The Oracle and the Firm
Then it needs to produce a result. To get models to solve harder and harder tasks that run for increasing amounts of time, you need to figure out how to scale that context management.
The Oracle and the Firm Calvin French-Owen / Writing Calvin French-Owen About Writing Bookshelf The Oracle and the Firm JUN 13, 2026 Like most of the internet, I've been diving into Fable 5 over the last 24h. And like most of the internet, I've been pretty blown away with the quality. But as I've been using both Fable and GPT-5.5, I couldn't help but notice there are clear differences in approach which make the two models behave quite differently. And we're seeing two very different training regimes play out. For any frontier model, accomplishing real work is an exercise in context management
saved by
related reading
- Reflections on OpenAIcalv.info
- How Codex Happenedevery.to
- Shipping at Inference-Speed | Peter Steinbergersteipete.me
- How Long Contexts Faildbreunig.com
- A Guide to Claude Code 2.0 and getting better at using coding agents – sankalp's blogsankalp.bearblog.dev
- Latency Scaling Differences for GPT and Claude Modelsepoch.ai
- The Extreme Inefficiency of RL for Frontier Models - Toby Ordtobyord.com
- Effective harnesses for long-running agents \ Anthropicanthropic.com
- Don’t Outsource Your Thinkingteltam.github.io
- Two different tricks for fast LLM inferenceseangoedecke.com
- Training Imperativesdan.io
- What it feels like to work with Mythos - by Ethan Mollickoneusefulthing.org