Weekly Top Picks #105
On recent AI progress / Meta's Llama 4 / Reasoning AI isn't reasoning / ARC-AGI 2 / AI experts vs general public on AI / Google owns Pareto frontier / AI 2027
The week in AI at a glance
Recent AI model progress feels mostly like bullshit: The public reports of AI performance don’t reflect its real-world value/competence.
Meta’s Llama 4 launch is a PR disaster: A rushed launch, an unreleased model, unconfirmed rumors of data contamination… but not all is bad.
Reasoning AI models might not be reasoning after all: A new study finds that AI often relies on pattern-matching and other heuristics rather than reason.
The man out to prove how dumb AI still is: The ARC-AGI 2 benchmark is giving top AI models a headache they won’t fix with more compute and memorization.
How the US public and AI experts view AI: Pew Research has found there’s a widening gap in AI approval rates between experts and the public.
Google owns the Pareto frontier (performance vs cost): Google has both the best models and the most cost-efficient ones.
A new well-researched “future of AI” report: AI 2027: Take a look at what could be the future of AI for the next two years.
The week in The Algorithmic Bridge
(PAID) Weekly Top Picks #104: Can't believe your eyes / Demons with pitiful masks / The biology of an LLM / Dwarkesh Patel's and Karen Hao's books on AI / Artists have accepted AI won
(FREE) What I Learned at Ghibli Day: We don’t know how to deal with abundance. To survive the semantic apocalypse, perhaps you need to will yourself out of it—and then perhaps fight it with us.
(PAID) AI Is Learning to Reason. Humans May Be Holding It Back: What if the way to build superhuman AI is to let it find its improved ways to reason instead of insisting on training it from our flawed premises?
(FREE) AI Died the Day It Was Born: “AI” is a bad name. Perhaps no other name would be better because naming things imprisons them. That’s why some people won’t bother to be bewildered by this marvelous thing.
(FREE) GPT-5, o3, and o4-mini Are Coming Soon: OpenAI has changed plans; o3 and GPT-5 will be launched separately. They also likely have o4 ready, with o4-mini launching within two weeks.