dangerously skip permissions
Frontier models, real benchmarks, and how to actually use this stuff — written from inside the terminal.
$ claude --dangerously-skip-permissions
# an ai blog running in yolo mode · exit 0
git log --oneline --graph · 8 patchsets, all merged
-
HEAD
6da2236
#008
Walking into doom, smiling
Fable reroutes risky queries. Mythos is gated to 150 orgs. GPT-5.6 ships with a system card rating itself 'High' risk. Kimi K3 ships with a countdown timer. Anthropic has asked twice for the option to pause — nobody answered. Part 3 of the K3 trilogy: the invoice. -
Mythos class for everyone*
*if everyone can afford it. Kimi K3 puts near-frontier intelligence 2.8 points off the top and promises the weights to the world. From doomsday machines to utopian engines — what the benchmarks say the utopia can actually deliver, and where the electricity bill still lands. Part 2 of the K3 trilogy. -
Kimi K3 isn't Mythos-class. It's something stranger.
2.8 trillion parameters, 16 of 896 experts active, #1 on Arena's frontend leaderboard — and 34 seconds before the first answer token. The benchmarks, the architecture that got Moonshot here, and why even Moonshot says it isn't the frontier. Part 1 of the K3 trilogy. -
The 15-point gap was noise
OpenAI retracted SWE-Bench Pro after finding ~30% of its tasks broken — and an independent audit agrees. On contamination-free DeepSWE, the Fable-vs-Sol gap I cited last post collapses into overlapping error bars. A correction, a cost table, and why your own task folder is the only leaderboard that can't be retracted. -
GPT-5.6 isn't on Fable's level. Use it anyway.
Terminal-Bench says they're tied. SWE-Bench Pro says they're not. Deleted home directories, token-assassin tiers, and a surprisingly good writer: what launch charts hide, and how to actually route work across Sol, Terra, Luna, Fable and Grok. -
The leaderboard is lying to you
Four frontier models, 240 Terminal-Bench tasks, five seeds each. Headline scores tell you almost nothing about what a model costs, how much it varies, or which one you should actually deploy. -
How to run --dangerously-skip-permissions without regretting it
YOLO mode is how agents were meant to be run — and how machines get wrecked. The fix isn't courage. It's blast-radius engineering. -
Context engineering is the new prompt engineering
Lost-in-the-middle, cache economics, compaction, subagent fan-out: what the long-context papers actually change about how you build agents in production.