Back to AI Notes

GPT-5.4 drops, Karpathy automates research, Claude Code gets voice, QMD 2.0

OpenAI shipped GPT-5.4 with native computer use on March 5. Karpathy released AutoResearch on March 6. Claude Code got voice mode. QMD 2.0 dropped with a stable library interface.

GPT-5.4

link

OpenAI

OpenAI's latest frontier model, launched March 5. The big deal: it's the first mainline model with native computer use - not a plugin, not a wrapper, the model itself can operate computers and carry out multi-step workflows across apps.

It also absorbed GPT-5.3-Codex's coding capabilities, so you're not choosing between "good at reasoning" and "good at code" anymore. 1M token context window. Most token-efficient reasoning model they've shipped.

Cool feature: mid-response planning. For complex queries it outlines its work first, and you can adjust direction mid-response without starting over. Sounds small but it makes long generations way more controllable.

Also shipped GPT-5.4 mini and nano for high-volume workloads at a fraction of the cost.

AutoResearch

link

Andrej Karpathy

Karpathy released a 630-line Python script on March 6 that lets an AI agent design, run, and analyze ML experiments completely on its own. No human in the loop. Just point it at a problem and let it cook.

You give it a codebase and a goal (like "make this language model train faster"), and it loops: modify the code, train for a bit, measure results, decide what to try next. Repeat forever. Runs on a single GPU.

He let it run for 2 straight days. 700 experiments. 20 optimizations discovered. 11% speedup in language model training. Then Tobi Lutke (Shopify CEO) ran it overnight - 37 experiments, 19% performance gain.

Karpathy called it "the final boss battle" for frontier AI labs. A human researcher tries 5-10 ideas a day. This thing does hundreds. Doesn't get tired, doesn't have bias toward certain approaches, doesn't stop at 6pm. Hit 30K GitHub stars in one week.

Claude Code Voice Mode + /loop

link

Anthropic

Voice mode rolled out March 3. Type /voice and hold spacebar to talk. Release to send. That's it.

It sounds simple but it changes the workflow. Instead of typing out "can you look at the error in the auth middleware and figure out why the token refresh is failing," you just say it. Way faster for explaining context, describing bugs, or giving high-level direction.

Push-to-talk, not always-listening. No ambient recording. The key is customizable.

/loop landed March 7. It turns Claude Code into a recurring monitor - define an interval and a prompt, and it runs automatically. /loop 5m check the deploy status. Lightweight session-level cron job. Great for watching CI pipelines or keeping an eye on a long-running process.

March also brought ultrathink mode, /effort for tuning reasoning depth, and MCP elicitation (MCP servers can now ask for user input mid-workflow).

QMD 2.0

link

Tobi Lutke

Shopify's CEO shipped QMD 2.0 on March 10 - a local search engine for your markdown files. BM25 full-text search + vector semantic search + LLM re-ranking, all running locally via node-llama-cpp. Nothing leaves your machine.

2.0 brings a stable library interface so you can embed it in your own tools. The CLI, MCP server, and HTTP interfaces all run on top of that library now.

The killer feature: it cuts token usage by 95%+. Without QMD, you stuff docs into the context window. With QMD, you search locally first, pull out just the relevant chunks, send only those to the LLM. Cheaper and more focused.

Has a built-in MCP server so Claude Code can search your notes directly. So many people at Shopify use it that they integrated it into their code monorepo for doc search.