AI Daily — 2026-05-24

English 中文

DeepMind Solves Nine Erdős Problems with LLM-Driven Agents · LongLive 2.0 Infra: NVFP4 Parallel L...

Covering 28 AI news items

🔥 Top Stories

1. DeepMind Solves Nine Erdős Problems with LLM-Driven Agents

Google DeepMind reportedly solved nine Erdős problems using autonomous Lean agents guided by large language models, with formal verification followed by human review. The approach signals AI’s advancing capacity for mathematical reasoning and could intensify competitive pressure among frontier labs, illustrating an autonomous loop of LLM-enabled agents that are formally checked before human oversight. Source-x

2. LongLive 2.0 Infra: NVFP4 Parallel Long Video Gen

NVlabs releases LongLive 2.0, a NVFP4-parallel infrastructure enabling long video generation with parallelized training and inference, achieving 45.7 FPS. It adds kv-cache compression with TriAttention (≈50% KV reduction, no quality loss) and adapts RoPE for KV-cache relative RoPE to support infinite-length videos, with ICLR-2026 acceptance and broader ecosystem updates. Source-github

3. Join OpenAI, Google, Meta, Anthropic/XAI for pretraining-for-AGI

The post argues that the widening compute gap means AGI-critical problems now demand massive compute, suggesting pursuing pretraining-for-AGI work at leading labs such as OpenAI, Google, Meta, or the Anthropic/XAI/Cursor collaboration. The message underscores industry-wide need for heavy compute to advance toward AGI. Source-x

📰 Featured

AI Safety

DeepMind Solves Nine Erdős Problems with LLM-Driven Agents — DeepMind reportedly solved nine Erdős problems using autonomous Lean agents guided by large language models, with formal verification followed by human review, illustrating AI’s advancing mathematical reasoning and potentially upping competitive pressure among frontier labs. Source-x

Open Source & Tools

LongLive 2.0 Infra: NVFP4 Parallel Long Video Gen — LongLive 2.0 enables NVFP4-parallel long video generation with 45.7 FPS and includes kv-cache compression and RoPE adaptations to support infinite-length videos, showcasing open-source hardware-software innovations in multimodal AI. Source-github
LongCat reveals MIT-licensed open-source talking-avatar model (SOTA) — MIT-licensed, state-of-the-art talking-avatar model with a Hugging Face Space demo enabling AI tutors, dubbing, talking-head agents, and related products. Source-x
Community report leads to ban after PR training in vLLM — A community report led to banning a PR submission in vLLM as part of a PR-training workflow for resume-building, highlighting the workload and trust costs of low-signal contributions in OSS. Source-x

Industry

AI Automation Drives Software Engineer Demand as GitHub Commits Surge 14x — Despite AI agents automating coding, demand for software engineers grows as codebases expand; a 14x YoY GitHub commits surge points to a productivity boom for bespoke software. Source-x
Join OpenAI, Google, Meta, Anthropic/XAI for pretraining-for-AGI — The compute gap implies AGI-critical problems require massive compute, urging pretraining-for-AGI efforts at top labs. Source-x
Generative AI Video Enters Industrial TV Production with Kling — Kling’s AI video tech moves from demos to real TV/film production, with a production reaching roughly 44 million global viewers and strong US Prime Video performance, signaling AI-generated video’s mainstream adoption. Source-x
Compute gap widens as GPUs land, accelerating AGI pretraining — A discussion argues the compute gap to AGI is widening as GPUs land, predicting accelerated progress and dominance by leading labs in pretraining-for-AGI. Source-x

Hardware & Compute

BitCPM-CANN Trains 1.58-bit LLMs on Ascend NPU — BitCPM-CANN demonstrates 1.58-bit quantization-aware training for LLMs on Huawei Ascend NPU, porting a GPU workflow to CANN and achieving near full-precision performance across several models, enabling end-to-end on-device training. Source-reddit

Model Fine-Tuning & Multimodal

Thinking Machines Fine-Tunes Qwen3.5-397B in Hours — Thinking Machines demonstrates rapid fine-tuning of Qwen3.5-397B with usable multimodal capabilities, hinting at personal AI and real-time human–AI collaboration. Source-x

Open Source

Community report leads to ban after PR training in vLLM — (see Open Source & Tools) Source-x

Note: Some items share themes across multiple topics; grouping reflects primary angle for the featured section.

Kling and other entertainment-focused items reflect broader industry adoption of AI-generated media.

Generated by AI News Agent | 2026-05-24

━━━━━━ End of Template ━━━━━━

⚡ Quick Bites

AI memes: Twitter reacts to viral AI demonstration — A rapid wave of memes highlights public reaction to AI demos. Source-x
Presenton: Open-Source AI Presentation Generator and API — Open-source tool and API for AI-generated presentations. Source-github
Qwen3.6-35B Uncensored Genesis with APEX-MTP Release — Uncensored Genesis release with APEX-MTP prompts broader safety and capability discussions. Source-reddit
Codex prompt reuses patterns across sessions for automation — Patterns reuse across sessions enables more efficient automation. Source-x
Ask Codex: Reuse Patterns, Build Smallest Automations — Guidance on building compact automations by pattern reuse. Source-x
Anthropic onboarding meme features Karpathy, Wemby, Michael Scott — Meme reflecting onboarding culture and pop references. Source-x
Open-source 754 Cybersecurity Skills for AI Agents Mapped to Frameworks — Open-source mapping of cybersecurity skills to AI frameworks. Source-github
Granite DocLing 2Stage Boosts OCR with Dynamic Layout Prompts — OCR enhancements via dynamic prompts for document understanding. Source-reddit
GPU VRAM limits small LLaMA models with llama.cpp: possible? — Discussion on VRAM constraints for small LLaMA models. Source-reddit
Choosing Abliterated Gemma 4 31B and 26B-A4B Versions — Guidance on Gemma model variants for local use. Source-reddit
Is NVIDIA Still the Default for Local LLMs in 2026? — Debates on NVIDIA’s dominance for local LLM deployments. Source-reddit
CEOs prone to AI hype, missing essential enterprise work — Critique of AI hype among executives. Source-x
Codex Is Open Source, Surprising to Many — Codex open-source status surprises the community. Source-x
Is There Value in Uncensored LLMs Outside Roleplaying? — Questioning the practical value of uncensored models beyond roleplay. Source-reddit
LlamaBench Fails With MTP, Speculative Decoding Questioned — LlamaBench issues raise questions about MTP and decoding strategies. Source-reddit
Generative Recursive Education Enables On-The-Fly Custom Textbooks — Generative methods enable on-demand, customized textbooks. Source-reddit
Can Someone Explain MCP and Its Privacy? — Inquiry into MCP and privacy implications. Source-reddit
Geoff Hinton’s Google title once listed as ‘intern’ — Anecdote about Geoff Hinton’s early Google title. Source-x

Generated by AI News Agent | 2026-05-24