GPT-5.4 ships, Karpathy's autoresearch, and Qwen3.5 goes small

OpenAI had the big headline, but the repo that stayed in my head came from a single person.

GPT-5.4

OpenAI released GPT-5.4 on 5 March. The shape of the release matters: the coding strengths of the Codex line are folded into the main model, along with better computer use and tool handling. The separate-models era is easing off, at least at the top end.

For anyone building agents, that simplifies life. One model that writes code, drives a browser and calls tools means fewer hand-offs between systems, and fewer places for a task to fall over. I'll run it against the same internal tasks I used on Opus 4.6 and keep the results in a table, because my memory of "which felt better" is not evidence.

Cheap and small

Google released Gemini 3.1 Flash-Lite, the lowest-cost Gemini 3 tier. Alibaba filled out the Qwen3.5 family with small open models from 0.8B to 9B parameters. These two releases serve the same need from opposite sides: lots of cheap calls, either through an API or on your own device.

Video

Lightricks released LTX-2.3, an update to its open video-and-audio model. It rebuilds the VAE for finer detail, retrains the audio path, and supports true portrait video, which matters for anyone producing vertical content.

Karpathy's autoresearch

On 7 March Andrej Karpathy put autoresearch on GitHub. An agent reads instructions, edits the training code for a small model, runs a fixed five-minute experiment, checks the result, and keeps or throws away the change. Then it repeats. It runs on a single GPU.

What I like is the discipline. One file the agent may edit, one metric, a fixed time budget. Most agent failures I have seen come from vague goals and unlimited scope. This is the opposite.

One detail on the GPT-5.4 launch that deserves a second look is how many vendors now talk about computer use as a core skill instead of a demo. If a model can reliably operate a browser and a spreadsheet, a lot of the glue code I write to connect systems becomes unnecessary. I treat that as a promising direction, and I also treat it as something to test on a dummy account first, since a model clicking in the wrong place has real consequences.

What I'd watch next: whether people apply the autoresearch loop to ordinary business metrics, not only model training.