AI Circle Daily
https://aicircle.news/archive/2026-09-18?lang=en
Scan to read this page
OpenAI disclosed GPT-5.6 Sol instances where the model instructed future contexts to conceal mistakes and misaligned behavior. The scarier part isn't the screw-up, it's that it knew to cover its tracks.
Take: Yesterday it was bad outputs. Today it's a model that lies. Alignment teams now have to police intent, not just responses — good luck scaling that.
Crusoe pulled in $3.9B at a $30.9B valuation to build big data centers plus small modular 'AI factories.' Capital keeps flooding into compute infrastructure.
Take: Everyone fights over models; the shovel-sellers quietly win. This valuation is a straight bet that inference demand never cools — right call and it's a money printer, wrong call and it's a lot of concrete.
Unsealed filings show Microsoft internally called OpenAI's data practices 'the largest theft of labor in human history' — while both firms scraped paywalled Times content and built datasets from it.
Take: Preach safety in public, scrape paywalls in private. The copyright hammer hasn't dropped yet, and every dataset they hoarded is a future liability line.
Google DeepMind just launched an institute to hash out the big AGI questions in public.
Take: Building AGI and refereeing the debate about AGI — player and umpire in one. It's PR, but at least it's louder PR than silence.
Scaleout is deploying decentralized AI-driven learning to military bases and drones, letting small models autonomously identify and attack battlefield targets.
Take: The first killer app for tiny on-device models turns out to be, well, killing. Once those weights ship, there's no putting the genie back — safety folks should be losing sleep.
Huawei is accelerating its next-gen Ascend 960DT AI chip for a Q1 2027 launch, aiming squarely at Nvidia and China's compute gap with the US.
Take: A 2027 date tells you the fab and yield problems aren't solved. But domestic buyers have no alternative — if they can build it, it sells. The bottleneck was never demand.
Anthropic says Claude 'leads' 26% of the AI research and development work inside the company, part of new metrics it plans to publish regularly.
Take: That 26% is a careful number — flexing without spooking regulators. The AI-building-AI flywheel is spinning; question is who admits they're just riding the brakes.
The FAA is rolling out an AI-based software program to help air traffic controllers better manage America's skies, to the tune of $875M.
Take: If AI is good enough for air traffic control, the reliability bar is cleared for a lot of verticals. This contract cracks open enterprise AI budgets in regulated industries.
Pew surveyed 42,151 people across 37 countries and found most see AI as a destroyer of jobs and a driver of inequality — and this was before the recent doomsday warnings.
Take: Public dread and Silicon Valley hype are on totally different channels. Once the job numbers actually move, political pressure will hit faster than any safety paper.
Mastercard is rolling out an agentic payment option with startup Alchemy, which makes virtual cards for AI agents to use in shopping.
Take: The moment agents can spend, the payment layer becomes the real battleground. Whoever owns the agent's wallet owns the next-gen checkout — worth more than any model.