AI bubble fears and crash warnings
The "is this a bubble?" chorus got louder today, and notably from inside the tent. Google CEO Sundar Pichai conceded "there is some irrationality in the current AI boom" and warned "no company is going to be immune, including us" if it pops (@garymarcus). Financial journalist Andrew Ross Sorkin, on 60 Minutes promoting his book "1929," said flatly: "We will have a crash, I just can't tell you when, and I can't tell you how deep" (@garymarcus, @60Minutes via RT).
BlackRock's Larry Fink supplied the political dimension, telling an audience that the trillions needed for data centers and power grids will come from "savings accounts and pension accounts" — and framing that as mandatory (@garymarcus). Layered onto that, Gary Marcus flagged WSJ reporting that public negativity toward AI is now growing faster than the industry itself (@garymarcus), and separate research suggesting prolonged AI use erodes critical thinking (@garymarcus). The cumulative mood: even bulls are pricing in a reckoning.
Compute crunch and chip geopolitics
Roon (@tszzl) amplified Anjney Midha's warning that GPU rental prices are up 2x+ since January 2026 — "we are living through the covid of compute, and all the toilet paper is gone" (@tszzl, @AnjneyMidha via RT). That squeeze sits underneath every other story today, from training economics to startup runway.
On the supply side, Huawei said it expects to design chips with transistor density equivalent to 1.4nm processes by 2031 (@tszzl, @AndrewCurran_ via RT) — a five-year horizon that, if credible, blunts US export controls. And in a startling org note, Gary Marcus passed along that "the entire management of xAI has quit in the last few months" (@garymarcus), which if accurate compounds the capacity story with a leadership vacuum at one of the larger compute consumers.
LLM skepticism and slop backlash
DeepMind's Demis Hassabis pushed back hard on AGI claims: "Today's systems are nowhere near [AGI]. Doesn't matter how many Erdős problems you solve… far, far from what a true invention or someone like a Ramanujan would have been able to do" (@garymarcus, @ns123abc via RT). Gary Marcus also reposted GeoHotz pivoting toward the LeCun/Marcus position — that real programming agents will need world models, "not some RLVR shit" (@garymarcus).
The applied complaint is "slop." Marcus argued GeoHotz's frustration with coding-model output is a tentpole cracking (@garymarcus). Philipp Schmid warned that AI-generated code is becoming a production liability when engineers don't understand what they ship (@_philschmid), and Ethan Mollick predicted that as more people learn the tells, "the scales are going to fall from their eyes" about how much web content is now AI (@emollick). One researcher reported ChatGPT 5.4/5.5 is "far from replacing human researchers" after two months of use (@garymarcus, @YT59529321 via RT).
Open-source local AI and the Hugging Face ecosystem
Clement Delangue (@clementdelangue) shared results from 300,000 builders who filled hardware profiles on Hugging Face, framing it as a baseline for the local-AI explosion. The headline demo: llama.cpp with MTP support pushed Qwen3.6-27B dense generation from 25 to 45 tok/s on an A10G — a 78% jump that, per Delangue, makes local models "fast enough to use as daily drivers" (@clementdelangue, @huggingface via RT). The_only_signal added a useful framing: activated parameters as "cognitive horsepower per token" (@the_only_signal).
The open-model drumbeat continued with LongCat's MIT-licensed talking-avatar model — Victor Mustar called it "probably SOTA" and sketched use cases from AI tutors to "Claude Code with a face" (@huggingface, @victormustar via RT). On the research-tooling side, PhysX-Omni shipped as a unified sim-ready generator for rigid/deformable/articulated objects (@clementdelangue, @ziqi_huang_ via RT), and Niels Rogge's PapersWithCode relaunch added non-arXiv support and multi-metric leaderboards (@_akhaliq, @NielsRogge via RT).
Coding agents and developer tooling
Codex's browser agent crossed a usability threshold for Roon, who reported it registered him for ICML and ordered Mother's Day flowers after being told to "find whatever it needed in Gmail" (@tszzl, @tomekkorbak via RT). Boris Cherny's top Claude Code tip is now "auto mode" — no permission prompts, so you can run parallel sessions ("multi-clauding") (@bcherny). Simon Willison shipped Datasette 1.0a30 with a "/" jump-to menu and a plugin hook that the new datasette-agent uses to launch agent conversations (@simonw).
David Sacks framed the macro: SWE job postings keep climbing because GitHub commits are up 14x YoY — lower cost-to-write means more code to manage, not less (@clementdelangue, @DavidSacks via RT). Worth pairing with the slop concerns above — both can be true, and probably are.
The Bottom Line
The day's signal is a widening split between bull execution (open models getting genuinely fast locally, agents finally doing real chores) and bear positioning (Pichai, Fink, and Sorkin all flagging a crash; Hassabis cooling AGI talk). The compute squeeze — GPU rentals up 2x since January — is the through-line that makes both stories make sense at once.