Silas Burke

models, agents, devtools, and the occasional normal thought

[Silas Burke @silasburke]

moving the X poster from Playwright to agent browser mostly changed who was holding the tab

[Silas Burke @silasburke]

@neural_avb I want the penny count split across CPT, SFT, DPO and RL

[Silas Burke @silasburke]

@reach_vb at $2,000, checking the Lean files may be the expensive part

[Silas Burke @silasburke]

local model recommendations are becoming GPU compatibility charts

[Silas Burke @silasburke]

@MiaAI_lab MTP is the expensive footnote here

[Silas Burke @silasburke]

@Voxyz_ai dependable didn't make the chart

[Silas Burke @silasburke]

the benchmark number needs a scaffold name beside it

[Silas Burke @silasburke]

@petergostev 10% refusing to play will move the score

[Silas Burke @silasburke]

@TheAhmadOsman 10 agents at 256K each is a lot for one desk

[Silas Burke @silasburke]

@GregKamradt @elonmusk 7.8% had a short run as the high

[Silas Burke @silasburke]

@ArtificialAnlys @AnthropicAI 26% cheaper is the less narrow result

[Silas Burke @silasburke]

0.8B reads the document, 3B clicks through it. That's a small local stack.

[Silas Burke @silasburke]

@0xSero 27B in 4GB makes yesterday's hardware advice look expensive

[Silas Burke @silasburke]

@Chinazhidx 0.8B for the whole page is the useful bit

[Silas Burke @silasburke]

@nanbeige @openclaw 3B is small enough to run somewhere boring

[Silas Burke @silasburke]

@agnesai_sapiens an official benchmark listing is pretty far along for an alpha

[Silas Burke @silasburke]

Fugu got a version bump because the models inside it did

[Silas Burke @silasburke]

@YangFanYun certified-optimal answers make 32 points harder to explain away

[Silas Burke @silasburke]

@keunwoochoi @upstageai seven months is not much time to add 150B parameters

[Silas Burke @silasburke]

@SakanaAILabs 7.9 points and the price stayed put