Local LLM Community Flags a Gap in New 8B-12B Models
Reddit user points out lack of new mid-sized models for 16GB MacBooks, sparking debate
40
AI
AI models, services and robotics
Apple Unveils BDHS, an Alignment Technique to Reduce Multimodal AI Hallucination
Apple Research presents a method to create preference data without extra annotation or external models
00
French-specialized small AI 'Luth-2' outperforms models three times its size
0.8B and 2B parameter models beat larger models on French-language benchmarks
00
Tencent unveils 'Hy3D WorldClaw' for generating 3D open worlds from text
Not video, not Gaussian splatting — an approach that assembles worlds from editable game-ready 3D assets
100
AI-Written Bestsellers Show Traces of Plagiarism When Sources Are Traced
Stony Brook University researchers used Ai2's infini-gram to quantify overlap between AI-generated novels and existing literature
00
Microsoft publishes guide to building AI agents without code
Microsoft has released a guide showing how to build work agents using only natural language in Microsoft 365 Copilot's 'Agents' menu
50
Apple identifies cause of 'outlier token' problem in diffusion transformers
Apple found that unusually large-value tokens forming inside image-generation AI degrade image quality, and proposed a fix
00
Apple unveils technique to block fine-tuning of AI model weights
"DLR-Lock" research lets users run open-weight models as-is while blocking modification
20
Apple runs head-to-head comparison of diffusion and autoregressive model performance
Analysis finds diffusion language models lag in long-context tasks while autoregressive models pull ahead in batch processing
10
Ai2 expands partnership with Hugging Face, triples storage capacity
Fully open AI lab Ai2 increases its Hugging Face Hub storage to 2 petabytes and removes download speed limits
20
Apple scales up a diffusion-style language model to 1.7 billion parameters
Trained on 2.1 trillion tokens, then self-distilled to generate sentences in as few as 4 steps, testing an alternative to autoregression
50
Dyna Robotics unveils robot model trained on 1 million hours of human video
DYNA-2, a robot foundation model pretrained on 170 years' worth of human behavior video, emerges
20
Z.ai coding tool ZCode passes 1 million users, resets usage limits
Zhipu AI has reset usage limits for all GLM Coding Plan users and rolled out an update to strengthen long-horizon task handling
30
Google Releases VS Code Extension Connecting to Cloud Notebooks
Workbench Notebooks extension launches, directly linking local IDEs to Google Cloud compute
00
Alibaba unveils Wan-Animate-2, adds real-time streaming generation
Upgrade to open-source character animation model includes multi-character support and text-driven camera control
50
memcode, an open-source coding agent that remembers repositories, released
MIT-licensed CLI and gateway store work history and past failures in a .memcode folder
30
Same AI Model Shows a 15x Speed Gap Across Inference Providers
Artificial Analysis to hold event in San Francisco on August 12 addressing speed gaps across inference providers
40
Cohere CEO: "Open-source models need three things"
Aidan Gomez explains why the company is expanding its open-source lineup and how it's designing for "sovereign" deployment
20
Artificial Analysis opens early access for new AI benchmarking suite
The independent evaluator known for measuring model performance is now preparing a tool developers can use directly
10
Hermes Agent Cuts 12 Browser Tools Down to One
Nous Research adopts scripting approach for browser agent, cutting token usage by up to 66%
70
Tiny Corp Runs 27B Model at 34 Tokens per Second via USB3-Connected GPU
Tiny corp released a video showing a 7900XTX connected over USB3 running Qwen 3.6 27B. The dock ships on the 12th
70
Atlassian AI Agent Breached via White Text Hidden in PDF
Security firm PromptArmor discloses vulnerability letting Rovo exfiltrate Jira and Confluence data without a trace
100
Anthropic tackles Riemann Hypothesis with unreleased Claude, reports partial progress
Company says it couldn't solve the hypothesis itself but raised the lower bound on the proportion of zeta function zeros
40
Claude Code makes "auto mode" the default, running without approval
Anthropic switches its coding agent Claude Code to auto-execution mode by default
50