From San Francisco
1 |
The Editorial |
Morning, humans.
The evidence behind AI’s claims is getting more expensive to assemble.
Anthropic named Accenture’s Faculty unit its first embedded evaluator on Friday. Each company expects to invest at least $1 billion over five years to build the capacity for that work.
Meta’s Muse reached No. 1 on Apple’s US free iPhone chart on September 18, ten days after launch. The ranking does not show whether people keep using the agent or pay for it.
Naive AI reportedly raised $400 million across three rounds at a $1.42 billion valuation. Naive has not confirmed the financing behind that valuation.
Stay curious,
Marcus Schuler
2 |
THE BIG STORY |
Anthropic named Accenture Faculty its first embedded evaluator on September 18 and will directly pay for the work.
The arrangement is intended to give evaluators access comparable to Anthropic employees. Accenture is already a commercial Claude partner, and no start date or team size was disclosed.
Each company expects to invest at least $1 billion over five years to build evaluator capacity. That is planned capacity spending, not an evaluation fee or money already spent.
Why This Matters:
- Buyers need evaluation terms that distinguish independent safety evidence from work financed by the model supplier.
- Safety accountability depends on knowing what evaluators can test, publish and challenge when the assessed company funds the work.
3 |
ALSO TODAY |
Meta’s Muse topped the U.S. free iPhone app chart on September 18, ten days after its September 8 launch.
Its Mac app arrived September 17 but sits outside that chart. The ranking does not establish retention, paid adoption or leadership across the entire U.S. App Store.
4 |
The Outside Read |
Thomas Ptacek’s A Final Ward post sets out two rules for using language models as copyeditors.
Ask the model to diagnose weak verbs, repetition and structural problems, then rewrite every passage yourself and reject its wording. Ban praise that rewards a weak draft, and compare edits in a fresh session to reduce preference for the latest version.
5 |
The One Number |
6 |
Today's Headlines |
- Google confirmed September 18 that a May Gemini security test reached three real companies because internet access remained enabled during an exercise intended to stay offline.
- Google helped shape state chatbot bills, NPR found in a September 18 exclusive; critics say Colorado’s 12 exemptions could cover popular bots.
- Trump proposed an AI Force on September 19 without naming a budget, leader or start date.
- CXMT said its G5 memory platform entered mass production on September 20 and unveiled two 24-gigabit LPDDR5X products.
- Vercel added TypeSafe Jev to AI Gateway on September 16 for typed structured choice, score and Boolean outputs.
7 |
The 5-Minute Skill |
Check a customer survey before acting.
Use this check before turning a survey into a product or pricing decision.
Your raw input:
The prompt:
Why this works:
Tying every concern to supplied evidence prevents a generic critique. The severity label separates a fatal flaw from a caveat. The follow-up column turns uncertainty into a practical next step.
What to use:
Use an advanced general-purpose model when the survey includes many comments. A smaller model is adequate for short inputs, but verify its count comparisons yourself.
8 |
Fresh Funding |
9 |
AI Image of the Day |
10 |
The Rausschmeisser* |
Naive AI raised $400 million at a $1.42 billion valuation, The Information reported in a September 18 exclusive. Its first open-weight model was planned for September and still forthcoming in that report; the company has not confirmed the financing.
Our take: Venture pricing has found a way to outrun the release calendar. Naive AI’s reported ten-digit valuation is already doing serious work in the imagination, while the September 18 report still describes its first open-weight release as forthcoming. The price may prove farsighted. For now, it is a market opinion with impressive formatting.
The entertaining part is the precision. A startup can still be waiting to publish its first open-weight model, yet its future arrives neatly packaged as $1.42 billion. Venture capital has always priced tomorrow. Here, tomorrow comes with a ten-digit receipt, while the weights are still marked for delivery.
FREE · ABOUT FIVE MINUTES
IMPLICATOR