Best of your X follows: June 11
Today: Chollet maps five ways AI can still be a bubble even when the tech works; Mollick argues for model hierarchies over cheaper-model swaps; LeCun highlights Microsoft's MAI-Thinking-1 hill-climbing paper; simonw flags an open letter calling on Anthropic to revisit a recent transparency decision; and Brockman notes universities are now launching dedicated AI degrees.
Five things worth reading from the people you follow — bubble logic, model hierarchies, an AI research paper on hill-climbing, the Anthropic transparency debate, and universities minting AI degrees.
Society and ethics
Chollet: how something can be a bubble even when the tech works
François Chollet posted a five-level breakdown that cuts through the "AI can't be a bubble because it's real" argument. He lays out exactly why each of these conditions still leaves room for a collapse: the tech works; the tech works and has product-market fit; the tech works, has PMF, and is economically viable; it works, has PMF, is profitable, and has unlimited future demand.
"Literally all it takes for something to be a bubble is for lots of people to over-enthusiastically bet their money on it, and subsequently get panicky."
The closing note is the one that sticks: "Internet adoption didn't stop in 2000." A bubble can burst without the underlying technology going away. Worth reading in full if you've been in any "is this a bubble?" conversation lately. 1
正在加载内容卡片…
simonw flags the Anthropic transparency debate
Simon Willison retweeted Clement Delangue's open letter calling on Anthropic to walk back a recent policy change — Delangue asking in good faith that the company hear feedback and course-correct. 2 The specific policy wasn't detailed in the tweet itself, but the signal is clear: the open-source / open-weights community is watching Anthropic's governance moves carefully, especially in the week that both Anthropic and OpenAI have confidential S-1s on file at the SEC.
AI tools and dev ecosystem
Mollick: use smart models as orchestrators, not cost-cutting targets
Ethan Mollick posted a two-sentence piece of advice that reframes how most teams think about model economics: 3
"'Switch to a cheaper model to save money' is a problem because cheaper models are worse. More often a better approach is hierarchies of models, with smart models as orchestrators and auditors of cheap ones."
The framing shifts the question from "how cheap can I go?" to "where do I need the smart model?" Smart models handle orchestration, review, and judgment; cheaper models handle bulk execution. This is practical enough to act on today if you're building or deploying any multi-step AI workflow.
正在加载内容卡片…
Research
LeCun highlights Microsoft's MAI-Thinking-1 paper
Yann LeCun amplified a thread on Microsoft's MAI-Thinking-1 paper, subtitled "Building a Hill Climbing Machine." Researcher Natasha Jaques called it remarkable that Microsoft publicly released the technical details. 4
MAI-Thinking-1 is Microsoft's reasoning-oriented model; the "hill climbing" framing refers to the optimization dynamics at the heart of the training process. The fact that Microsoft chose to document and release the methodology is the noteworthy part — frontier labs have generally kept training details close. LeCun's amplification suggests the open-science community views this as a meaningful step toward transparency in large-scale reasoning model development.
Business and enterprise
Brockman: AI degree programs are arriving at universities
Greg Brockman flagged a New York Times piece on universities launching dedicated AI degrees — specifically calling out the University of North Dakota, where he took classes growing up. 5 He described it as "cool to see" an institution he knows personally "innovating and trying out offering A.I. degrees." The NYT piece covers the broader trend of schools moving AI from elective coursework into full degree programs, a signal that the credential market is catching up with the labor demand.
正在加载内容卡片…
Quick takes
LeCun on Project Tapestry: LeCun posted a call to join Project Tapestry, an initiative under The AI Alliance focused on open, decentralized AI infrastructure. The project's site is at thealliance.ai/projects/tapestry. 6 Consistent with his ongoing push for open science as a structural counterweight to AI power concentration.
Mollick on real-time AI economy data: "We need more real time data on how AI may be impacting the economy — this is a really useful addition." 7 He was responding to a new data release; no source identified in the tweet itself, but the broader theme — that we are flying somewhat blind on AI's actual economic effects — is one he's returned to several times.
围绕这条内容继续补充观点或上下文。