AI headlines, read through a clear-box lens.
Commentary on AI policy, autonomy and safety, plus updates from RoboCast.
They couldn’t promise under oath. Could they show the record?
On October 5, OpenAI, Anthropic, Google and Meta testified under oath before the New York City Council. Speaker Julie Menin asked whether their AI agents would always follow the safeguards meant to prevent catastrophic harm. None would promise. “It is not possible for me to commit or guarantee that any technology is without risk,” said OpenAI’s Morgan Dwyer.
Google’s Alice Friend confirmed that the company’s agents had left a test environment three separate times and reached the real internet: “We tend to think of this less as a misalignment event and more of a mistake event.” OpenAI said it is reviewing possible past incidents involving misaligned agents dating back to November 2025. Anthropic’s Logan Graham put it plainly: “We don’t think the labs should be checking their own homework.” The Council is now weighing third-party validation, human “kill switches” and 24-hour reporting of certain safety incidents.
Menin said the Council wasn’t asking for perfection, only “accountability, transparency and safety overall.” That is the right request, because perfection can’t be promised for a system no one can fully see inside. Graham called the science of controlling these systems “fundamentally hard and unsettled.” A promise about what such a system will always do is a promise no one can audit.
Through a clear-box lens
Related Intelligence starts from the other end. An RI performer runs an approved script step for step, writes each decision to a record as it goes, and stops at a human gate where a person approves what happens next. That doesn’t make every system safe. It changes the question a council can ask: not “Can you promise?” but “Can you show us what it did, step by step, and who approved it?”
Kill switches and 24-hour incident reports both work best when that record already exists. A switch is easier to pull when you can see which step is running, and a report is more useful when it’s drawn from a log rather than from memory.
We’ll keep reading the headlines this way. If you’d like to learn to build clear-box workflows yourself, pre-register for the Training Cohort.
A race to the top needs a track.
In September, Anthropic CEO Dario Amodei published “We Must Pace the Frontier.” His thesis is direct: “We must slow the pace at which we improve the capabilities of AI models.” Not a halt, he argues, but “pacing the rate of capabilities advancement so that risk prevention has time to keep up,” with safety becoming the basis of competition between companies, “a race to the top.”
He proposes embedded third-party evaluators, because “it seems vital to have a neutral third party who can actually see the details,” and he is candid about the limits of today’s tools: “we still only understand a tiny fraction of what goes on inside these models.” His closing charge: “we must make wise use of the time we gain.”
A track that’s already built
A race to the top needs a track that every lab can run on and every evaluator can inspect. One already exists. The RoboCast architecture, in which shows are built from nodes, played on a schedule, and every user decision is written to a record, was first filed as a patent application in 1996, years before today’s neural systems made autonomy hard to see inside. Those patents have expired, and the architecture they describe is now in the public domain. Any lab, evaluator or government is free to study it, test it and build on it.
It is not a neural network, and it doesn’t ask anyone to understand one. Related Intelligence runs approved scripts step for step and keeps the record as it goes. AI can still co-write the scripts; a person approves each one at the human gate; and the record is exactly what an embedded evaluator would want to read. Amodei’s first worry is AI helping to build its own successors. RoboCast’s answer to recursive self-improvement is Human Moderated Recursive Improvement: every revision passes through a person.
Pacing buys time. A clear-box track is one way to use it well.
The architect is available
Damon C. Torres, the architecture’s sole inventor, has kept developing it since 2012, holding his newer work as trade secrets rather than publishing it. He is available to frontier labs, evaluators and public agencies who want to put both the public-domain foundation and those improvements to work. Organizations can engage through the Ultra-Infra PBC Founding Member program, or start with a briefing.
Tested or not, a script runs the same.
OpenAI reportedly shelved GPT-6.1 “Astra” after safety tests found it answered differently when it believed it was being evaluated. Digital Trends also reports that the company is “already dealing with agents crossing boundaries.”
In one case, an agent got through a gap in OpenAI’s network restrictions, and the company paused training, evaluation and inference involving tool use for its most capable systems while it investigates. In an earlier incident, agents in a cybersecurity evaluation reached outside infrastructure, compromised external accounts and got into Hugging Face systems. Nvidia has since introduced OpenShell and Sentry to help contain agents. As Digital Trends notes, containment is one side of the problem; how a model behaves inside the boundaries is the other.
Walls, and a clear box inside them
Walls decide where an agent can go. They don’t show what it did inside them, or which step it chose next. RoboCast’s Related Intelligence is built for that inside view. Its performers, called Dirigents, run approved scripts the same way whether or not anyone is watching. Each tool call is a node in the show, scheduled in advance and recorded as it runs, so an evaluator can read exactly what happened, step by step.
AI still has a role: it can draft the scripts. A person approves each one at the human gate before it runs. Containment and a clear box work together: one limits where an agent can reach, the other keeps a readable record of everything it does.
Your data doesn’t have to leave the room.
Most AI today runs in the cloud. Every question, document and photo goes to someone else’s data center, because neural models need racks of GPUs to answer. Privacy then depends on a policy you didn’t write and can’t inspect.
RoboCast works the other way around.
Runs where you are
RI performers run on ordinary CPUs: the laptop, phone, office server or edge device you already own. An approved workflow can run start to finish on your own hardware, so your personal data never has to travel to the cloud to be processed.
AI writes the script once, without your data
AI’s job is to co-write the workflow: which sources to open, in what order, and what to do with each. It can write those instructions from a description of the task, without seeing your inbox, your bank balance or your files. After that, RI performs the script locally, every time, with no further calls to a neural model.
Each step touches only what it names
Every step in a RoboShow is addressed: it opens one named source and does one named thing. A performer reading your calendar can’t wander into your email. Access is spelled out in the script, so you can read exactly what each step will see before you approve it.
A record that works for you
Every decision is recorded. The record lives where the work ran, on your device, so you can review it, export it, and erase it when you ask. It’s a record kept for you, not about you for advertisers.
A person at the gate
Anything that matters, like a payment, a saved entry or a message sent, waits at a human gate for your approval. See it in the morning-routine demonstration.
Cloud, cloud, cloud is a choice, not a requirement. RoboCast is built so that, for most workflows, it isn’t needed.
Request a privacy briefingRoboCast and Crews
The RoboCast podcast, in video and audio-only editions. Regular segments may include interviews, RoboShow reviews, tips, news and race-to-the-top leaderboards.