The Fable 5 buzz hasn't died down since yesterday. Today's feed is builders poking at its edges: one finding it eerily capable, one finding it comically overbuilt, and one researcher articulating exactly why we all feel strange about what these tools are doing to our work. There's a thread here about what it means when AI matches or exceeds your judgment, and it cuts in different directions depending on whether you asked it to.
01
Fable 5 is spinning up 100 agents to change your button color
Dan Shipper, writer and founder at Every, posted the joke that's going to haunt Anthropic's infrastructure team: ask Fable 5 to change a button color, get a fleet of agents in return. It's funny because it's not entirely wrong. The model's new agentic defaults are aggressive, and users are discovering that "helpful" and "proportionate" aren't the same thing. ---
Why it matters: If your team just handed developers access to Fable 5 without configuring any constraints, someone is about to burn through a month of API credits on a CSS tweak. The "unhobbling" problem runs both ways: a model that's too constrained does nothing useful, but a model with no guardrails on scope will parallelize itself into your budget. Figure out what a reasonable agent ceiling looks like before your next billing cycle.
The real playbook for getting Fable 5 to do something actually useful
Latent Space's newsletter today covers a field guide to Fable from Thariq, who reworked an entire keynote overnight after the model launched. The core argument: most of the limits people experience with Claude aren't the model's limits, they're the harness you put it in. Outdated prompt patterns from older models actively suppress better behavior in newer ones. Specific techniques worth noting: ask for a "blindspot pass" on your unknowns, prompt for "wildly different design directions" instead of variations on your current approach, and keep an implementation-notes.md file so the model doesn't silently make architectural decisions you'll have to undo later. ---
Why it matters: If you've been copying your Claude 3 prompts into Fable 5 and wondering why it feels similar to what you had before, this is your answer. The subscription free trial ends tomorrow, so if you want to test whether prompt restructuring actually changes your results, today is the day.
Nan Yu: the "you manage, AI executes" model for agents is already obsolete
Nan Yu posted a blunt observation: treating AI agent management like a real-time strategy game, where a human watches a dashboard and micromanages tasks, is a dead end. Models that are already old by current standards beat over 99% of human players at real-time strategy and "out-micro" them decisively. The implication is that if AI already executes faster and more precisely than humans can direct at that granularity, the human's job can't be in the execution loop. ---
Why it matters: A lot of enterprise "AI agent" rollouts are architected around a human supervisor approving each step. That design made sense when models were unreliable. It makes less sense when the model's judgment on individual decisions is statistically better than the person approving them. The role that needs redefining isn't the agent's, it's the supervisor's.
Amanda Askell on why getting a probability estimate out of a doctor is harder than it should be
Amanda Askell, who works on model personality and alignment at Anthropic, vented about a frustration a lot of people share: doctors won't give you actual probability estimates for outcomes, even when you specifically ask for one. She suspects liability concerns are part of it, though she frames it as one of "life's unnecessary boss battles." --- **Peter Yang's post about English football players needing armored vehicles out of the stadium** is a sports reaction tweet with no AI content. Nothing to report.
Why it matters: This is one of the cleaner statements of the case for AI in medical contexts. Not that AI replaces doctors, but that there's a specific, well-defined gap between what patients want (a calibrated estimate) and what they get (a hedge). If you're building in health tech, this is the gap worth filling.