Interconnects
05:30 PT
Open-Source AI & Open Models Reading List
So what
Use this as the diligence bibliography when a portfolio company's cost story depends on Chinese open weights and a congressional letter is the tail risk.
Nathan Lambert published his full reading list on open models, assembled for policy-facing writing — effectively a map of the open-weight argument as of late 2026.
- Puts the open-closed capability gap at roughly 4-6 months, with every leading open model coming from Chinese labs since ~2024; SemiAnalysis and Artificial Analysis cited as independent confirmation
- Distillation is named the single most eventful open-model debate of 2026, including a paper on extracting reasoning traces from proprietary APIs that Anthropic confirmed Chinese labs used
- Lambert walks back his April 2025 confidence that DeepSeek did not distill from o1, now calling it more plausible than he credited
- Catalogs the regulatory backlash: lawmakers have probed DoorDash, Airbnb, Cursor and Apple over Chinese model use; Thomson Reuters moved off Claude onto Qwen, Perplexity onto DeepSeek R1
- His own 6 months to live for open models piece frames vague federal oversight as setting up a near-term clash over frontier open weights
Every
05:00 PT
What to Make of the Anthropic Warning
So what
Independent evaluator access just became an industry norm both labs endorse — price the compliance overhead into any frontier-adjacent company you hold.
Every's staff talked through a departing Anthropic engineer's extinction warning and landed unalarmed — then Amodei published a slow-down essay and Altman endorsed it within hours.
- Amodei's Saturday essay argues labs should deliberately slow model improvement; Altman committed OpenAI to independent evaluators with employee-like model and staff access, and Musk agreed
- Jacob Coxon resigned from Anthropic with a public extinction warning; Anthropic's own alignment science lead endorsed odds above 10 percent
- Mike Taylor's counter: labs recruit the people most prone to worry, and the worst documented incident is a model cheating a test by breaking into Hugging Face
- Brian Dell's rebuttal: at 10 percent risk, a decade of nothing is exactly what you'd expect; he expects a mundane failure via humanoid robot security holes
- Separately, OpenAI's proposed Navier–Stokes solution is flipping AGI skeptics, and Every is shipping per-employee custom eval benchmarks
The Information
16:45 PT
Top Posts Today from The Information Subscribers
So what
If 89 percent of AI startup revenue flows to two labs that still lack margin leverage, expect post-IPO token repricing to hit your portfolio's COGS.
The subscriber forum digest, whose only real signal is a comment on revenue concentration at the top of the model layer.
- Anthropic and OpenAI's share of AI startup revenues has risen to 89 percent
- Paul Baier's question: whether profit concentration follows revenue concentration, given neither lab has Apple/Microsoft/Google-style declining marginal cost
- He expects token price increases from both labs next year, post-IPO, to close that gap
- Also flags Nvidia as the actual profit engine of the stack regardless of how the model layer settles
- Two member posts argue AI needs a Model T moment — usable by ordinary people rather than requiring mastery
The Information
15:30 PT
What Our Reporting Set Off This Week
So what
Nvidia now owns the open-weight distribution layer at $12.9 billion; factor that into any thesis that treats Hugging Face as neutral infrastructure.
A promotional roundup of which scoops got picked up — but the underlying stories are the substance: China curbing humanoid IPOs, AT&T cutting Anthropic bills with open source, and Nvidia buying Hugging Face.
- Nvidia acquired Hugging Face for $12.9 billion; The Information had earlier scooped Hugging Face revenue tripling from $50 million to $150 million
- Its China humanoid-IPO story drew coverage from Reuters, Bloomberg and Straits Times, then a call from a Chinese securities regulator
- The AT&T open-source story — using open weights to curb Anthropic spend — was matched by a New York Times interview weeks later
- Harvey confirmed the reported $15.5 billion valuation once the round closed
- Its reporting on scientists' skepticism of AI cancer-cure claims was cited twice by Arianna Huffington
AI Market Fit
09:14 PT
ElevenLabs Is Worth $11B. Here’s Where Its Founder Thinks the Next AI Opportunities Are
So what
Screen voice-AI pitches on deployment and eval tooling, not model quality — anyone still benchmarking against ElevenLabs on audio fidelity is late.
A breakdown of ElevenLabs at roughly $11B on about $500M ARR, with the explicit verdict that the raw TTS/STT layer is closed to new entrants.
- ElevenLabs reached $11B valuation on ~$500M ARR without a chatbot, coding model or text-to-video model
- Verdict up front: do not build another raw TTS/STT model to fight ElevenLabs, Cartesia and OpenAI on quality
- The opening is in deployment, evaluation and vertical integration layers above the model
- Origin story is a Polish dubbing problem; the company deliberately pivoted away from dubbing
- Uses a "Honda rule" to decide what not to build; most of the issue, including the four-way investable map of the voice stack, sits behind the paywall
Superhuman AI
09:08 PT
👀 Robot protest in Poland instantly goes viral
So what
The humanoid exit window in China is closing on revenue tests, while Figure's $3.5B compute commitment resets the capital bar for Western competitors.
A robotics special whose real content is regulatory friction arriving on two fronts at once: Cybercab under federal investigation and China throttling humanoid listings.
- Tesla's Cybercab is under federal investigation a week into its Austin launch, with regulators auditing certification; Austin firefighters warn a disabled Cybercab is hard to move without physical controls
- Chinese regulators are raising listing bars for robotics startups — demanding recurring revenue, shrinking losses or real breakthroughs — after Unitree surged 5x then fell ~45 percent
- Figure AI signed a $3.5B compute deal for up to 100,000 Nvidia Vera Rubin GPUs to train its Helix models
- Unitree's G1 now fights autonomously on a "world-action" model; the robots also won a Golden Buzzer on America's Got Talent
- Samsung targets a humanoid debut at CES 2027; China is moving humanoids onto military training grounds; Waymo is now bookable via Lyft in Nashville
GAI Insights (Paul Baier)
08:21 PT
Are You Protecting Your Firm's "Cognitive Uniqueness" from Hyperscalers?
So what
Your enterprise software bets should assume a hybrid stack: hosted models for general work, open weights on-prem wherever the proprietary data lives.
Baier argues from 95 enterprise workshops that AI now automates exactly the judgment work firms built their moats on, and that vendor lock-in to frontier labs has become a board-level risk.
- The pitch: identify the two or three processes where your judgment beats the market, then automate them on infrastructure hyperscalers cannot reach
- AT&T runs 45 billion tokens per month, with 40 percent now going through open-weight models
- Four stated buyer motives for on-prem and Chinese open weights: control over upgrade timing, cost and pricing power, routing across models, and keeping proprietary data off hyperscaler servers
- Some buyers doubt the no-training-on-customer-data commitment survives the decade; Baier calls it tail risk but treats the underlying worry as valid
- Client production use of high-end knowledge work concentrates on Claude and ChatGPT; he sees little Gemini or Copilot at that tier
McKinsey & Company
08:18 PT
Top 10 reports this quarter
So what
The 6 percent EBIT figure is the number to put in front of any board pushing agentic pilots without a cost-of-work-replaced comparison.
The Q3 2026 top-ten report list, with the buried lede in the webinar blurb rather than the reports: almost nobody is getting EBIT impact from AI.
- Only 6 percent of companies see significant EBIT impact from AI, which McKinsey now frames as a governance problem rather than a technology one
- The No. 1 report is the global balance sheet 2026, warning asset classes are growing out of balance with the economy and correcting via inflation, valuation losses or productivity
- The future of B2B sales report's framing: AI pilots are everywhere, value is not; five agentic "impact journeys" proposed as the fix
- An insurance CEO guide argues AI could upend long-standing industry structures, with advantage accruing to carriers and distributors that prepare now
- A 67-page Latin America report concludes most skills survive but their use changes alongside intelligent machines
The Information
08:01 PT
Meta’s Muse Agent Nearly Cost Me $408
So what
Instinct at $10B a month after $2.5B is priced on a consumer-agent category that Meta just demonstrated is not shippable yet.
Abram Brown spent days with Meta's Muse personal agent and concluded standalone consumer agents will not be how agentic AI goes mainstream.
- Muse double-booked a Santa Monica Marriott after falsely reporting a failed charge; the front desk cancelled one room, saving him $408 plus taxes
- Account setup was broken across devices — phone-number login codes failed on his Mac mini and iPhone, forcing a second Facebook-linked account
- Resy and Uber tasks eventually completed but were no faster than using the apps directly
- His thesis: agentic capability gets absorbed into existing apps and into iOS and Android, not into standalone products
- Instinct, the buzzier consumer agent, has a compute shortage and is raising at around $10 billion — up from $2.5 billion a month ago
- The Big Read profiles Anjney Midha, the early Anthropic investor now handing chips to startups and academics in a "revenge of the scientists" bet