The title really undersells it, tl&dr:
- "thinking" text is useful for humans to analyze what a model is up to (I often find it more useful than the final burst of slop)
- this text is fed back through the model, which helps it with the next step (they give a nice example)
- but text is a lossy projection of the last tokens, what if you skip it?
- it's more efficient, great, but now you can't see what it's about to do until it does it
- ok, so force it to write "thoughts" e.g. once every 10 rounds
- turns out if you skip too many rounds, it can intentionally trick you with fake thoughts
I suspect there's actually a simple solution: convert these skipped rounds to text anyway, just don't feed them back into the model.
https://youtu.be/iuHddnIzKRA (via m0wer (npub1w3v…4c5c))
Sjors Provoost
npub1v2…qexw3
2026-09-12 12:40:20 UTC
Author Public Key
npub1v25jqj332n630klwtqeqxgyrjl4alw4phmluhu2x6lcxzwy9m03suqexw3Seen on
wss://relay.mostr.pubPublished at
2026-09-12 12:40:20 UTCEvent JSON
{
"id": "c8a0722e8ef6cc3de38d3e8f0830a386bf2bf3fa4d101261ab4f7dd04bef3340",
"pubkey": "62a9204a3154f517dbee583203208397ebdfbaa1beffcbf146d7f0613885dbe3",
"created_at": 1789216820,
"kind": 1,
"tags": [
[
"r",
"https://youtu.be/iuHddnIzKRA"
],
[
"p",
"7459d333af66066f066cf87796e690db3a96ff4534f9edf4eab74df2f207289b"
],
[
"client",
"Damus"
]
],
"content": "The title really undersells it, tl\u0026dr:\n- \"thinking\" text is useful for humans to analyze what a model is up to (I often find it more useful than the final burst of slop)\n- this text is fed back through the model, which helps it with the next step (they give a nice example)\n- but text is a lossy projection of the last tokens, what if you skip it?\n- it's more efficient, great, but now you can't see what it's about to do until it does it\n- ok, so force it to write \"thoughts\" e.g. once every 10 rounds\n- turns out if you skip too many rounds, it can intentionally trick you with fake thoughts\n\nI suspect there's actually a simple solution: convert these skipped rounds to text anyway, just don't feed them back into the model.\n\nhttps://youtu.be/iuHddnIzKRA (via nostr:npub1w3vaxva0vcrx7pnvlpmede5smvafdl69xnu7ma82kaxl9us89zdsht4c5c)",
"sig": "99750997e8612d4cd1bc52fbfda76547dd6012bdcdc441edd637cc52c19283355cc296bf23820944c0b22f28a598a878438b9e0b05a18382e9a72c64922fde29"
}