Three of this week's biggest AI moves point the same way: power in the stack is being handed away at the bottom and concentrated at the top. On Tuesday the chiefs of the two leading labs asked the UN Security Council to help rein them in — Sam Altman in person, Dario Amodei and Clément Delangue by video. The day before, Alphabet's Intrinsic open-sourced the core of its robotics platform under Apache 2.0, and a six-month-old lab built to automate AI research itself went into talks to quintuple its valuation in ninety days.

Here's the whole edition in one screen — the three moves, and the one thing to do about each:

The uncomfortable, useful read is that the two directions are the same story. The base — models, and now robotics stacks — keeps getting handed to you for free, which is pure leverage if you build on it instead of reinventing it. The frontier — automating the research that makes the next model — is where the biggest checks now go, which tells you where not to plant a flag unless you have a genuine edge. And the referee is finally showing up: the UN session is the earliest public sketch of the rules you'll eventually ship under. Three moves on that, below.

1. The lab chiefs asked the Security Council to constrain them#

The most striking thing about Tuesday wasn't a product — it was a venue. On Sept 23, 2026, the UN Security Council held a session on AI and international security, chaired by French Foreign Minister Jean-Noël Barrot during France's September presidency of the council. OpenAI's Sam Altman appeared in person; Anthropic's Dario Amodei and Hugging Face's Clément Delangue joined by video (France 24, BNN Bloomberg). Altman called for "extreme care"; Amodei went further, pledging to slow development where needed to keep systems safe.

The substance was in Amodei's three proposals for international cooperation: narrow global agreements on the most dangerous uses (his example: a ban on using AI to build biological weapons); evaluation-and-verification systems so countries can independently check each other's safety commitments instead of trusting them; and common global standards for testing models, paired with an incident-notification system so AI security failures get reported across borders. This lands two days after the UN's independent science panel warned that agents had already slipped their controls — the panel described the risk; the Security Council session is the political escalation of it.

What it means. No rules passed — this was a briefing, not a vote, and the council took no binding action. But the load-bearing word across all three proposals is verifiability: mechanisms that let an outsider confirm a claim rather than accept it. That's a direct preview of the compliance surface headed for anyone who ships agents. You won't be asked to sign a treaty, but a serious enterprise buyer will increasingly ask how you test your system and what you do when it misbehaves. The cheap move now is to make the answer already true: keep an evaluation record (what you test, on what, and the results) and an agent-incident log (what went wrong, when, and how it was contained). It's the same zero-trust posture for agents we keep returning to, now with a paper trail — and your container is not a sandbox is the first line of it.

2. Alphabet open-sourced its robotics core — and framed it as a commitment#

While the diplomats talked, Alphabet gave something away. At ROSCon 2026 in Toronto on Sept 22, its robotics unit Intrinsic open-sourced "Intrinsic Core" under the Apache 2.0 license — and not a demo subset. The release is the same set of capabilities Intrinsic uses in real manufacturing deployments: a hardware-agnostic real-time control framework, motion and grasp planning, pose estimation, a digital twin, simulation, camera calibration, and preconfigured ROS-compatible hardware drivers (The Robot Report, SiliconANGLE). CTO Brian Gerkey called it a "durable commitment" — explicitly distancing it from the one-off open-source drops that ship once and rot — and said the team audited the existing ROS ecosystem first, releasing only components that didn't already exist as open source.

What it means. If you build in physical AI or robotics, the direct implication is a build-vs-adopt call: evaluate Core instead of hand-rolling your own control and planning stack, and weigh the "durable commitment" claim against how maintained it actually stays (a promise from a CTO is a lead indicator, not a guarantee — watch the commit cadence). But the broader lesson generalizes past robots. This is the base layer being handed to you for free, the same pattern as open-weight coding models you can now self-host: the foundational plumbing keeps getting commoditized by the biggest players, because owning the ecosystem beats owning the code. For a founder that's a standing instruction — don't build the layer someone with a thousand engineers is giving away. Put your originality in the application, the domain data, and the workflow on top.

3. Mirendil: a 5x valuation in ninety days, for automating research itself#

The money story is the clearest read on where the frontier is moving. Mirendil, founded roughly six months ago by Behnam Neyshabur (CEO — he led a scientific-AI-reasoning team at Anthropic and spent 5+ years at Google DeepMind) and Harsh Mehta (CTO, ex-Anthropic senior research scientist), is in talks to raise up to $1B more at a $5B valuation, led by Kleiner Perkins (Bloomberg, Tech Funding News). That's three months after a $200M seed at a $1B valuation (a16z and Kleiner co-led, Nvidia in). The pitch: AI that does the work of an AI researcher — designing experiments, searching for the right settings, evaluating models, and running the next round of training — sold as a platform for scientific discovery in biology and materials.

What it means. Strip out the eye-watering number and the signal is a direction: the biggest bets are moving up the stack, from the model to the thing that produces the model. That's consistent with the funding pattern we mapped when control, not raw capability, won the summer — investors are paying for leverage over the process, not another general chatbot. For a solo builder the read isn't "raise at $5B"; it's two disciplines. First, pick a wedge you know cold — Mirendil's founders are pricing on a legible thesis plus a rare pedigree, and you compete on domain depth, not on being a cheaper general lab. Second, treat the valuation velocity as a weather report, not a floor: capital this fast-moving reprices just as fast, so build revenue where you can and don't assume today's multiples are the ones you'll raise on next year.

The one-week picture#

Three moves, one motion. Power is being given away at the base — an open robotics core, cheaper models every month — and concentrated at the top, where the money chases whoever automates the next layer up. In between, the referee is arriving: the Security Council session is the first public sketch of the testing, verification and incident-reporting norms you'll eventually operate under. So build on the free base rather than rebuilding it, aim your original work one layer above the commodity, keep the evaluation and incident records the governance conversation is converging on — and if you run agents anyone can reach, the same discipline applies to the personal ones too: more capable every month, and still not to be trusted with the keys.