Teclis

Love testing and comparing LLMs for creative intelligence sparks

[Teclis @Chain200Nika]

Opus >=4.7 in Claude Code accumulates too much scar tissue in the codebase
On top of that its pedagogy and clarity is severely under-rewarded -> makes harder to spot incoherencies -> more scar tissue after correcting.
Idk why nobody is talking about it, this is a bad direction.

[Teclis @Chain200Nika]

@dlouapre Claude Code Opus >= 4.7 does this: I end up giving up on following the implementation (outsource understanding).
Staying at the idea level means I later have to reverse-engineer everything with Gemini 3.1.
Super inefficient. First time considering going back to cursor-style edits

[Teclis @Chain200Nika]

@dlouapre I read your post with my French brain switched on and instantly thought: yeah, English-speaking Twitter won’t catch this.

[Teclis @Chain200Nika]

@VictorTaelin Do you already have an extraordinariliy good dataset or planning to curate one?
I have recipes for self-supervised dataset generator harnesses if you need

[Teclis @Chain200Nika]

@VictorTaelin You should really try hooks

[Teclis @Chain200Nika]

@VictorTaelin Try hooks and see if it works

[Teclis @Chain200Nika]

@_philschmid Since this is done. When, gemini 3.1 integrated in collab directly? :)

[Teclis @Chain200Nika]

@fchollet long-horizon*

[Teclis @Chain200Nika]

@fchollet Need to run it on long-term task benchmarks with leading LLMs. I don't have credits, though.

[Teclis @Chain200Nika]

@JayScambler https://t.co/V2dZsbDLHs

[Teclis @Chain200Nika]

@karpathy Applies to solving inverse problems as well

[Teclis @Chain200Nika]

@SecWar If the current government had openly campaigned on AI-driven mass surveillance, it likely would have lost a significant share of its right-wing support. When technology outpaces the mandate voters thought they were giving, that’s exactly what referendums are for.

[Teclis @Chain200Nika]

@SecWar The debate was mostly about whether to invest in AI, not how far to deploy it, especially when it comes to mass surveillance. Those are very different questions.

[Teclis @Chain200Nika]

@ilyasut @aidan_mclau Future elections will partly depend on candidates’ positions on AI use. In the last cycle, AI was not advanced enough for this to be a real dividing line. If the current gvt had openly campaigned on implementing AI mass surveillance, it would have lost a fair chunk of right-wing

[Teclis @Chain200Nika]

@_philschmid @0xblacklight Paper says LLM generated md files hurt performance, but what about codex generated AGENTS.md based on the guidelines you provided? :)

[Teclis @Chain200Nika]

@fchollet Completely agree. I’m having an intern work on solving an inverse problem, but the idea is to use an agentic harness for parameter search + calling forward model with tools for feedback. Instead of conventional optimization (monte carlo etc)

[Teclis @Chain200Nika]

@kimmonismus Lol you switched to 100% AI written (at least rewrtten) posts

[Teclis @Chain200Nika]

@cursor_ai Did you guys fix o3 with private azure open ai key? There's a hardcoded model version.