# AI Predictions **URL:** https://discuss.criticalfallibilism.com/t/ai-predictions/2157 **Category:** Unbounded **Tags:** ai **Created:** [March 18, 2026, 4:53am UTC](https://discuss.criticalfallibilism.com/t/ai-predictions/2157 "2026-03-18T04:53:45Z") **Posts on this page:** 1 **Showing post:** 1
### Author: ![Max](https://discuss.criticalfallibilism.com/user_avatar/discuss.criticalfallibilism.com/max/32/97_2.png) [@Max](https://discuss.criticalfallibilism.com/u/Max) #### Post date: [March 18, 2026, 4:53am UTC](https://discuss.criticalfallibilism.com/t/ai-predictions/2157/1 "2026-03-18T04:53:45Z")
With the AI discussions going on, I thought I’d post some predictions I have about AI and where it will go or how it will affect us. Feel free to post your own, too. Mostly this is focused on coding since we are at the start of the S curve ramp and I’ve been thinking a lot about it. I have some more (including some related to CF) that aren’t in this list. Will reply with them in the future. AI Predictions: (I bolded some words to make scanning easier) - Coding as a practice moving towards error correction processes, higher level architecture, operating on a codebase in parallel, etc. Software dev becomes much more accessible but quality drops. - Longer term: language design shaped by LLMs (not all langs though). example: emphasis on very fast, highly parallel type checking / linting for saved files, easy and explicit typing so that compilers catch errors (but the compiler needs to be fast), smarter testing frameworks that handle more tests better (eg via memoization and caching depending on whether the explicit code paths for those tests have changed), more emphasis on functional purity to avoid side effects and hard-to-predict breaking changes. - Coding becomes largely **remote** (from the machine with the actual files). Running multiple agents on a _large_ local codebase _chugs_. Unlike normal dev, you end up needing to run N\times the compilers, the tests, the LSPs, etc. Design will improve but with resources constrained for consumers (eg RAM situation), it only makes sense to move to ephemeral cloud instances or something. - **Software licenses are dead** if you can point an LLM at an example repo, port it to another language, and publish under a different license. - **Software proofs** become a lot more common. A common workflow is a human directing feature implementation + background agent automatically constructing proofs of behavior and patching issues. I’ve already had some success with formal verification (hard to tell if it matters yet, but current SOTA models can write proofs in coq/roqm and use with OCaml) - **Rehab** specifically for vibe coding by the end of 2026 (and other psychological/addiction based products) - Quality Human Time, Personal AI assistants & remotely directing coding agents: People will **constantly be attending to small things with their agents** , even while watching a movie with friends. It will be pervasive. Constant notifications and associated anxiety. correct: - software license are dead: now there’s a startup (This took like less than a month from me having the idea to someone made a fully working product) [Primagen Video](https://www.youtube.com/watch?v=cahSKUYjuTE) discussing it and [Malus (featured)](https://malus.sh/) wrong: - AI coding agents being as good as they are. I thought there’d be more issues as things got more complex and that my job (programming on the more complex end) as it used to exist would be secure for longer. I was saying this 6-8 months ago. --- _[View the full topic](https://discuss.criticalfallibilism.com/t/ai-predictions/2157)._