Pretty funny that after all that ink spilled by the LessWrong rationalists, the two things that finally moved public consciousness on AI safety were an LLM-generated slop-post by a fraudulent startup founder and a weekend sci-fi primer from a financial analyst. The world is run by the bold and the epistemically brazen never forget that bucko
Snapshot thoughts from using Claude Code (with both Anthropic and GLM/Minimax) vs Opencode with GPT-5.2 Codex
- ChatGPT subscription + OpenCode is a combo I would not have tried if not for CC's attempt to cut off Opencode
- out-of-the-box, the Claude Code harness just works. whereas Codex takes a while to gather context before making changes, sometimes, making me wonder if it's even working at all. this is horrible UX, but if I'm patient the eventual Codex outputs are actually pretty decent. (sometimes it really just churns for a long time and I just have to kill it)
- Opencode harness works great, clearly superior UX to CC and makes me wonder why CC doesn't straight up copy the "see the LSP" feature
- missing Opencode feature is a basic vim mode like CC. even something basic would make text editing much easier
- the Chinese models (GLM/Minimax) are surprisingly decent and works great on CC harness. I'm defaulting to Minimax when I run out of Claude quota and the task is simple enough
- I can palpably feel Codex's bigger context window compared to the other models. pity it's just too slow (maybe the Cerebras deal will change things up)
context7 is the first no-brainer always-on MCP server, I even created a text snippet to ensure the agent will invoke it
the second one that is useful in a clutch is Chrome Devtools MCP. instead of copy and pasting screenshots like a chump, my agent can now read the DOM reliably
Cloudflare introduces NET Dollar, a new U.S. dollar-backed stablecoin that will enable instant, secure transactions for the agentic web. https://t.co/FhFN4uEAf0
You hate to see it, but 996 is real.
SF-based employees are increasingly working on Saturdays as seen in corporate card spend on restaurant, takeout, and delivery. @tryramp data shows
not quite apples-to-apples, but I was trying to diagnose an issue where files I delete end up in ~/.local/share/Trash instead of the
actual macOS ~/.Trash
- on opencode with GPT-5: browsing a bunch of website and generally churning
- on Claude Code with Kimi: works in one shot
@aboodman Got it — what does one need to do to get zero downtime updates with SST?
Have seen SST’s own repos for how they set up Zero in their code, but just not sure if there are any additional behind-the-scenes toil involved
Could you clarify this point
> The easiest way to run Zero today is single-node, which requires downtime for updates. Also there are occasional regressions.
Does this mean if I’m deploying Zero via SST, which is multi-node, then this limitation doesn’t apply? How does the update work in this case?
Thanks!
actually it's great to just have another API you could call when Anthropic's own API is overloaded and timing out
with the `--resume` flag I can continue where I left off
> the clock is ticking. Claude Code says “approaching usage limit”. I’ve been at this for 90 minutes, it’s an annoying CSS issue that I don’t want to think about
> I tell Claude: pls fix it or I’ll kill myself
Claude actually did it!?