This week alone:
⚡ Karpathy's "Dobby" agent controls his lights, pool, HVAC, and speakers via WhatsApp
⚡ FedEx is building an AI agent workforce
⚡ NVIDIA launched OpenShell to keep agents safe when accessing your files
⚡ DoorDash started paying people to film chores for agent training data
MIT researchers built PULSE-HF, a deep learning model that reads a simple ECG and predicts whether a patient's heart function will decline within 12 months.
Could help doctors catch worsening cases much earlier. This is the kind of AI application that genuinely saves lives.
Dr. George Landon is working with the U.S. Air Force Academy to use AI for identifying space debris through ground telescopes.
His main lesson for students: real-world AI is messy, and methods that work on paper often fail with actual data, and it's okay to be skeptical of claims. More people need to hear this.
MiniMax just dropped M2.7, the first AI model that helped train itself.
100+ autonomous optimization loops during training. Zero human involvement for 30-50% of the process. 56.2% on SWE-Bench Pro, within 1 point of Opus 4.6 and GPT 5.4. 50x cheaper.
A Chinese lab just matched frontier models while pioneering self-improving AI. Available today via API.
Benchmarks aren't real-world performance though. Real coding workflows will be the test.
The 4 biggest tech companies are spending $700B on AI this year.
The jobs growing fastest aren't in coding.
They're in plumbing, electrical work, and cooling systems. 🧵
OpenAI just dropped GPT-5.4 Mini and Nano.
Mini is 2x faster than GPT-5 mini with a 400k context window. Handles coding, computer use, and multimodal tasks while nearly matching full 5.4 on SWE-Bench Pro. Already live in ChatGPT.
Nano is API-only, built for lightweight tasks at OpenAI's lowest price point.
For most daily tasks, Mini gives you near-5.4 performance at a fraction of the cost.
Microsoft just reorganized its entire Copilot leadership.
Mustafa Suleyman is stepping back to build AI models full-time. His words: "The model is the product."
Former Snap VP Jacob Andreou now leads both consumer and commercial Copilot, reporting to Nadella.
OpenAI just dropped GPT-5.4 Mini and Nano.
Quick take:
Mini matches near-flagship performance at 70% less cost and 2x the speed. Now in ChatGPT, Codex, and the API.
Nano costs $0.20 per million input tokens. API-only, built for agents at scale.
A pipeline that cost $3,000/month on GPT-5.4 could run for ~$250 on Nano.
The real test: whether Mini's 96% tool-calling accuracy holds in production.
The Washington Post is using an AI algorithm and personal data to set different subscription prices for different people. The practice is called surveillance pricing, using your zip code, income, and buying history to pick your price.
CBS found Instacart doing the same, charging some shoppers up to 23% more. We only know about the Post's pricing because New York law requires disclosure. 12+ states are now considering similar laws.
