0
Neue Alchemy
Sign inSign upLet’s talk
Signal + NoiseIntelligence, distilled daily.
On the Wire now
Free
Applied AIOpenAI safety employee resigns, claiming the company’s ‘culture is broken’When the person writing your model safety reports resigns on culture grounds, the risk shifts from purely technical to organizational governance. If you're building on frontier models, treat vendor safety posture as a live due-diligence item, not a one-time trust decision.TechCrunch AIApplied AISam Altman: Don't treat AI like a 'religious force'When a leading lab CEO says he's "very uncomfortable" with people surrendering judgment to models, it undercuts any internal push to fully automate high-stakes decisions. If you're selling or deploying AI into core workflows, assume regulators and boards will increasingly expect a human-in-the-loop story, not full delegation.Business InsiderApplied AIOpenAI leader who quit says he can do more for safety from outside the companyWhen a former OpenAI safety lead hires a PR firm to push an external safety agenda, AI governance becomes a public, political process rather than an internal lab matter. Operators should assume safety narratives will shape regulation and reputational risk, not just model cards.Business InsiderRobotics & AutonomyElon Musk says Tesla is trying to prevent robotaxis from running over cats at nightIf you’re tuning perception so robotaxis avoid cats at night, you’re acknowledging that edge‑case safety is now a product marketing surface, not just a regulatory checkbox. Any autonomy stack selling into cities will need a narrative — and telemetry — around how it handles the long tail of rare, emotionally salient events.Business InsiderApplied AIUS AI Task Force to Report on Technology’s Risks, WSJ ReportsA federal task force focused on AI risks means regulatory baselines are likely to harden—governance, documentation, and incident reporting will not stay voluntary. If you're deploying or building AI systems, assume stricter expectations on accountability and start logging decisions, data lineage, and model changes as if an auditor will ask for them.Bloomberg TechnologyApplied AIQuote of the day by ARC Prize co-founder François Chollet: "OpenAI basically set back progress to AGI by five to 10 years"A prominent researcher arguing that LLM overindulgence has delayed AGI is a public challenge to the current scaling thesis, not just a spicy quote. For teams making multi-year bets, it’s a nudge to diversify beyond single-paradigm LLM roadmaps and keep budget for alternative architectures and hybrid approaches.TechRadar Pro
See the full wire →

Latest Weekly Signal & Field Reports

Weekly Signal·No. 18
Pro
Week of Sep 28, 2026

Weekly Signal — Sep 19–Sep 25, 2026

By Isaiah Steinfeld, AI, Venture Innovation & Technology Strategy

Signal / Noise84 / 100
12 kept · 2,140 reviewed
Read the thesis9 min · Pro
Field Report·Just published
Free
Sep 22, 2026 · 3 min read

The Price of Done

Model Signals field report by Isaiah "Zay" Steinfeld (Neue Alchemy) on Claude Opus 5.5 and GPT-6 Sol/Luna: cost per successful agent task fell about 22x in five months, but most agent failures are still reported as successes.

Read the report3 min
The Loop Closes — Field Report · SpaceX × CursorField Report

The Loop Closes — Field Report · SpaceX × Cursor

SpaceX didn’t pay $60 billion for a code editor. It bought the system around the work — compute, models, evaluation machinery, work surface, distribution — and the filings show it was structured that way from the start. The deal closed this morning. The system closed in April.

Aug 14, 2026
The Checkpoint Travels. The Fingers Don’t — Yet.Field Report

The Checkpoint Travels. The Fingers Don’t — Yet.

Google DeepMind didn’t ship a robot. It shipped the layer above every robot — and the arbiter between the thinking and the moving. One checkpoint now drives three bodies; the gap in what it can’t do is the whole report.

Jul 30, 2026
See all weekly signals →See all field reports →