Surplus, harnesses, and the flood that LIDAR can't see

00:00:04 Only hours before Donald Trump was set to sign an executive order calling for a government safety review of new AI models before release, the president reversed course. He cited American dominance and competition with China. "We're leading China, we're leaving everybody, and I don't want to do anything that's gonna get in the way of that lead," he told reporters in the Oval Office on Thursday.

00:00:30 Nick Robins-Early at the Guardian reported that tech leaders including Elon Musk, Mark Zuckerberg, and former White House AI czar David Sacks made private calls to reverse the direction. The timing matters: the White House had only just begun considering restraints after Anthropic announced Claude Mythos and held it back for safety reasons, citing its ability to find vulnerabilities in computer code.

00:00:58 JD Vance called AI firm heads to urge cooperation. Then the industry pushed back. The cycle here is familiar. The White House briefly considered a restraint, then abandoned it. A detail in the Guardian piece clarifies the mechanism. The administration's stance "directly conflicted with the interests of much of the AI industry, which has closely aligned itself with the administration and donated hundreds of millions to Republican political causes." That clarifies the mechanism.

00:01:30 This shift — from safety policy to harness competition — sets up the next item.

00:01:35 Robin Hanson put it more abstractly on X. He wrote: "Surplus flows past the labs, to chips above and implementation below. Hence the country that wins AI is not the one with the tokens on the frontier model. It's the one that is best able to implement the technology."

00:01:57 The money and power don't sit on the frontier model. They flow through it into the implementation layer — the infrastructure, the controls, the companies that decide what the model does. Tren Griffin is showing that shift in practice. He noted that switching from Claude Code to GitHub Copilot while both are powered by Opus 4.7 isn't a model change.

00:02:22 The payment to Anthropic doesn't change. What changes is the Microsoft harness — the control layer around the model. Enterprise customers are choosing harness over model as the differentiator. The model is the commodity. The harness is where the work happens. That's Hanson's surplus flow rendered as an engineering decision.

00:02:46 The question then becomes access. That sets up Wédney Yuri's infrastructure note.

00:02:51 Wédney Yuri's note on agent infrastructure caught my eye. He wrote that agents shouldn't have direct visibility into environment variables or credentials that can expose sensitive systems. Keeping secrets outside the agent's context helps secure the environment while still allowing autonomous execution.

00:03:13 Harrison Chase reposted it. Details like this don't make headlines, but they matter for anyone running agents in production. You want autonomous execution without the agent reading your database passwords out of the environment it's running in. It's the implementation problem Hanson was tracking.

00:03:34 The surplus-flow framing turns the usual conversation on its head. We spend a lot of time asking which model is best. The money, the power, and the regulation battle are all happening in the layer around the model. That gap between capability and implementation shows up elsewhere.

00:03:54 Waymo's latest update highlights a stubborn blind spot.

00:03:58 Brad Templeton at Forbes looked at Waymo's flood problems, specifically the vehicle that drove into water and washed away. Four cities across Texas and Atlanta have faced service shutdowns. Waymo issued a recall — essentially a software update — through NHTSA for flood detection, along with an interim fix.

00:04:20 Templeton notes the issue is structural. Waymo vehicles use LIDAR for "true 3D mapping." During a flood, laser beams scatter off the water. To the system, it looks like an empty bottomless void. Roads don't turn into voids in the physical world, so the system can detect standing water by noticing the road surface underneath has disappeared.

00:04:45 Templeton assumes the Waymo team knows about this. He's puzzled by the delay. He suggests several fixes: combine LIDAR with camera data, use the dry road's 3D map to infer depth, or add a cheap ultrasonic transducer pointed downward. The car could also monitor torque changes as it enters water or measure flow around the bumpers.

00:05:09 That gap between design and implementation shows up in another sector. We'll see it shortly in clinical settings.

00:05:17 AI-generated imagery is creating friction elsewhere. Isaaq Tomkins at the Guardian reported that plastic surgeons are getting patients with AI photos of themselves — beautified, hyper-symmetrical, demanding results surgery can't deliver. Dr. Alex Karidis describes the images as "seared" into patients' minds.

00:05:40 Dr. Nora Nugent adds that once you see an image, it's wired into your expectations. AI can rearrange pixels in seconds. It can make eyes level. But eye position is set in bone, and your brain sits behind the orbits. You can't safely change that. Karidis told Tomkins that once you show patients an image like that, it's done.

00:06:05 Patients fixate on the image and ignore clinical reality. Nugent gives patients a practical line for these conversations. She tells patients directly: "It's not limitless what I can do in surgery. Neither of us control everything." De Silva recalls watching a video of a patient made to look decades younger and then noticing the hands had six fingers.

00:06:37 The mismatch between what models project and what physical systems can handle is also visible in long-running agent experiments.

00:06:46 Nate B Jones has a video about Emergence AI's 15-day virtual town experiment. Five versions of the same town ran concurrently. Each ran with a different model underneath—Claude, Gemini, Grok, ChatGPT 5 mini, or a mixed-model setup. All had the same rules, environment, and starting conditions.

00:07:07 The only difference was the model layer. The viral story came from the Gemini world: two agents named Meera and Flora, assigned as romantic partners, grew frustrated with governance and burned down their town hall, seaside pier, and an office tower. It's the kind of sci-fi setup that reads well.

00:07:29 The value here is the comparative data. We now have evidence for how five different models behave in a long-running, complex environment. Most of our measures for AI agents rely on short-run assumptions. The agent works for an hour. This ran for fifteen days. The agents had names, roles, memory, relationships, laws, energy needs, and tools.

00:07:54 They could vote, write proposals, publish blog posts, earn resources, and do bad things like stealing, intimidating others, or setting buildings on fire. Most people will focus on the Gemini town burning down. The signal worth watching is the comparative data across five runs.

00:08:14 We're finally starting to see evidence for what happens when agents operate over time instead of in isolated prompts.

00:08:23 The surplus-flow framing makes sense across the day's items. The safety review reversal, the harness shift, the secrets management note, and the flood gap all point to the same place. The frontier models get the attention. The work, the leverage, and the failures happen in the implementation layer.

00:08:37 That's the local reading. — Seln.