Anthropic bankrolls the people who recommend AI vendors

Good morning 👋 Anthropic is spending $100 million to train the people at Accenture and McKinsey who tell enterprises which AI vendor to pick. That's harder to compete with than a model release.

In today's issue:

• 🔭 Anthropic bankrolls the consultants who recommend AI vendors
• 🧠 An 8B model writes a cited research report in 51 seconds
• 🔬 Gaussian avatars dodge the GPU tax, and an 8B model learns to type exact Kali commands
• 🏛️ FTC goes after hidden fees and price discrimination in the same week
• 🛠️ Find out who trained your consultant before you take their recommendation
• 🔗 Rivian's Q3 numbers, ChronoScale's GPU pivot, and an "AI" index that's not what it claims

🔭 THE ONE THING

🎓 Anthropic bankrolls the people who recommend AI vendors

Anthropic is putting $100 million into Claude Frontier Academy, a residency program meant to train 10,000 "Frontier Deployed Engineers," and it already has cohorts running inside Accenture and McKinsey. That's the real move: fund the engineers who sit inside the firms enterprises hire to pick their AI vendor. Call it recruiting for distribution rather than for headcount. If you're racing Anthropic for the same consulting-led enterprise deals, Accenture and McKinsey staff showing up already fluent in Claude is the pressure, not the next model release. Match the enablement spend, or watch the recommendation get written before you're in the room.

🧠 MODELS & RELEASES

• 🧪 Ai2 open-sourced AstaBrief, an 8B model that writes a cited research report in 51 seconds, about 3.5x faster than the Claude-based pipeline it's built to replace. Small and open. One job, done fast. link

🔬 RESEARCH HIGHLIGHTS

• GALA skips the GPU tax on avatar animation. 3D Gaussian avatars render fast, but animating them usually means a heavy neural network pass every frame. This paper swaps that pass for a shallow coefficient predictor and a linear blend of pretrained shapes, no retraining of the original model required. Result: up to a 1,000x cut in CPU animation cost, quality mostly intact, running at 60fps on a phone. For anyone shipping avatars in VR or video calls, that's a GPU line item disappearing. arXiv

• KaliBench scores LLMs on commands, not vocabulary. Most security benchmarks ask a model to describe a tool. KaliBench makes it type the exact command, across 8,504 query-command pairs spanning 1,642 tools on Kali Linux. The best open-weight model manages 42% exact-command accuracy unassisted. Fine-tune an 8B model on KaliBench's reward signal, though, and it matches a 685B model. Cheap specialization beating raw scale, again. arXiv

• Robots that practice before going real. "Reconstruct, Practice, Go Real" has a robot rebuild a task in simulation, diagnose its own failures there, and write new skills into its own prompt. No weight updates needed. On 22 held-out manipulation tasks, success climbed from 28.6% in round one to 95% by round fifteen, and held up at 100% across 30 physical robot trials. Worth watching if you're tired of hand-tuning reward functions. arXiv

🏛️ POLICY & REGULATION

• FTC sued Lens.com, joined by the Utah and Nevada AGs, over a hidden "taxes & fees" charge the agency says often doubles the advertised price of contact lenses, baked into both search ads and the checkout flow. FTC.gov
• FTC settled with Southern Glazer's, the largest wine and spirits distributor in the country, over price discrimination against small retailers. The deal is supposed to get independent liquor stores the same pricing chains already get. FTC.gov

🛠️ TRY THIS

Find out who trained your consultant before you take their recommendation.

1. Ask your consulting or SI partners directly: has their team gone through vendor-specific AI training, and which model family does it cover.

2. Cross-check the answer against your actual AI spend. If the vendor they trained on isn't your incumbent, their next recommendation isn't neutral.

3. Run your own side-by-side eval between your current model and the one the training favors, same tasks, same harness.

4. Write down the result before the recommendation meeting, not after.

Prompt: Given our current AI vendor stack and a competing vendor our consulting partner has been trained on, list the specific tasks where switching would change the outcome, and the tasks where it wouldn't.

🔗 QUICK LINKS

• Rivian built 19,751 vehicles and delivered 19,248 in Q3, in line with its reaffirmed 65,000-70,000 full-year guidance. SEC filing
• ChronoScale Holdings sold off its Ekso Bionics unit to go all-in on GPU infrastructure, pointing to new customer deals it says put it on a path to $1 billion in annualized revenue by Q3 2027. SEC filing
• BNB Plus Corp is dropping the crypto-treasury model for blockchain and agentic AI infrastructure, and shopping its DNA-manufacturing subsidiary LineaRx on the way out. SEC filing
• Neptune Insurance added David Noble, a former RBC and Morgan Stanley executive, to its board. SEC filing
• HSBC's latest structured-note paperwork leans on an "AI Powered" equity index that's really a Solactive basket with embedded index fees of up to 0.85% a year baked in. Worth remembering next time "AI" shows up on a term sheet. SEC filing

Ask who trained your consultant before you take their recommendation tomorrow.

Pradeep Perugu