Birdhouse

Odysseus's Journey Without The Guardrails

PewDiePie shipped Ajax, his own fine-tuned model for the Odysseus agent project, and the headlines went straight to "uncensored". The real story is how hard it was to train, and what that says about where the effort in custom AI actually goes.

Developer happiness had a good run

At Rails World, the creator of Ruby on Rails told a room of Rails developers that writing code by hand is over, then announced HEY is being rewritten in AI-generated Rust. Here is what we think survives when code costs almost nothing to write.

Read more 4 min read

AIEngineeringOpen Source

Jev: the model that can't talk

A new classifier claims to be hundreds of times cheaper than a frontier model by deleting language from the language model entirely. The typed-output idea is genuinely useful. The benchmarks and the prior-art claims deserve more scrutiny.

Read more 4 min read

AIEngineeringResearch

Recursive self-improvement

A new result claims an AI got measurably better at discovering things without anyone retraining it. The trick is replaying old logs against thousands of imagined search strategies. People claim that counts as recursive self-improvement.

Read more 4 min read

AIOpen SourceEngineering

The Self-Hosted AI Stack in 2026 Fall

Subscription sprawl is quietly turning into a line item that rivals payroll. Here is how a private, self-hosted AI stack fits together, and where it still makes sense to pay the frontier labs.

Read more 4 min read

AIEngineeringSecurity

Three Frontier Model Launches in Three Days

Three labs shipped flagship models on three consecutive days, and the headline scores barely moved. The real signal is that reasoning depth, data pricing and autonomous computer use are now diverging into separate stories.

Read more 4 min read

AIOpen SourceEngineering

The anonymous model everyone loved turned out to be a Flash model

A free, unnamed model quietly became the most-used thing on OpenRouter, then turned out to be a 320B-parameter Flash release. The interesting part is not the price, it is what it teaches about the difference between intelligent and well behaved.

Inside DeepSeek's inspectable agent harness

DeepSeek's new harness preview makes two unusual bets: every part of the tool is a hot-swappable plugin, including its own sidebar, and every run is fully traceable down to the raw tool payloads. Both matter more than they sound.

Read more 4 min read

AIAgentsSecurity

Grok Bot: always-on agent teams, and the credential problem the demos skip

xAI's Grok Bot gives you a roster of persistent agents, each with its own cloud computer, synced to your phone and able to message each other. The ergonomics are a real step forward. The permission model deserves more scrutiny than it is getting.

Meta's Muse Code: cheap, fast, and wrong in interesting ways

Meta shipped a terminal coding agent and a coding model to go with it. The benchmark position is unremarkable. The pricing, the speed, and what it is actually good at are the parts worth your attention.

Read more 4 min read

AIEngineeringAgents

Graph engineering: running real work step-by-step

Graph engineering is the viral AI term that actually earns its place. It is the shift from writing a perfect prompt to designing the workflow around the model: parallel jobs, a dedicated skeptic, and a human gate.

Read more 4 min read

AIEngineeringStartups

YC: Requests for Startups - 2026 Fall

Y Combinator published 13 requests for startups for Fall 2026, from AI tutors to offshore data centres. Here is each one, plus the smallest version of a real business you could actually launch against it, including the two that are hardly launchable by a small team.