Sakana AI Brings Recursive Self-Improvement to 'Physical AI' for Robots That Learn on Their Own
Japanese AI startup Sakana AI says it is expanding its RSI Lab to pursue world-model-based physical AI research
30
AI
AI models, services and robotics
Meta AI's 'Muse Glimmer' 30B passes local repo code review test
AI influencer runs it offline, finds it understands existing code and even generates new examples
230
Meta releases 30B-parameter open model for local agents
Muse Glimmer is an agent-specialized model that runs on 24GB-class consumer GPUs with 4-bit quantization
180
Multiverse Computing Unveils Techniques to Cut LLM Knowledge Distillation Costs
Caching teacher model logits and a memory-efficient KL loss enable long-context distillation on a single GPU
10
Claude Code switches to auto mode by default
Anthropic to enable auto-approval mode by default for Pro, Max, and Team accounts starting August 14
40
DeepSeek-V4-Flash-0731 reported to stall during long-context tasks
Generation halts past 100K tokens when run locally with OpenCode; resuming with a command restores work
90
LG AI Research unveils 750B-parameter K-EXAONE 2.0 FP8 model
MoE architecture with 37B active parameters, FP8 quantized version uploaded to Hugging Face
20
DeepSeek v4 Flash 0731 run locally with RTX 4090 and Tesla P40 combo
Adding two GPUs to 128GB of DDR4 yields 176GB total, with 2–3 tokens per second depending on quantization
70
Microsoft Unveils Video-Capable Vision-Language Model Mage-VL
Multimodal model handling images, video, and text released open-source on Hugging Face
30
xAI unveils image generator Imagine 2.0, ranks second behind GPT-Image-2 on Arena benchmark
New version adds editing tools and templates, Elo score trails OpenAI's model by a narrow margin
30
DynamoDB Adds Native Vector Search Support, Scales to Trillions of Vectors
AWS unveiled a native search feature that manages operational data and embeddings together without requiring a separate vector database
30
AWS adds native web search to Amazon Bedrock
Server-side built-in tool grounds model responses in up-to-date web knowledge, activated with a single parameter
20
AWS unveils 'Runtime Instances,' always-on compute for agents
Sessions lasting up to 14 days with GPU support expand production AI agent infrastructure
00
NVIDIA Releases Open Model 'Alpamayo 2 Super' for Robotaxis Under Commercial License
Fully open under OpenMDW license, tops LingoQA benchmark
60
NVIDIA Unveils NOOA, a Framework That Builds Agents as Python Classes
Research preview argues harness design determines benchmark performance, released as open source
10
NVIDIA unveils 'Ising Calibration 1.5' VLM for automated quantum computer calibration
31B-parameter vision-language model interprets qubit diagnostic results to automatically generate calibration values
20
AWS unveils "Agentic Catalog" preview for Amazon Quick
A workflow that finds top-level catalog assets via natural language and auto-generates datasets and topics
30
NVIDIA Details How to Build Fully Isolated K8s Tenants on Shared GPUs
Combining KAI Scheduler and vCluster lets multiple teams use a single GPU as if it were their own independent cluster
00
GPT-5.6 model pricing on Bedrock cut by up to 80%
Amazon Bedrock has adjusted on-demand inference pricing for its OpenAI GPT-5.6 model lineup
10
AWS Unveils Automated Improvement Feature for Bedrock Automated Reasoning Policies
New policy refinement engine diagnoses failed tests and proposes formal-logic fixes
20
F1 cuts data onboarding time from 8 weeks to 40 minutes with agentic AI on AWS
Formula One built a Data Accelerator with AWS to automate operations on its marketing data platform
20
Google unveils on-device offline voice translator 'Gemma Translator'
Open-source project built on Gemma 4 and LiteRT-LM, with a frontend for small displays
90
NVIDIA Unveils 34B-Parameter Reasoning Model for Autonomous Driving
Alpamayo 2 Super unifies trajectory generation, reasoning, and auto-labeling into a single model
70
NVIDIA proposes robot policy trained on video instead of vision-language models
Cosmos 3-based World Action Model suggested as a replacement for VLA robot policy architecture
10