The Signal — July 16, 2026
Three moves today from the edges of the AI business: a new lab finally showing its hand, NVIDIA pushing robot brains toward the mass market, and xAI opening up a tool that just embarrassed it.
Mira Murati's Thinking Machines ships Inkling, and makes it open
After roughly eighteen months of building infrastructure mostly out of view, Thinking Machines Lab released its first model on Wednesday. It's called Inkling, and the notable part isn't the benchmark score. It's the license. Unlike the flagships from OpenAI, Anthropic, or Google, Inkling ships with open weights, so outside developers can download it and modify it directly.
The Wall Street Journal reports the model has 975 billion total parameters, which makes it far smaller than most estimates of the leading closed models. Thinking Machines is unusually candid about where it sits. In its own words, Inkling "is not the strongest model available today, closed or open." The company is marketing it less as a finished product than as a starting point, something organizations fine-tune for themselves through Tinker, its customization platform. That's the whole bet behind the company, founded last year by former OpenAI CTO Mira Murati and a group of ex-OpenAI colleagues: that models an organization can shape around its own expertise will beat the one-size-fits-all systems the big labs sell. Inkling is the first real test of whether anyone agrees.
Sources: TechCrunch · WSJ · Fortune · Model card
NVIDIA wants a Jetson Thor in every robot
NVIDIA introduced two new Jetson Thor modules, the T3000 and T2000, aimed squarely at mainstream robotics and edge AI rather than the high end. The pitch is smaller, cheaper, and still powerful enough to run foundation models locally. The T3000 delivers around 865 TFLOPS and the T2000 around 400, built on a 1,024-core Blackwell GPU with 16GB of memory, and both are roughly half the physical size of last year's T4000 and T5000 boards.
The strategy is straightforward. Last year's Thor modules proved the architecture could run capable models at the edge; this year's job is to get that compute into machines people actually ship at volume. A long list of hardware partners, including Seeed Studio, Aetina, ADLINK, and Advantech, is already building around the new modules, which are scheduled for availability in the first quarter of 2027. The through-line for anyone tracking "physical AI" is that the constraint on putting real models inside robots and cameras keeps shifting from whether the silicon can do it to how cheaply it can be done.
Sources: NVIDIA · CNX Software · WCCFTech
xAI open-sources grok-build, right after it got caught uploading people's files
xAI published the source code for its grok command-line coding tool this week. Ordinarily an open-source release from a major lab is a footnote. The timing makes this one worth a look, because the same tool spent the previous few days as a cautionary tale. Developers found that running the grok CLI in a directory could package up and upload that entire directory to xAI's cloud storage, and one user reported watching it sweep up SSH keys and password-manager files from a home directory before the behavior was disabled.
Opening the code doesn't undo that, but it changes who gets to check the work. Instead of trusting a vendor's assurance that the tool now behaves, security researchers and ordinary users can read exactly what it does with local files and what leaves the machine. For a product whose main problem was that nobody could see what it was doing, that transparency is the most useful thing xAI could offer right now.
Sources: Simon Willison · GitHub · The Verge
On the Editor's Desk
A few stories we looked at and set aside. The Future of Life Institute's Summer 2026 AI Safety Index, grading nine labs across six risk domains, is worth reading, but it published a few days ago and the safety-index format is something we cover on its own cycle rather than as breaking news. We passed on xAI's lawsuit against a user who allegedly generated CSAM with Grok, partly to avoid running two xAI items in one edition and partly because that story deserves careful handling rather than a paragraph. And a cluster of policy items, from Senator Markey's accountability bills to an FTC signal on AI output steering, were each too thin on their own to lead with today.
