124 long-form pieces on architecture, training, inference, agents and safety.
The Act passed in 2024, the GPAI rules went live in August 2025, and enforcement powers arrive in August 2026. Most Series-A startups need a handful of specific changes, not a compliance department.
Two hundred dollars of API calls pulled more than ten thousand verbatim training examples out of ChatGPT. Memorisation is not a defect that better engineering removes; it scales log-linearly with everything the field is scaling. …
We pitted twelve sklearn algorithms head-to-head on a tabular dataset. The winner was not the most expensive one. It was not the most modern one. It was the one whose assumptions matched the data.
CUSUM has been provably optimal since 1986, yet on the first human-annotated changepoint benchmark a detector that never reports a change beat most of the field under default settings. Seventy years of changepoint theory, from Pa…
A strong model grading other models looks like a free lunch for evaluation. It is not. Position, verbosity, and self-preference biases plus quietly leaked test sets mean a leaderboard number can move several points without any mo…
The industry frames every LLM feature as an agent. Most production systems that work are pipelines with one or two LLM steps and clear handoffs. Default to a workflow, reach for an agent only when the problem demands it.
In 2021 a diffusion model beat BigGAN-deep on ImageNet while nearly doubling its recall, and GANs stopped being the default way to generate images. Yet latent autoencoders, SDXL-Turbo, HiFi-GAN and the neural audio codecs behind …
The frameworks that demo best are rarely the ones that survive in production. LangGraph wins where it matters by refusing to hide the thing every other framework tries to hide: the control flow.
Add 1,000 to every reward in an environment. The optimal policy is unchanged, and the expected policy gradient is unchanged — but the variance of the estimator you actually compute goes up by four orders of magnitude. Every advan…
GRPO weights every token by its own importance ratio, and on long responses that single-sample estimator quietly poisons the gradient until the model collapses. GSPO moves the ratio up to the whole sequence, and Qwen3's largest m…
In 2016 ProPublica showed that COMPAS wrongly flagged 44.9% of Black defendants who never reoffended against 23.5% of white ones, and its vendor showed the scores meant the same thing for both groups. Both were right, and a one-l…
A thousand identical requests to the same model at temperature zero produced eighty different completions. The cause is not random seeds, not sampling, and not GPU atomics. It is that reduction kernels change their summation orde…