Blog.

Google Research just built a glucose model with only 0.72 million parameters, and it beats every bigger model that came before it. It reads the continuous glucose monitor people already wear and predicts diabetes risk, insulin resistance and beta-cell dysfunction, with no new blood test. The lesson is bigger than health: the signal you want is usually already in a stream you are already collecting. You are just reading it as spot values.

Two researchers turned down more than $2 billion tied to Jeff Bezos' new AI venture — then came out of stealth with a model that does not do language at all. It simulates physics in four dimensions. We read it as confirmation of a CYSTEMS conviction: prediction, not conversation, is what intelligence is made of.

Three names are dominating the AI conversation right now: GPT-6, something called Astra, and a mystery model from the most secretive lab on the planet. Depending on who you listen to, one or all of them are landing any day now. Here is what is actually confirmed, what is speculation, and what it means if you are building a business on top of AI.

The most important number in AI right now is not a model score. An older, cheaper model wrapped in a loop solved 95 percent of a coding benchmark that the newest model, asked once, solved 67 percent of. Architecture beat intelligence. If you are still hunting for the perfect prompt, you are optimizing the wrong layer.

80 percent against 65 and 52 — a free coding model with no author, no company and no terms you would sign beat both of the models we actually pay for. Three days after it appeared, trillions of tokens of real production work were already flowing through it. The capability is real. The flock is the problem.

Nobody budgeted for a line item that grows every time the team gets better at using it. Intelligence went from scarce to summonable, and the cost followed the same curve electricity did: down per unit, up in total. The operators who win the next decade will not be the ones spending the least.

MCP solves a problem our agents stopped having the day we handed them a terminal. The protocol is a bridge to capability the model already carries in its hands.

Most operators bolt AI onto old workflows and call it adoption. The rebuild starts when you admit the workflow itself was designed for scarce intelligence.