GPTS24AI 情报

Headline

Research · 09-08

DiffusionNFT on MiniMax H3: Open-Source RL Training for Joint Audio-Video Generation

The VeRL-Omni team documents silent failures that derail training without triggering an error

OpenAI Chief Scientist Warns AI Monitoring Is Eroding Before RSI Arrives

Artificial Intelligence · 09-08

RSI: AI Accelerating AI: The Bottleneck Has Shifted From Capability to Verification

Research · 09-07

Briefs

AI Chatbot Safety Guardrails Miss 60% of Delusional Conversations, Benchmarks Show

Artificial Intelligence · 09-07

Google Makes Lyria 3.5 AI Music Available in Gemini App and Developer API

Artificial Intelligence · 09-07

Meta Muse Voice Transcribe Tops Streaming Speech Benchmark but Its API Still Splits the Model

Artificial Intelligence · 09-07

Local LLM Inference Costs More Than You Think: Hardware, Not Electricity, Drives the Bill

infra · 09-06

More stories

View all
Artificial Intelligence · 09-06

AI Agents Rewrote All of Bun in 11 Days: The Verification Problem Now Defines Software

Jarred Sumner orchestrated 64 parallel Claude agents to port 960K lines of Zig to production Rust.

robotics · 09-05

Touch Becomes Standard: Why Robot Tactile Sensing Hits Its Industrial Inflection in 2026

New AI architecture, simulation tools, and 60% adoption mark a turning point for robot touch.

infra · 09-05

FreeToken Rewrites Local MoE Serving: 2x Faster Than Ollama on the Right Hardware

UC Berkeley's edge inference engine outpaces Ollama by 2x on models too large for VRAM

Research · 09-05

Anthropic's Claude Produces First Complete Computer-Checked Proof of Fermat's Last Theorem

Dozens of agents proved 29,500 theorems; Kevin Buzzard says autoformalization has matured

robotics · 09-05

Tesla Cybercab Launches Commercial Rides With Unfinished AI and Self-Certified Safety

FSD v15's 10-billion-parameter rewrite is 40% complete; NHTSA is probing the same camera-only architecture

Research · 09-05

Global Workspace Survives Recurrence: Interpretability Study Maps How Looped Transformers Differ

J-lens study finds workspace in Ouro and Huginn but shows loop boundaries block linear transport

Artificial Intelligence · 09-03

GPT-6 Astra Launches as OpenAI's First Critical Cybersecurity AI With a Monitoring Regression

Alignment behavior improved significantly, but the tool used to verify it has become less reliable

Artificial Intelligence · 09-03

Meta Muse Spark 1.3 Leads DeepSWE Coding Benchmark While Undercutting Rivals on Price

The benchmark-leading score required max reasoning mode, which had not yet shipped at launch

Research · 09-03

Moderna's mRNA Cancer Vaccine Beats Keytruda Alone in Phase 3 Melanoma Trial

The science is proven for melanoma: cost, scale, and tumor biology will decide the rest

Artificial Intelligence · 09-02

Gemini 3.8 Flash Hits Near-Frontier Coding Scores at a Fraction of Flagship Prices

No new pretraining: post-training alone narrowed the coding gap with Anthropic's flagship

Artificial Intelligence · 09-02

OpenAI Astra Reportedly Uses Recurrent Depth as AI Safety Researchers Warn of Monitoring Gap

Research confirms looped MoE efficiency gains as safety experts warn of reduced reasoning visibility

robotics · 09-02

TwinDEX Solves Robot Manipulation's Data Problem at the Hardware Design Stage

Co-designed wearable and robot hardware eliminates the alignment gap that locked training data behind real robots.

Browse all 72 stories →