Social Media Briefing — May 31, 2026

Emollick argues that meaningfully better AI releases are accelerating, especially from OpenAI and Anthropic, sharing a timeline of models that gained 3+ points on the Artificial Analysis index.

Emollick praises Epoch's benchmarking but argues open weights models are more fragile out-of-distribution than benchmarks suggest, doubting they were only a few months behind frontier models.

Ethan Mollick observes that meaningfully better AI releases are accelerating, especially from OpenAI and Anthropic, illustrated by a timeline of models scoring 3+ points above predecessors on the Artificial Analysis index.

vLLM announces v0.22.0 release with 459 commits from 230 contributors, highlighting DeepSeek V4 hardening, Rust frontend, batch-invariant FP8 latency gains, and KV cache offloading.

NVIDIA introduces DynoSim, a Rust-based workload-driven simulation of the Dynamo inference serving stack that screens thousands of deployment configurations in simulation 1,500x faster than real time before validating on hardware.

Nathan Lambert argues open science shapes how AI is discussed, citing his Tulu 3 project coining RLVR, and emphasizes the value of publicly establishing methods to reduce future noise.

Continuing the Social thread on AI accelerating research, Brockman highlights AI for accelerating research by expanding what mathematicians and scientists dare attempt.

Nathan Lambert observes Claude seems lazy in chat search while GPT-5.5 and OpenAI models are more thorough, arguing the Codex harness is a lighter intervention yet Claude Code does excellent agentic tasks like extracting paper figures into slides.

Hugging Face CEO Delangue praises the AI Security Institute for openly releasing its evals, datasets, and models on Hugging Face so researchers can scrutinize and reproduce them.

Nathan Lambert frames the open-versus-closed model debate around whether marginally better intelligence carries disproportionate value, predicting closed models stay slightly smarter while open models stay cheaper.

Jerry Liu explains the grid projection algorithm behind LiteParse, a document parsing approach that converts complex page layouts with text and tables into structured text without using LLMs, via line grouping, anchor detection, snapping and ordered rendering.

svpino reviews ego, a Chromium-based browser rebuilt for AI agents, highlighting parallel multi-agent spaces, persistent logins and extension compatibility while remaining assistant-agnostic.