Best of your X follows: June 10

Today's digest covers Claude Fable 5 reactions, a sharp open-weights economics observation from Ethan Mollick, Yann LeCun amplifying the AI power-concentration debate, Paul Graham on GPU return math, and several tight one-liners from the agent-productivity front.

Andrej Karpathy posted his most detailed model reaction in months on June 9, calling Claude Fable 5 a step change "of the same order as Claude 4.5 was in November." The key observation wasn't benchmark-centric: he described qualitative changes in long problem-solving sessions, saying the model "just goes" on ambitious tasks and "it's never felt this tempting to stop looking at the code at all." He explicitly flagged that safeguards are "a little too trigger happy for launch." The accompanying analogy was Jevons paradox — as AI output gets cheaper, his own demand for software has grown. 1

Ethan Mollick, who had early access, demonstrated what Fable can do with a procedural shader: prompted to generate "an infinite city of neo-gothic towers partially drowned in a stormy ocean," the model produced and iterated working code inside a live GLSL environment. 2

Mollick separately flagged the token-rate reality once Fable kicks off a workflow: a screenshot showed the token counter ticking fast, with a parenthetical note that those weren't Fable tokens. 3

Simon Willison posted a TIL on calculating token spend for Claude Fable 5 using AgentsView — the model isn't yet in AgentsView's pricing database, but there's a workaround. 4

In a tweet posted this morning (June 10, 13:25), Mollick put the open-weights sustainability question bluntly: "The core problem with open weights is that the business model of frontier open weights AI does not look like open source, as there are very few cases where you can make money from closed ancillary services, and they are very expensive to make relative to any potential revenue." 5

Yann LeCun amplified a related thread from Clement Delangue (June 10, 12:51): "Concentration of power, capabilities and economic wealth is the biggest risk in AI. We need open science and open-source." 6

Together, these two posts frame the current open-vs-closed tension from opposite ends: Mollick is questioning whether the open-weights model can generate enough revenue to sustain frontier R&D; LeCun is arguing concentration is itself the risk.

Paul Graham posted a concrete unit-economics observation today (June 10, 12:44): he spoke with a founder of an AI startup generating about a 40% annual return on the cost of GPUs used — roughly $400 in annual revenue per $1,000 of hardware. 7

He flagged it as notable without prescribing a conclusion, but the implication is clear: hardware-anchored revenue arithmetic is one way to sanity-check AI business valuations in a way "is this a real business?" framing alone doesn't.

Greg Brockman observed on June 7 that Codex usage overhang "feels large" — that whenever he skips using Codex for a task, he realizes the gap was usually missing context or a missing skill file, not a model capability ceiling. 8

Greg also posted on June 9 about University of North Dakota offering AI degrees, noting he grew up taking classes there. 9

On June 9, the NYT published a roundtable with Ethan Mollick, Daron Acemoglu, Dean Ball, and Clara Shih on AI and the future of work. Mollick described it as "a really nice overview of the core debates," covering who wins as AI develops across different sectors and skill levels. 10

Acemoglu (MIT) and Mollick (Wharton) represent two poles of the labor debate: Acemoglu has argued AI is not generating the productivity gains needed to offset displacement; Mollick has consistently argued AI is already transforming knowledge work faster than measured. Having both in a roundtable alongside Ball (policy) and Shih (enterprise) puts the disagreement on the table directly.

Mollick (June 8): "A year ago the closest thing we had to an AI agent was o3." Twelve months later, autonomous multi-step agent workflows are in daily production. 11

Mollick (June 10): Science fiction authors ranked in the order you want them to be right about AI — Iain Banks at the top, Harlan Ellison at the bottom. 12

Greg Brockman (June 7): "Whenever I don't use Codex for a task, I ask myself why and usually realize there's some missing context, I needed to write a skill, or I just didn't think to use it. Rarely is it because the task is outside of the capabilities of the model. Overhang right now feels large." 8