May 13, 2026 · Wednesday

Perplexity Publishes Inference Study of Qwen3 235B on Blackwell GPU

New research quantifies how GB200 NVL72 racks deliver throughput gains over Hopper for serving large MoE models, marking a pivotal shift beyond training workloads.

Perplexity AI's inference stack running Qwen3 235B on NVIDIA GB200 NVL72 Blackwell racks.

Perplexity AI published research demonstrating that serving the post-trained Qwen3 235B model on NVIDIA GB200 NVL72 Blackwell platform significantly boosts throughput compared to Hopper. The company detailed how GB200 transforms the prefill and decode disaggregation pipeline when serving large MoE models like Qwen, providing a quantified view of real-world inference performance on next-generation hardware. CEO Arav Srinivas noted that the new architecture changes the fundamental approach to disaggregated serving, with throughput gains making Blackwell a pivotal step beyond Hopper for inference workloads. The study highlights that GB200 is not merely a training platform but a major upgrade for high-throughput inference on large mixture-of-experts models, reshaping how AI labs approach deployment economics.
@demishassabis · ISOMORPHIC LABS

Isomorphic Labs Secures $2.1 Billion to Accelerate AI Drug Discovery

Demis Hassabis announced that his AI drug discovery company Isomorphic Labs raised $2.1 billion in new funding, aiming to solve all diseases and making AI for human health a primary application. The raise turbocharges the mission that began with AlphaFold to reimagine drug discovery from first principles using artificial intelligence, positioning the company to scale its computational chemistry platform across multiple therapeutic areas.
@sundarpichai · GOOGLE

Google Unveils Gemini Intelligence for Android Multi-Step Automation

Google CEO Sundar Pichai announced Gemini Intelligence at the Android Show, enabling multi-step task automation across apps and Chrome. The system can fill out forms in a single tap and turn spoken thoughts into polished text with Rambler, bringing the best of Gemini to advanced Android devices for seamless cross-application productivity.

“The answer is not to fear the future. The answer is to guide it wisely.”

Jensen Huang, NVIDIA CEO · Carnegie Mellon University 2026 Commencement

StepFun Launches Step Image Edit 2, Tops KRIS-Bench

Step Image Edit 2: a 3.5B-parameter model achieving state-of-the-art on instruction-based image editing.

Stepfun introduced the 3.5B parameter image editing model Step Image Edit 2, ranking first overall and in factual and conceptual categories on KRIS-Bench, outperforming models with five to six times more parameters. The model requires only 0.7 seconds for inference, making it one of the most efficient instruction-based image editing systems available. The results underscore the growing capability of compact, purpose-built models to match or exceed much larger generalist systems at specialized tasks.
@op7418 · MODEL RELEASE

Ex-OpenAI CTO Mira Murati's Thinking Machines Launches Multimodal Real-Time Interactive Model

Thinking Machines unveiled an interactive model that continuously ingests native multimodal inputs — audio, video, and text — and thinks, responds and acts in real time. Unlike traditional agent scaffolding that stitches together multiple models, this system processes all modalities within a single unified architecture, marking a departure from the prevailing orchestration-heavy approach to multimodal AI systems.
Research & Model Releases May 13
@huggingface · PAPER

ModernColBERT Nearly Solves BrowseComp-Plus with Just 149M Parameters

Reason-ModernColBERT crushes state-of-the-art on BrowseComp-Plus, outperforming models 54 times larger and demonstrating that retrieval-augmented small models can rival giant reasoning systems.
@huggingface · MODEL RELEASE

Meta Quietly Releases Sapiens2 High-Resolution Human Model Family

A family of high-resolution models trained on 1 billion human images, designed for tasks such as pose estimation, marks Meta's latest contribution to computer vision foundation models.
@fchollet · RESEARCH VIEW

Chollet: Symbolic Learning Is a New Scalable Learning Substrate

Chollet clarified his concept aims to replace gradient descent and neural nets with a general, highly scalable new learning paradigm, not to replace coding agents.
@sedielem · PAPER

Self-Conditioned Embedding Diffusion Rivals Autoregressive Text Generation

A continuous embedding space text diffusion model that matches autoregressive generation quality while promising more efficient inference on accelerator hardware.
@_akhaliq · PAPER

Tongyi Releases Qwen-Image-2.0 Technical Report

Alibaba's Tongyi Qianwen image model technical report shared, expected to include architecture details and benchmark performance data for the next generation image generation model.
@_akhaliq · PAPER

TMAS Scales Test-Time Compute via Multi-Agent Synergy

New research leverages multi-agent collaboration to scale test-time compute and improve reasoning capabilities in large language models.
@_akhaliq · BENCHMARK

Soohak: A Mathematician-Curated Benchmark for Research-Level Math in LLMs

A new benchmark curated by professional mathematicians assesses large language models on research-level mathematical competence.
@huggingface · PAPER

Physics-Intern Agent Framework Boosts Theoretical Physics Performance

The physics-intern agent framework lifted Gemini 3.1 Pro accuracy from 17.7% to 31.4% on the CritP benchmark for theoretical physics reasoning.
@AndrewYNg · OPINION

Andrew Ng Dismisses AI Job Apocalypse, Calls Panic Irresponsible

Andrew Ng tweeted that AI will not cause mass unemployment, calling fear-mongering irresponsible and harmful, arguing technology waves ultimately create more jobs.
Products & Tools May 13
@OpenAIDevs · PRODUCT

OpenAI Codex Can Control Computers Across Apps in Background

Codex's Computer Use feature clicks, types and works continuously across apps without taking over the Mac.
@OpenAIDevs · PRODUCT

OpenAI Symphony: Dedicated Codex Agent for Every Task

Symphony assigns a running Codex agent to each to-do, enabling end-to-end automation.
@elonmusk · PRODUCT

Elon Musk Announces Grok Now Has Skills Feature

Grok now includes a skills feature, suggesting new plug-in or capability modules for xAI's assistant.
@cursor_ai · PRODUCT

Cursor Integrates Claude Opus 4.7 Fast Mode

Cursor editor now supports Claude Opus 4.7 fast mode, offering 2.5x speed for rapid coding tasks.
@op7418 · PRODUCT

Claude Code Adds Agent View for Background Agent Management

Users can now manage multiple running agents in terminal tabs and launch background tasks.
@dotey · PRODUCT

Anthropic Launches Claude for Legal with 12 Legal Plugins

Specialty library with role-based plugins and 20+ MCP connectors for legal ops, M&A and privacy.
@LumaLabsAI · PRODUCT

Luma Agents Launch Ad Optimization and Creative Generation

Set optimization directions and generate higher-performing ad versions across creative and copy.
@llama_index · TOOL

LlamaIndex Releases Self-Hosted Document Parsing Server Liteparse

Open-source, fully self-hosted service for parsing PDFs, Office files and images with local privacy.
@ViduAI_official · PRODUCT

Vidu Q3 Turbo Mode: 5x Faster Video Generation

ViduTurbo delivers generation speeds up to five times the industry average for rapid creative iteration.
@op7418 · LEAK

Google May Release Veo 4 Video Generation Model

Reportedly features reference video editing and slightly better text rendering quality than Seedance 2.
@huggingface · MODEL

DeepSeek V4 GGUF Released by antirez for Local Execution

GGUF files enable running DeepSeek V4 locally on devices, with M3 Max users reporting strong results.
@nvidia · APPLICATION

NVIDIA Earth-2 Extends Hailstorm Warning Time from Minutes to Hours

Colorado State University uses Earth-2 and PhysicsNeMo to pair generative AI with high-resolution radar.
@huggingface · MODEL

Unsloth Publishes Quantized GGUF for Qwen3.6-35B-A3B

MTP-enabled quantized GGUF versions simplify local inference for the latest Qwen model generation.
@huggingface · PRODUCT

OpenMed Agent Preview Released, Built on Hugging Face

Clinical extraction and medical terminology processing agent now available for healthcare AI tasks.
@zai_org · FRAMEWORK

Slime: The Most Elegant RL Training Framework for LLMs

Deep analysis explores how Slime redefines reinforcement learning training for large language models.
@hardmaru · MODEL

Sakana AI Releases KAME Real-Time Voice Architecture

Cascading fast speech model with backend LLM injects knowledge into real-time voice conversations.
@AravSrinivas · INFRA

Perplexity CEO Explains How GB200 Serves Large MoE Models

GB200 transforms prefill and decode disaggregation, with stack details showing throughput gains over Hopper.
@_akhaliq · PAPER

Rebellious Student: Self-Distilled RLVR with Reversed Teacher Signals

A new reinforcement learning method explores reasoning by inverting teacher signals for deeper exploration.
@_akhaliq · PAPER

Pixal3D: Pixel-Aligned 3D Generation for SIGGRAPH 2026

A new method generates high-fidelity 3D assets from images with pixel-aligned precision.
@emollick · COMMENTARY

Mollick: AI Labs Believing in ASI Will Disband Deployment Teams

Jobs are safe as long as human integration and change management remain necessary, argues Ethan Mollick.
@clementdelangue · INFRA

Hugging Face Infrastructure Scale Amazes CEO, Urges Migration

Clement Delangue recommends users on S3 or R2 migrate to HF for better performance, price and security.