Search the archive

STATEMENT ON SUPERINTELLIGENCE

Over 20,0000 signatories, including AI pioneers Geoffrey Hinton and Yoshua Bengio, Apple co-founder Steve Wozniak, Richard Branson, Prince Harry, Steve Bannon, Glenn Beck, five Nobel laureates, and Pope Francis's AI advisor, called for a prohibition on developing superintelligence until proven safe and controllable with strong public buy-in.

CHATGPT ON WHATSAPP ENDING

OpenAI has announced the end of WhatsApp access for ChatGPT users and encouraged linking to a ChatGPT account to continue conversations elsewhere.

ANTHROPIC IS CATCHING UP WITH OPENAI

Anthropic quietly went from $1B to nearly $7B in run-rate revenue this year, a six-fold jump that puts it within striking distance of OpenAI.

UNSEEABLE PROMPT INJECTIONS IN SCREENSHOTS: MORE VULNERABILITIES IN COMET AND OTHER AI BROWSERS

Indirect prompt injection is a systematic challenge that the entire category of AI-powered browsers faces.

Get access to the full library, free →

CHATGPT ATLAS

OpenAI introduced ChatGPT Atlas, an AI-powered web browser that integrates ChatGPT into the browsing experience. Atlas includes a sidecar chatbot panel with contextual awareness of what's on screen and tracks browsing history to personalize responses.

ANTHROPIC, GOOGLE IN TALKS ON MULTIBILLION-DOLLAR CLOUD DEAL

Anthropic is in early-stage talks with Google about a deal that would provide it with additional computing power valued in the high tens of billions of dollars. The plan involves Google providing cloud computing services to Anthropic, giving Anthropic access to Google's tensor processing units, chips custom-designed to accelerate machine learning workloads. Google is a previous investor in and a cloud provider for Anthropic.

BLACKROCK AMONG BIGGEST INVESTORS IN META'S GIANT DATA-CENTER DEBT DEAL

Meta and private-credit firm Blue Owl raised $27 billion to finance the buildout of a data center in Louisiana. BlackRock was one of the biggest investors in the deal. It purchased more than $3 billion in bonds last week to finance the deal. The project is 80% owned by Blue Owl, with Meta owning the remaining stake. The deal is the largest private-debt offering so far.

ON DWARKESH PATEL'S PODCAST WITH ANDREJ KARPATHY

THE CONTINUAL LEARNING PROBLEM

Memory layers are a natural architecture for models that are always training and continually learning from experience over time. Finetuning memory layers enables learning without forgetting much more effectively than LoRA. Continual learning and memory is still a huge and rich space to explore. We are only just starting to reach the point where we can imagine what it might look like to have models that learn online in the real world.

MINI MODELS BATTLE: CLAUDE HAIKU 4.5 VS GLM-4.6 VS GPT-5 MINI

Claude Haiku 4.5, GLM-4.6, and GPT-5 Mini all fall within a similar price bracket. They are mid-tier models optimized for cost and speed rather than maximum capability. GPT-5 Mini delivers the strongest concurrency safeguards, Haiku 4.5 is the fastest with the most features, and GLM-4.6 produces multi-file architecture but requires disabling reasoning mode for tool calls. Each model makes different trade-offs.

GOOGLE REVAMPS AI STUDIO WITH NEW FEATURES FOR VIBE CODING

Google has revamped its user experience for AI Studio. The update, targeted at developers, startups, and non-coders interested in rapid AI prototyping, helps users build AI-powered applications with less friction from prompt to production. Apps can be easily deployed to Google Cloud Run via a single click. An animated preview of the new interface is available in the article.

BOLT - HOW MURA WROTE AN IN-HOUSE LLM EVAL FRAMEWORK

BOLT is an evaluation system built by Mura. It has become critical infrastructure for how the company ships improvements, migrates between models, and maintains accuracy as it grows. Mura is a seed-stage startup that helps commercial HVAC providers get paid faster without hiring more back-office staff. It turns messy inputs into structured data using AI and orchestration, and then presents suggestions for users to approve. This article talks about how Mura built BOLT.

THE NOT-SO BITTER LESSON

Maybe AGI will take over everything, but we're clearly not there yet and probably won't be anytime soon. Software engineers still have a lot of time before they become obsolete. Right now, developers should focus on building scalable systems that leverage search and compute, build harnesses instead of solutions, and define problems instead of hard-coding solutions.

OPENAI HAS AN ARMY OF EX-INVESTMENT BANKERS MAKING FINANCIAL MODELS TO TRAIN CHATGPT

OpenAI has hired over 100 former investment bankers to help teach OpenAI's models. They are getting paid $150 an hour to do the grueling work of junior bankers, including tweaking PowerPoint slides and billing financial models in Excel. This shows that OpenAI is looking for opportunities in the business world. The company's consumer subscriptions currently account for 70% of its $13 billion annual recurring revenue.

AIRBNB CEO SAYS CHATGPT ISN'T READY

Airbnb's CEO, Brian Chesky, said that he didn't integrate Airbnb's travel app with OpenAI's ChatGPT because ChatGPT's connective tools aren't quite ready yet. Airbnb will monitor the development of ChatGPT's app integrations and may consider adding them in the future. Chesky, who is close friends with OpenAI CEO Sam Altman, said he advised OpenAI on its new capability for third-party developers to make their apps available within ChatGPT. Airbnb on Tuesday announced new AI tools to let customers take more actions without the need for a live representative. The company's AI customer service agent now displays action buttons and links that can help people complete reservation changes and cancellations.

Subscribe now.

YOUTUBE ROLLED OUT AI LIKENESS PROTECTION (8 MINUTE VIDEO)

YouTube has launched its likeness-detection technology for eligible creators in the Partner Program, allowing them to request the removal of AI-generated content that uses their face or voice.

GOOGLE SKILLS OFFERS PRACTICAL AI LEARNING PATHS

US CHIP PLANT INVESTMENT TO OUTPACE CHINA, TAIWAN, AND SOUTH KOREA FROM 2027

US semiconductor investment will surpass China, Taiwan, and South Korea by 2027 due to rising AI demand and government localization efforts.

CLAUDE CODE ON THE WEB

Anthropic has launched a browser-based version of Claude Code. The beta allows users to connect GitHub repositories and delegate multiple tasks simultaneously across different repositories, with each session running in isolated sandboxes.

THIS IS HOW MUCH ANTHROPIC AND CURSOR SPEND ON AMAZON WEB SERVICES

It is estimated that Anthropic spent $2.66 billion on AWS through September against $2.55 billion in estimated revenue, before accounting for Google Cloud costs or operations. Cursor's AWS bills doubled from $6.2 million to $12.6 million in June after Anthropic introduced Priority Service Tiers that sharply increased caching costs. Anthropic's costs appear to scale linearly with revenue - the path to profitability might have to be dramatic price increases.

META AI'S APP DOWNLOADS AND DAILY USERS SPIKED AFTER LAUNCH OF ‘VIBES' AI VIDEO FEED

Anthropic launched a web app for Claude Code yesterday. It is now rolling out to Pro and Max users. Anthropic is attempting to evolve Claude Code beyond a command-line interface tool. It hopes developers will spin up AI coding agents in more places now that Claude Code is on the web. Claude Code has grown 10x in users since its broader launch in May. It now accounts for more than $500 million of Anthropic's annual revenue.

BUBBLE, BUBBLE, TOIL, AND TROUBLE

Pro-AI-bubble arguments are remarkably weak, given the amount of bubble talk. The recent series of circular deals done in close succession has raised concerns. If there is a massive financial crisis or bubble pop that affects AI, it will be mostly for mundane economic reasons rather than the technology not making rapid progress. Such a crisis will have, at most, modest effects on timelines.

CAN YOU INFINITELY LEARN WITH ONLINE RL?

Geospot Infinity is a photo-to-GPS model. It retrieves 10 candidate coordinates for each upload and ranks them for the user. Users then pick the closest. The model attempts to learn from interactions via online reinforcement learning. It turned out that over 65% of users would just click the first guess regardless of accuracy, effectively nullifying the model's ability to learn anything.

BERT IS JUST A SINGLE TEXT DIFFUSION STEP

BERT-style models are essentially just text diffusion models trained on one masking rate. Masked language models can be repurposed into full generative engines by interpreting variable-rate masking as a discrete diffusion process. The standard masked language modeling objective can be turned into a step-by-step generation procedure by gradually corrupting text with mask tokens and training the model to iteratively denoise at increasing mask intensities. A fine-tuned RoBERTa can generate coherent-looking text after slightly modifying the training objective, even without architectural changes.

DO AIS THINK DIFFERENTLY IN DIFFERENT LANGUAGES?

Testing ChatGPT, Claude, and DeepSeek across English, Chinese, Arabic, Hindi, French, and Spanish revealed striking consistency in liberal, secular values regardless of language. All models strongly disagreed that sons deserve education priority over daughters, despite many humans worldwide answering yes. AIs seem to share a mostly cohesive, single underlying worldview across languages rather than adapting to local cultural values.

DEEPSEEK-OCR

DeepSeek-OCR was created to investigate the feasibility of compressing long contexts via optical 2D mapping. It consists of DeepEncoder and DeepSeek3B-Moe-A570M. DeepEncoder serves as the core engine, while DeepSeek3B-Moe-A570M acts as a decoder. DeepSeek-OCR can generate training data for LLMs and VLMs at a scale of over 200,000 pages per day. Experiments show that when the number of text tokens is within 10 times that of vision tokens, the model can achieve a decoding precision of 97%.

LOLMIL: LIVING OFF THE LAND MODELS AND INFERENCE LIBRARIES

Researchers built malware that uses Windows' built-in AI model to autonomously find and exploit security vulnerabilities without connecting to external servers. The technique only works on systems with NPUs or GPUs since CPU inference would be prohibitively slow, limiting deployment to CoPilot+ PCs and high-end computers until AI hardware becomes standard.

OMNIVINCI ADVANCES OMNI-MODAL LLMS

OmniVinci is a multi-modal architecture with novel techniques for aligning and encoding image, video, audio, and text data. The 9B-parameter model outperforms Qwen2.5-Omni across perception benchmarks while requiring just a fraction of the training tokens.

CLAUDE FOR LIFE SCIENCES

Anthropic is pushing into life sciences following OpenAI and Google with connectors to Benchling, PubMed, BioRender, and genomics platforms, alongside dedicated Agent Skills for tasks like single-cell RNA sequencing quality control. The company aims to support the entire research workflow from early discovery through regulatory compliance and commercialization, rather than just individual tasks like code writing or paper summarization.

ADOBE'S CUSTOM AI FOUNDRY FOR ENTERPRISES

Adobe has released AI Foundry, a new service that allows enterprises to train Firefly-based generative models on proprietary branding data. The offering supports output across text, images, video, and 3D, and shifts pricing to usage-based billing for greater flexibility.

OPENEVIDENCE, THE CHATGPT FOR DOCTORS, RAISES $200M AT $6B VALUATION

OpenEvidence raised $200 million at a $6 billion valuation three months after raising $210 million at $3.5 billion.

PINTEREST HAS INTRODUCED GENAI FEED CONTROLS

Pinterest rolled out new tools that let users adjust how much generative AI content appears in their feed.

AGENTIC AI'S OODA LOOP PROBLEM

AI agents embed untrusted actors within themselves in their training, so fixing hallucinations is insufficient because even if an AI accurately interprets its inputs and produces corresponding output, it can be fully corrupt.

SOLVING VIRTUAL MACHINE PUZZLES: HOW AI IS OPTIMIZING CLOUD COMPUTING

LAVA is a scheduling algorithm that continuously re-predicts and adapts to the actual lifetimes of virtual machines to optimize resource efficiency in large cloud data centers.

Implement MCP Auth with WorkOS →

GROUNDING WITH GOOGLE MAPS: NOW AVAILABLE IN THE GEMINI API

Google Maps integration is now available in the Gemini API, enabling developers to incorporate Maps data into geospatial-aware AI applications. The tool enhances use cases in travel, real estate, retail, and logistics by providing precise, location-based responses. The tool's combination with Google Search further boosts response quality, offering structured data and timely information.

APPLE EMPLOYEES HAVE 'CONCERNS' OVER SIRI PERFORMANCE IN EARLY IOS 26.4 BUILDS

Apple engineers testing the early iOS 26.4 build have concerns about Siri's performance despite extended development time. Internal teams are working on two Siri models, one using on-device models and the other powered by Google Gemini. Concerns persist about Apple's AI strategy, potentially leading to reliance on external solutions.

ALIBABA CLOUD CLAIMS TO SLASH NVIDIA GPU USE BY 82% WITH NEW POOLING SYSTEM

Alibaba Cloud's new Aegaeon system reduces Nvidia GPU usage by 82%, cutting the number of GPUs needed for large AI models from 1,192 to 213. Beta tested in their model marketplace, Aegaeon addresses resource inefficiency by pooling GPU power. This pooling allows one GPU to serve multiple models, optimizing concurrent LLM workload management.

SUMMONING GHOSTS WITH ANDREJ KARPATHY (146 MINUTE VIDEO)

In a wide-ranging conversation, Andrej Karpathy discusses AGI's economic impact, the limits of reinforcement learning, lessons from self-driving research, and his vision for the future of education.

ANDREJ KARPATHY ON AGI TIMELINES, ANIMALS VS GHOSTS, AND MORE

Andrej Karpathy was recently on a podcast where he discussed AGI timelines, Animals vs Ghosts, Reinforcement Learning, AI agents, and job automation. Some of the explanations on the podcast were botched. This post expands on the topics that Karpathy discussed on the podcast to clear things up. It includes links to previous discussions Karpathy has had on these topics.