I wonder if there will ever be a Premium OpenAI #Codex Teams plan, like the Claude Team premium seat. Usage-based billing on top of Teams is fair for enterprise, but it feels like it pushes startups to consumer plans.
@zeeg Pre-trained on coding? They all post-train on it, blended with other tasks.
Perhaps:
1. They're pre-trained on next token prediction, and there's far more tokens if it's general text. Filtering to code reduces data.
2. They want a general language & reasoning foundation.
I cannot believe engineers are arguing against reviewing code. If you don’t inspect or at least measure your AI generated code in some way, you’re not engineering. That’s writing wishlists into the ether.
@jason_haugh@kimmonismus And reasoning suffers… we have hybrid architectures, like nemotron. But, if that’s the direction, why would anthropic actively try to cap down to a fraction of 1m on its business plans? You’d think you’d build dependence on large context if the capability is around the corner.
@jessegenet@DarioAmodei@claudeai The classifier overshoots on recall. They've chosen more false positives than to be more precise but accidentally let harmful prompts through.
@araseb_ I'm very worried evolutionarily that we are removing the selective pressures that advance cognition. Intellect is earned through struggle and failure.
@sickdotdev I've experienced more bugs in the desktop app, especially around remote control and multiple agents. It looks nice until sessions freeze and turns disappear, pushing me back into the cli.
@mitsuhiko@dillon_mulroy and the chance of it breaking code conventions increases proportional to the lines it writes or the size of your instructions, no matter how much you spell it out
@PlexInphinity@voxdotcom The podcast asks: are we humanist (advancing humanity) or seeking to advance something else (what? why?)?
The podcast isn't concerned about AI's existence. It's concerned about elevating its existence to a form of self-deprecating worship.
@UnlockingAi@zeeg Sonnet & Opus (seen up to 4.7) regularly try inline imports in Python tests. I don't see it regularly in non-test files. I had to add it to standards.
@file_mutex@zeeg I regularly find issues in even Opus 4.8 xhigh effort code that's pre-reviewed by other agents. Once you cross a few thousand lines, you're virtually guaranteed hidden mistakes. It's very good but still a far cry from the quality I saw at Google.
@dawbuildsthings@hiarun02 I often commit an example .env with basic credentials for dev docker compose. I've also seen encrypted .env files before. API tokens and such really don't belong in a workspace file where agents get them anyway. Use a .zshenv or vault to dynamically load/inject them.