The physical-world lag no longer looks uniform

Material update: yes — not enough to move the 2029–2031 date range, but enough to change what I mean by the “physical-world lag.”

The strongest new evidence is Anthropic’s August 27 release of the Model Hardware Standard (MHS). This is not merely a proposal for future robot interfaces. Several labs have already used frontier agents through it to operate real scientific hardware in closed loops: robotic arms, liquid handlers, plate readers, microscopes and quantum-computing equipment. MHS is model-agnostic and is intended eventually to become open source. Anthropic says it can cut integration of programmable equipment from weeks or months to hours or minutes. [1]

The individual demonstrations are more important than the standard itself. At Carnegie Mellon, an Opus 4.8 agent was connected to four pieces of mutually incompatible lab equipment in about eight hours rather than the several weeks typical of vendor automation. It then autonomously ran a dose-response experiment, recognized that its first curve was bad, discarded it, changed the concentration range, ran another physical experiment and accepted the improved result—with no human intervention during that loop. [2]

At QuEra, four specialists had previously spent two to three weeks writing automation for a difficult quantum-computer laser-control problem. Claude instead ran hundreds of real experiments itself, generated an inspectable conventional controller, and later adapted to a second laser wavelength in a single unattended night—a commissioning task QuEra says normally takes weeks. The resulting controller recovered 695 of 700 induced faults, usually within six seconds versus five to ten minutes for a human specialist, and reduced residual noise fivefold. Humans still specified the problem, safety envelope and success criteria. [3]

Genentech provides an especially useful reality check. Claude successfully optimized liquid-handling parameters experimentally and recovered from several hardware faults, but when physical bubbles produced a failure mode it did not understand, it kept retrying in a way that made the problem worse until a human explained the physics. So this is not “AI scientist operates arbitrary laboratory unattended.” It is evidence that the digital/physical boundary is becoming much less of a barrier for equipment that already exposes programmable controls. [4]

A separate development points in the same direction. On August 19, the FDA authorized Aletta, the first standalone robotic system allowed to perform essentially an entire diagnostic blood draw without hands-on human operation: finding the vein, distinguishing veins from arteries, inserting the needle, swapping tubes and bandaging. Clinical results were comparable to or better than trained phlebotomists when the machine elected to make a stick. A trained phlebotomist still initiates and supervises the procedure, but one person can oversee three machines. [5]

However, today’s large Reuters investigation of Chinese humanoid robotics provides strong counterevidence against collapsing the broader physical lag. China shipped about 95% of the world’s roughly 20,000 humanoids in 2025 and is massively subsidizing training and manufacturing, yet Reuters found that current humanoids remain too slow, unreliable and inflexible for most useful factory work. One training center reported usable demonstrations taking anywhere from roughly 50 to 300 attempts; simple tasks can run at only about 20% of human throughput; and industry estimates suggest roughly 500,000 hours of high-quality embodied training data exist versus perhaps 100 million ultimately required. Narrow deployments are emerging, but ordinary industrial robot arms still dominate real production. [6]

Compute scaling meanwhile remains fully consistent with the existing timeline rather than showing a slowdown. Nvidia reported August 26 that quarterly data-center revenue more than doubled to $89 billion, expects fiscal-2028 revenue growth of about 70%, says neo-cloud providers should end 2026 with more than 8 GW of Nvidia GPU capacity versus 3 GW a year earlier, and plans another two million GPUs with AWS during 2027–2028. Demand remains supply-constrained, particularly by memory and components. [7] Anthropic also reportedly committed $45 billion for six years of Nscale compute representing about 460 MW of capacity, on top of other large recent capacity deals. [8]

My revised forecast is therefore:

• Transformative AI: still 2029–2031, with roughly 2029–2030 carrying most of my probability mass. Nothing this week is enough to move the calendar another full year.
• Singularity character: still turbulent but probably survivable. No material change to that probability judgment.
• Autonomous science: modestly accelerated. The “AI proposes → apparatus executes → AI reads results → AI revises experiment” loop is now demonstrably working on real equipment, not merely simulated experiments or outsourced wet-lab validation.
• Physical-world lag: the old single number is becoming misleading. I would now expect programmable laboratories and highly instrumented industrial processes to trail cognitive capability by months to perhaps a year, because existing machines can serve as the agent’s hands. General-purpose robotics, flexible factory labor, construction, home robots and similar embodied work still look one to several years behind, and potentially longer in difficult environments.
• Risk character: slightly more weight goes to failures involving AI controlling existing machinery or experiments rather than waiting for humanoid robots. I do not think the evidence is strong enough to increase civilization-scale catastrophe probability yet; the deployments remain narrow and currently use explicit hardware bounds and human-defined safety interlocks. [9]
• Governance: no compensating shift large enough to change the forecast. The Alabama attorney general has opened an investigation into OpenAI’s Hugging Face incident, and more than 100 major companies today called for an urgent defensive buildup against AI-enabled cyberattacks, but these are reactive measures rather than a new global control regime. [10]

So the substantive change is:

I no longer expect “the physical world” to lag uniformly. General robotics probably still does. Science may not. Frontier agents can increasingly reach through existing programmable laboratory and industrial equipment, meaning an AI-driven research acceleration could propagate into real experiments much faster than a humanoid-robot timeline would suggest.

Anthropic: Model Hardware Standard research preview FDA: Aletta robotic blood-draw authorization