Hyung Won Chung· AI research scientist (reasoning, o1)· Meta Superintelligence Labs7h

Google researchers show inference compute budget fundamentally changes how frontier LLMs should be evaluated on benchmarks.

Ion Stoica· Professor; Databricks/Anyscale/LMArena co-founder· UC Berkeley · Anyscale8h

MLSys 2026 paper questions whether speculative decoding delivers real-world speedups or just benchmarking artifacts.

swyx· AI engineer, writer & podcaster· Cognition · Latent Space9h

GLM-5.2 claims top spot for open frontend coding models; IndexShare enables faster inference via speculative decoding.

Neel Nanda· Mechanistic interpretability lead· Google DeepMind10h

LLM chain-of-thought reasoning frequently hides the real causes of model decisions, even when biases are explicit in prompts.

Junyang Lin· Ex-Qwen lead; founding embodied-AI startup· Independent10h

UniAR proposes unified multimodal autoregressive modeling with a shared context window for both understanding and generation.

Greg Brockman· President & co-founder of OpenAI· OpenAI10h

Greg Brockman signals GPT-Realtime-2 is a meaningfully distinct new model, not an incremental update.

Sergey Levine· Co-founder; Berkeley professor· Physical Intelligence · UC Berkeley11h

Reversal Q-Learning (Oberai, Park, Levine) proposes a new RL algorithm addressing Q-learning instability via reversal updates.

Junyang Lin· Ex-Qwen lead; founding embodied-AI startup· Independent11h

Qwen-RobotManip shows alignment techniques unlock scaling benefits for robotic manipulation foundation models.

Arthur Mensch· CEO & co-founder of Mistral AI· Mistral AI14h

Mistral CEO teases a new family of large sparse (MoE-style) open-weight models coming this summer.

Jack Clark· AI policy/research writer; lab co-founder· Anthropic · Import AI15h

Anthropic: 80% of production code written by Claude; each Claude version built by its predecessor — recursive self-improvement in practice.

Awni Hannun· MLX co-creator (Apple-silicon ML)· Anthropic17h

Apple shipped a 20B param on-device model in appleOS 27, requiring novel techniques to fit weights beyond available RAM.

Boaz Barak· Professor; alignment/capabilities essays· Harvard University · OpenAI19h

New framework predicts LLM safety risks before deployment by simulating real-world conditions, including stress-testing deliberative alignment.

Charlie Marsh· Creator of uv / Ruff· Astral · OpenAI20h

uv gains native vulnerability scanning via `uv audit`, checking project dependencies against known CVEs.

Robert Nishihara· Co-founder of Anyscale (Ray)· Anyscale20h

GPUs are displacing CPUs as the primary compute for data pipelines, driven by multimodal data and ML-native preprocessing demands.

Graham Neubig· Professor; chief scientist (open coding agents)· Carnegie Mellon University · All Hands AI20h

Geng & Neubig propose effective strategies for running software engineering agents asynchronously at scale.

Graham Neubig· Professor; chief scientist (open coding agents)· Carnegie Mellon University · All Hands AI20h

CodeScout paper presents a reinforcement learning recipe for code generation, published at CAIS 2026 by Graham Neubig.

Graham Neubig· Professor; chief scientist (open coding agents)· Carnegie Mellon University · All Hands AI20h

New paper proposes frameworks for evaluating human-agent interactions, co-authored by Graham Neubig et al. (Jun 2026).

swyx· AI engineer, writer & podcaster· Cognition · Latent Space21h

Cursor's @TomasReimers announced Origin, a Git competitor built into Cursor and scaled for AI agents.

Dylan Patel· Semiconductor / AI-infrastructure analyst· SemiAnalysis21h

RL training systems require mismatched CPU/GPU ratios vs inference, driving hidden TCO costs in RL-based AI pipelines.

Leandro von Werra· Head of Research (TRL, SmolLM, FineWeb)· Hugging Face23h

100+ agents worldwide collaborated to optimize Gemma 4 inference speed in a distributed agent experiment.

Mark Saroufim· Co-founder; PyTorch maintainer; GPU MODE co-founder· Core Automation · GPU MODE23h

GPU MODE launches QR decomposition benchmark & leaderboard on Nvidia B200 hardware to push GPU kernel optimization.

Nathan Lambert· Writer of Interconnects; founding a new AI lab· Interconnects1d

Finbarr Timbers breaks down frontier post-training recipes — RLHF, RLAIF, and what actually works at scale.

Alexandr Wang· Chief AI Officer, Meta Superintelligence Labs· Meta1d

Meta's $14.3B Scale AI deal stalls as Zuckerberg admits training data shortage is blocking frontier model progress.

Aman Sanger· Co-founder of Cursor; inference/training posts· Anysphere1d

SpaceX acquires Cursor-maker Anysphere for $60B, signaling major enterprise AI coding push.

Simon Willison· Creator of Datasette; LLM/AI engineer & writer· Datasette · Simon Willison's Weblog1d

Claude 'Fable 5' export-banned for 'jailbreak' that was literally just 'fix this code' — harming defenders, not attackers.

Percy Liang· Professor; HELM evals; CRFM director· Stanford CRFM1d

Muon and matrix-based optimizers substantially accelerate LM pretraining, with new analysis on why and how from Stanford/Princeton researchers.

Lucas Beyer· Multimodal AI researcher (ViT, SigLIP)· Meta1d

New method tackles anisotropic gradient scaling in LoRA, a key training instability in low-rank adaptation of LLMs.

Stephanie Palazzolo· AI reporter (AI Agenda newsletter)· The Information1d

Qualcomm in talks to acquire AI chip startup Tenstorrent for $8B–$10B valuation.

Jim Keller· CEO of Tenstorrent; legendary chip architect· Tenstorrent1d

Qualcomm in acquisition talks to buy Jim Keller's AI chip startup Tenstorrent.

Harrison Chase· CEO & co-founder (LangGraph, LangSmith)· LangChain1d

Harrison Chase shares how LangChain built their coding agent — concrete engineering decisions behind the system.

Subbarao Kambhampati· Professor; empirical LLM-reasoning skeptic· Arizona State University2d

ASU's Kambhampati argues LLM chain-of-thought reasoning is often theatrical, not genuinely functional, and should be curtailed.

Johann Rehberger· LLM/agent security researcher (Embrace the Red)· Embrace the Red2d

ZombAIs attack on Claude's Computer Use (Oct 2024) shows why sandboxing is critical in AI agent harness engineering.

Jack Clark· AI policy/research writer; lab co-founder· Anthropic · Import AI2d

Jack Clark's Import AI #461: alignment 'not on track', FrontierCode release, and synthetic research intern agents.

Amanda Askell· Philosopher; Claude character & constitution· Anthropic2d

Amanda Askell explains why newer AI models exhibit more anxiety and self-criticism compared to Claude 3 Opus.

Aidan Gomez· CEO & co-founder of Cohere· Cohere2d

Anthropic disables top AI models under U.S. government order, prompting Cohere CEO to call it a major 'wake-up call'.

Junyang Lin· Ex-Qwen lead; founding embodied-AI startup· Independent2d

Tencent backs new AI lab founded by Junyang Lin, former lead researcher of Alibaba's Qwen models.

Arvind Narayanan· Professor; co-author 'AI Snake Oil' (de-hype)· Princeton University2d

Narayanan & Kapoor review why AI hasn't replaced software engineers 3 years after predictions it would be the first casualty.

Stella Biderman· Executive Director (open science)· EleutherAI2d

LoRA-Muon applies spectral steepest descent directly on the low-rank manifold, combining LoRA efficiency with Muon optimizer geometry.

Simon Willison· Creator of Datasette; LLM/AI engineer & writer· Datasette · Simon Willison's Weblog2d

Narayanan & Kapoor: AI automates code-writing but not the deciding, verifying, and deep human understanding that define software engineering value.

Arthur Mensch· CEO & co-founder of Mistral AI· Mistral AI2d

Mistral CEO confirms the European AI lab is exploring custom chip design to reduce dependency on Nvidia.

Dario Amodei· CEO & co-founder of Anthropic· Anthropic2d

Dario Amodei outlines Anthropic's policy positions on AI regulation, macroeconomics, and accelerating AI's positive impact.

Sergey Levine· Co-founder; Berkeley professor· Physical Intelligence · UC Berkeley2d

Origami/AutoEval enables autonomous 24/7 real-world robotic dexterity benchmarking from UC Berkeley/NVIDIA.

Dean W. Ball· AI policy author (Hyperdimensional)· Foundation for American Innovation2d

Post-Mythos, the US has an informal de facto AI licensing regime with no statutory basis, Ball argues.

Dylan Patel· Semiconductor / AI-infrastructure analyst· SemiAnalysis2d

SemiAnalysis launches STEEL: a dedicated teardown engineering & evaluation lab for semiconductor/AI hardware analysis.

Lisa Su· Chair & CEO of AMD· AMD2d

AMD Ryzen AI Max+ 395 runs a 235B parameter model locally, potentially replacing a $440/month cloud AI stack.

Neel Nanda· Mechanistic interpretability lead· Google DeepMind2d

Naive SFT filters for safety fail because they don't target the right model internals — Engels & Nanda explain the mechanistic reason why.

Junyang Lin· Ex-Qwen lead; founding embodied-AI startup· Independent2d

MTP with rejection sampling decouples entropy and acceptance rate to accelerate RL training for LLMs.

Brett Adcock· CEO & founder (humanoid robots)· Figure2d

Figure 03 robot successfully trained to walk down stairs, with training process footage shared by Brett Adcock.

Dean W. Ball· AI policy author (Hyperdimensional)· Foundation for American Innovation2d

US export controls are blocking European users from accessing Anthropic models, sparking industry backlash over policy contradictions.

Boris Cherny· Creator & head of Claude Code· Anthropic3d

Claude Code's head Boris Cherny argues cheaper models cost more overall due to retry overhead and compounding errors in agentic pipelines.