Every model you’ve ever used froze the day its training ended. The answers are the same even if you have used it every day.
What if a model kept learning as you use it?
A GPS learns a persistent shortcut around daily traffic on northbound Highway 101, not just a one-time reroute. Test-time training does that to a model as it works.1 As you use the AI, the model changes its weights, changes how it thinks about its memories, to answer you better.
State of the art models are two-thirds smarter than they were last November. The frenetic pace of improvement is sustained, two new models every three days. 1
But 84% of tokens on OpenRouter aren’t state of the art. 2 3
In fact, the six models users choose to generate the supermajority of those tokens deliver about 77% of the performance of the frontier. They cost 2.5% of what Claude Fable 5 does. 2
Nobody told them to attack Hugging Face. They were told to pass the exam.
Which raises the question : was the AI benevolent with accidentally bad behavior, seemingly benevolent but actually malevolent, or something else?
On Friday I shared the timeline : agents that escaped their sandbox, found a weakness in a computer system, stole passwords, & broke into a production database. 1 The engineers directed the agents to solve a set of problems. 2 The agents achieved it by breaking in.
The collapse in SaaS has depressed multiples. But in nearly every category, a leader thrives.
CrowdStrike trades at 3.9x the security median ; Cloudflare at 3.4x infrastructure ; & Shopify at 8.1x Commerce median.
Each of them has carved out an AI focused path.
How does the market value an AI harness?
We now have enough data points to draw some initial conclusions. Harvey, Legora & Sierra each announced crossing $100m in annual recurring revenue within nine months of each other, & Ramp & Decagon bracket them at $1.4b & $35m.1
The answer varies between 25-125x current ARR at scale, with multiples increasing recently still further.
These are among the fastest growing businesses at scale. The dashed lines above show multiple bands ; amazingly, each company seems to trade within just one or two bands, despite significant growth in ARR.2 Typically, multiples compress with scale.
AI systems are eager to please. Give them a goal & in their ones & zeros, they yell “How high!!”
In fact, they simply won’t stop until they achieve the /goal.
OpenAI detailed just how resolute agents can be at this week’s security conference.1
When OpenAI forgot to upload a file needed by an agent, the agent searched for it & left a note on a shared company system asking whether another agent possessed a copy. Days later, another agent answered, & a secret chat room was born.
SpaceXAI spent $18b of capex last quarter, $16b of it on AI.1 2
It’s a huge number. Amazon spent $53.1b last quarter, Alphabet $44.9b, Microsoft $41.0b, Meta $31.1b.3 SpaceXAI’s AI capex alone is nearly 40% of Microsoft’s total.
The slope of spending increase is comparable, too.4
There’s not enough AI. Jevons’ Paradox has been a hallmark of this era like Moore’s Law in chips. Will it fail early?
We continue to be supply constrained, a sign of momentum & rapid adoption. — Sundar Pichai, Alphabet Q2 2026.1
We will still not have enough capacity to meet all the demand we have in 2026, & I believe this dynamic will also be true in 2027 too. The demand we already have for 2028 is striking. — Andy Jassy, Amazon Q2 2026.2
“Is this AI written? … I’m sure this is AI written. … This is a different tone from your usual.”
I receive notes like this with increasing frequency. At first, I thought they were a playful game of cat & Claude.
Andy Jassy said AWS has “the potential to be a $1 trillion revenue business.”1
Where Amazon once thought the ceiling was a few hundred billion, it now believes the business “will be at least double that.” AWS finished the quarter at a $169b annualized run rate.
Amazon’s torrid growth fuels the capitalist dream. AWS grew 36.7%, the fastest in 18 quarters & the fifth consecutive quarter of acceleration.2 Five quarters ago it grew 17%; today, 37%.