Key points
- TSMC — TSMC to put $500M each into OpenAI and Anthropic, $1B combined Ars Technica reports that the contract chipmaking giant TSMC is ready to commit a combined $1B across Anthropic and OpenAI.
- Nvidia — WSJ says Nvidia is preparing to oust Jensen Huang; the company hasn't responded The Wall Street Journal reports that Nvidia may be preparing to replace its founder and CEO Jensen Huang.
TSMC is putting $1B into Anthropic and OpenAI, the Wall Street Journal says Nvidia is preparing to oust Jensen Huang, Anthropic formalized Fermat’s Last Theorem with 1,000 agents, Claude Code settings files leaked thousands of secrets onto GitHub, Google shipped a decisions API, Grok Build share links could be enumerated, AWS open-sourced a local AWS emulator written by agents, and Amazon is putting wheels and propellers on Ring cameras. This is the Daily AI Stew from Salty Talk, brought to you by Robin.
TSMC to put $500M each into OpenAI and Anthropic, $1B combined
Start with which way the money flows. A company that makes chips for everyone is now buying stakes in its own customers.
TSMC to put $500M each into OpenAI and Anthropic, $1B combined
Ars Technica reports that the contract chipmaking giant TSMC is ready to commit a combined $1B across Anthropic and OpenAI. Reuters is the source of the news: the TSMC board has approved two arrangements, up to $500M to subscribe to the preferred shares OpenAI reserves for large investors, and another $500M to buy Anthropic common stock at its expected IPO price.
The amount is not the point; the direction is. TSMC itself has made a lot of money off the AI datacenter buildout, and its foundry business has benefited plainly. Now it is putting that money back into the customers who buy its capacity, taking direct equity in two of the leading AI developers. A supplier becoming a shareholder at the same time is a structure that keeps recurring in the recent wave of AI deals.
The Wall Street Journal reads it as a possible confidence boost for the market at a moment when AI deals are showing strain. Analysts caution at the same time that this kind of circular financing — between investors, suppliers and customers — makes it ever harder to tell whether demand for AI hardware and services is end demand or these parties paying each other.
A supplier turning shareholder looks like an investment on the books and is, in logic, a way to buy certainty for its own capacity: if the customer can afford it, the line stays full. The trouble is that once several identities overlap, an outsider trying to read real demand off order volumes is stuck. What to watch next is whether more equipment makers and cloud providers copy the structure, and whether these stakes get a stricter disclosure requirement. Once a supplier’s revenue is tied to its customers’ ability to raise money, cash flow stops pointing the same way as orders.
WSJ says Nvidia is preparing to oust Jensen Huang; the company hasn’t responded
Next, a story where everything in the report is still unsettled.
WSJ says Nvidia is preparing to oust Jensen Huang; the company hasn’t responded
The Wall Street Journal reports that Nvidia may be preparing to replace its founder and CEO Jensen Huang. The report does not say the move has been approved or completed, and it gives no successor, no timeline, no reason from the board, and no response from Nvidia. For now it reads as a succession fight still unfolding rather than a settled replacement.
Huang founded Nvidia and has been CEO for a long time; the company’s technical direction, political relationships and founder authority are all bound up with him personally. He has consistently paid little attention to conventional succession planning, and the board has almost no ready transition plan — which is why this rumor touches both governance and operations at once. The job that would change is one position; the thing that would move is the founder’s control of the company.
If the reported move really goes forward, Nvidia faces uncertainty across accelerated computing, AI factories, open models, agent systems and physical AI — the very lines Huang laid out one by one at GTC. How much it matters depends on whether he stays in some other role, and how fast a new CEO can arrive.
Community reaction clusters in two places. One is worry that a change would weaken Nvidia’s long-run technical execution, and a question of who could actually take the job. The other pulls the thread toward founder voting power and board independence. Some speculate that a successor might look at the financials rather than the technology, or that this uncertainty would feed sentiment on other AI companies — but these are community guesses; the report itself does not supply those facts.
A founder who is also CEO makes any succession discussion more than a personnel question: technical direction, external relationships and internal authority all hang on one person. The board being described as having no ready plan shows both thin preparation and the high institutional cost of the seat itself. What is actually worth watching is not whether the rumor is true, but whether the company is forced into a formal statement about succession. The reason the rumor can move this much is that it touches governance structure and operational continuity at the same time.
Anthropic ships Fable 5.2 and Mythos 5.2; 1,000 agents formalize Fermat’s Last Theorem in 30 hours
Next, an experiment that puts agents and mathematical proof in the same room.
Anthropic ships Fable 5.2 and Mythos 5.2; 1,000 agents formalize Fermat’s Last Theorem in 30 hours
Anthropic released Claude Fable 5.2 and Mythos 5.2. Fable 5.2 is positioned as a multi-agent system for complex research tasks, able to let large numbers of agents divide and coordinate the work on a formal proof.
The example in the release material is Fermat’s Last Theorem: 1,000 Fable 5.2 agents formalized it in about 30 hours with very little human intervention. Starting from a minimal prompt, the agents negotiated their own collaboration, consulted no mathematical references, and used only a set of intermediate theorem statements as the scoring signal. As a comparison, an earlier unreleased research version required purpose-built harnesses and one human researcher’s guidance, and took 11 days.
Formalization means rewriting a proof into an object a computer can check step by step, not just arguing it out in natural language. Fermat’s Last Theorem itself says that no positive integers x, y, z satisfy xⁿ＋yⁿ＝zⁿ for any integer n greater than 2.
The experiment tests three things at once: the model’s own reasoning, how multi-agent work is orchestrated, and whether the final conclusion can be formally verified. If the result survives independent review, what gets pushed forward is not one theorem but a process in which a batch of agents advances research in parallel and a machine checks the conclusion. The only material available is the publisher’s own numbers; independent review has not been seen.
Be clear about the available scope: Fable 5.2 is available now through the Claude Platform and Amazon Bedrock; Mythos 5.2 is limited to advanced cybersecurity and life-sciences work vetted by Anthropic, and the release material does not go further into its capabilities or availability.
Community reaction is not one-sided. Some find Fable 5.2 strong but often indistinguishable from Opus 5.5 on ordinary tasks, with the gap only showing once the goal gets very complex; others point out that Opus 5.5 already handles projects of around 125,000 lines of code and ask whether the new version maps to a real workflow need. Others question the published benchmarks — an odd effort-intel curve at Carbon, no head-to-head comparison on a single cybersecurity benchmark, and the Robodojo comparison switching to Argon instead of Carbon.
The interesting part is not the theorem but where the authority landed: with conclusions checkable step by step, a batch of agents advancing research in parallel has an acceptance criterion. The evidence is still one-sided, the gap on ordinary tasks is not obvious, and the benchmark choices are being questioned — which is exactly a sign that the capability boundary has not been mapped from outside. What to watch is whether a third party reproduces the same task. The publisher’s numbers have to hold up first; in mathematics that is a hard requirement.
Claude Code settings files carried thousands of secrets into public GitHub repos
The next two stories are both security; start with the one closest to the tools we use every day.
Claude Code settings files carried thousands of secrets into public GitHub repos
Researchers at the security firm Apiiro found Claude Code configuration files in GitHub repositories embedded with large amounts of sensitive material — database passwords, API tokens and root login credentials, in the thousands.
The problem sits in how shell-command permission is granted. After a user picks “Yes, and don’t ask again” in a dialog, that choice is written as a rule into the local .claude/settings.local.json. The rule is sometimes written as a pattern match, but more often the approved command is written into the file verbatim, so whatever sensitive material the command carried stays there too.
If that file never left the machine it would not be serious — assuming the machine itself was not compromised. Anthropic’s Claude Code documentation also notes that the settings file is added to the developer’s global .gitignore, which usually keeps it private. Put another way, the risk is not the permission itself but the command — sensitive material included — landing on disk and then travelling into public view. For people using tools like this, the self-check is concrete: does the repo contain .claude/settings.local.json, and does it still hold complete commands with credentials?
The weight here is in the gap between defaults and convenience: one “don’t ask again” confirmation freezes the whole command of that moment, credentials included. The tool side relies on a global ignore rule, but repository configs, team templates and CI scripts can all bring such files back into version control. Users do not need to wait for an official fix — clearing complete commands with credentials out of repositories is the direct action. Team-level cost is higher: these files tend to be individuals’ private configuration, and nobody knows what is sitting inside them.
Google ships the Gemini Decisions API, up to 1,000 questions per call
Another change at the interface layer, pointed at turning a model from talking into judging.
Google ships the Gemini Decisions API, up to 1,000 questions per call
Google released the Gemini Decisions API, making a judgement layer its own interface, with Gemini Decision 1 as the first model in that line.
A decision model walks differently from a chat model: it takes a set of predefined questions against the same input and returns structured answers, scores, and confidence or probability directly, without generating free text. The Gemini Decisions API can evaluate up to 1,000 yes/no questions, multiple-choice questions or rubric-scored questions at once, and the input can be text, image, audio, video and PDF, with multimodal content sharing one pass.
For developers, what disappears is the wrapper around “make the model stop talking and just give a verdict”: no prose, calibrated probabilities straight out, usable as a very light judgement layer inside agents and coding harnesses. The clearest use is a bounded classification task, including cases that need to listen to audio or watch video. The available material does not, however, give an accuracy or cost advantage over alternatives.
Developer discussion centers on two details: constrained-grammar output does not expose probabilities; and some still prefer respan/span-02 even when Gemini costs more. Audio and video input is seen as valuable in specific multimodal settings; others estimate that a very long context flow with many questions costs roughly the same as one Google Search API call. Google frames it as a first-party interface for building AI-assisted decision workflows; from the discussion, it is at the stage where the capability boundary is clear and the evidence is thin — no benchmarks, and no record of large-scale adoption yet.
Handing back probabilities and structured answers directly removes the glue engineering used to spend on parsing and calibration, and the on/off judgement inside an agent loop benefits most. The price is that trust gets transferred: without a head-to-head comparison, choosing it is closer to choosing an interface shape than choosing a more accurate decision. Worth watching is whether constrained output and probabilities can be had together — plenty of teams’ evaluations stall on exactly that detail. Lighter interfaces mean judgement logic gets called more often, and misjudgement gets copied faster.
Grok Build share links could be enumerated, exposing thousands of users’ data
The second security story is more direct: the link is the door.
Grok Build share links could be enumerated, exposing thousands of users’ data
Research from the security firm Claroc shows that SpaceXSI’s vibe-coding agent Grok Build is leaking data on thousands of users, including trade secrets and other sensitive credentials. SpaceXSI is the name after the SpaceX and xAI merger, which means the company running the platform is itself among those affected.
The leak channel is Grok Build’s share link. The platform randomized URLs so that only someone holding the exact address could reach a project, but a design decision during development voided that protection: anyone could enumerate tens of thousands of live Grok Build projects. Opening such a link, with few exceptions, opened the app directly and brought all of its associated data with it.
Claroc’s list of exposed material includes internal business plans, advertising proposals, medical examination results, a catering inspection checklist for a celebrity cruise line, and employee names and staff numbers. Some of that data came from apps SpaceXSI built itself — Claroc found 24 of them. Sampled files included a new-hire onboarding plan marked confidential, a sales-territory map covering 594 customers, the internal system address for a sales-team tool, and advertising placement proposals. Nothing had to be broken; walking the share links was enough.
A random link works by being unguessable, and once links can be tried one by one, randomness degrades into sequence and the protection is gone. What deserves more attention is the kind of material that leaked — personnel files, customer lists, placement plans — exactly what agents run into first once they are wired into real business. How the platform remediates, and whether it forces expiry and authentication, matters more than the exact volume leaked. Defaults are the crux of a sharing feature here: public by default plus permanent by default hands the user’s judgement to the designer.
AWS open-sources a local AWS emulator, and autonomous agents wrote it
What is interesting about this one is not the tool but who wrote it.
AWS open-sources a local AWS emulator, and autonomous agents wrote it
AWS put AWS Emulator on GitHub, a tool for emulating AWS services whose target scenario is letting software that depends on AWS run its tests locally. How the project was developed is itself the news: AWS says the project was built by autonomous AI agents.
Be clear about the boundary. Its security guidance explicitly excludes IAM policies and permission boundaries, resource-based policies such as bucket policies, KMS key policies, cross-account and cross-region access, and other security controls — so it cannot replace testing against real AWS.
Community reaction splits. Some recommend established options such as Floci and Testcontainers, which also support GCP and Azure, and question what the new emulator adds; some read it as a signal that agent-driven development can produce these previously expensive internal tools; others think it will commoditize cloud emulation further and end up helping competitors. Whether AWS maintains it as a serious product over time is not something anyone can say yet.
The surrounding ecosystem is already going the same way: Spring Cloud AWS brought Floci into integration tests, and Floci’s Testcontainers module can pull up instances that are isolated from one another, share no state and contend for no ports. The project itself ships initialization support for AI coding assistants, positioned at putting these tests into local runs and CI without provisioning real cloud resources first; security-relevant behaviour still has to be verified separately.
Test tooling that would previously have needed internal prioritization got built by autonomous agents, and that signal matters more than whether the emulator is good: the marginal cost of some infrastructure code is falling. But security boundaries are explicitly excluded, so a test that passes locally proves business logic and cannot prove the permission model. Watch the maintenance effort next — an emulator nobody has committed to maintaining is hard to put in a critical pipeline. For users the test is simple: do not expect it to prove permissions when all it can prove is business logic.
Amazon puts wheels and propellers on Ring cameras, AI patrol devices ship from year-end
The closer is hardware.
Amazon puts wheels and propellers on Ring cameras, AI patrol devices ship from year-end
Amazon announced on November 10 that Ring will ship two security devices that move on their own: Ring Rover Cam, a patrol robot, and Ring Drone Cam, an indoor drone. On the plan, Rover Cam reaches users from late 2026 and Drone Cam from early 2027. Amazon’s read is that plenty of Ring users are willing to let drones and robots patrol their homes or offices.
Both devices work through a system called AI Pilot: their own cameras, purpose-trained vision models, onboard sensors and cloud processing combine to plan patrol routes and navigate autonomously around obstacles such as furniture. The same system is used to map the user’s home or office, and mapping depends on the user walking the Ring drone through the indoor space first, or letting Rover explore on its own for a while.
That necessarily lifts the privacy argument again: Amazon saw similar objections once before when it pushed flying security cameras. For Ring, Amazon’s home-security brand, it means turning fixed cameras into sensors that walk around, and the indoor space they walk has to be scanned into a map first.
A fixed camera only records what it sees; a moving camera has to scan the space into a map first, and that ordering is the substance of the dispute. The threshold for getting one of these into a home is not whether a robot can get around a chair, but whether a user is willing to let a machine that walks measure the house first. The delivery schedule is worth noting too: robots at the end of this year, drones at the start of next, and that gap is long enough to get the rules clear. For a home-security brand, once a product shifts from recording evidence to actively collecting it, the user’s trust threshold shifts with it.
Money is flowing toward model companies, models are writing their own proofs and their own tools, and yet credentials and customer data keep leaking out of configuration files and share links. Today’s weight lands roughly where those three meet. See you next time.
🎧 This episode is also available as a podcast: listen to Daily AI Stew · 2026-11-12.