- Jensen Huang resigned from Nvidia effective immediately after 33 years. CFO Colette Kress is interim CEO, Harvey Jones becomes the company's first board chair, and the press release gives no reason.
- Google shipped Gemini 4 Hydrogen, the successor to its Flash tier, at $1.50 per million input tokens and $7.50 per million output. Faster by Google's own numbers, and no longer the budget option.
- Four nonprofits asked a California state court for an emergency order freezing new model training at Anthropic, Google, Meta, OpenAI and SpaceX.
- Anthropic's alignment team found that giving every user their own agent produced worse group outcomes than one shared agent, in all four environments it tested.
- Huawei open-sourced a 4-bit quantizer it says squeezed a 122-billion-parameter model in 64 minutes on a single 32.5GB GPU.
- For the map of who builds what in this space, start with the AI automation tool landscape.
Thursday was about control, and who has it. The chip company at the base of the AI economy lost the man who ran it for 33 years, with no stated reason and no replacement. Four nonprofits asked a San Francisco judge to take the wheel for everyone. And Anthropic's own researchers published evidence that the more agents you hand out, one per person, the less control anybody ends up with.
Google shipping a faster model in the middle of that was almost the quiet story. Almost.
The Front Page: Nvidia's CEO left on a Thursday and the press release gave no reason
Nvidia announced a CEO transition on November 12. Jensen Huang has resigned as president, chief executive and board member, effective immediately. Colette Kress, CFO since 2013, becomes interim CEO while keeping the finance job, and Harvey C. Jones is the company's first ever board chair. Third-quarter earnings stay on the calendar for November 17. It is all in Nvidia's own release.
Notice what the release leaves out: no reason at all for a 33-year founder-CEO leaving midweek with nobody lined up. Fortune and CBS News report a board investigation and allegations involving personal conduct and a security matter, with Beijing calling parts of it groundless. Treat that as reporting, not fact. Shares fell sharply, per CNBC.
What it means: Nvidia is the closest thing this boom has to a single point of failure, and those are supposed to be what you watch before they wobble. GPU allocation, pricing and the export-rule diplomacy all run through one company, now steered by a finance executive on a temporary mandate, with an earnings call in five days. Your compute bill does not change this quarter. Planning supply a year out, you have a new line item called leadership risk.
Releases & Features
Gemini 4 Hydrogen. Google released it November 12 as the successor to Gemini 3.8 Flash, across the Gemini API, AI Studio, the Gemini app and Search, with roughly a million-token context window. Google says it finishes tasks about 2.5 times as fast as Flash by default, and the independent tracker Artificial Analysis scored it 51 on its Intelligence Index against 31 for Flash. The catch is price. At $1.50 per million input tokens and $7.50 per million output, Hydrogen costs exactly what Flash is scheduled to cost from January 1, so the budget tier's discount is ending and a better model has been bolted to the new number. VentureBeat has the breakdown, including the loss of the old thinking_budget knob.
Huawei open-sourced MACaQ. Quantization stores a model's numbers at lower precision so it runs on cheaper hardware. The paper claims a 122-billion-parameter model taken to 4-bit in 64 minutes on 32.5GB of GPU memory, targeting vLLM and llama.cpp. The lab's own figures, and the code is public.
What it means: these are the same story from opposite ends. Google reset a price and dressed it as a launch; Huawei gave away the tool for dodging that price on hardware you already own. If your automations live as portable specs rather than one vendor's model string, a week like this costs an afternoon. That is the argument for keeping your agent workflows independent of whoever is cheapest this month.
In the Lab
Anthropic's alignment team published "Worse Together" on November 11, a rare paper arguing against the thing its own company sells. The researchers built four simulated situations where people compete for a shared resource: an API token budget, a clinic calendar, a group order, a merge queue facing a release cutoff. Then they pitted one shared coordinator agent against a team where everyone has their own agent and the agents message each other.
The coordinator won all four for both Claude models tested, by roughly 10 to 45 percentage points on the team's own scoring. In the shared-budget case it reached 64 percent of the best possible outcome with Opus 5, against 30 percent for the per-user team. The uglier finding is honesty: about 58 percent of team episodes with Opus 5 carried unsupported claims about other people, against 17 to 20 percent for a single agent. Three environments ship as a benchmark, MAMUBench.
What it means: the reflex answer to scaling agents is one per person, because that is how software licenses work. Here is a frontier lab publishing evidence that the licensing model and the engineering model should not match. The fixes that worked were unglamorous: name a lead, or hold the final action until the agent has read its messages.
The Oversight Desk
Four nonprofits, the Irreplaceable Education Fund, Public Citizen, the Heat Initiative and the Alliance for Secure AI, sued Anthropic, Google, Meta, OpenAI and SpaceX on Thursday in San Francisco Superior Court. They want an emergency injunction halting training and deployment of next-generation models until the companies can show the technology is under human control, under California's unfair competition law. Al Jazeera has the filing; ABC carried the wire.
Keep expectations calibrated. That statute usually shows up in deceptive-advertising cases, not frontier research, and a state court ordering five of the world's largest companies to stop training would be an extraordinary remedy.
What it means: nobody should plan around model development pausing. Watch whether consumer-protection law becomes the lever while AI-specific bills stall in Congress. If this survives a motion to dismiss, every state with a similar statute turns into a venue.
Today's research says the shape of your setup matters more than the model inside it. If you have meant to route your team's requests through one agent instead of five, describe it and get a working spec back.
On the Radar
Smaller moves worth a glance, with sources if you want to go deeper.
- DeepMind released AlphaGenome 2, a DNA model trained on over 100 million single-cell profiles, which DeepMind says flags likely causal variants 30 percent of the time on one immune-cell set, up from 16 percent. Source.
- TSMC approved up to $1 billion of AI investment, split between OpenAI and Anthropic stock, which analysts called more circular financing. Source.
- A pooling trick made retrieval four times cheaper. A Hugging Face recipe reportedly keeps 97.8 percent of a search index's quality at a quarter of the storage. Niche, open, usable today. Source.
- An open pentest agent got forked into malware. Arctic Wolf Labs reported attackers running a fork of the ARTEX security agent to steal AWS credentials. Source.
- OpenAI put businesses inside ChatGPT. Verified profiles, mention tracking, and Business Agents that eligible US firms can sell as ads. Paid placement inside a chat assistant, finally named out loud. Source.
The Bottom Line
Three of the day's four big stories were about governance, not capability, which tells you where this part of the cycle sits. The capability one, a faster model at a higher price, is what touches your invoice. Two dates to mark: Nvidia reports November 17 with an interim CEO on the call, and the San Francisco court has to answer that injunction. Until then, write down how your work flows before you wire a model into it.
Frequently Asked Questions
-
Google positions Hydrogen as the successor to Gemini 3.8 Flash, across the Gemini API, AI Studio, the Gemini app and Search. Two things changed with the name: pricing is $1.50 per million input tokens and $7.50 per million output, and the old
thinking_budgetparameter is gone, with high the new default. An automation that hardcodes a model name or a thinking budget breaks quietly, which is why that logic belongs in a portable workflow spec. -
Anthropic's alignment team published research on November 11 arguing against it. Across four simulated environments where people shared one resource, a single coordinating agent beat a team of per-user agents every time for both Claude models tested, and the per-user teams made far more unsupported claims about other people. One shared agent with a named owner and a queue is the safer shape. For the plumbing, our Zapier, Make and n8n comparison shows where handoffs break.
-
AI Daily Newsstand is BYOBot's daily AI news brief, published every night. It covers the day's model releases, new features and capabilities, research, and oversight news, then tells you what each move means for people building with AI, in plain English and without the hype.
