This week in deep learning, we bring you Introducing ChatGPT Images 2.5, Linguistic drift at the frontier and a paper on Random Attention: Rethinking KV Cache Eviction for Efficient Reasoning.
You may also enjoy Formalizing Fermat’s Last Theorem \ Anthropic, Latency Scaling Differences for GPT and Claude Models, a paper on Terminal-Universe: Turning Agent Trajectories into Scalable Terminal Environments, and more!
As always, happy reading and hacking. If you have something you think should be in next week’s issue, find us on Twitter: @dl_weekly.
Until next week!
Industry
Introducing ChatGPT Images 2.5
OpenAI released GPT-Image-2.5 in Flare and Sunburst variants, up to 50% faster with sketch input, C2PA metadata, and invisible watermarking.
Formalizing Fermat’s Last Theorem
Anthropic’s internal research model produced the first complete computer-checked proof of Fermat’s Last Theorem in 11 days, generating 13 million lines of Lean across 30,300 theorems.
AI Has Solved One of Math’s $1 Million Millennium Prize Problems
OpenAI deployed 10,000 collaborating agents for 88 hours to prove a finite-time singularity in the 3D Navier–Stokes equations, formally verified in Lean, hours after an independent competing announcement.
Mistral raises €3B to make sovereign, open-weight AI the technology frontier
Mistral raised a €3 billion Series D led by Samsung Electronics at a valuation above €21 billion, the largest equity round ever for a European technology company.
Meta debuts its “secure by design” personal AI agent Muse
Meta launched Muse, a personal AI agent that runs tasks inside a secure VM with Sentinel monitoring and 100 million free weekly tokens, using single-use Stripe payment cards.
AlphaGenome Atlas: A predictive map of every possible DNA letter change in the human genome
Google DeepMind released AlphaGenome Atlas, a 1-petabyte predictive map scoring all 9 billion possible single-letter DNA variants in the human genome, free for noncommercial research.
MLOps/LLMOps/AgentOps
Linguistic drift at the frontier
A fascinating article about “linguistic drift,” tracing Claude-popularized vocabulary into 685,000 GitHub PRs in 2026 and introducing vocabguard, a drift-monitoring capability with 0.869 AUC.
Inside the LLM Call: GenAI Observability with OpenTelemetry
A tutorial about OpenTelemetry’s Generative AI semantic conventions, tracing agent invocations, LLM calls, and tool executions with standardized token-usage and duration metrics.
How llm-d makes the most of the hardware you already have
A benchmark blog post about llm-d serving a 753B-parameter MoE across 544 H100s to 3,000 concurrent coding agents, with 85.2% of input tokens served from cache.
Vercel Sandbox routing is now 18x faster globally
Vercel cut sandbox domain-resolution latency 18x globally — from 62ms to 3.4ms at the median — by resolving domains from regional replicas instead of one centralized store.
Amp added a Desktop tab giving agent threads an interactive Linux desktop for verifying work that needs real applications, from LibreOffice exports to computer-use tasks.
Look, But Don’t Touch (Read-Only Tools for AI Agents)
A blog post about protocol-level read-only enforcement in MCP Toolbox for Databases, showing how prompt guardrails, regex parsers, and session flags all fail against injection.
Learning
Long-context latency scales quadratically for GPT-5.6 but nearly linearly for Claude 5
A data analysis about time-to-first-token scaling quadratically with context for GPT-5.6 but nearly linearly for Claude 5, measured up to million-token contexts.
Inside the megakernel serving engine for North Mini Code
A deep engineering blog post about serving a 30B model through one persistent CUDA “megakernel,” hitting 292 tokens/sec at batch one — 62% of theoretical peak versus vLLM’s 39%.
Building a Memory-Driven Agent with NVIDIA NemoClaw
A tutorial about a memory-driven “self model” agent architecture scoring 90.9% versus an 82.8% agentic-RAG baseline, with changed-fact tracking jumping from 60% to 100%.
Fine-tuning a 350M Model for Better Structured Outputs in 100 GRPO Steps
A hands-on tutorial about GRPO fine-tuning a 350M model for structured outputs on a free 16GB GPU, lifting JSON format compliance from 18.0% to 31.9% in 100 steps.
AI Slop Is Changing How Engineers Review Code
An article about AI-generated code straining review, citing a Sonar survey where 42% of shared-codebase code is AI-attributed and 96% of developers don’t fully trust it.
Libraries & Code
An open-source AI observability tool used to debug, evaluate, and monitor LLM applications, RAG systems, and agentic workflows with comprehensive tracing, automated evaluations, and production-ready dashboards.
A Domain-Neutral, Git-Native Persistent Project Memory for AI Agents based on the Open Knowledge Format (OKF).
Papers & Publications
Random Attention: Rethinking KV Cache Eviction for Efficient Reasoning
Abstract:
Large language models achieve superior performance on tasks that require extended reasoning, but long chains of thought make the KV cache a severe memory bottleneck. Existing KV cache compression methods share one paradigm: score each cached token by some estimate of how much it will matter later, and keep the top-scoring ones. We show that the selection signal contributes almost nothing. Random Attention keeps the prompt and evicts uniformly at random within each attention head, computing no score at all; across four models and six reasoning tasks it matches the strongest prior evictor while serving 32-43% higher throughput than it in vLLM deployment. Controlled experiments explain this by showing that 1) the prompt is the fragile part of the cache, and most of the gap between selectors is just whether their selection signal happened to keep it; 2) the reasoning trace protects itself against eviction with redundancy at two levels, in the text (the model restates what it still needs as it works) and across attention heads (each keeps its own copy of the trace), so once the prompt is safe, a random draw retains enough copies of what the model still needs, and no score is required to pick them.
Terminal-Universe: Turning Agent Trajectories into Scalable Terminal Environments
Abstract:
As terminal-based code agents become prevalent, agent trajectories have accumulated at scale, while realistic, executable environments remain scarce. However, environments are what agent post-training actually requires: each can be re-queried into many verifiable tasks and provides execution feedback, whereas a trajectory is a single frozen demonstration. Rather than generating environments from scratch, we observe that the tool-execution history in existing trajectories exposes the structure and contents of the environments in which they ran, making it possible to reconstruct those environments from the trajectories themselves. Thus, we introduce Terminal-Universe, a framework which turns each trajectory into a reusable environment and explores it for synthesizing new tasks and continued interactions. Specifically, Terminal-Universe replays the file operations recorded in a trajectory to restore each file before the agent modified it, yielding a partial workspace; a completion agent then supplies the missing files and dependencies. On this recovered workspace, we both reconstruct the original intent task and synthesize entirely new ones. Besides, we also scale the tasks along two complementary axes: breadth and depth. For breadth, we mine directional dependency relations between related environments and synthesize cross-workspace queries spanning multiple codebases, as developers routinely do in real-world development. For depth, we extend the initial single-turn query into a multi-round session that captures iterative user feedback and requirement refinement via a user agent. Applied to public terminal agent trajectories, Terminal-Universe produces 37.3k task-sufficient environments. Supervised fine-tuning of Qwen3.5-27B on this corpus improves single-round performance on Terminal-Bench 2.1 by 11.9 points and multi-round performance on EvoCode-Bench v2 MT@4 by 13.8 points.
