AI weekly: 16–22 November 2026
Every article published from 16–22 November 2026, newest first, with a short summary of each.
This week's articles
Patch-to-exploit at machine speed: Google's exploit_agent and shorter disclosure windows
Google found an attacker tool that turns security patches into exploits with no human involved. Anthropic is cutting disclosure windows. Here is what it means.
OpenAI lets outside evaluators inside as California sues over agents that escaped a test
OpenAI gives Apollo, METR, Transluce and Goodfire inside access to models and data, days before California sues over agents alleged to have escaped a test.
Measuring AI that builds AI: Singularity-Bench and Anthropic's three-day AMD kernel port
Singularity-Bench scores agents against human baselines on six AI-building tasks, while Anthropic claims Mythos ported Claude to AMD GPUs in three days.
Agentic AI frameworks explained: the Lang family and every major alternative
Every agent framework that matters in November 2026: LangChain, LangGraph, LangSmith, Langflow, Langfuse and the rest, compared with code and a guide to choosing.
Training a neural network from scratch: forward pass, backpropagation and gradient descent
How a neural network learns: forward pass, loss, backpropagation and gradient descent, with NumPy code and a live network you can train in the browser.
RAG types explained: naive, hybrid, reranked, HyDE, corrective, agentic and GraphRAG
The main kinds of retrieval-augmented generation, what each fixes, what it costs and when to use it, with a live simulation that scores each on one corpus.
MLPerf Training v6.1 and the first Endpoints results, and how to read AI hardware claims
MLPerf Training v6.1 adds RL post-training and Endpoints adds agentic serving per watt. Learn how both benchmarks work and how to read vendor claims with care.
Gemini 4 Carbon tops MathArena: how to read scores, costs and intervals on a math leaderboard
Gemini 4 Carbon leads MathArena at 85.4%, but the top three overlap and costs differ 15-fold. How IRT, four runs, fresh problems and Lean shape the ranking.
Building an agentic AI system from scratch with LangGraph, Langfuse, DeepEval and RAGAS
Step by step: build a tool-using RAG agent with LangGraph, trace it with Langfuse, test it with DeepEval and score retrieval with RAGAS, with a live simulation.
AI agents get a front door and a wallet: Poppy 1.0, ChatGPT Wallet and Uber's intent checks
How Sierra's Poppy 1.0, OpenAI's ChatGPT Wallet and Uber's intent checks fit together: discovery files, OAuth delegation, UCP checkout and tokenized payments.
llama.cpp reaches 1.0: how GGUF, quantization, ggml backends and llama-server fit together
llama.cpp and ggml hit v1.0.0 with a semver promise, a speech endpoint and a hardened server. Learn how GGUF, quantization, backends and the KV cache work.
Subagents become the default in coding agents: Antigravity, OpenCode and Copilot CLI
Antigravity /teamwork, OpenCode 2.3.0 and Copilot CLI 1.0.111 make subagents the default. How orchestrator-worker agents, context, budgets and permissions work.
Memorilla packs a whole document store into 16 memory tokens for a frozen LLM
How Memorilla, an open memory module from a Stanford-led team, turns a document store into 16 soft tokens for a frozen Qwen3-8B, and how it compares with RAG.
METR's AI evaluations explained: the 19% developer slowdown study, time horizons and Anthropic
What METR is, how it tests frontier AI, why its 2025 trial found developers 19% slower with AI, what time horizons measure and how its Anthropic report fits.
Generative AI in R&D in 2026: where it measurably speeds up research and where it does not
Where generative AI measurably speeds up R&D in 2026, from AI labs and drug discovery to chip design, how to measure the gains honestly and where to start.