Loading activity…
@kalomaze
ML researcher, speculator • extremely silly jester
AI
A user says their logit-lens probe showed arcs as `(` in the original vocabulary.
AI
A user argues that obvious ways to get more training tokens undercut concern about a “data wall.”
AI
A user says varying power-of-two spans from 128 to 512 for 20% of samples showed apparent gains.
AI
A user says weight decay in LoRA looks more like regularizing toward the base model’s original weights than penalizing absolute weight magnitudes.
AI
Google announces a 1 million-token output limit, up from 64,000, for its new model. Initial access goes to trusted cyber defenders through the Fairwind Program.
AI
A commenter finds evolution strategies and genetic algorithms uninteresting as alternatives to backpropagation. They point instead to synthetic gradients when naive backpropagation isn't tractable without backpropagation through time.
AI
A poster says they've figured out how to use zeroth-order optimization for transformer pretraining and argues that many core assumptions in optimization research are wrong.
AI
A user calls Gemma 4 12B underappreciated, praising its fully linear, standardized multimodal embeddings and arguing that its size makes a dense design less questionable than at 30B.
AI
A user suggests trying causal next-token prediction over discrete cosine transform (DCT) coefficients to explore what happens with a less language-like dataset.
AI
A user comparing GPT-5.6 with GLM variants attributes “a scary amount” of the variance to model failures exposing hidden serving issues tied to chat templates.
AI
For multi-epoch training, a project participant says the optimal number of loops increases with the compute budget and looping has a useful regularizing effect.
AI
OpenRouter says Typesafe's Jev returns structured decisions with probabilities rather than generated text. A browser-agent creator reports using it with Browser Use to find flights in 7 seconds for $0.0039.
AI
A user describes filtering for synthetic environments that Qwen fails on almost every time but GLM5.3 can solve reasonably well.
AI
A user calls V4.1 “very brittle,” claiming it can see “massive gains” when AGENTS.md instructions are optimized for a particular scenario.
AI
A user points to chain-of-thought reasoning as the rationale: problems break down into reusable steps, and strategies can work well across very different domains.
AI
One post challenges the idea that AI’s uneven strengths depend on how verifiable a field is, pointing instead to “information asymmetry” as the more general ingredient.
AI
The user says the setup uses a version of the game's message format that controls text pacing, size and villager mood—not just the words displayed.
AI
Reply identifies author as ML researcher at Prime Intellect working on LLMs.
AI
Pseudonymous commentator posts screenshot claiming GPT-6 Astra finished the game.
AI
Tweet from Prime Intellect engineer contrasts two vision systems on spatial tasks.
