Pinned post
Don't wait for the future, co-author it.
Latest posts 5
On Saturday at the American Heart Association meeting, Cleerly presented PARAMOUNT, the first randomized trial of AI-read coronary CT against the usual stress test for people with chest pain.
530 patients at 21 US sites, all with LDL cholesterol of 100 or more and no known heart disease. One group got a stress test. The other got a CT scan where AI measured the plaque in every coronary artery.
At six months the AI group came out ahead in 46.7% of head-to-head patient comparisons, against 23.4%. Use of cholesterol-lowering drugs went from 31% to 75% in the AI group, versus 32% to 47% with stress testing. Among patients sent for an invasive angiogram, 38% turned out not to need a stent or bypass in the AI group, compared with 60% in the stress-test group.
A stress test looks for an artery already narrow enough to starve the heart. Plaque is there years earlier. Measure it and you start treating years earlier.
57% of the patients were women, which almost never happens in cardiology trials.
The data are preliminary until the paper comes out in early 2027.
530 patients at 21 US sites, all with LDL cholesterol of 100 or more and no known heart disease. One group got a stress test. The other got a CT scan where AI measured the plaque in every coronary artery.
At six months the AI group came out ahead in 46.7% of head-to-head patient comparisons, against 23.4%. Use of cholesterol-lowering drugs went from 31% to 75% in the AI group, versus 32% to 47% with stress testing. Among patients sent for an invasive angiogram, 38% turned out not to need a stent or bypass in the AI group, compared with 60% in the stress-test group.
A stress test looks for an artery already narrow enough to starve the heart. Plaque is there years earlier. Measure it and you start treating years earlier.
57% of the patients were women, which almost never happens in cardiology trials.
The data are preliminary until the paper comes out in early 2027.
A small Y Combinator team called Standard Physics just showed what AI agents can do in engineering when every idea has to survive real physics.
Their system, Archimedes, takes a goal and has swarms of agents write designs as code. Every design runs through OpenFOAM, a real fluid-dynamics solver. In one run, six teams of agents on 12 cloud machines produced 1,251 propeller designs in 12 hours. The winner flies and swims: a single rotor that reaches about 90% of a specialist propeller in air and in water, in simulation, for roughly $160 of compute.
They also reworked the cooling of an AI data hall: 10.9% less electricity, about $227,000 a year.
What I like most is that it reports its own mistakes. Its first data-hall winner assumed old cooling coils could take warmer water than they can. A finer re-check caught it, and the search ran again with that limit built in.
The simulation agent, Reynolds, is open source under an MIT license.
Engineering has always been limited by how many ideas you can afford to test. That limit just moved.
Their system, Archimedes, takes a goal and has swarms of agents write designs as code. Every design runs through OpenFOAM, a real fluid-dynamics solver. In one run, six teams of agents on 12 cloud machines produced 1,251 propeller designs in 12 hours. The winner flies and swims: a single rotor that reaches about 90% of a specialist propeller in air and in water, in simulation, for roughly $160 of compute.
They also reworked the cooling of an AI data hall: 10.9% less electricity, about $227,000 a year.
What I like most is that it reports its own mistakes. Its first data-hall winner assumed old cooling coils could take warmer water than they can. A finer re-check caught it, and the search ran again with that limit built in.
The simulation agent, Reynolds, is open source under an MIT license.
Engineering has always been limited by how many ideas you can afford to test. That limit just moved.
China Telecom, the fiber maker YOFC and ZTE sent 128 terabits per second over 1,007 km of hollow-core fiber, 64 Tbps each way on a single strand.
In hollow-core fiber light travels mostly through air instead of glass, so it arrives about 30% sooner. One way over 1,000 km took 3.43 milliseconds.
This matters for AI. Big models are increasingly trained and served across several data centers, and every millisecond between them is time expensive GPUs spend waiting. A hollow-core line already carries financial traffic between Shenzhen and Hong Kong.
The speed of light in a cable was supposed to be a fixed cost. It isn't.
In hollow-core fiber light travels mostly through air instead of glass, so it arrives about 30% sooner. One way over 1,000 km took 3.43 milliseconds.
This matters for AI. Big models are increasingly trained and served across several data centers, and every millisecond between them is time expensive GPUs spend waiting. A hollow-core line already carries financial traffic between Shenzhen and Hong Kong.
The speed of light in a cable was supposed to be a fixed cost. It isn't.
Esto es absurdo. Un ataque al español y a la calidad de la medicina.
Imagina. Te dejas los codos en bachiller para sacar un 13 y entrar en Medicina. Seis años de carrera, uno más encerrado preparando el MIR y otros cuatro de residente, cobrando como mileurista con 30 años.
Y mientras, el gobierno de la gente homologa en 2025 los títulos de 30.303 médicos de fuera. Y sin examen.
Tan a lo loco que Sanidad ya estudia medidas para que no colapse el MIR de 2027. JEJEJE. El chiste del día.
PD: mira si te llegan mis correos. Si no te llegan, apúntate otra vez aquí: capitanbitcoin.net/comunidadView quoted post on X
