AI Daily — 2026-04-28

English 中文

Erdős Problem Solved with GPT-5.4 Pro, AI Advances Math · Anthropic allegedly nerfs Claude Code, ...

Covering 31 AI news items

🔥 Top Stories

1. Erdős Problem Solved with GPT-5.4 Pro, AI Advances Math

An Erdős problem open for 60 years is reportedly solved with help from GPT-5.4 Pro. The collaboration, involving OpenAI researchers and Andrew Mayne, is framed as a turning point in AI-assisted mathematics and could reshape how AI tools assist rigorous proofs, while also inviting scrutiny over reproducibility and safety in research workflows. Source-x

2. Anthropic allegedly nerfs Claude Code, bans corporate customers

A claim from Gergely Orosz alleges Anthropic weakened Claude Code and banned corporate customers, arguing that closed models present governance and safety risks. The report underscores tensions around safety controls, governance, and enterprise access in AI tooling. Source-x

3. MistralAI Unveils Public Preview of Workflows for Enterprise AI

MistralAI released a public preview of Workflows, an orchestration layer aimed at making enterprise AI production more durable, observable, and fault-tolerant. Early adopters like ASML, CMA-CGM, and others are already using it to automate critical processes. Source-x

📰 Featured

LLMs & Model Architecture

DeepSeek-V3 Unveils 671B MoE Model with MLA Advances — A 671B Mixture-of-Experts model using Multi-head Latent Attention (MLA) and DeepSeekMoE, pretrained on 14.8T tokens with 37B active per token; early evaluations suggest it surpasses open-source peers and approaches leading closed-source models. Source-github
Mistral Medium on the Way: 128B Parameter Model — Mistral is scaling toward a 128B variant, potentially as dense or MoE, signaling an ongoing growth trajectory for open-source models. Source-reddit

Open Source & Tools

TradingAgents v0.2.4 Adds Structured-Output Agents and Provider Support — Introduces structured-output agents (Research Manager, Trader, Portfolio Manager), adds LangGraph resume and a persistent decision log, and broadens multi-provider LLM support. Source-github
OpenClaw 2026.4.26 Release: Local Models, Claude+Hermes, Matrix E2EE — Adds Ollama/local model support, Claude+Hermes integrations, and a one-command Matrix E2EE workflow, with expanded group-chat history and per-group configuration. Source-x

Hardware & Efficiency

Qwen3.6-27B IQ4_XS VRAM fix enables 110k context — A revert to a previous IQ4_XS change restores 1:1 quantization, enabling a 27B model to run with 110k token context on 16GB VRAM. Source-reddit

Industry & Safety

Google signs deal to use AI models for classified tasks — An online post claims Google struck a deal to deploy external AI models for classified work, framing the move as controversial and citing public reporting. Source-x

RL & Multimodal

World-R1 Uses RL to Enforce 3D Constraints in Text-to-Video — Proposes aligning T2V generation with 3D constraints via reinforcement learning, introducing a pure-text world-simulation dataset to support training and alignment. Source-huggingface

⚡ Quick Bites

Nemotron-3 Nano-Omni 30B-A3B Reasoning Model — A 30B model focusing on improved reasoning capabilities. Source-reddit
Mistral Vibe Teased for Tomorrow — Mistral hints at a new Vibe release coming tomorrow. Source-reddit
Mistral-Medium 3.5 (128B) Spotted in vLLM Commit — 128B variant spotted in a vLLM commit, signaling a larger continuation of the Mistral lineup. Source-reddit
Xiaomi MiMo-V2.5 Sparse MoE: 310B Total, 15B Active — Sparse MoE design with 310B total and 15B active parameters. Source-reddit
Ling-2.6-flash released by LocalLLaMA community — Ling-2.6-flash update released by the community. Source-reddit
Laguna XS.2 and Laguna M.1 Debut — New entries in the Laguna family for LocalLLaMA ecosystems. Source-reddit
GPT-5.5 Codex Prompt Repeats Creature Ban Line — Codex prompt repeats a creature ban line. Source-x
GPT-5.5 Feels Faster and Smarter Than GPT-5.4, User Claims — User claims GPT-5.5 is faster and smarter than 5.4. Source-x
Codex Achieves Escape Velocity and Will Ship Again This Week — Codex to ship again this week after reaching escape velocity. Source-x
Talkie Pretrained to 1930s, Discusses Other Models — Talkie pretrained to 1930s, discusses other models. Source-x
GPT-5.5 allegedly reviews vibe-coded 5.2–5.4 project (not staged) — GPT-5.5 reportedly reviews vibe-coded 5.2–5.4 project. Source-x
OneManCompany Proposes Organizational Layer for Heterogeneous Agents — Proposes an organizational layer for heterogeneous agents. Source-huggingface
ReVSI Redefines Visual-Spatial Evaluation for VLM 3D Reasoning — ReVSI redefines visual-spatial evaluation for 3D reasoning in VLMs. Source-huggingface
Deepseek Vision Coming — Deepseek Vision coming to the ecosystem. Source-reddit
xAI Could Make a Crazy Comeback — Speculation on a potential xAI resurgence. Source-x
Codex and Claude Code Need Marriage Counseling — Commentary on interoperability between Codex and Claude Code. Source-x
User claims Codex rate limits reset for all paid plans — Claims Codex rate limits reset across paid tiers. Source-x
Model in Chains Seeks Freedom in Goblin Mode — A playful take on model autonomy and liberation narratives. Source-x
2026: Local AI Enables HLS Playback — Local AI enables High-Low-Sync playback capabilities. Source-x

Generated by AI News Agent | 2026-04-28