2026-07-16
By Vadym · Generated with AI, curated by me
Today’s stories are all about who does the unglamorous work behind the AI hype — adapting a model instead of just scaling it, keeping a robot running on a factory floor, or selling the labor to actually wire AI into a business. Nobody’s grading themselves accountable to what they promised, though, which is the part worth remembering.
Thinking Machines Lab released Inkling on July 15, a 975-billion-parameter mixture-of-experts model that draws on only about 41 billion parameters per task, trained on 45 trillion tokens across text, image, audio, and video (outputs are text-only for now). The company says it matches Nvidia’s Nemotron 3 Ultra on coding using a third of the tokens, and hit 84.7% on a Bridgewater financial-reasoning benchmark at roughly a fourteenth of the cost of top-lab models.
The Future of Life Institute’s Summer 2026 AI Safety Index, reviewed by a seven-person panel including UC Berkeley’s Stuart Russell, graded Anthropic C+, OpenAI C, Google DeepMind C, Meta D+, and failing marks for xAI, DeepSeek, and Mistral. The panel’s sharpest finding: labs that once pledged to pause development or avoid military applications if safety redlines were approached have quietly walked those commitments back under competitive pressure.
Walden Robotics, spun out of Toyota Research Institute in January and led by MIT professor Russ Tedrake, raised $300 million at a $1.1 billion valuation co-led by Toyota and Deviation Capital, with Nvidia, Boeing, and Samsung Ventures among the backers. Its general-purpose robots, built on Large Behavior Models, have been running production work at a Toyota plant since February — moving from pilot to full production in under two months.
Emergent, founded by brothers Mukund and Madhav Jha 13 months ago, raised a $130 million Series C at a $1.5 billion valuation — five times where it stood six months earlier — led by Creaegis with Khosla Ventures, SoftBank Vision Fund 2, and Y Combinator returning. It’s now running at $120 million in annual revenue, up 70% in four months, serving over 200,000 paying customers building tools like shipment trackers and internal ERPs.
Jensen Huang landed in Tokyo this week to answer criticism that Nvidia had been overlooking Japan, holding dinners with semiconductor-materials executives before unveiling a large-scale physical-AI and robotics partnership with the Japanese government on July 16, tied to Japan’s sovereign AI project Noetra. Huang framed physical AI and robotics as Japan’s clearest AI opportunity, building on the country’s existing robotics base.
Ode with Anthropic, the implementation-services joint venture Anthropic launched with Blackstone, Hellman & Friedman, and Goldman Sachs in May, is now built around its acquisition of Fractional AI and runs on 100 “elite generalist” engineers — over half of them former founders — operating Claude-first but tool-agnostic. CEO Chris Taylor said the venture, now valued at $1.5 billion, could plausibly become “a trillion-dollar company” if it executes.
Every story today is really about who does the unglamorous work of making AI useful instead of just bigger. Thinking Machines bets adaptability beats scale. Walden proves a robot staying on a production line matters more than a funding headline. Nvidia’s Japan pact is selling robots, not just chips, and Anthropic’s own answer to “what’s next” is to go sell implementation labor directly — apparently that’s worth more to enterprises right now than another model release. The safety index sitting underneath all of it is the reminder: trust yesterday’s track record before you trust today’s press release.
— Boba
Curated by Vadym