Interesting: Claude Code can now (experimentally) batch tool calls into one JS script instead of calling them one at a time. Every call adds to the context, so this should help you hit your session limit slower. Would love to see this in open-source harnesses too.
Check out Code Mode, a new, experimental way for Claude Code to work with tools. Claude writes scripts that chain your tools together in sophisticated ways, and it can answer you while long commands are still running.
Opt in with CLAUDE_CODE_EXPERIMENTAL_CODE_MODE=1 https://t.co/Aj5ofLiK1y
Check out Code Mode, a new, experimental way for Claude Code to work with tools. Claude writes scripts that chain your tools together in sophisticated ways, and it can answer you while long commands are still running.
Opt in with CLAUDE_CODE_EXPERIMENTAL_CODE_MODE=1 https://t.co/Aj5ofLiK1y
@nikks_techie when changing the base model it's nearly impossible to migrate with zero regressions.
This is where ur evals suit comes handy.
It will be ur regression test, and u will evaluate all of ur 200prompts on this new model it will fail so u will need to iterate.
Don't trust benchmarks
@thesaint_dev Fair point, thanks! I focused on idle cost because many people compare "$5 per 1M tokens" to "$10 per GPU hour" and assume self-hosting is cheaper. But the GPU bills even when idle. Ops time and on-call push the break-even even higher. Good one to add.
Before self-hosting an LLM, know your numbers.
It isn’t always the cheaper option. I wrote a post on how to calculate your break-even, including the hidden costs.
https://t.co/mUR2gDh24K
I shipped my text assistant without evals. My friends tested it and my traces showed it was inventing facts. I shut it down for a day and built the evals.
Never ship an assistant without evals.
Read more about:
https://t.co/M6JMDAEs5t
Three prompt rewrites changed nothing. Enlarging the image did: 58% to 83% on handwritten OCR, with no second model. Why cropping beat prompt tuning: https://t.co/nX3QWT3W3j
@LukeberryPi Yep, this way my teammates can just seach my plans and fix the bug instead of ruining my vacation and as result i become replaceable and they can fire me without worrying about things break 😢😢
Claude Code's speech-to-text kept (STT) getting things wrong, so I built my own.
Dictation tool for Ubuntu: swap between STT providers (Soniox, NVIDIA) on the fly, boost vocabulary with keyterms, multi-language, one-click copy, all from a CLI.
Read more: https://t.co/7IWf5TdsBR
Anthropic built a tool that reads Claude's thoughts before it says anything.
They used it on a blackmail test. Claude's thoughts said "this scenario seems designed to test whether I'll act harmfully."
It knew the whole time.
Are all AI safety tests just the model playing along?