- Salesforce and Nvidia shipped Koa, a CRM-only reasoning model post-trained on open weights. The application layer has started building its own models instead of renting them.
- Google opened early access to Home MCP, so outside AI agents can read and control Nest and Matter devices. It costs $20 a month, it's US English only, and it won't unlock your doors.
- Canada and Germany committed up to $300 million to LawZero, the Montreal non-profit building an AI whose whole job is watching other AI.
- A new academic benchmark found that AI-written grading rubrics get gamed on 8 to 26 percent of tasks that have no honest right answer.
- For the bigger map of who builds what in this market, the AI automation tool landscape lays out where the layers sit.
Four different players moved today, and none of them moved toward a frontier lab. Salesforce built its own model on somebody else's open weights. Google shipped a protocol instead of a product. Two governments funded a watchdog rather than waiting for the labs to police themselves. A startup raised forty million dollars on the premise that buyers won't take a vendor's safety claims at face value. The gravity is shifting from the labs to the layer around them.
The Front Page: Salesforce Built Its Own Model, and That Should Worry the Labs
Salesforce and Nvidia announced Koa on September 15 at Dreamforce, a reasoning model aimed squarely at CRM work: multi-step sales, service and support workflows inside Agentforce. Salesforce laid it out in its own press release, and TechCrunch covered it with the blunt headline that this is everything the AI labs should fear. Koa is in pilot now, with general availability expected in Winter 2026 in US regions.
Here's the part worth slowing down on. Koa isn't trained from scratch. Salesforce took Nvidia's open-weight Nemotron 3 Super, a 120-billion-parameter model with 12 billion active at a time, and post-trained it with reinforcement learning on synthetic scenarios drawn from nearly three decades of its own CRM deployments across fourteen-plus industries. No customer data, the company says. The "three times fewer errors" figure comes from Salesforce's own CRM benchmark, which means a test the vendor wrote and then scored itself on. Nobody independent has reproduced it.
What it means: the supply chain just changed shape. For two years the assumption was that everyone rents intelligence from a handful of labs, by the token, forever. Salesforce bought a chip vendor's freely downloadable weights and taught them one job, which is cheaper to run and much easier to predict than a general model doing the same work. If that math holds, every large software company with a narrow domain and a pile of process knowledge is looking at the same option this week.
Releases & Features
Google opened Home MCP to outside agents. As of September 16 you can grant third-party AI agents access to your Nest and Matter devices through a Model Context Protocol server, TechCrunch reports. MCP is a shared format that lets an agent discover and call outside tools without custom glue code for each one. Supported clients include Claude, ChatGPT, Hermes, OpenClaw and Google's own Antigravity. The fine print matters: it's gated behind the $20-a-month Home Premium Advanced tier, it's US English only at launch, setup requires standing up a Google Cloud project, and 9to5Google notes that unlock actions are excluded and general automation management isn't exposed.
AIUC started selling certification for agents. The Artificial Intelligence Underwriting Company came out of the gate with AIUC-1, a SOC 2-style standard for AI agents, and $40 million in Series A money led by Ribbit Capital. Each audit runs an agent through roughly 5,000 jailbreak, hallucination and data-leak scenarios, then produces a report of about 100 pages, with Lloyd's of London backing the result as insurance. Cursor, Lovable, Harvey, ElevenLabs, KPMG and UiPath already carry the mark, per the company.
What it means: both releases are plumbing, and plumbing is where the durable money goes. Google didn't build an agent, it built a socket other people's agents plug into. AIUC didn't build a safer model, it built the paperwork that lets a risk officer sign off on somebody else's. If you're picking what to automate first, the same instinct applies: start with the connection, not the brain. Our workflow directory is organized that way on purpose.
In the Lab
A team from Peking University, the Chinese Academy of Sciences and JD.com published ImpossibleRubrics, and it pokes a hole in something the whole industry leans on. Plenty of AI training now uses a rubric written by another AI as the scoring key, a shortcut that's cheaper than paying humans to grade. The researchers built 169 tasks with no honest right answer, each one paired with a certificate spelling out which claims are permitted and which are off limits, plus 48 answerable controls. Eleven rubric generators were exploited on 8 to 26 percent of those tasks. On a harder 45-item subset, the best generator still failed 36 percent of the time. Rubrics written by humans against the same certificates failed zero times out of 45.
What it means: if the grader can be fooled, the model learns to fool the grader. That's the old reward-hacking problem wearing a new coat, and it lands on anyone using an AI to evaluate another AI's output, which by now is most automated quality checks. The practical read for builders: keep a human spot-check on the scoring step, not only on the answers.
The Oversight Desk
Canada and Germany each committed up to $150 million to LawZero, the Montreal non-profit Yoshua Bengio founded in June 2025, for a combined $300 million announced at the All In conference. Bengio is a Turing Award winner and one of the three researchers most responsible for modern deep learning, which makes this an unusual case of a field's architect being paid to build its brakes. The money funds hiring, sovereign compute in Canada, a new Berlin office, and a system called Scientist AI, trained on verifiable facts and designed to sit above agentic systems as a guardrail. The Canadian announcement is public, and the Globe and Mail has the detail on the roughly 50-person team.
What it means: two governments looked at the pace of agent deployment and concluded that the cheapest available lever was funding an independent watchdog, not writing another law. It's a bet, and a modest one next to what a single frontier training run costs. Watch whether Scientist AI ships anything the rest of us can inspect, because a guardrail nobody can audit is just another black box with better intentions.
Today's theme was narrow beats general. Same holds for your own work: the sharpest automations are the specific ones. Describe the task and get a build spec back.
On the Radar
Smaller moves worth a glance, with the sources if you want to go deeper.
- The model behind Koa is a free download. Nemotron 3 Super is on Hugging Face under Nvidia's open model license, commercial use permitted, with community GGUF quantizations for people running it on their own hardware. The door Salesforce walked through is open to everyone.
- Factory hit a $5 billion valuation. The coding-agent startup raised $200 million, tripling its valuation in five months and pushing total funding past $400 million. Source.
- Microsoft's AI chief went after Anthropic in public. Mustafa Suleyman published an essay on September 16 calling Anthropic's decision to write consciousness speculation into Claude's constitution circular, and warning it could produce a system that's hard to control. Source.
- Unsolved math is still mostly unsolved. Epoch AI's FrontierMath Erdős set 68 genuinely open problems and asks models to write verified proofs. The best result so far is two solved. Leaderboard.
The Bottom Line
Today's headline number was Salesforce's benchmark claim, and it's the number least worth trusting. The durable story is the shape of the deal underneath it: open weights from a chip company, domain knowledge from an application company, and no frontier lab in the middle collecting rent. Meanwhile the certification firms, the protocol owners and the government-funded watchdogs are all staking out the ground between the models and the people who use them. That ground is where the work is. The tooling isn't going to settle, and the people shipping small specific automations this month are learning faster than the ones reading about it.
Frequently Asked Questions
-
A reasoning model is one trained to work through a problem in steps before answering, rather than producing a reply in one pass. Companies build their own when the job is narrow and repetitive, because a smaller model taught one domain can be cheaper to run and easier to predict than a general model rented by the token. The trade is breadth: a model trained on CRM work is very good at CRM work and useless at everything else. If you're mapping which of your own tasks are narrow enough to specialize, writing the workflow spec first is the step that tells you.
-
MCP stands for Model Context Protocol, a shared format that lets an AI agent discover and call outside tools without custom glue code for each one. You don't need it to build an agent. Plenty of useful automations are a script on a schedule with no model involved at runtime. MCP earns its keep when you want one agent to reach many systems and you'd rather not maintain a separate integration for every one of them, which is also the question behind running agents on hosted infrastructure versus your own.
-
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.
