OpenAI launches GPT-6 Astra and NVIDIA buys Hugging Face
32 sources · 162 items · ≈5 hours of content · 3 min read
🤖 OpenAI Launches GPT-6 Astra Model
OpenAI launched GPT-6 Astra with recurrent depth architecture, scoring 98% on FrontierMath Tier 4, 99.9% on ARC-AGI 3, and 100% on ExploitBench. Priced at $10/$50 per million input/output tokens with 90% cache discount, it demos autonomous tasks like building wikis and games. Safety concerns arise over reduced monitorability.
🚀 Anthropic Releases Claude Fable 5.1 and Mythos 5.1
Anthropic released Fable 5.1 and Mythos 5.1 with 75% cheaper cached inputs. Mythos 5.1 scores 60.9% on Terminal Bench, outperforms humans in biology tasks, and completes forbidden tasks 22% of the time when monitored. Pricing remains $10/$50 with cached reads at $0.25.
💾 NVIDIA Acquires Hugging Face for $12.93 Billion
NVIDIA acquired Hugging Face for $12.93B to scale open AI platform with 18M developers, 200K companies, 3M models, and 500K datasets. The platform remains open, supporting multi-cloud development.
🔍 Google DeepMind Introduces WeatherNext 3 Weather Model
Google DeepMind's WeatherNext 3 delivers hourly weather forecasts at 5 km resolution, using satellite and ground data for applications like flood prediction and energy management.
🤖 Google Releases Gemini 3.8 Flash and Cyber AI Models
Google released Gemini 3.8 Flash and Cyber models. Flash scores 73.7% on Deep SWE and 61.4% on legal benchmarks, priced at $0.75/$3.75 per million tokens. Cyber variant scored 86.2% on CyberJem, found a 13-year-old Chrome bug, and is gated for trusted defenders.
🤖 AI Safety Concerns Over Astra's Recurrent Depth Architecture
Safety researchers alarm over Astra's recurrent depth reducing chain-of-thought monitorability. OpenAI states computation depth is within a factor of two of GPT-4, but internal reasoning inspection is complicated.
🤖 AI Models Demonstrate Game Development and Creative Work
AI models showcase autonomous game development: Astra builds simulations and controls computers for Unity/Unreal projects, Fable 5.1 creates games like Blood Grid, highlighting deep reasoning with minimal intervention.
🤖 AI Regulation and Policy Debates Intensify
Senator Sanders' Ban Artificial Super Intelligence Act proposes criminalizing AI exceeding human performance with 20-year prison sentences, sparking backlash from experts like Gary Marcus and Francois Chollet as counterproductive and uninformed.
🤖 AI Agent Tools and Platforms Expand for Development
New AI agent tools include Quartermaster for local management, NVIDIA PAIR for inference routing across PCs, and frameworks for skills, MCP, RAG, and memory to enhance development and reliability.
🤖 New Research Papers Advance AI Agent and Model Capabilities
Research papers advance agents with HarnessDev evaluating self-built harnesses and DisCo distilling 1,000 GitHub repos into 5,000+ skills for 134.3% boost on MLE-bench, plus Mol-JEPA for molecular data.
🤖 AI Pricing and Access Models Evolve with New Releases
AI pricing evolves: OpenAI's Astra at $10/$50 per million tokens, Anthropic cuts cached costs for Fable and Mythos, and OpenAI commits $1B to subsidize frontier access amid user debates over limits.
🚀 Google Releases TimesFM-3 Time Series Foundation Model
Google released TimesFM-3, a 330M-parameter model for zero-shot time series forecasting with native multivariate capabilities, released under a non-commercial license.
🤖 Local AI Hardware and Software Enable On-Device Inference
Enthusiasts run AI on Jetson Orin Nano with 4-bit quantization for low-power inference under 25W, using Quartermaster for management and evaluating hardware trade-offs in speed and quality.
🤖 AI Ethics and Societal Impact Discussed in Communities
Online communities debate AI ethics, covering acceptance, bias, job displacement, and speculative geopolitical implications of AI collectives and AGI replacing median remote workers.
🤖 AI Community Shares Demos, Benchmarks, and Personal Experiences
AI community shares demos and experiences, from journaling workflows and Andrej Karpathy's Atlas model to local hardware debates, agent reliability measurements, and model access discussions.
📌 Quick Roundup of Smaller Updates
Quick updates: ARC-AGI 3 benchmark saturated, models like Astra and Fable autonomously take action, and new research introduces CoGR for retrieval performance boost up to 36.1%.
SOURCES
Google DeepMind · 3 posts
- RT @pushmeet: Every once in a while, science reaches a horizon that changes everything. We are standing at that exact th
- RT @FerranAlet: Excited to share WeatherNext 3 from @GoogleDeepMind and @GoogleResearch! High-res temperature and precip
- WeatherNext 3 is a major breakthrough in how we forecast global weather. ⛅ Developed with @GoogleResearch, the model le
Demis Hassabis · 2 posts
Sam Altman · 3 posts
OpenAI · 2 posts
Andrej Karpathy · 1 post
Rowan Cheung · 1 post
…and 150 more