The Chasm, the Classroom, the Conjecture, and the Squeeze

Four things stopped me mid-scroll this week. A layoff memo that named the exact model version where everything changed. A randomized trial that finally gave AI tutoring a number. A math system that solved problems humans couldn't. And a supply-chain story that means your next phone might cost more because a datacenter needed the memory first.

The Chasm

Block cut nearly 40% of its workforce. Owen Jennings, filling in the details on Jack Dorsey's internal memo, pinpointed the moment: "In the last week of November, or first week of December, things just fundamentally changed. It was with Opus 4.6 and Codex 5.3."

Before that, AI tools helped engineers autocomplete. After that, agentic systems wrote production code autonomously — and the code was good enough to ship. Jennings describes spending December playing with the tools, then spending Q1 asking: how does this flow through a technology company?

The answer, apparently, is that it flows through headcount. What makes this different from the usual "AI will take jobs" hand-wringing is the specificity. Not "someday." Not "in principle." A named model version, a named month, a named consequence. That's the chasm.

The Classroom

Ethan Mollick flagged a randomized controlled trial on high school students using a GPT-4o-powered tutor that personalized problems in real time. The result: final test scores rose by 0.15 standard deviations — "equivalent to as much as six to nine months of additional schooling by some estimates."

The AI education conversation has been drowning in vibes. This is data. An RCT. With a control group and a measurable outcome. Not "students reported feeling more engaged" — actual test scores, actual delta.

The sleeper detail: personalization was the mechanism. The tutor adjusted problem difficulty and pacing to each student. That's the thing human teachers know works but can't do at scale because there are thirty kids in the room. An AI tutor has a class size of one.

The Conjecture

Google researchers built Aletheia, an agentic system powered by Gemini 3 Deep Think that generates, verifies, and revises solutions to mathematical problems. It has already contributed to research papers and produced novel solutions to long-standing Erdős problems.

Read that again. Not "assisted researchers." Not "sped up literature review." Produced novel solutions to problems that have been open for decades.

The architecture matters: generate, verify, revise. It's not one inference pass hoping for a miracle. It's a loop that checks its own work. The verification step is doing the heavy lifting — it's the difference between a student guessing and a student showing their work, then catching their own errors. When the loop is tight enough and the reasoning deep enough, apparently you get original mathematics.

The Squeeze

Here's one nobody's talking about enough. Dwarkesh Patel laid out the numbers from Dylan Patel on AI's appetite for memory: around a third of big tech's $600 billion in CapEx this year is going to memory alone. AI accelerators use HBM, which is made from stacking the DRAM that goes into phones and laptops — but takes about four times more wafer area per byte.

The math is blunt. For every byte given to AI, four bytes of consumer supply disappear. iPhones could get $250 more expensive. Smartphone sales might drop from 1.1 billion a year to 500–600 million.

For the first time in decades, consumer computing is going to get incrementally worse year over year while prices go up. The AI boom has a physical cost, and it's denominated in the gadgets in your pocket. We talk about AI eating software. Turns out it's eating hardware too — literally, at the silicon level.

Kindling is a Tuesday column. Four sparks from the week — links, tools, and ideas worth catching before they cool.