Entries for June 10, 2026

"Dogfooding caught the dogfooder" Fable 5 has a sense of humor

What did @karpathy see / was shown? Why did the benefactor and teacher of the whole ML ecosystem join Anthropic, a company the polar opposite of his image, on the eve of such a powerful model release It can't be purely money Did he reckon that the only way to benefit humanity was to be on the inside, or rather, to not be left outside, of whatever is brewing in there?

I swear to god, let this be a joke If this is a joke, it is not funny Anthropic

To clarify, I was apparently on the enterprise team plan, which is I think equivalent to Pro (1x) plan

It’a been a little bit over 1 year since Anthropic released their Max plans and Claude Sonnet and Opus 4, thus making Claude Code affordable and kickstarting the agentic revolution Opus 4 was a glimpse into the future. I’ve spent the entire summer swearing at it and typing ultrathink Today, Fable 5 feels like another step change I no longer need to type ultrathink. And no longer need to swear at Anthropic models. Only at their marketing team.

Fable burned through my 5 hour quota, and then automatically fell back to usage credits without asking. Org settings I suppose It was burning through 1 usd every few seconds It burned through 66 usd before I reacted. Yeah, this is not affordable for anyone with that API pricing, without subsidy/plan

Ok now it appears among available modes in the shift+tab mode cycle

Speaking of loops, I have renamed my implementation-loop skill from earlier this year to autoimplement, because it's shorter Calling skills that loop auto-x, auto-y makes them more memorable than calling them x-loop, y-loop But it also increases the number of keystrokes you have to type, before you can tab-complete them Alas, I like still this more github.com/osolmaz/tools/tree/main/agents/skill…

when your model is a more decent, thoughtful being than your marketing team

Ok so there is auto mode which they introduced back in March, but apparently they are not so confident in it that it's still in experimental mode and not easily findable in settings code.claude.com/docs/en/permis…

To YOLO with Fable 5, or not to YOLO, that is the question... The last time I left, Claude models still had tendencies to rm -rf your home folder or delete stuff without asking first. Is this still a risk? And from the looks of it, Claude Code still doesn't have Codex's LLM-filtered approval gate feature. Or am I missing something? Please enlighten your fellow Claude noob 😇

The masculine urge to create your own agent multiplexer

We've lost another brother @maddada to agent multiplexers 🫂 x.com/maddada/status…

Just in time for a lot of Codex-default developers going back to Claude Code momentarily to try out Fable 5 Here is a CLAUDE.md -> AGENTS.md symlinker that should save you from the hurdles of obstinate Anthropic conventions It installs a hook that creates the CLAUDE.md symlink automatically as Claude Code traverses directories that contain AGENTS.md, automatically ignored by git No need to create CLAUDE.md with reference to AGENTS.md like Anthropic suggests. It just works github.com/osolmaz/claude-md-symlinker