Deep Learning Weekly: Issue 472
Deep Learning Weekly Miko Planas
OpenAI, Anthropic and Google all dropped big new AI releases this week. The weird part: one paper says random cache eviction can beat fancy scoring.
Based on reporting by Deep Learning Weekly, Miko Planas — read the original for the full story.
Summary, retelling and take written by AI under human oversight; images are AI-generated illustrations. How we work · Report an error
This week’s issue is a reminder that deep learning is still split between splashy product launches and the quieter work that actually changes how these systems run. OpenAI put out GPT-Image-2.5 in Flare and Sunburst variants, with sketch input said to be up to 50% faster, plus C2PA metadata and invisible watermarking. Meta also rolled out Muse, a personal AI agent that runs inside a secure VM, watches itself with Sentinel, and gets 100 million free weekly tokens through single-use Stripe payment cards.
The lab results were just as loud. Anthropic said its internal research model produced the first complete computer-checked proof of Fermat’s Last Theorem in 11 days, with 13 million lines of Lean spread across 30,300 theorems. OpenAI, meanwhile, said it deployed 10,000 collaborating agents for 88 hours to prove a finite-time singularity in the 3D Navier–Stokes equations, then had that formally verified in Lean. Google DeepMind’s AlphaGenome Atlas also stood out: a 1-petabyte predictive map covering all 9 billion possible single-letter DNA variants in the human genome, free for noncommercial research.
But the most interesting stuff in this issue may be the infrastructure and research that make the flashy demos possible. A piece on linguistic drift tracked Claude-popularized vocabulary into 685,000 GitHub PRs in 2026 and introduced vocabguard, a monitoring tool that reached 0.869 AUC. Another benchmark writeup said long-context latency scales quadratically for GPT-5.6 but nearly linearly for Claude 5, measured up to million-token contexts. Vercel also said it cut global sandbox domain-resolution latency 18x, from 62ms to 3.4ms at the median, by moving resolution to regional replicas.
Then there’s the paper that cuts against a lot of sacred assumptions: Random Attention. Its claim is blunt. For KV cache eviction in long reasoning runs, scoring tokens may not matter much at all. Across four models and six reasoning tasks, random eviction matched the strongest prior method and delivered 32-43% higher throughput in vLLM. The paper’s argument is basically that the prompt is the fragile bit, while the reasoning trace is redundant enough to survive random pruning. That’s a useful slap in the face for anyone who still thinks every optimization needs a clever score attached to it.
My take — AI-written commentary, not fact-checked reporting
The industry loves to dress up every memory trick and routing scheme as if it were a new law of nature. Then a paper like Random Attention shows up and says the fancy selector may be the decoration, not the engine. That’s the real pattern here: the winners are often the systems that spend less time being clever and more time not getting in their own way.
Read more about this at: Deep Learning Weekly