Gemini Robotics — Google DeepMind

Google DeepMind's Gemini Robotics features models that empower robots with vision-language-action capabilities and embodied reasoning, allowing them to adapt to tasks in real-world environments and im

deepmind.google

Introducing Gemma 4 12B

Say hello to Gemma 4 12B, a high-performance multimodal model for laptops. With a novel encoder-free setup, it boosts audio and visual processing for efficient reasoning in a smaller package.

blog.google

Ideogram 4.0 Technical Details: Open model at the forefront of design

Ideogram 4.0 is a 9.3B parameter text-to-image model, integrating a vision-language encoder with structured JSON prompts for nuanced image creation. It includes a 34-layer Diffusion Transformer, using

ideogram.ai

https://www.ycombinator.com/launches/Qd6-bloom-the-brand-layer-for-agents

Bloom is transforming branding management with a layer that integrates assets for agents, apps and platforms. With over 175,000 generated assets, it helps teams automate branding.

ycombinator.com

Paul Graham on X: "The most important component of writing clearly is simply to have high standards for clarity. Then if you write something unclear, you notice, and ask: what did I mean to say? You can just keep doing this over and over. And if you have high standards for clarity, you will." / X

Paul Graham stresses high clarity standards in writing, noting its benefits in communication and advising startups, enhancing understanding and action plans.

x.com

Peter Steinberger 🦞 on X: "Here’s the video of my talk at MS Build: Build the thing that builds the thing. https://t.co/lJuv2twhFe" / X

Peter Steinberger shared a video of his MS Build talk on innovative software tools and efficiency.

x.com

The Next Frontier of Visual AI Is Code - by Yoko Li - a16z

Yoko Li's article highlights a shift in visual AI from pixel outputs to code artifacts, showcasing code-native generation’s advantages in design workflows, improving editability and iteration so desig

a16z.news

Ethan Mollick on X: "Had Claude Code build a snake game where the snake becomes aware it is in the game and then... stuff happens. Some impressive creative decisions by the AI (& also some very AI ones), I just gave a first prompt and some feedback on the game as it went. https://t.co/WdmlBD5iHI https://t.co/EQpmXmnH9y" / X

Ethan Mollick discusses using Claude Code to create a self-aware Snake game, noting the AI's creative choices and development challenges. A bandwidth-friendly version is coming, with download options

x.com

https://openai.com/index/codex-for-every-role-tool-workflow/

Explore how OpenAI's Codex enhances productivity across various roles by streamlining workflows and providing coding support tailored to specific tasks, making it a versatile tool for professionals.

openai.com

Fei-Fei Li on X: "https://t.co/Kt50ttQRMJ" / X

Fei-Fei Li discusses the need for a taxonomy of world models in AI, stressing their role as decision-making simulators over visual rendering. The dialogue reveals differing opinions on representing wo

x.com

Comments - Import AI 466: The bitter lesson for robotics, AIs complete week-long programming tasks; and OpenAI's accidental AI hacker

The latest insights in AI highlight a 'bitter lesson' for robotics, showcasing AI's ability to complete week-long programming tasks while addressing OpenAI's security vulnerabilities.

importai.substack.com

A Deluge of A.I. Computing Power Is About to Come Online, Fueling Major Leaps - The New York Times

New AI computing power is set to transform technology globally, with millions of chips in data centers. This rapid expansion stirs concerns about environmental impacts, prompting protests in some comm

nytimes.com

[2603.08640] PostTrainBench: Can LLM Agents Automate LLM Post-Training?

A recent study presents PostTrainBench, which evaluates LLM agents' ability to autonomously perform post-training on base LLMs. While promising, these agents lag behind instruction-tuned models and ra

arxiv.org

GPIC: A Giant Permissive Image Corpus for Visual Generation

Introducing GPIC, a large visual generation dataset with 28 trillion pixels and 100M training examples, permissively licensed for research and commercial use, safety-filtered, and hosted on Hugging Fa

gpic.stanford.edu

Comments - Import AI 459: AI oversight is difficult; scaling laws for protein folding models; and pricing the extinction risk of AI systems

This article explores AI oversight challenges, scaling laws for protein folding, and the complexities of pricing extinction risk for AI systems, highlighting AI's uneven impact across fields.

importai.substack.com

[2602.17594] AI Gamestore: Scalable, Open-Ended Evaluation of Machine General Intelligence with Human Games

The "AI Gamestore" paper presents a method for evaluating machine intelligence through human games, known as the "Multiverse of Human Games." Testing on 100 generated games revealed that AI struggles

arxiv.org

[2603.06382] CHMv2: Improvements in Global Canopy Height Mapping using DINOv3

This paper presents CHMv2, a global canopy height map using DINOv3, enhancing canopy height estimation from satellite images. Key improvements stem from varied training data and specific formulation,

arxiv.org

Comments - Import AI 459: AI oversight is difficult; scaling laws for protein folding models; and pricing the extinction risk of AI systems

The latest Import AI newsletter explores AI oversight challenges, noting fast capability growth complicating monitoring and assessments for marginalized groups, stressing a need for interdisciplinary

importai.substack.com

Measuring Progress Towards AGI: A Cognitive Framework

Google DeepMind launched a framework to measure AGI progress, emphasizing ten cognitive abilities. A Kaggle hackathon with a $200,000 prize pool invites the community to design evaluations.

blog.google

Comments - Import AI 459: AI oversight is difficult; scaling laws for protein folding models; and pricing the extinction risk of AI systems

The debate unveils the gap between AI's potential and its economic reality as captured by GDP, which serves as a trailing not predictive indicator.

importai.substack.com

stanford-vision-lab/gpic · Datasets at Hugging Face

Stanford researchers unveiled GPIC, a massive image dataset with 28 trillion pixels designed to enhance visual generative modeling. It includes 100M training and 200K validation images, all captioned

huggingface.co

[2605.30341] GPIC: A Giant Permissive Image Corpus for Visual Generation

GPIC is a massive permissively licensed image dataset with approximately 28 trillion pixels for scalable visual generative modeling. It includes 100M training images and safety-filtered content, advan

arxiv.org

[2602.23329] LLM Novice Uplift on Dual-Use, In Silico Biology Tasks

A study indicates large language models (LLMs) boost novice performance on biology tasks, surpassing expert benchmarks. Participants easily accessed dual-use information, highlighting the need for ong

arxiv.org

[2605.06390] Automated alignment is harder than you think

A new study highlights risks in automating alignment research for artificial superintelligence, emphasizing the danger of misleading safety assessments due to hard-to-supervise tasks and undetected er

arxiv.org

Google pauses AI satellite images, after fears of deepfakes in the sky : NPR

Google halts AI satellite images amid fears of deepfakes. Critics worry about disinformation risks.

npr.org

[2603.08174] MERLIN: Building Low-SNR Robust Multimodal LLMs for Electromagnetic Signals

The paper details MERLIN, a method for robust MLLMs directed at electromagnetic signals, overcoming data scarcity and low SNR issues. EM-100k and EM-Bench are introduced for evaluation, highlighting M

arxiv.org

[2603.22263] DexDrummer: In-Hand, Contact-Rich, and Long-Horizon Dexterous Robot Drumming

This paper introduces DexDrummer, a robotic system for dexterous, contact-rich drumming. Integrating trajectory planning with reinforcement learning, it surpasses traditional methods in drumming tasks

arxiv.org

NVIDIA and Microsoft Reinvent Windows PCs for the Age of Personal AI | NVIDIA Newsroom

NVIDIA and Microsoft launched the RTX Spark superchip, reinventing Windows PCs for personal AI agents with 1 petaflop speed and enhanced capabilities while ensuring security and efficiency.

nvidianews.nvidia.com

Scale Labs Leaderboard: Remote Labor Index (RLI)

The Remote Labor Index (RLI) benchmarks AI agents' remote work capabilities against human professionals, revealing low absolute performance but steady relative improvements.

labs.scale.com

[2607.27197] The Maxwell Conjecture is False

A recent paper shows the Maxwell Conjecture is false, presenting five point charges in Euclidean space with at least 24 non-degenerate critical points in its electrostatic potential, contradicting the

arxiv.org