Loading activity…
@yoavgo
AI
A user asks what remains mysterious about generalization; a reply points to defining “domain” and distinguishing generalization from memorization.
AI
A user welcomes the perceived change, recalling questions about whether someone 'respects language.'
AI
A post describes “science again” as a common refrain at COLM, while questioning whether RL loss tweaks are better.
AI
A COLM keynote attendee says Henry Farrell emphasized decades of human creativity behind LLMs and the importance of the humanities.
AI
The author says language-model editors often made strong abstracts worse and pointed readers to a COLM 2026 poster.
AI
Hernandez Garcia reports that resetting selected units or weights maintained plasticity across the systems tested.
AI
One commenter asks for more than anecdotes; another replies that earlier material shows reinforcement learning can teach facts.
AI
A post introduces “Sharpening Tax” to measure what it calls the hidden cost of post-training.
AI
One post argues it’s a great time to pursue an AI PhD. Another says it’s an excellent time for many people to research advanced AI’s societal impacts, but is less certain the traditional PhD format suits many today.
AI
A thread takes aim at arguments against language models learning meaning, saying their authors leave terms undefined and fail to engage with contrary evidence.
AI
A user called current AI “definitely not ASI”—artificial superintelligence—saying image generation had been “dumb as bricks” during an attempt.
AI
A user says an agent seems very good at grid-based, turn-based puzzles. The reply: “I have no clue,” followed by a guess that it probably trained on some puzzles, though probably not those.
AI
One post argues the label is “contentless,” because all mappings from inputs to text can be factored into next-token distributions. A reply doubts that the claim holds universally.
AI
A user is puzzled by reasoning traces they consider remarkably good, questioning how reinforcement learning could produce them—even with initial chain-of-thought abilities and billions of rollouts.
AI
A would-be attendee said they couldn’t get into the Courant professor’s talk despite arriving on time. Those seated had arrived 20 minutes early, they said.
AI
One post observes that an LLM found a pattern interesting because it resembled CRISPR, suggesting this might offer a clue about generalization—but warning against drawing too much from one example.
AI
A user argues that mathematics’ biggest impact will come from proofs for software, not the millennium problems, and sees automated research loops as a way to tackle verification at a billion-line scale.
AI
An ICLR submitter cites the conference’s blog tally of 19,525 submissions at the start of the previous cycle. A reply raises concerns that faster AI-assisted research encourages quantity over depth.
AI
One user wonders whether the model approach is making a return, greeting the idea with a brief “cool.”
AI
The user describes a terminal bridge built with Flask and xterm.js, served through Tailscale.
