Comment by orliesaurus
12 hours ago
Anthropic is on a roll:
- best harness overall (well maybe until like a month ago when gpt5.5 and codex came out)
- acquires bun
- acquires stainless for SDKs
- deal with Elon for compute
- karpathy
what else did I miss?
they invented MCP, Skills, made these standards open so anyone could build the harness around them.
MCP is barely an invention. It's a fuzzy spec detailing a pretty obvious design pattern.
I'm trying to think of a standard this doesn't apply to
LLM is barely an invention. It's an auto complete we had years before.
3 replies →
this pre-IPO is gonna be incredible
It's going to be dramatic - it's unclear how much of their DAUs are organic and how much is through their PE usage deals. There's a large amount of organic usage certainly, it's a useful tool, but there are quite a few of the tell tale signs that they have an internal number they want their user acquisition to be at and they're failing to meet that through organic growth.
1 reply →
To be fair, MCP and Skills do not have any source code. It is fundamentally impossible to release either standard without making it open.
1. Best harness? It ranks the worst with Opus in terminalbench: https://www.tbench.ai/leaderboard/terminal-bench/2.0?models=...
2. Mixed for the entire bun ecosystem, especially with the Rust, Anthropic-focused rewrite
3. Good, because Anthropic's SDK was one of the worst ones to use.
4. Deal with the guy that has a shit ton of compute around wasting money because no-one uses Grok and was frequently calling Anthropic "Misanthropic".
https://i.redd.it/kp4uy1egspjg1.png
5. Glorified marketer whose probably greatest achievement in pushing AI forward was instructing on CS 231n and coining the term vibe coding.
Yeah, on a roll.
You're not entirely wrong but your snide tone is annoying and unsuitable for this platform. Anyway,
1. Claude Code is widely used and beloved despite not benchmaxxing on the terminalbench like these harnesses that nobody has ever heard of or uses.
5. Karpathy's contributions are way more than CS 231n and coining vibe coding. In terms of pedagogy, his "zero to hero" videos, nanoGPT, etc, are all great. For actual work, he also built a great org at Tesla.
NTA but Claude Code is everything but beloved. It's incredibly meh, very buggy (to that extend that customers were literally losing money), heavily vibecoded and all around just... bad. I appreciate it for kickstarting the whole terminal agent thing and I would still use it but only because Anthropic mandates it for using Claude with your subscription.
3 replies →
I've never heard of any of those other harnesses
Tim Abbott the Zulip guy.
Having crappy limits, lots of down times, 4.7opus isn't all that.