the year.

  6. Scientific discoveries will happen with AI+human with the majority of the work being done by the AI with guidance from the intuitions of the humans. Terry Tao and other top mathematicians will be working directly with companies like Aristotle and Math Inc to be part of this “formalization of Math.” Outside of mathematics we can expect to see major advancements in new drug discovery and cures to diseases. To say the least, the Overton window in 2026 for AI will be stretched to the max. From “slop” generator to discovering novel science while also automating that job you have been grinding to get for years. It’s hard to say how people will feel.

Society:

  1. It will take one company to price in when software engineers become mostly obselete. Whether it’s less job openings or lower wages, both will send seismic shocks through society. A once highly regarded profession, that pays a lot, and required a strong technical degree will fall off the face of the Earth. See my substack article to understand why this hasn’t already started to happen (tldr: until the tool>=tool+human humans can leverage the tools and be more productive/paid more/more jobs). And as more information work gets automated like email/excel/powerpoint jobs with Codex like experiences everyone will realize that the writing on the wall is for them. Capabilities don’t need to be at the level of full automation for there to be global panic about the situation at hand. The coming years will be tough for “nothing ever happens” bros.

Benchmark predictions EOY 2026:

  • SWE Bench: 95%
  • Frontier Math: 90%