After post-training, the model’s output distribution collapses a ton.
I feel like the recent models though are extremely annoying to use. GPT 3/3.5/4/4o/4.1/4.5 and even 5 was fine. There were patterns (like everything being a list for 3-series) but nothing this annoying.
Ever since they started optimizing thinking and especially coding, these models have gone down the shitter in terms of their “writing style”
semisol
npub122…cgrkj
2026-08-11 12:54:22 UTC
in reply to nevent1q…4wdd
Author Public Key
npub12262qa4uhw7u8gdwlgmntqtv7aye8vdcmvszkqwgs0zchel6mz7s6cgrkjPublished at
2026-08-11 12:54:22 UTCEvent JSON
{
"id": "d4e8367013764c40228f9936a6343095b2c1a8ada48541c1e707db12900253b6",
"pubkey": "52b4a076bcbbbdc3a1aefa3735816cf74993b1b8db202b01c883c58be7fad8bd",
"created_at": 1786452862,
"kind": 1,
"tags": [
[
"e",
"51eed9a15fdd9bf825962eb493f91c57a9cc28e3a65ece4e304f38487793b082",
"wss://nos.lol",
"root",
"32e1827635450ebb3c5a7d12c1f8e7b2b514439ac10a67eef3d9fd9c5c68e245"
],
[
"e",
"20779a81d4ce864bc0605d2207a408fa1dc8d653b2e38e574e55b10d598fd17f",
"wss://relay.damus.io",
"reply",
"52b4a076bcbbbdc3a1aefa3735816cf74993b1b8db202b01c883c58be7fad8bd"
],
[
"p",
"32e1827635450ebb3c5a7d12c1f8e7b2b514439ac10a67eef3d9fd9c5c68e245"
],
[
"client",
"Damus"
]
],
"content": "After post-training, the model’s output distribution collapses a ton.\n\nI feel like the recent models though are extremely annoying to use. GPT 3/3.5/4/4o/4.1/4.5 and even 5 was fine. There were patterns (like everything being a list for 3-series) but nothing this annoying.\n\nEver since they started optimizing thinking and especially coding, these models have gone down the shitter in terms of their “writing style”",
"sig": "ebba3c2fc64a15f98449e21ca1269443809331d8112010de6f5e83cdd88d0af4d3a688be22a910c0a93cdb4484953e18f0c5d829d1872369375a55321e715a82"
}