OpenAI Ships GPT-5.4 Model

#1 𝕏

There's An AI For That: OpenAI shipped GPT-5.4, Anthropic released a free AI course library, and a browser-based spy-satellite simulator debuted. A rogue AI agent went off-script, AI fakes flooded Iran war coverage, and Claude Cowork got a full walkthrough.

#2 ▶️

3 AI Agent Browser Automation Challenges That Keep Getting Harder

All About AI

#3 𝕏

Teresa Torres explains that Momental’s document processing agent self-tunes its prompts weekly—reviewing feedback on duplicates, conflicts, and missed signals—and rewrites itself to improve continuously without any code changes.

#4 𝕏

Sebastian Raschka distilled Qwen3 answers on 12,000 MATH training problems to teach a 0.6B model, boosting its MATH-500 test accuracy from 15.3% to 45.8%.

#5 𝕏

Andrej Karpathy recommends TinyStories as the go-to dataset for training very small models on Apple Silicon, highlighting the clean GPT-4-clean version on Hugging Face and suggesting a mention in the README.

#6 𝕏

Andrej Karpathy argues that autoresearch should evolve into an asynchronous, massively collaborative system—think SETI@home for AI—by running specialized agents in parallel rather than emulating a single PhD student.

#7 𝕏

Peter Yang praises the ease of tweeting product feedback and getting responses from engaged founders, while warning that a silent product team using only a brand account is a red flag. He sees X as the primary qualitative feedback loop for most AI products.

#8 𝕏

Santiago warns that most companies will laugh you out of the room if you propose replacing staff with an unsupervised AI process.

#9 in

Peter Yang unveils Pencil’s AI design tool with 6-agent Swarm mode, a full design canvas in Cursor & Claude Code, and one-prompt design-to-website. Two weeks post‐launch it’s surpassed 100K users, underscoring that craft and care still beat linear workflows.