potential customers: “hey can you get back to us?” the shit me and claude are going through:

2026 is the year of continual learning And we are getting some amazing papers towards that This paper introduces Self-Distillation Fine-Tuning (SDFT): on-policy continual learning from expert demonstrations, with no explicit reward inference or engineering The trick here is:

Okay this is definitely the coolest way to setup a personal website

represent all the domain specific objects in code, then put those in a REPL agent!

Software engineers in 2026 be like:

We set a new ARC-AGI-2 SotA: 85.28% using an Agentica agent (~350 lines) that writes and runs code.

I'm joining OpenAI to bring agents to everyone. @OpenClaw is becoming a foundation: open, independent, and just getting started.🦞 steipete.me/posts/2026/ope…

train of starlink satellites over Seaview, Isle of Wight Starlink you shouldn’t leave your headlights on

Such a great use case I’ve been doing the same but for analysing benchmark traces to get a swarm of agents to summarise “what difficulties did the agents encounter with their execution environment” In my case the agent had to hierarchically merge the reports because all 120

pov: how it feels when Al can't solve your problem so you switch to documentation

backpressure: nice to put a word to the concept I will be saying that a lot

See symbolica.ai/agentica for an open source, model agnostic version of this where you can inject your own Python objects and modules into the repl (and it’s done securely with wasm, not heavyweight containers)

first puzzle solved! ARC-AGI-3 here we come.

continual learning, only continual learning, and nothing other than continual learning, is what's missing right now I couldn't care less about saturating benchmarks, getting +3% in SWE Bench or whatever will not make these tools much better than they are, for as long as they

Imo this is the most important benchmark to be watching in AI right now …and it just got saturated :( NEED MORE EVALS!

Based take We should deffo do a dedicated article on Claude Code vs Agentica

My money was on world models being the solution to computer use Maybe all we needed was a more efficient video encoder?

We've released the code used to solve all 3 publicly available ARC-AGI-3 games. github.com/symbolica-ai/A…