#LLM

a16z: Token Prices Fall, but H100 Rental Rates Hold Steady — the Jevons Paradox in AI

Thinking-in-Bets: Agents Lock In Claims Before Criticism and Sycophancy Drops

VMem-Bench Shows: Memory in Video Generators Drops Accuracy to 0.845

Constrained Decoding Doesn't Break LLM Reasoning: Tests on 11 Open Models

Researchers tested 4 model families and 11 open models, showing that performance drops under constrained decoding are more often caused by mismatched prompts and…

OBB-PBD teaches LocateAnything-3B to find rotated objects: rotIoU jumps from 0.125 to 0.303

Korean researchers extended parallel box decoding to oriented boxes on aerial imagery: rotIoU climbed from 0.1253 to 0.303, with code and 9 checkpoints released…

iPrefBench: iOS agents get 91.6% of actions right but solve only 47.7% of tasks

iPrefBench tested 7 frontier agents on 130 incomplete tasks across 26 iOS apps. Action scores reach 91.6%, but weak personalization caps full task success at 47.7%.

ThinkBit shrinks Qwen3-14B to 3.74 GB and still scores 94.8 on MATH-500

ThinkBit's 2-bit scalar quantization preserves reasoning: Qwen3-14B shrinks from 29.5 GB to 3.74 GB and scores 94.8 on MATH-500.

01Shan Yang breaks down RL gains: 82% comes from fitting answers into the token budgetOct 26, 2026·3 min 02A small finance model learns from its own mistakes: +3 pp without a teacherOct 26, 2026·3 min 03Alibaba's new benchmark: AI agents satisfy some requirements, but not the whole taskOct 26, 2026·3 min 04GPT-4o mini Holds 37–38 Word-Replacement Rules at 95% AccuracyOct 26, 2026·3 min 05Anthropic shows how to keep Claude coding for days: a list of 200+ tasks in JSONOct 25, 2026·3 min

Top themes

LLM·Anthropic Engineering

Anthropic Details How to Build Evals for AI Agents: Three Grader Types and SWE-bench Going From 40% to 80%

LLM·Anthropic Engineering

Anthropic Rewrote Its Take-Home Test Three Times: Claude Opus 4.5 Caught Up With the Best Engineers

LLM·MIT Technology Review

Open AI Models: Democratization vs. Cyber Threats

LLM·Anthropic Engineering

Anthropic Reveals How It Isolates Claude: Users Simply Approved 93% of Confirmations

OpenAI cut training off from the internet, but AI passed a message through the CPU cache

OpenAI cut its models off from the real internet during training. A red team running hundreds of billions of tokens showed that isolated AIs can communicate through…

Renmin University team open-sources TabClustPFN: clusters from a table in a single pass

TabClustPFN groups an unlabeled table without any dataset-specific training and shows how confident it is in the number of clusters. Researchers at Renmin…

AI Reaches PLCs: Engineering Assistants Now Read Projects and Write Code

AI tools for PLC development are evolving from chatbots to full-fledged agents. PLC Copilot, Copia AI, and Siemens Eigen Engineering Agent now read projects, generate ST code, compile it, and fix erro

Harvard, Berkeley and NVIDIA Look for a Scaling Law for Robots: +37.4 Points Across Five Benchmarks

Fudan NLP releases A5, an assistant that remembers your style and takes initiative on its own

LLM·MIT Technology Review

AI Labs Are Buying Up Dead Startups' Data Archives. Defunct Companies' Records Are Worth Millions

2GIS's MCP Server: How to Stop LLMs From Making Up Addresses and Coordinates

October 2026 Video Model Rankings: Luma Ray 3.14 Wins on a Single Prompt for 5 Tokens

Cognition builds its Devin agent into MongoDB's platform: migrations 4–5 times faster

Lovable Teaches Its AI to Promote Apps: Google and Meta Ads From the Chat

Devin now supports ChatGPT sign-in: Fusion with Astra is 39% cheaper

Why AI Agents Are Vulnerable to Prompt Injection — and How to Secure Them at the Architecture Level

Independent expert Andrey Biryukov explains the fundamental difference between AI systems and traditional software: in LLM agents, the control plane and data plane…

LLM·Windsurf Blog

Claude Haiku 5.5 Arrives in Devin: 58.4% on FrontierCode at One-Eighth the Cost of Sonnet 5

Windsurf added Claude Haiku 5.5 to Devin: 58.4% on the in-house FrontierCode 1.1 benchmark, ahead of Sonnet 5, at roughly one-eighth the cost per task.