there's a lot of alpha in eliminating all logits for a specific word and then asking a model to say it

New blog post where I explain why I disagree with this, and why I have slightly longer timelines to AGI than many of my guests. I think continual learning is a huge bottleneck to the usefulness of these models, and extended computer use may take years to sort out. L-nk below.

for the last 60~ years, a team of Russian scientists located at the Institute of Cytology and Genetics in Novosibirsk, Siberia have been running an experiment to domesticate foxes, specifically the wild red fox. at each generation of foxes, they picked the top 10% calmest, most

The race for AGI is borderless. AI spending has surpassed consumer spending for contributing to US GDP growth in H1 2025 Datacenter capex is exploding

Who invented convolutional neural networks (CNNs)? 1969: Fukushima had CNN-relevant ReLUs [2]. 1979: Fukushima had the basic CNN architecture with convolution layers and downsampling layers [1]. Compute was 100 x more costly than in 1989, and a billion x more costly than

perlin noise algorithm felt like magic in the teens

Our open models are here. Both of them. openai.com/open-models

this new chat template "harmony" may be the single best thing to happen for agents

Next to Qwen3 of comparable size: Looks like gpt-oss is a wide (vs deep) model

Also according to the NYT, Claude and Gemini had similar response patterns if you simulated dropping them into the middle of that conversational context. That said, it was ultimately Gemini that called bullshit on the whole thing and snapped the guy out of it.

GRPO vs GSPO, or DeepSeek vs Qwen - a workflow breakdown of the main Chinese reinforcement learning algorithms ➡️ Group Relative Policy Optimization (GRPO): Learning by comparison GRPO is tailored for reasoning-heavy tasks where relative quality matters more than absolute

vik "apparently heroin is bad so now its time to try crack"

The prevailing wisdom is that compute is the most important factor for frontier AI training. We think this is wrong: data is the most costly and important component of AI training. We collected estimates of revenue for major data labeling companies and compared them with the

What makes the HRM model work so well for its size on ARC Prize? We ran ablation experiments to find out what made it work Our findings show that you could replace the "hierarchical" architecture with a normal transformer with only a small performance drop We found that an

Huawei’s datacenter and chip strategy is centered on achieving AI hardware self-sufficiency through a vertically integrated, full-stack ecosystem. -It involves in-house development of processors (Ascend NPUs and Kunpeng CPUs), a proprietary interconnect protocol (Unified Bus),

after leading a few projects, i've found that once you've set up the evals + experiment harness and make it easy to tweak config and prompts with 1-click run + eval, teams enjoy running experiments and hill climbing those numbers, and progress comes quickly. but setting up that

Why does Claude sometimes claim to have lived in San Francisco and married a Japanese woman? Why did Grok briefly love Hitler? Models infer their personas from cultural cues in their fine-tuning data. Article linked in the replies.

Big drama today in the Tor community. Conrad Rockenhaus, a Tor operator based out of Michigan, United States, was arrested in 2020 after refusing to cooperate with the United States Federal Bureau of Investigation Rockenhaus, a disabled United States military veteran, ran the

OpenAI hasn’t even deployed TPUs yet and they’ve already saved ~30% on their entire lab wide NVIDIA fleet. This demonstrates how the perf per TCO advantage of TPUs is so strong that you already get the gains from adopting TPUs even before turning one on. The piece covers a lot

As a bonus, Erdős problem #729 has also been fully autonomously resolved by GPT-5.2 Pro with Aristotle. Note that literature review is still ongoing.