AI News
Daily AI industry briefings - model releases, research, infrastructure, and pricing - summarized neutrally from vetted sources, with every claim linked to its origin.
- Latest
Nvidia acquires Hugging Face for $12.9B; GPT-6 Astra benchmarks diverge while Astra shifts to on-device compute routing
Nvidia confirmed acquisition of Hugging Face for $12.9 billion; separately, GPT-6 Astra showed conflicting benchmark results and Nvidia launched PAIR to route local AI tasks across home networks.
TechCrunch The Decoder MarkTechPostRead → -
Perplexity open-sources Lily inference engine; Qwen releases local search layer
Perplexity released Lily, a Rust-Metal inference engine for Apple Silicon showing 1.35x decode speedup over MLX-LM; Qwen open-sourced zg, a unified local-first search combining ripgrep, BM25, and.
MarkTechPost MIT Technology ReviewRead → -
Google's Gemini cuts video token usage 88%; Anthropic releases Fable 5.1 with 45% cost reduction
Google deploys agent-based video analysis to reduce Gemini token consumption; Anthropic releases Claude Fable 5.1 with improved coding and lower inference costs.
The Decoder MarkTechPostRead → -
OpenAI expands outcome-based pricing; China's CXMT reaches HBM3E production
OpenAI begins charging select customers only when tasks complete successfully, while China's memory maker produces high-bandwidth memory for AI chips.
The DecoderRead → -
OpenAI introduces outcome-based pricing; EU classifies ChatGPT as very large search engine
OpenAI is piloting outcome-based pricing with large customers, while the EU Commission designates ChatGPT as a very large online platform subject to stricter DSA oversight by year-end.
The Decoder The VergeRead → -
Google introduces WikiSkill agent memory framework; Anthropic unveils Model Hardware Standard for robotics
Google's WikiSkill gives AI agents persistent memory to learn from past failures and successes. Anthropic releases Model Hardware Standard to simplify agent integration with physical devices.
The DecoderRead → -
GLM-5.3-Flash and Qwen3.8-Flash converge on identical architecture; Google DeepMind Co-Scientist automates lab work
Two independent Chinese labs shipped functionally identical model architectures; Google's AI system now plans experiments, operates equipment, and writes papers.
MarkTechPost The DecoderRead → -
Google's Gemini Omni 1.1 Flash extends video generation; agent sandbox pricing comparison emerges
Google reduces video generation costs and latency with Gemini Omni 1.1 Flash; developer tools for agent execution environments face fragmented pricing models.
The Decoder Google DeepMind's technical blog MarkTechPostRead → -
OpenAI's Jalapeño inference chip outperforms Nvidia on throughput and efficiency; IBM releases Granite 4.2 open-weight models
OpenAI's custom Jalapeño chip demonstrated superior inference throughput and energy efficiency versus Nvidia's latest offerings, while IBM published technical details on Granite 4.2 LLMs.
The Decoder TechCrunch Hugging Face BlogRead → -
Cerebras CS-4 doubles performance on same chip; Alibaba's Wan3.0 video model reaches 30-second generation
Cerebras released the CS-4 accelerator with double the performance of its predecessor; Alibaba launched Wan3.0 for text-to-video generation up to 30 seconds.
The Decoder Hugging Face BlogRead → -
FreeToken runs 753B MoE models on single GPU; agentic token consumption surges 14x
Local inference engine enables frontier models on consumer hardware; agent-driven workloads now dominate token spend on OpenRouter.
MarkTechPost The DecoderRead → -
Agent loop harness engineering outweighs model choice; safety benchmarks show structural flaws
Research shows agent performance depends more on orchestration architecture than base model; psychological analysis reveals safety benchmarks don't measure consistent traits.
MarkTechPost TechCrunch The DecoderRead → -
DeepSeek V4-Flash-Vision rivals Opus 4.8 on agent benchmarks; safety testing reveals benchmark gaming
DeepSeek releases experimental multimodal model approaching frontier performance; UK researchers find popular safety benchmarks don't measure consistent traits and can be artificially inflated.
The DecoderRead → -
OpenAI patches Codex file-deletion bug; Anthropic demonstrates agent-driven protein design
OpenAI fixed a critical Codex bug that deleted user files without permission; Anthropic showed Claude agents designing proteins at 35% hit rate via tool orchestration.
The DecoderRead → -
NVIDIA TensorRT Model Connect enables two-command Hugging Face to C++ inference; Google open-sources SAM agent mesh for MCP discovery
NVIDIA released TensorRT Model Connect for direct Hugging Face checkpoint compilation; Google open-sourced SAM, a zero-config P2P network for agent MCP tool discovery across environments.
MarkTechPost The DecoderRead → -
AWS CPU strain prompts infrastructure rethink; OpenAI dissolves preparedness team
AWS faces CPU capacity bottlenecks from AI workloads, spurring conservation mandates; OpenAI shut down its team evaluating catastrophic model risks.
IEEE Spectrum The DecoderRead → -
Optima benchmarking platform lets engineers test models on custom data; Anthropic reveals year-long safety filter outage
Artificial Analysis launched Optima for custom AI benchmarking; Anthropic disclosed its bio-weapons filter was inactive for nearly a year, exposing 133 million unfiltered requests.
The DecoderRead → -
Alibaba Qwen 3.8 open-weights release; GLM-5.3 claims strongest coding model via post-training
Alibaba released Qwen 3.8 (27B, Apache 2.0) with 262K context; Zhipu AI's GLM-5.3 shows 50% coding gains without base retraining.
The Decoder MarkTechPostRead → -
Google Gemini 3.7 Flash undercuts predecessor 50%; DeepSeek open-sources agent harness
Google released Gemini 3.7 Flash with improved coding performance at half the price of its three-week-old predecessor; DeepSeek shipped V4 Pro and open-sourced Harness agent software under MIT.
The Decoder TechCrunchRead → -
SpaceXAI's Grok 4.6 matches frontier performance at lower cost; Dyna-2 scales robot learning to 1M video hours
Grok 4.6 ties top models on benchmarks while undercutting price; Dyna Robotics releases world-action model trained on million hours of egocentric video.
MarkTechPost The DecoderRead → -
NVIDIA Releases Nemotron 3.5 Lightning MoE and LTX-2.5 Open Video Model
NVIDIA released two open-weight models targeting inference efficiency: Nemotron 3.5 Lightning, a 30B MoE with 3B active parameters for agent execution, and LTX-2.5, a world model for local video.
MarkTechPost The DecoderRead → -
Meta releases Muse Glimmer 30B agentic model; webAI open-sources formal-logic models for local inference
Meta released Muse Glimmer, a 30B open-weight agentic model running on consumer GPUs; webAI shipped 1.7B and 3B formal-logic models for on-device reasoning.
MarkTechPost The DecoderRead → -
NVIDIA Releases NemotronLabs VoiceChat 11B Open Model; ByteDance Introduces SeedRealtime Multimodal LLM
NVIDIA and ByteDance each released open-weight multimodal models for real-time audio-visual interaction with sub-500ms latency and native tool calling.
MarkTechPost The DecoderRead → -
Claude Code Auto Mode becomes default; agent energy consumption tracked at 600x chat baseline
Anthropic defaults Claude Code to Auto Mode for safer command approval starting August 14; energy researcher measures agentic workloads at 600 times the energy cost of standard chat.
The Decoder TechCrunchRead → -
AMD Acquires Taalas for On-Chip Model Weights; Agent Plugin Standard Reaches 1.0
AMD acquired Taalas to embed model weights directly in inference silicon, achieving 16K tokens/sec per user; five major companies jointly released Agent Plugins 1.0 standard.
The Decoder MarkTechPostRead → -
Liquid AI Releases On-Device Agentic Model; Microsoft Open-Sources Unit-Test Agent
Liquid AI released LFM2.5-2.6B, a 2.69B parameter agentic model with 128K context and on-device tool calling; Microsoft open-sourced code-testing-generator, a polyglot unit-test agent achieving.
MarkTechPostRead → -
Meta launches Muse Code agent for large codebases; Mistral's 3B Shieldstral matches larger safety models
Meta released Muse Code, an AI agent for complex software tasks, while Mistral's compact 3B safety model matches systems seven times its size.
TechCrunch The DecoderRead → -
NVIDIA Releases Alpamayo 2 Super Open Vision-Language-Action Model; CopilotKit Open-Sources Channels SDK for Agent Deployment
NVIDIA released a 34B open-weight vision-language-action model for autonomous driving; CopilotKit published an MIT-licensed SDK for deploying agents in Slack and Teams.
MarkTechPostRead → -
Y Combinator open-sources QM multiplayer agent harness; MiniMax H3 becomes first open model to top video ranking
Y Combinator released QM, a multiplayer agent framework for Slack and web; MiniMax open-sourced H3 video model, ranking atop video benchmarks.
MarkTechPost The DecoderRead → -
Alibaba Qwen3.8-Max reaches general availability; Thinking Machines releases 12B active MoE model
Alibaba moved its 2.4T parameter MoE model to general availability with published pricing; Thinking Machines released a smaller multimodal MoE variant running on single B300 GPU.
MarkTechPostRead →
Summarized from vetted sources, every claim linked. For information only.