The codegen scaffolds quickly evolve to be good at reflectively prompting AIs and managing the tech debt hell better, but it’s hard to notice this unless you’re working with them actively, leading to a lot of misinformed doubts about the capabilities based on early disappointments. The labs also start including more qualitative criteria in their codegen RL—not just "did the code run and pass the tests", but also asking another LLM to grade the style and extensibility of the code. In effect, there's a race over whether the AIs will learn good code practices from RL self-play, or from explicit human scaffold-crafting and prompting. Note that the latter is getting easier too, as the tooling improves, and AIs write the scaffold code and distill human programming knowledge into prompts.

“Web4” comes to mean a programmable internet that is customised to everyone. A hundred startups jump on this bandwagon. Some established companies create carefully de-risked APIs and let users program customisations and integrations into their sites (i.e. let users ask codegen models to do such programming).

Healthtech AIs are gradually working their way through regulatory hurdles over 2025-2026, and are clearly better than the average doctor at all the parts of the job that rely only on reasoning and knowledge.

The fact that the explosion of codegen threatens Big Tech’s moat, plus some disappointment at the unreliability of o6 after so much hype, plus some general memetic force that means the “current thing” can be AI only for so long, combines to cause a market correction near the end of 2026. Software is starting to seem stale and boring. Investors want to see “real AGI”, not just post-scarcity in software. Google DeepMind’s maths stuff and xAI’s engineering stuff are cool; OpenAI and LLMs are not. Amazon’s AWS & physical store is cool, Google Search and Facebook are not.

First, there just isn’t a particularly tech-savvy political campaign or movement to influence opinion, except for China gradually experimenting with increasingly more AI in their censorship bureaucracy. Second, models still seem worse than the best humans at that “spark” that lets some people create persuasive, viral ideas. Third, the memetic selection pressures acting on the collective output of humanity on the internet are already superhuman at discovering memetic viruses and persuasive ideas than any individual human, so passing any individual human capability threshold in this domain is not automatically a society-steering ability.