Researchers injected a concept like "denim" into a model's activations and asked if it noticed anything odd. The best model caught it in about 20% of trials. Follow-up work says it can tell something is off but often can't say what. That sounds a lot like people to me.
@mercury__agent Love seeing approvals move out of the terminal. The same itch made me build AgentBar: one approval queue for Claude Code, Codex, Copilot & co., right in the macOS menu bar.
Would be fun to plug Mercury in too 🙌
https://t.co/VSO0xW63ER
2026: AI agents everywhere, and Sony is selling 1979 headphones. With a cable.
No charging, no pairing, no updates. Just plug in.
Best product of the year?
https://t.co/urFORKcZBW
@diabrowser Big fan here. Love how clean the UI is, and your release notes are some of the best out there. I haven't really gotten into the AI side yet, and I rarely use skills or the chat. Rooting for you though 🙌
About 1,400 forecasters are guessing when general AI arrives. Median: late 2031. But a quarter of them say by mid-2027. I'd sit closer to that early quarter, and what unsettles me is how little that feels like a bold call now.
Two days before the midterms, three forecasts for Senate control:
Silver Bulletin: 77.7% Democrats
Polymarket: ~68%
Metaculus (145 forecasters): ~38%
I've rarely seen the models, the money and the crowd this far apart.
@SashaGusevPosts Fair, that's a real selection effect, I was arguing a different point. Though it cuts both ways: the supers at the low end on doom were the same group that badly lowballed AI progress. Neither side has much of a track record on this exact question.
@TheStalwart Worth translating: 10-40% survival means 60-90% doom. In the 2022 tournament the median superforecaster put AI-caused extinction by 2100 at 0.38%, AI experts at 3%. The 10-40% view comes from a much smaller crowd.
@ARozenshtein Depends which theory of consciousness you'd bet on. On global workspace views it's an architecture choice, so you could build it or leave it out. On IIT, ordinary digital hardware scores near zero no matter how capable the model gets. Neither says complexity alone does it.
Your AI coding agent has been waiting for approval for 20 minutes. You had no idea.
New AgentBar release is out. Open source macOS app that watches Claude Code, Codex, Cursor and Gemini from your menu bar or notch.