Mark Beall

How might conservatives grapple with AGI? americanmind.org/salvo/a-conser…

Every year, I share this video of French caretakers who take sand from Omaha Beach in Normandy, and scrub them into the letters to give them the gold coloring. They do this for all 9,386 US soldiers who died. France also gave us this land as American soil. #MemorialDayWeekend

All of my favorite things

That water clarity is an engineering decision, and the math behind it is wilder than the video. Roman aqueducts ran on gravity alone. No pumps, no pressure systems. Engineers carved channels with a gradient so shallow it borders on absurd. The Pont du Gard in southern France drops 2.5 centimeters over 275 meters. That's roughly the thickness of a coin over the length of three football fields. They surveyed that accuracy with plumb lines and wooden leveling instruments. The clarity you're seeing is a direct product of flow velocity. Too steep and the water erodes the channel walls, picks up sediment, turns brown. Too flat and it stagnates. Roman engineers targeted a slope of about 20 centimeters per kilometer, which kept the water moving fast enough to stay fresh but slow enough to stay clear. Before the water reached the city, it passed through multi-chamber settling tanks where velocity dropped near zero. Suspended particles sank. Clean water flowed out the top into the next chamber. Repeat three or four times. Pliny specified the minimum slope in writing. Vitruvius published the exact mortar ratio for hydraulic cement: one part lime to two parts volcanic ash for underwater work. The pozzolana from Pozzuoli reacted with water to form a calcium-aluminum-silicate compound that actually gets stronger the longer it sits submerged. Modern concrete degrades in water. Roman concrete bonds with it. Scale the whole system and it gets harder to process. Eleven aqueducts fed Rome at its peak. Combined output: roughly 1 million cubic meters of water per day. That works out to about 250 gallons per person for a city of one million. Modern New York delivers about 125 gallons per person per day. Ancient Rome had access to double the per capita water supply of the largest city in the United States, running entirely on slope and stone. The Trevi Fountain in Rome is still fed by one of them. Two thousand years, same source, same gravity, same water.

Last month we launched Project Glasswing, our collaborative AI cybersecurity initiative. Since then, we and our partners have found more than ten thousand high- or critical-severity vulnerabilities in essential software.

Maybe we need that ASI

The Empire Strikes Back

Humans using Mythos as seen by Mythos

Fascinating info trickling out about LLMs in active military ops in Iran, e.g. Air Force AI tools converting intelligence from top secret -> secret. Means info gets to pilots & drone operators in "seconds" rather than 20-30 min, saving "a lot of aircraft" from incoming threats.

🇺🇸

It’s finally happening 🇺🇸

Today, we share a breakthrough on the planar unit distance problem, a famous open question first posed by Paul Erdős in 1946. For nearly 80 years, mathematicians believed the best possible solutions looked roughly like square grids. An OpenAI model has now disproved that belief, discovering an entirely new family of constructions that performs better. This marks the first time AI has autonomously solved a prominent open problem central to a field of mathematics.

This work is the culmination of years of effort on AI evaluation science and third-party risk assessment and disclosure mechanism design. It feels like a big milestone for METR. We designed this new procedure with an eye toward “showing by doing” how we think evaluations for the AI loss-of-control threat model should work: laying out a process that can be done periodically, not just immediately pre-deployment, and holistically assessing risk inside of an AI lab, rather than just an individual AI system. The exercise also involved significantly deeper access than we've previously had, including raw chains-of-thought from the developers' best models and info about private model training & control protocols. The report is long, with a bunch of new evaluation results and documentation of our process. Please, check it out, or at least the executive summary!

Today @NSF announced a major investment of $1.5B towards a new model for scientific research🚨 NSF X-Labs will fund independent teams of researchers, engineers, and entrepreneurs to pursue bold milestone-based scientific challenges. This is how we revitalize America's scientific engine for the 21st century outside of traditional institutions, conducting science in a way that actually reflects the modern R&D ecosystem ⬇️ nsf.gov/news/nsf-annou…

After agreeing to discuss AI guardrails with China, Trump is now #1 on polymarket to win the nobel peace prize.

More incredible thought-leadership on AI x national security for @SenatorBanks

🧵🧵🧵 1/ Now to dispel myths propagated by the tech bro oligarchs. Myth 1: “The people pushing mandatory AI safety testing are mostly leftist ‘doomers.’” Reality Check 👇 h/t @FamStudies

Tim Apple ☠️

China did not hack Anthropic. Did not bribe Anthropic. Did not invade Anthropic. They sent a think tank emissary to Singapore and quietly asked Anthropic for access to its most powerful AI model, the system behind Mythos. When Anthropic refused, the White House did not treat it as a minor diplomatic courtesy. It treated it as a warning. That reaction alone tells you more about the new Cold War than any Pentagon briefing. We are not talking about tanks or carriers anymore. We are talking about model weights, inference servers, and which country gets to see the next leap in frontier AI first. This was not a failed espionage plot. It was a discreet probe. Beijing is testing whether it could acquire what it cannot reliably build on its own. The fact that they did not even try to disguise it as academic collaboration says a lot about how openly strategic this game has become. Behind the scenes, the US government is already scrambling to regulate how powerful models like Mythos and OpenAI’s latest cyber-focused GPT 5.5 variants are deployed, because the same AI that can hunt for vulnerabilities in software can also be used to find them in air defense networks, power grids, and missile-sensing infrastructure. In other words, the AI race is no longer an abstract tech debate. It is a national security race where access to a single model can shift the balance of power, and one quiet note from Anthropic in Singapore just became a geopolitical event. If you are not watching model licensing deals, back-channel outreach, and export control foot traffic around the big AI labs, you are still fighting the last war.

Exponential on exponential

One of the things that made the Mythos release hard to interpret is that Anthropic held back details on most vulns they found, to give defenders time to patch. 1 month later, info from orgs with access to Mythos is starting to trickle out, e.g. this post from Mozilla today: