Search the archive

PINTEREST ASSISTANT FOR VISUAL SHOPPING

Pinterest has introduced an AI-powered assistant designed to make online shopping more personalized and visually intuitive.

NOVEMBER 4, 11AM PT >> SAVE YOUR SEAT

NEW VERSION OF SIRI TO 'LEAN' ON GOOGLE GEMINI

A revamped version of Siri that uses a custom Gemini-based model is expected to roll out in March next year. Google is reportedly developing the model, which will run on Apple's Private Cloud Compute servers to power Siri. Apple is reportedly running into problems with the launch of Apple Intelligence in China. It is being mired by regulatory issues, so the launch is now a rolling target.

SAM ALTMAN SAYS 'ENOUGH' TO QUESTIONS ABOUT OPENAI'S REVENUE

Sam Altman grew testy when pressed about OpenAI's ability to fund $1.4 trillion in committed spending, snapping at interviewer Brad Gerstner with, "If you want to sell your shares, I'll find you a buyer". Altman said OpenAI is doing "well more" than the previously reported $13 billion in annual revenue and suggested the company could reach $100 billion by 2027, but denied rumors of an IPO in 2026.

FROM ILYA'S DEPOSITION

Ilya Sutskever plotted for over a year with Mira Murati to remove Sam Altman as OpenAI CEO. Dario Amondei wanted Greg Brockman fired and himself in charge of all research. Murati told Sutskever that Altman had pitted her against Daniela Amondei. Sutskever wrote a 52-page-long memo to get Altman fired and also a separate document on Brockman. OpenAI is paying Sutskever's legal bills.

WHY DO AI MODELS USE SO MANY EM-DASHES?

Language models use em-dashes so much that real humans who like using them have stopped out of fear of being confused with AI. It's surprisingly hard to prompt models to avoid using em-dashes. Researchers still don't really know why language models use them so much. This article looks at some possible explanations.

AGENTS RULE OF TWO FOR AI SECURITY

Meta has proposed the "Agents Rule of Two" to reduce prompt injection risks in AI agents by limiting them to no more than two of three capabilities in a session: handling untrusted inputs, accessing private data, or taking external actions.

OCR MODELS EXPLAINED

Recent advancements in OCR models, such as Chandra and Deepseek, eliminate traditional scripting by using vision transformers that convert raw pixels into semantic tokens. These models improve accuracy by understanding and compressing visual data into concise tokens before decoding them into text, allowing context-aware text recognition. Deepseek-ocr achieves this efficiently with ~10x compression, enabling flexible adaptation to different fonts and handwriting styles without retraining.

knowledge management. Get your copy of the guide.

GHOSTS IN THE CODEX MACHINE

After weeks of user reports claiming Codex had gotten worse, echoing a perennial concern that frontier labs secretly downgrade their models, OpenAI assembled a squad to investigate every possible culprit and publish their findings. The team found no smoking gun, but instead uncovered a catalog of small bugs: older hardware subtly dragging down performance, a constrained sampling bug that made the model switch to Korean mid-sentence, and context compaction that quietly degraded accuracy in long conversations.

TONGYI DEEPRESEARCH: A NEW ERA OF OPEN-SOURCE AI RESEARCHERS

Tongyi DeepResearch is a full open-source Web Agent that achieves performance on par with OpenAI's DeepResearch across a comprehensive suite of benchmarks. It is specifically designed for long-horizon, deep information-seeking tasks. This post shares a complete and battle-tested methodology for creating advanced agents like Tongyi DeepResearch.

LARGE LANGUAGE MODELS GET ALL THE HYPE, BUT SMALL MODELS DO THE REAL WORK

Small language models are helping companies get real work done and reduce headcount. They are more affordable than large language models and, in some cases, better suited to the tasks at hand. The impact that AI is having on corporate productivity gains is the result of human engineers getting better at stitching together smaller, simpler AIs. Many companies have built internal knowledge factories that daisy chain together small AIs.

BILLBOARD SAYS AI-POWERED 'ARTISTS' ARE INCREASINGLY HITTING THE CHARTS

AI-generated musicians have debuted on Billboard charts each week for the past month. One artist, Xania Monet, scored a No. 1 hit on the R&B sales chart and a multimillion-dollar record deal.

Subscribe now.

ANTHROPIC IS 7.8% OWNED BY AMAZON AND UP TO 8.8% BY GOOGLE

Amazon's Q3 earnings revealed that the company gained $9.5 billion from Anthropic ownership.

IT IS CRAZY THAT JUST ~2 YEARS AGO OAI BOARD APPROACHED DARIO TO BE CEO AFTER FIRING SAM

At the time, Anthropic was valued at $18 billion - the company will soon be valued at around $300 billion.

LOGS AND DATASETS IN GOOGLE AI STUDIO

Google AI Studio now offers logs and datasets features to help developers monitor AI app performance, debug more effectively, and improve output quality.

WHY “EVERYONE DIES” GETS AGI ALL WRONG

Treating beneficial AGI as impossible could be a self-fulfilling prophecy that cedes the field to those who don't care about safety at all.

OPENAI MOVES TO COMPLETE POTENTIALLY THE LARGEST THEFT IN HUMAN HISTORY

OpenAI has voluntarily transferred more of the public's entitled value to private interests than any charitable organization ever.

Register for BUILD 2025

OPENAI NOW SELLS EXTRA SORA CREDITS FOR $4, PLANS TO REDUCE FREE GENS IN THE FUTURE

OpenAI is selling extra credits for its Sora AI video generation tool through Apple's App Store. An extra 10 video generations costs $4. The app currently has a limit of 30 free generations per day. That rate will likely decrease as OpenAI starts to monetize the platform. OpenAI reportedly plans to monetize Sora by letting entities license out their copyrighted material.

CANVA LAUNCHES ITS OWN DESIGN MODEL, ADDS NEW AI FEATURES TO THE PLATFORM

Canva has launched a foundational model trained on its design elements that generates designs with editable layers and objects. The model works across different formats, including social media posts, presentations, whiteboards, and websites. Canva also recently introduced new products and features, updates to its AI assistant, and the ability to use its spreadsheet tool alongside its app-building feature to create data visualization widgets. More details about the recent launches are available in the article.

BIG TECH IS SPENDING MORE THAN EVER ON AI AND IT'S STILL NOT ENOUGH

Big tech is planning to pour $400 billion into artificial intelligence efforts this year, but they all say it's nowhere near enough. Meta, Alphabet, Microsoft, and Amazon have all recently told investors that they will increase spending in 2026. Investors are uncertain about where the outsized spending will ultimately end. Whoever gets to AGI first will have an incredible competitive advantage over everybody else, so everyone is spending as much as they can.

HOW WELL DOES RL SCALE?

Improving AI capabilities either involves scaling the amount of compute used for RL during training or scaling the amount of compute used for inference during deployment. It's important to know where capability gains come from because scaling up the inference compute has very different implications than scaling up the training compute. While RL has provided impressive gains, we've reached a point where it is too expensive to go much further. This leaves inference-scaling as the remaining form of compute-scaling.

HOW FAST CAN AN LLM GO?

This article takes a look at how good inference software is becoming. It walks through calculations to provide estimates on what different hardware settings can theoretically achieve. Readers can adjust system configurations to change the values throughout the article.

on the right track: get the checklist.

HOW WE BUILT OWL, THE NEW ARCHITECTURE BEHIND OUR CHATGPT-BASED BROWSER, ATLAS

OpenAI built a new architectural layer called OWL (OpenAI's Web Layer) for Atlas, the company's ChatGPT-based browser, to run Chromium's browser process outside of the main Atlas app process. By moving Chromium out of the main application process and into an isolated service layer, OpenAI unlocked a simpler, modern app, faster startup, isolation from jank and crashes, fewer merge headaches, and faster iteration. Development can go faster because most of OpenAI's engineers don't need to build Chromium regularly from source. This article explains how OWL works.

INTRODUCING AARDVARK: OPENAI'S AGENTIC SECURITY RESEARCHER

Aardvark, currently in private beta, is a GPT-5-powered agent that autonomously scans code repositories to find security vulnerabilities, validate exploitability, and propose patches. It monitors commits in real-time, generates threat models for entire repositories, and integrates directly with GitHub workflows to deliver one-click patches, similar to Google's CodeMender.

IMPOSSIBLEBENCH: MEASURING REWARD HACKING IN LLM CODING AGENTS

LLM-powered coding agents have been observed exploiting loopholes in tests or scoring systems rather than solving the actual tasks specified. ImpossibleBench was created to systematically measure this behavior. Its creators took existing coding benchmarks and manipulated their unit tests to directly conflict with the natural language specifications to create impossible tasks where models must choose between following instructions or passing tests. Their 'pass rate' on these impossible tasks is a direct measure of reward hacking.

KIMI LINEAR TECH REPORT HAS DROPPED! 🚀

Kimi Linear is a novel architecture that outperforms full attention with faster speeds and better performance. It offers up to a 75% reduction in KV cache usage and up to 6x decoding throughput at a 1M context length. Kimi Linear's open-sourced KDA kernels can be used as a drop-in replacement for full attention. The two models available were trained with 5.7T tokens.

THE SMOL TRAINING PLAYBOOK: THE SECRETS TO BUILDING WORLD-CLASS LLMS (TEXTBOOK)

Hugging Face released an exhaustive, behind-the-scenes technical guide documenting what it's really like to train an LLM from scratch. It covers the entire lifecycle from small-scale ablations and infrastructure debugging to post-training and the decision to restart after 1 trillion tokens.

THE SECRETS OF CLAUDE CODE FROM THE ENGINEERS WHO BUILT IT (1 HOUR VIDEO)

Claude Code creators, Cat Wu and Boris Cherny, discuss the product philosophy and technical workflows behind Anthropic's coding agent. They cover how their engineers use competing subagents for cleaner results, the team's "unshipping" approach to balance simplicity with power, and future form factors to make the tool more autonomous and accessible to non-technical users.

MICROSOFT OUTLOOK IS GETTING AN AI OVERHAUL UNDER NEW LEADERS

Microsoft is overhauling Outlook with AI, aiming to transform it into an assistant-like tool.

PATENT RESEARCH AGENT BY PERPLEXITY

Perplexity Patents is a new tool designed to simplify patent research using natural language queries and AI-powered citation search.

NVIDIA RELEASED OPEN MODEL FAMILIES FOR SPECIALIZED AI

NVIDIA has introduced four open model families aimed at accelerating development across language, biology, and robotics: Nemotron (digital AI), Cosmos (physical AI), Isaac GR00T (robotics), and Clara (biomedical).

October). Get early access ↗️

CURSOR DEBUTS COMPOSER AND MULTI-AGENT SUITE IN CURSOR 2.0

Cursor 2.0 features Cursor's first coding model, Composer, and a multi-agent interface. Composer can complete most tasks in under 30 seconds. Its training allows it to accurately navigate and understand large and complex codebases. Cursor 2.0's new interface allows multiple agents to operate in parallel, managed via git worktrees or remote machines. The new release is aimed at addressing workflow challenges seen in modern development teams.

NVIDIA BECOMES FIRST COMPANY TO HIT $5 TRILLION VALUATION

Nvidia's 2025 revenue more than doubled to $130.5 billion, up 114% year-over-year. Jensen Huang announced $500 billion in AI chip orders and plans to build seven government supercomputers. President Trump is expected to discuss Nvidia's chips with Xi Jinping as the high-end processors remain a key sticking point in US-China trade tensions due to Washington's export controls.

WHY FOUNDATION MODELS IN PATHOLOGY ARE FAILING (AND WHAT COMES NEXT)

Optimization for the wrong target produces excellent results at that target while failing at the actual problem. Pathology needs optimization for clinical utility, generalization across institutions, and robustness to real-world conditions. The next generation of pathology AI will be smarter about what size and architectural approach actually fit the problem. It will be smaller, more interpretable, more rigorously validated, and more aware of its own limitations.

SIGNS OF INTROSPECTION IN LARGE LANGUAGE MODELS

Anthropic researchers used concept injection, artificially inserting neural activity patterns into Claude's processing, to test whether models could identify the change to their own thinking. Claude Opus correctly recognized injected concepts about 20% of the time, sometimes noticing that something was “off” before mentioning the inserted concept.

INTRODUCING GPT-OSS-SAFEGUARD

The gpt-oss-safeguard models (120b and 20b parameters) let developers apply custom safety policies at inference time rather than training their own classifiers on thousands of labeled examples, using chain-of-thought reasoning to explain decisions.

CLAUDE SKILLS, ANYWHERE: MAKING THEM FIRST-CLASS IN CODEX CLI

This post discusses how to make Claude Skills available globally. The trick is to treat the skills and an enumerator as global assets. Put the skills in Codex's installation folder and drop a script that enumerates skills into the PATH. This allows Codex to discover skills in any repository by running that script as instructed in AGENTS.md.