J-Lens barely sees what subliminal learning transmits
J-Lens, an interpretability tool that reads a model's internal token rankings, detects the concept a subliminal-learning teacher transmits in only about 1.5% of its number completions, barely above chance, even though students trained on those numbers adopt the preference. Selecting the completions the lens flagged sharply strengthened the learned preference, and the weak readings persisted across models and lens variants. — lesswrong.com
GPT swarms outplay Claude swarms in competitive matches
In Minecraft and Mindustry testbeds, agent swarms spontaneously developed leaders, roles, rationing systems, and communication styles. In competitive matches, GPT-powered swarms beat Claude-powered teams, with GPT agents communicating faster and reading game source code for an advantage. — labs.ycombinator.com
FLUX.2 Klein image model ships as open GGUF build
Black Forest Labs' fast klein image model now ships as an Unsloth GGUF quantized build that runs on consumer hardware with as little as 13GB of VRAM. — huggingface.co
Looped world model games run on a Raspberry Pi Pico
A tiny looped-transformer world model and an actor network run on a Raspberry Pi Pico 2 W, with the model predicting each next frame of a retro game while the actor chooses actions inside the model's own dream. — github.com
Kitten TTS Nano debuts as open realistic voice model
Kitten TTS is an open-source, realistic text-to-speech model with 15 million parameters and a compact file size. — huggingface.co