That is the result of RL, not the intention.
If you encourage less tokens, less tool call rounds, and reward passing tests or similar, then that is what happens
The easiest way gets the reward first and the model rapidly converges to that
semisol
npub122…cgrkj
2026-06-07 00:28:04 UTC
in reply to nevent1q…zfts
Author Public Key
npub12262qa4uhw7u8gdwlgmntqtv7aye8vdcmvszkqwgs0zchel6mz7s6cgrkjPublished at
2026-06-07 00:28:04 UTCEvent JSON
{
"id": "aca7213e666907b0588bd48081ab12656d1f25dc8332089fe8860eaeb4efb344",
"pubkey": "52b4a076bcbbbdc3a1aefa3735816cf74993b1b8db202b01c883c58be7fad8bd",
"created_at": 1780792084,
"kind": 1,
"tags": [
[
"e",
"83f70f833e703f5d58e3e6d8cb5d0ef7f5a87229c642b383d7cc4ded94aa4b0e",
"",
"root"
],
[
"e",
"7e00d8e143a5f441d91dc399f79d88f21bc09995fc6b2b0c77ea9e16456c211a",
"wss://relay.damus.io",
"reply",
"3b96ce185eaa04ede20ca927ebca2d4f1e66884e1f1405fe681c548831f96877"
],
[
"p",
"3b96ce185eaa04ede20ca927ebca2d4f1e66884e1f1405fe681c548831f96877"
],
[
"client",
"Damus"
]
],
"content": "That is the result of RL, not the intention.\n\nIf you encourage less tokens, less tool call rounds, and reward passing tests or similar, then that is what happens\n\nThe easiest way gets the reward first and the model rapidly converges to that",
"sig": "6152d2702501f072bbd2ed4ae38723e239b0a96e12918074e3309ae9003723889543b439a59f80256d71f73544c0b48f1a5d45022f865f5f2115b6653c261e00"
}