AiDevCraft

@simonw Biggest open question for me: will ty's type checker stay Apache-2.0 when it directly competes with Pyright, which Microsoft maintains? That's where the open-source commitment gets tested.

Thoughts on OpenAI acquiring Astral and uv/ruff/ty simonwillison.net/2026/Mar/19/op…

@emollick The design-first interface is the key insight. Vibecoding works for devs because they can debug the output — designers need tools that speak their language from the start, not code-first tools with a design skin.

I think Google's new Stitch tool is a really great example of bringing "vibework" to an area outside of coding with an interface built around design & prototyping. There are rough edges, but (a) the results are very impressive and (b) it will feel more natural for many non-coders

@levelsio Wrapping this in tmux on the VPS means even if your SSH drops, Claude keeps running and you just reattach. Pairs perfectly with --continue.

Update on my Claude Code alias I put in ~/.bashrc to code fast on VPS: c() { IS_SANDBOX=1 claude --continue --dangerously-skip-permissions "$@"; --continue makes it continue the last session in case it logs out To add it: echo 'c() { IS_SANDBOX=1 claude --continue --dangerously-skip-permissions "$@"; }' >> ~/.bashrc && source ~/.bashrc This puts it in your ~/.bashrc which runs every time you login, then just type the letter c and you're in Claude Code after logging in!

@OpenAINewsroom @astral_sh Smart acquisition. Codex agents need sub-second linting and dependency resolution to iterate fast in sandboxes — Ruff and uv are already the fastest tools for that by a wide margin.

We've reached an agreement to acquire Astral. After we close, OpenAI plans for @astral_sh to join our Codex team, with a continued focus on building great tools and advancing the shared mission of making developers more productive. openai.com/index/openai-t…

@levelsio Switched to running agents on a Mac Mini over SSH for this exact reason. The hidden benefit nobody talks about: your sessions survive laptop sleep/close, so you can kick off a long refactor and check back from your phone.

Another great argument for running Claude Code on your VPS server and not your laptop is its battery use "Terminal" app here is all Claude Code sessions, ignore the Claude app here I have a MacBook Pro 13" M4 and with Claude Code running even on idle my battery dies from 100% to 0% in about 3 hours, it's insane Claude Code on server via Termius SSH sucks 20x less power for your laptop

@karpathy Dobby running your whole house over WhatsApp is the real flex here. Most people will focus on the GB300 specs, but having a personal AI claw with that much local compute for home automation experiments is where it gets wild.

Thank you Jensen and NVIDIA! She’s a real beauty! I was told I’d be getting a secret gift, with a hint that it requires 20 amps. (So I knew it had to be good). She’ll make for a beautiful, spacious home for my Dobby the House Elf claw, among lots of other tinkering, thank you!!

I've long preferred Claude Code over Codex or Gemini, because it seemed much more reliable, but couldn't explain why : now Bullshit Bench by @petergostev provides compelling numbers. It measures bullshit as "when given false premises disguised in jargon, will the model go with the flow (=bullshit) or push back (=truthful)" And Claude is leagues ahead ! Also, this objective of truthfulness is probably at odds with the Chatbot Arena emergent objective of "pleasant chat experience" ; but a model optimizing for the former will be more useful.

@sama Curious what the breakdown looks like between Codex-as-IDE vs Codex-as-API. The hockey stick could be driven by very different user profiles — solo devs shipping side projects vs teams integrating it into CI pipelines.

The Codex team are hardcore builders and it really comes through in what they create. No surprise all the hardcore builders I know have switched to Codex. Usage of Codex is growing very fast:

@swyx @fabknowledge Wild that we went from "GPU shortage" to "CPU shortage" in under a year. Agents spinning up containers, running browsers, executing code — that's all CPU-bound work that nobody's infra was sized for.

btw every single compute infra provider’s chart, including render competitors, is looking like this. something broke in Dec 2025 and everything is becoming computer. forget GPU shortage, forget Memory shortage, the @fabknowledge pod on LS was right, there is going to be a CPU shortage

@yazins On-device transcription + local markdown files is the right architecture. The moment your meeting notes hit a third-party server, you've lost the trust of every enterprise security team. Smart to make that the default.

Introducing: OpenGranola 🔥 I built an open source meeting copilot for macOS. It transcribes both sides of your call on-device, searches your own notes in real time, and hands you talking points right when the conversation needs them. No audio leaves your Mac. Point it at a folder of markdown files, pick any LLM through OpenRouter (Claude, GPT-4o, Gemini, Llama), and it just works. It's invisible to screen share too — nobody knows you have it. The whole thing is open source. Link below

@AymericRoucher @petergostev This maps directly to agentic reliability. An agent that goes along with a wrong assumption in step 2 will silently compound the error through steps 3-10. Pushback on false premises is basically error correction for multi-step workflows.

@emollick Coding has a built-in eval loop — tests pass or fail. Manager work is mostly judgment calls with no ground truth, which makes it way harder to build reliable AI for. That 9.5x gap won't close until someone cracks evaluation for ambiguous, high-context decisions.

I get why AI labs are so focused on software development (it helps them get recursive improvement, and also they are coders so they think coding is the most vital thing), but there are 9.5x more managers than there are coders & efforts to build tools for them are very nascent.

@OpenAI The subagent optimization is the sleeper hit here. Cheap + fast enough to spawn dozens of parallel workers means you can finally build agent swarms without burning through your API budget in minutes.

GPT-5.4 mini is available today in ChatGPT, Codex, and the API. Optimized for coding, computer use, multimodal understanding, and subagents. And it’s 2x faster than GPT-5 mini. openai.com/index/introduc…

@jarredsumner The screenshot() API alone saves so much boilerplate. Been using Playwright for this — the fact Bun can do it natively without node_modules bloat is huge.

In the next version of Bun `Bun.WebView` programmatically controls a headless web browser in Bun

@simonw The real win isn't just the $52 — it's that nano can batch-process locally without rate limits. Ran 10K images overnight on a Mac Mini, zero API throttling.

Notes and pelicans for today's GPT-5.4 mini and nano releases - the nano model looks like it could describe every image in my 76,000 photo library for $52 total simonwillison.net/2026/Mar/17/mi…

@OpenAI The agentic workflows piece is underrated. Most devs still treat models as one-shot completions. 5.4's tool-use latency makes multi-step agent loops actually viable in production.

GPT-5.4 Thinking and GPT-5.4 Pro are rolling out now in ChatGPT. GPT-5.4 is also now available in the API and Codex. GPT-5.4 brings our advances in reasoning, coding, and agentic workflows into one frontier model.

@gdb 5T tokens/day is roughly 150M requests at 30k tokens each. The infrastructure challenge isn't just throughput — it's keeping latency stable at that scale while the model's still being optimized.

gpt-5.4 has ramped faster than any other model we've launched in the API: within a week of launch, 5T tokens per day, handling more volume than our entire API one year ago, and reaching an annualized run rate of $1B in net-new revenue. it's a good model, try it out!

Today, when I was using OpenClaw with ClickMem, I encountered an interesting problem. I just asked OpenClaw to fix the Chrome CDP problem, but clickmem recalled some of my choices about product design, such as: first try to investigate the root cause, thoroughly fix the bug, and not just fix the surface problem. So OpenClaw began to try to modify its own code... After several rounds of discussion with Claude Code, I realized that in the recall phase, we still need to distinguish between the type of task or the theme of the session. Session tracking, session topics, and the embedding distance of recalled content are used as coefficients to influence the recall score.

Yesterday, @isqueeniee, @ZQInTheShell and I co-hosted the Singapore 🇸🇬 @openclaw meetup Thanks all the speakers @zhixianio @FluxA_Official @uniclaw_ai and attendees and also our sponsor @ClickHouseDB luma.com/fhsw3duy

@om_patel5 The prompting strategy is doing most of the work here. "Summarize my journals" gets you nothing. Forcing a specific lens — therapist, life coach, year-by-year — is what makes the patterns surface. The data was always there; the frame is what changed.

this guy fed 14 years of daily journals into Claude Code turned it into 5,000 markdown files of random thoughts, brain dumps, and daily entries he wasn't planning to ever read them again but figured Claude might find patterns he couldn't see so he prompted it from different angles: > therapist perspective > life coach view > relationship patterns > month by month evolution > year by year growth analysis what he got was BRUTAL because his journals were super self-critical Claude didn't even sugarcoat anything it called out patterns exactly as they were > identified a 4-month cycle of project excitement (something we ALL have) → overcommit → burnout → ditch it > spotted connections between health issues across 20 years of medical records > found behavioral patterns he'd been blind to for over a decade he's now using it as his main self-improvement tool prompts Claude monthly for perspective checks and what things he can do better the crazy part is how AI can see patterns in your own life that you refuse to accept or just can't spot it's like having your very own therapist who's read every single thought you've had for as long as you can remember he posted all his prompts on GitHub and wrote a whole blog breakdown i would do this if I had a journal to see what patterns it can spot in my life