Search the archive

LLMS GET LOST IN MULTI-TURN CONVERSATION

LLMs perform significantly worse in multi-turn conversations, with an average 39% drop in task performance due to unreliability and early, incorrect assumptions.

A GENERALIST ROBOT POLICY FRAMEWORK

UniVLA enables policy learning from unlabeled video across diverse robot embodiments by inferring task-centric latent actions.

JOURNALISTS REVEAL NUANCED APPROACHES TO AI INTEGRATION

A survey of media professionals from Reuters, The Washington Post, VentureBeat, 404 Media, and others shows newsrooms are adopting AI selectively, embracing it for transcription, data analysis, and translation while rejecting AI content creation. While Reuters reports that AI now generates about 25% of its code, many journalists remain skeptical, prioritizing maintaining audience trust and journalistic integrity over technological efficiency.

FIGMA WEBSITE BUILDER

Figma has introduced Figma Sites, allowing users to design, build, and publish responsive websites directly from Figma, streamlining the design-to-production workflow.

BAMBA-9B-V2

IBM, Princeton, CMU, and UIUC released Bamba v2, a Mamba2-based model that outperforms Llama 3.1 8B after training on 3 trillion tokens. Bamba v2 leverages Mamba2 architecture for 2-2.5x faster inference and excels in L1 and L2 benchmarks. The team plans further optimizations and invites community collaboration to enhance model performance.

TRUMP ADMINISTRATION SCRAPS BIDEN-ERA AI CHIP EXPORT CONTROLS

The Trump administration has canceled Biden's "AI diffusion rule," which would have restricted American technology exports.

MANUS EXPANDS FREE ACCESS

Manus has eliminated its waitlist, offering broader access to its virtual desktop AI agent with one free daily task for all users and a one-time bonus of 1,000 credits and significantly lowering the barrier to entry for this previously hyped automation tool.

A MODULAR AND MULTILINGUAL LLM

Helium 1, a 2 billion parameter LLM, excels in European languages and is optimized for on-device use.

Start learning today with Google Cloud Skills Boost

Analyze and Reason on Multimodal Data with Gemini

Enhance Gemini Model Capabilities

Implement Multimodal Vector Search with BigQuery

Learn. Explore. Gain the badges to show you've done it.

META'S NEW FAIR DIRECTOR

Meta has appointed Robert Fergus, previously a research director at DeepMind, to head its FAIR lab after a period of leadership changes and talent departures.

OPENAI AND MICROSOFT NEGOTIATE PARTNERSHIP TERMS TO ENABLE FUTURE IPO

OpenAI and Microsoft are renegotiating their multibillion-dollar partnership in a critical deal that could determine the AI leader's future. Microsoft, which has invested over $13 billion in OpenAI, is offering to exchange some of its equity stake for access to technology beyond their current contract's 2030 expiration.

DEEP RESEARCH NOW SUPPORTS GITHUB REPO ANALYSIS IN CHATGPT

ChatGPT's Deep Research agent can now scan GitHub repositories, reading source code and PRs to deliver detailed, cited reports. Users can query repositories directly through the Deep Research → GitHub integration.

OPTIMIZING GEMM WITH THREAD BLOCK CLUSTERS

Thread block clusters and 2-SM UMMA instructions on Blackwell GPUs enable higher arithmetic intensity and more efficient memory transfers in GEMM workloads using CUTLASS.

CHATGPT'S CARBON FOOTPRINT IS NEGLIGIBLE COMPARED TO DAILY ACTIVITIES

ChatGPT queries consume far less energy than commonly believed. Updated estimates suggest just 0.3 Wh per typical query - ten times lower than previous calculations. While AI's aggregate energy demand requires attention, individual text-based AI usage shouldn't trigger climate guilt, especially compared to higher-impact activities like transportation or heating.

OPEN SOURCE PROJECT CURL BATTLES WAVE OF AI-GENERATED FALSE VULNERABILITIES

curl's project founder, Daniel Stenberg, has reached a breaking point with the flood of AI-generated false vulnerability reports submitted through platforms like HackerOne. While HackerOne maintains that AI can enhance reporting quality when used responsibly, Stenberg is pushing for stronger infrastructure and tools to combat what he describes as effectively a DDoS attack on maintainers' time and attention.

META ASSETGEN 2.0

Meta's AssetGen 2.0 is a set of new diffusion-based models for generating detailed 3D meshes and textures from text and image prompts that improve consistency, fidelity, and view-aware texture resolution over its predecessor.

FLOW-GRPO FOR RL-TUNED DIFFUSION MODELS

Flow-GRPO integrates reinforcement learning into flow matching by converting ODEs into SDEs and using denoising reduction to enhance sample efficiency and alignment.

DEERFLOW

Bytedance's DeerFlow is an open source research assistant that employs a multi-agent architecture with search engines, web crawlers, and Python capabilities to deliver Deep Research-style reports and podcasts.

GEMINI 2.5 VIDEO UNDERSTANDING

Gemini 2.5 Pro has shown state-of-the-art performance on video benchmarks like YouCook2 and QVHighlights, outperforming GPT-4.1 and rivaling fine-tuned specialist models under equal evaluation settings.

CANADIAN PHARMACIST EXPOSED AS KEY FIGURE BEHIND WORLD'S LARGEST EXPLICIT DEEPFAKE SITE

Investigative journalists have identified Toronto-area pharmacist David Do as a central figure behind MrDeepFakes.com. The site permanently shut down following the revelations. It had 650,000 users and generated over 2 billion views since 2018. The site hosted tens of thousands of non-consensual AI-generated explicit videos of celebrities, influencers, and private citizens. While non-consensual deepfakes remain legal in Canada, Prime Minister Mark Carney has pledged to criminalize it following the practice of countries like the UK and Australia.

SINGLE-IMAGE TO 3D AVATARS

SVAD combines video diffusion and 3D Gaussian Splatting to generate high-quality animatable avatars from a single image, achieving real-time rendering.

OPENAI TO ACQUIRE CODING TOOL WINDSURF FOR $3 BILLION

OpenAI has reached an agreement to acquire AI-assisted coding platform Windsurf (formerly Codeium) for approximately $3 billion.

SOUNDCLOUD'S QUIET TERMS UPDATE

SoundCloud reportedly changed its terms of service to permit AI training on user-uploaded content, prompting concerns over transparency and consent.

FUTUREHOUSE RELEASES AI TOOLS IT CLAIMS CAN ACCELERATE SCIENCE

FutureHouse, backed by Eric Schmidt, has launched AI tools to aid scientific research, including Crow, Falcon, Owl, and Phoenix.

MISTRAL MEDIUM 3

Mistral Medium 3 was introduced to deliver strong enterprise performance at significantly lower cost. It is optimized for deployment flexibility and coding tasks.

FIDJI SIMO JOINS OPENAI AS CEO OF APPLICATIONS

OpenAI has appointed Fidji Simo to lead its Applications division, reinforcing its focus on scaling product and operations as it expands beyond research into global deployment and infrastructure.

ANTHROPIC LAUNCHES WEB SEARCH API FOR CLAUDE

Anthropic launched a new Web Search API for its Claude models. This API allows Claude to access and use real-time information from the web. The capability enhances Claude's ability to provide current and accurate responses.

IS THERE A HALF-LIFE FOR THE SUCCESS RATES OF AI AGENTS?

AI performance on long tasks follows a simple model with a constant failure rate, creating an exponential decline in success. Each AI agent can be characterized by a "half-life" that estimates success over varying task lengths. This model suggests failures arise from complex sets of subtasks.

SEPARATING FACT FROM FICTION: HERE'S HOW AI IS TRANSFORMING CYBERCRIME

AI in cybersecurity primarily enhances existing techniques rather than creating new threats and lowers entry barriers for cybercriminals. At the recent RSA Conference, experts highlighted AI's role in automating tasks and facilitating advanced cybercrime models like AI-as-a-Service. Future defenses will require AI-driven strategies and international collaboration to effectively combat evolving threats.

THE LEADERBOARD ILLUSION

Chatbot Arena's benchmarking exhibits bias due to undisclosed private testing and selective data access. Providers like Google and OpenAI dominate data access, while open-source models receive significantly less. These dynamics lead to overfitting rather than genuine model improvement.

OSMOSIS SELF-IMPROVEMENT VIA REAL-TIME REINFORCEMENT LEARNING

Osmosis is a platform for AI self-improvement via real-time reinforcement learning. The team has open sourced a lightweight model that works as well as SOTA for MCP. It can be run locally.

ACTOR-CRITIC LEARNING WITH OFFLINE DATA

A new actor-critic RL algorithm has achieved near-optimal sample efficiency using offline data and targeted exploration, addressing long-standing challenges in hybrid RL settings.

AMIE GAINS VISION: A RESEARCH AI AGENT FOR MULTIMODAL DIAGNOSTIC DIALOGUE

Google Research and DeepMind collaborated on this research, with contributions from many experts across various teams.

GOOGLE LAUNCHES 'IMPLICIT CACHING' TO MAKE ACCESSING ITS LATEST AI MODELS CHEAPER

Google's new "implicit caching" feature in its Gemini API reportedly offers 75% cost savings on repetitive context for its Gemini 2.5 models. Unlike the previous manual explicit caching method, this automatic feature could ease developer concerns about unexpected API costs. However, developers should monitor its effectiveness, as Google hasn't provided third-party verification for the cost savings claims.

AI-GENERATED CODE COULD BE A DISASTER FOR THE SOFTWARE SUPPLY CHAIN. HERE'S WHY

AI-generated code often includes non-existent library references, exposing systems to supply-chain attacks via dependency confusion. A study found 19.7% of dependencies from tested LLMs were fabricated, creating security risks. Open-source LLMs hallucinate more frequently than commercial ones, with JavaScript showing higher error rates than Python.

HUGGING FACE RELEASES A FREE OPERATOR-LIKE AGENTIC AI TOOL

Hugging Face's Open Computer Agent is a cloud-hosted AI agent capable of basic tasks but struggles with complex requests like flight searches. Despite limitations and queue times, it showcases the potential of open AI models powering workflows and is part of a growing trend in agentic technology investments. A KPMG survey shows 65% of companies are experimenting with AI agents. The market is expected to grow significantly.

PERPLEXITY EXPANDING AI-POWERED LEARNING

Perplexity has teamed up with Wiley to integrate textbook content into its AI search experience, allowing students and institutions seamless access to curriculum materials and real-time explanations.

FREEPIK RELEASES AN 'OPEN' AI IMAGE GENERATOR TRAINED ON LICENSED DATA

Freepik's F Lite is an AI image model developed with Fal.ai and trained using 64 Nvidia H100 GPUs that utilizes licensed, safe-for-work images.

META ENTERS THE TOKEN BUSINESS, POWERED BY NVIDIA, CEREBRAS AND GROQ

Meta showcased its capability to rival ChatGPT at LlamaCon by collaborating with Cerebras and Groq for faster inference processing.

ANTHROPIC API SUPPORTS WEB SEARCH

Claude has been updated with web search support via API, enabling developers to build applications that incorporate real-time, up-to-date information from the internet.

AMAZON VULCAN ROBOT

Amazon's Vulcan is the company's first robot. It has a sense of touch, enhancing handling precision and marking a new phase in physical AI capabilities.

AI SEARCH IN SAFARI

Apple is considering integrating AI search engines like Perplexity and Claude into Safari, reflecting growing momentum behind LLM-powered search.

ANTHROPIC ECONOMIC INDEX: AI'S IMPACT ON SOFTWARE DEVELOPMENT

AI systems like Claude are dramatically impacting coding by automating significant portions of programming tasks. Startups are leading in the adoption of AI coding tools like Claude Code, which is heavily used for front-end development, while enterprises lag behind. As AI capabilities grow, the nature of developer roles may shift towards managing AI systems, potentially accelerating technological advancements.

PATHOLOGY WITH DINOV2

The Mahmood Lab, using Meta's DINOv2, has developed open-source AI models for pathology, improving disease detection and diagnostics.

CAN O3 BEAT A GEOGUESSR MASTER?

Researcher Sampatt tested O3, an AI agent, against a GeoGuessr expert in the challenging street view geography game. O3 demonstrated strong reasoning and geographic inference but still fell short of consistently outperforming the human master. The experiment highlights both AI's growing spatial understanding and its current limitations in complex real-world tasks.