GPTing While You Sleep; Measuring AI Doing Your Job; Selling Your Phone Convos
ChatGPT Gets Proactive With New Pulse Briefs
As a young product makers 20 years ago there are two things I told a reporter I wanted in "bots": Do things on consumers' behalf and deliver results - e.g. scan job boards and dating sites and apply / reach out on my behalf - and do it while I sleep. Welp, we're almost there...
OpenAI rolled out ChatGPT Pulse, a preview feature that assembles 5–10 personalized 'cards' over night - news, meetings, travel, to-dos - by looking at your chat history and, if you connect them, Gmail and Google Calendar.
It’s Pro-only for now (with plans to expand to Plus) and notably stops after a few cards (“that’s it for today”) to avoid endless scrolling.
The bigger story: Pulse is OpenAI’s cleanest swing at agentic, “do it while I sleep” assistance - a curated morning brief that learns from your thumbs-up/thumbs-down and specific requests, then adjusts tomorrow’s summary.
Read more: The Rundown AI, AI Breakfast
xAI Sues OpenAI Over Trade Secrets
Elon Musk’s xAI filed (another) suit alleging systematic employee poaching and source-code exfiltration, citing eight former staffers, Signal chats with recruiters, and copied repositories.
OpenAI called it “the latest chapter in Musk’s harassment.” Beyond the courtroom drama, it spotlights the escalating IP risk in today’s AI talent wars: breakthroughs can walk out the door with a job switch if controls aren’t tight.
Read more: The Rundown AI
🧪 Research & Benchmarks
OpenAI's GDPval Measures Models Doing Real-Work Tasks
OpenAI introduced GDPval, a benchmark comparing models against 1,320 professional tasks across 44 jobs and nine sectors.
Headlines: Claude Opus 4.1 tops win-rate (≈47.6%) and shines on presentation polish; GPT-5 leads on technical accuracy.
OpenAI says model performance tripled from GPT-4o to GPT-5 in ~15 months - suggesting near-term leaps in office-grade work. The takeaway? Parity shows up on some tasks, not everywhere, but the trajectory is steep.
Read more: The Rundown AI, AI Breakfast
🧭 Policy, Platforms & Ecosystem
Spotify Cracks Down On AI Spam And Clones
Spotify unveiled AI safeguards: a spam filter, an impersonation policy, and AI disclosure credits - after removing 75M+ spam tracks. The goal is to keep AI creativity without drowning authentic artists in junk.
Read more: The Rundown AI
Microsoft Eyes Payouts For News Used By AI
Microsoft is piloting a Publisher Content Marketplace that would pay outlets based on usage inside Copilot and other AI products - testing a royalty-style model instead of one-off licensing.
Read more: The Rundown AI, AI Breakfast
U.S. Government Access To Grok
xAI said U.S. agencies can access Grok under a new agreement priced at $0.42 per organization through 2027 - framing a low-cost path to try agentic tools in government settings.
Read more: The Rundown AI
⚙️ Dev, Data & Robotics
Meta’s “Code World Model” (CWM) Simulates Code Before It Runs
Meta released CWM, an open-weights research model that simulates program execution to catch issues like infinite loops and reason about runtime behavior - a fresh angle beyond pure code generation.
Read more: The Rundown AI, AI Breakfast
Robots Repair Laptops Instead Of Scrapping Them
A Danish team trained a robotic refurbisher that replaces cracked laptop screens using vision + dexterous manipulation. If generalized, it flips e-waste economics (from ~$10 scrap to ~$200 resale) and pressures planned-obsolescence models.
Read more: AI Secret
🔐 Privacy & Society
Neon Pays People To Record Calls For AI Training
Neon Mobile (now #2 on the U.S. App Store social chart) pays users to record phone calls and licenses the audio to AI companies. It promises to capture only “your side,” but grants a broad, transferable license.
For labs, it’s a data pipeline; for users, it’s a privacy trade with long-tail risk. And something companies like my own mytrainingdata.com and others will be chasing.
The site was offline for a time September 25 after an apparent data breach - oops - but appears to be back up and still very popular in the app store.
Read more: AI Secret
🧰 Tools & Launches
- Wan2.5-Preview (Alibaba) - Multimodal image→video with audio. A creator-friendly path from photos + prompts to video. Useful for product demos, teasers, and quick social edits; still “preview,” so expect quirks.
- VibeSDK (Cloudflare) - Roll your own vibe-coding platform. One-click deploy; lets teams experiment with agentic coding under their own governance rather than third-party sandboxes.
- Code World Model (Meta) - Execution-aware coding research model. Simulates how code will run, adding a “world model” for software - strong complement to standard LLM coding.
- Perplexity Search API - Bring Perplexity-style answers to your app. Lets builders wire search+answer behaviors into products; good for in-product knowledge features without reinventing retrieval.
- Ardent AI Agentic Data Engineer - Autonomous pipeline maintainer. Targets the “boring but critical” layer: creating/fixing data pipelines across cloud & on-prem; a pragmatic agent use case with clear ROI.
- Meta Vibes - AI video feed for create/remix/share. Lightweight creation meets infinite scroll. Expect fast remix culture, new IP wrinkles, and a testing ground for video-native prompts.
Today’s Sources
The Internet
AI Breakfast;
AI Secret;
The Rundown AI.