Search the archive

NEURAL MOTION SIMULATOR FOR EMBODIED AI

MoSim introduces a world model for motion dynamics that improves skill acquisition and enables zero-shot learning, effectively turning model-free RL into model-based.

VIDEOCHAT R1

An inference time compute captioning method for video question answering. It uses RL to improve the overall reasoning of the model which leads to a 30% boost in object tracking tasks.

CONCEPT ATTENTION

A novel method to view concepts in the attention map of neural networks.

MAMMAL POSE ESTIMATION

KITPose is a novel keypoints-interactive model designed for general mammal pose estimation.

OUR VISION FOR ACCELERATING CREATIVITY AND PRODUCTIVITY WITH AGENTIC AI

Adobe is integrating agentic AI across its product suite, including Acrobat, Photoshop, and Premiere Pro, to enhance creativity and productivity by automating repetitive tasks and providing smart recommendations. Agentic AI aims to empower users at all levels by allowing them to focus more on creative endeavors and less on technical execution. This approach will help streamline workflows, enabling professionals to achieve more with their tools.

US ENGINEERS' AI CONVERTS SIMPLE TEXT INTO WALKING ROBOTS IN A DAY

Duke University's Text2Robot enables non-experts to create functional 3D robots from simple text inputs using a generative AI framework. This technology democratizes robotic design by eliminating complex barriers previously requiring extensive technical knowledge.

ANNOUNCING THE AGENT2AGENT PROTOCOL (A2A)

Aiming to advance AI collaboration and automation in enterprise environments, Google Cloud has launched the Agent2Agent (A2A) protocol to enable seamless interoperability between AI agents across different platforms and vendors, with support from over 50 tech partners including Atlassian, Salesforce, and PayPal. A2A provides a standardized framework for secure communication and task coordination among agents, enhancing productivity and reducing costs for enterprises. The open-source protocol also supports varied modalities and long-running tasks.

SCULPTOR: CATCH AND FIX ISSUES AS YOU CODE

Sculptor is an early access coding agent environment that embeds software engineering best practices by running code in sandboxed environments.

WORDPRESS AI WEBSITE BUILDER

A new AI website builder creates complete WordPress sites from user input, ideal for entrepreneurs, freelancers, and bloggers.

TRUMP ADMINISTRATION BACKS OFF NVIDIA'S H20 CHIP CRACKDOWN

The Trump administration reversed plans to restrict Nvidia's H20 AI chip exports to China after Nvidia CEO promised U.S. investments.

Watch the recording here →

GOOGLE IRONWOOD TPU

With improvements on podsize, bandwidth, memory, and overall flops, the 7th generation of Google's TPUs is a step change improvement over any chip on the market.

CLAUDE'S MAX PLAN

The new Max plan for Claude offers expanded usage with priority access to features and models, supporting users with demanding projects. It provides two levels: Expanded Usage for $100/month with 5x usage and Maximum Flexibility for $200/month with 20x usage. The plan is ideal for those who frequently work with substantial documents and complex data.

GROK API

Grok 3 has an API for its models, including multiple reasoning modes. This link includes pricing, which is reasonable given the assumed size of the model itself.

OMNI SVG

By treating SVGs as a foreign language, a pretrained Qwen model can generate novel SVGs based on text and images. It is easily the state-of-the-art model. An open release is coming soon.

OLMOTRACE

There is a debate in the language modeling world about how much a model actually learns and how much it just memorizes. This feature, in the AI2 Playground, searches through billions of input documents, in real time, to figure out if the output of a model is just regurgitated or if it is novel. It includes the source of the utterance from multiple documents.

AI SCIENTIST V2 (PDF)

Sakana AI had a research paper accepted to an ICLR workshop that was fully generated, executed, and written by a language model system. They improved their system by using VLMs, more general purpose search, and more.

EFFICIENT MOE INFERENCE

HybriMoE is a new framework for hybrid CPU-GPU inference on Mixture of Experts models that tackles instability and overhead with smarter scheduling and caching strategies.

DYNAMIC KNOWLEDGE CIRCUITS

This research explores how LLMs structurally internalize new knowledge by analyzing computational subgraphs. It reveals patterns in knowledge acquisition, optimization phases, and potential strategies for improving continual pre-training.

PROTEIN BACKBONE GENERATION

ReQFlow sets a new benchmark in protein backbone generation, delivering state-of-the-art results while being significantly faster than existing models. It is 37x faster than RFDiffusion and 62x faster than Genie2 for sequences of length 300.

COGITO V1 PREVIEW: INTRODUCING IDA AS A PATH TO GENERAL SUPERINTELLIGENCE

The Cognito team has released open-source LLMs ranging from 3B to 70B parameters, all of which outperform top open models of similar sizes, with the 70B model surpassing Llama 4 109B MoE. These models are trained using Iterated Distillation and Amplification (IDA), support both direct and reflective answering, and are available on Hugging Face, Ollama, Fireworks AI, and Together AI.

GENERATIVE MEDIA UPGRADES IN VERTEX AI

Google Cloud has added enhanced generative tools to Vertex AI, including new capabilities in text-to-music (Lyria), video editing (Veo 2), voice customization (Chirp 3), and image inpainting (Imagen 3).

INTRODUCING FIREBASE STUDIO

Firebase Studio is launching as a cloud-based development environment for building, testing, and deploying AI applications. It combines tools like Project IDX, Genkit, and Gemini within a unified platform, offering rapid AI prototyping and a customizable coding workspace. Developers can leverage features such as natural language prototyping, instant device previews, and easy collaboration to accelerate app development.

fuel explosive growth. Last day at $0.55 per share.

HUGGING FACE AND CLOUDFLARE PARTNERSHIP

Developers can now easily build real-time voice and video apps using FastRTC, with enterprise-grade WebRTC infrastructure provided by Cloudflare and simple access via Hugging Face tokens.

GITHUB COPILOT INTRODUCES NEW LIMITS, CHARGES FOR 'PREMIUM' AI MODELS

GitHub Copilot users will face rate limits on AI models beyond the base GPT-4o model, with premium requests capped for paid tiers.

SHOPIFY CEO DECLARES AI PROFICIENCY A BASELINE EXPECTATION FOR EMPLOYEES

Shopify CEO Tobi Lütke announced that reflexive AI usage is now a fundamental expectation for all employees.

Applications in the Enterprise [4]" here.

STATE OF AI INDEX 2025

A great, high level, and comprehensive summary in 10 charts of the current state of AI across models, investment, and costs.

WAYMO'S PRIVACY CONCERNS

Waymo's draft privacy policy reveals it may use in-cabin camera footage to train AI models and personalize ads, prompting fresh concerns over data use in autonomous vehicles.

AI STARTUPS IN HEALTHCARE

Google for Startups has launched its 2025 AI for Health cohort, supporting 25 startups across EMEA with mentorship, product support, and AI best practices.

BENCHMARKING OPEN SOURCE MODELS FOR OCR

OCR is the task of recognizing text in images. It is challenging in the long tail and immensely valuable when done right. Many closed models such as the Gemini series are exceptional at this task. The new suite of Llama 4 models pushes the state of the art for open models forward substantially.

DEEPCODER

Together AI has trained a coding model that is competitive with closed source reasoning models. It has released data, code, training recipes, and showcased the model's strong long context ability.

A NEW BATCH NORMALIZATION

This paper proposes a new batch normalization method for SPD manifolds that uses a learnable Generalized Bures-Wasserstein metric.

HI DREAM IMAGE GENERATION MODEL

A strong 17B parameter image generation model that uses 4 different text encoders for generation. It has strong overall performance and is permissively licensed.

A FRAMEWORK FOR DYNAMIC MULTI-PRODUCT PRICING

This paper introduces a new framework for dynamic multi-product pricing that uses a censored multinomial logit model where buyers only consider products below their valuations.

MOTIFBENCH FOR PROTEIN DESIGN

MotifBench is a benchmark designed for computational protein design that focuses on the motif-scaffolding problem by identifying protein structures that stabilize target motifs.

HOW STUDENTS USE CLAUDE IN EDUCATION

Anthropic analyzed one million student conversations to understand AI usage in education. STEM students dominate usage, primarily leveraging Claude for content creation, technical problem-solving, and higher-order learning tasks.

ARABIC AI BENCHMARKS

Inception and MBZUAI have launched a unified Arabic AI evaluation platform featuring updated AraGen benchmarks and a new instruction-following leaderboard built on the Arabic IFEval benchmark.

17K REASONING TRACES FROM R1 (HUGGING FACE HUB)

A great set of reasoning traces from R1 that can be used as training data to distill a smaller reasoner or kick start the RL process.

HOW GOOGLE BUILT THE PIXEL'S ADD ME FEATURE

The "Add Me" feature on Pixel devices leverages advanced image segmentation and AI for personalized video experiences.

MICROSOFT'S COPILOT CAN NOW BROWSE THE WEB AND PERFORM ACTIONS FOR YOU

Microsoft's Copilot AI chatbot now performs tasks on popular websites, remembers user preferences, and analyzes real-time video.

ELEVENLABS RELEASES OFFICIAL MCP SERVER FOR AI-DRIVEN AUDIO PROCESSING

ElevenLabs has launched an official Model Context Protocol server for Text-to-Speech and audio processing that is compatible with clients like Claude Desktop and OpenAI Agents.

INSIDERS FROM GOOGLE, META AND AMAZON HAVE INVESTED AND NOW, YOU CAN TOO.

Shares are currently $0.55 per share, but not for long. On Apr 10, the share price changes - so nows the time.

Jump in at the current share price of $0.55

META RESPONDS TO LLAMA 4 RUMORS

Meta's VP of Generative AI has denied claims that Llama 4 models were trained on benchmark test sets, pushing back against allegations of inflated performance scores.