Social Media Briefing — March 6, 2026

Delivering on OpenAI's Social tease from two days ago, Sam Altman announces GPT-5.4 launch: available in API, Codex, and rolling out in ChatGPT. Features include improved knowledge work, web search, native computer use, mid-response steering, and 1M token context.

Delivering on OpenAI's Social tease from two days ago, OpenAI officially announces GPT-5.4 Thinking and GPT-5.4 Pro rolling out in ChatGPT, API, and Codex. Described as unifying reasoning, coding, and agentic workflows into one frontier model.

Delivering on OpenAI's Social tease from two days ago, OpenAI details GPT-5.4 capabilities: fewer tokens, faster speed, improved deep web research, better context retention during extended thinking, and a new mid-response steering/interruption feature rolling out on Android and web.

Tri Dao releases FlashAttention-4 paper after a year of work. On Blackwell GPUs, attention now runs as fast as matmul. Describes novel algorithmic innovations: exponential emulation with polynomials, new online softmax avoiding 90% of rescaling, and 2CTA MMA instructions for shared memory traffic reduction.

Karpathy announces nanochat now trains GPT-2 capability in 2 hours on 8xH100 (down from 3 hours). Biggest gain from switching to NVIDIA ClimbMix dataset. He's now running AI agents that autonomously iterate on nanochat — 110 changes in 12 hours, reducing val loss. He's spending more time optimizing the agent workflow than the repo itself.

Alongside yesterday's Social buzz about GPT-5.4's reasoning capabilities, OpenAI publishes an evaluation suite and research paper on Chain-of-Thought Controllability, finding GPT-5.4 Thinking shows low ability to obscure its reasoning, supporting CoT monitoring as a useful safety tool.

Adding to Reddit reports of Anthropic's explosive growth, Mike Krieger (Anthropic CPO) announces that more than 1 million people are now signing up for Claude every day.

Delivering on OpenAI's Social tease from two days ago, Ethan Mollick reports having early access to GPT-5.4 and GPT-5.4 Pro, saying stats are 'very good' and demonstrating progress by comparing a creative coding prompt (Piranesi as p5js 3D space) between GPT-4 (2024, needed corrections) and GPT-5.4 Pro (single prompt success).

Delivering on OpenAI's Social tease from two days ago, Matt Shumer's massive GPT-5.4 review: declares it the best model in the world 'by far'. Claims coding is 'essentially solved', standard mode beats previous Pro modes, uses fewer reasoning tokens. Weaknesses: frontend taste behind Opus 4.6 and Gemini 3.1 Pro, misses real-world context, sometimes stops short. Over 1M views.

Karpathy discusses AI memory systems: argues memory ops could be introduced as 'tools' in RL to improve LLMs. Notes current compaction/memory implementations are early and crappy. Speculates humans do weight-based updates (possibly during sleep) and more exotic long-term memory approaches that change weights are exciting but outside established production stacks.

Allie K Miller provides extensive 21-point field report from sold-out OpenClaw meetup in NYC. Key findings: nobody thinks their setup is secure; security expert says assume all data will leak; everyone runs multiple named agents with personalities; ex-finance guy built stock trading platform making $300/day; token usage reaching ~1B/day costing $1-2K/month; excitement for proactive AI; people report being 'joyful and stressed'; sleep quality declining; agents not reliable enough alone (lie about task completion); prompting is dead (replaced by context/harness engineering); AI-led interviews preferred over detailed specs; agents upskilling other agents; severe gender diversity gap (~5% women vs 25-30% industry standard); model welfare not prioritized.

Mollick shares economist Alex Imas's analysis showing that macro-economic data is now suddenly reflecting the AI productivity gains previously seen only in micro-level research studies.