Search the archive

show up with work that actually lands. MEET OPAL

INTRODUCING AMAZON NOVA FORGE: BUILD YOUR OWN FRONTIER MODELS USING NOVA

Amazon Nova Forge is a service for building frontier models using Nova. It is the easiest and most cost-effective way to build your own frontier model. Customers can start their development from early model checkpoints, blend their datasets with Amazon Nova-curated training data, and host their custom models securely on AWS. The service is designed for organizations with access to proprietary or industry-specific data who want to build AI that truly understands their domain.

AGENT SAFETY BENCHMARK BY PERPLEXITY

BrowseSafe is a real-time content detection model and benchmark suite designed to protect AI agents from prompt injection in web browsers. The open-source BrowseSafe and BrowseSafe-Bench enable developers to scan HTML for hidden malicious instructions without slowing down performance, offering layered defenses for safer agentic browsing.

RAPTOR

RAPTOR (Recursive Autonomous Penetration Testing and Observation Robot) is an autonomous offensive/defensive security research framework based on Claude Code. It combines traditional security tools with agentic automation and analysis to deeply understand code, prove exploitability, and propose patches. The tool empowers security research with agentic workflows and automation. RAPTOR was vibe-coded and is still in an early release stage.

A PRACTICAL APPROACH TO VERIFYING CODE AT SCALE

AMP, INC.

Amp is spinning out of Sourcegraph to become an independent research lab. Its goal is to let software builders harness the full power of AI. The company believes that the way we develop software will be challenged fundamentally and drastically. Spinning out into its own company will give it more freedom to focus on the frontier.

THIS AI STARTUP WANTS TO REMAKE THE $800 BILLION CHIP INDUSTRY

Ricursive, a startup founded by ex-Google researchers, aims to build software that can automate the design of cutting-edge chips. This could allow every company to build their own chips from scratch. The company expects to release its first product next year. It recently raised $35 million and is currently valued at $750 million.

alternative? Get a free copy of the report from Miro

OPENAI TESTS MEMORY SEARCH AS GPT-5.2 LAUNCH RUMORS GROW

OpenAI's Memory search feature for ChatGPT lets users quickly find stored info via direct queries.

CLAUDE CODE HITS $1B RUN-RATE, ACQUIRES BUN

Anthropic announced it has acquired Bun, a fast JavaScript runtime.

HUMAN ART IN A POST-AI WORLD SHOULD BE STRANGE

The creative process is limited by human scale, but we're getting close to leaving that world.

OPERA ROLLS OUT GEMINI-POWERED AI FEATURES ACROSS ITS BROWSERS

Opera has expanded its AI capabilities with Google's Gemini models across its browsers.

THE BIGGEST DECISION YET

Humanity will have to decide by 2030 whether to risk letting artificial intelligence systems train themselves to become more powerful.

Plans start at just $49/month. Get a demo

DEEPSEEK-V3.2

DeepSeek's V3.2 matches GPT-5, and its higher-compute V3.2-Speciale variant rivals Gemini-3.0-Pro and earned gold medals at IMO, IOI, and ICPC 2025.

INTRODUCING RUNWAY GEN-4.5: A NEW FRONTIER FOR VIDEO GENERATION

Gen-4.5 topped the Artificial Analysis text-to-video benchmark, ahead of Veo 3 and Sora. The model emphasizes physical accuracy (realistic momentum, fluid motion, and material coherence). Runway has acknowledged persisting issues with object permanence.

EARLY LOOK: GEMINI'S CHATGPT-STYLE 'PROJECTS' ARE TAKING SHAPE

Google has been working on a new 'projects' option for Gemini. The feature will allow users to sort files and chats around particular topics. Projects can be pinned for easy access. There is a 10-file limit, but paying users may receive a higher ceiling.

A PRACTICAL APPROACH TO VERIFYING CODE AT SCALE

OpenAI trained an agentic code reviewer for Codex since noisy safety tools are inevitably bypassed. The system now handles 100k+ external PRs daily, providing repo-wide context and execution access. Internally, it has caught launch-blocking bugs and protected high-stakes experiments.

THE END OF THE TRAIN-TEST SPLIT

Models don't know what is 'out of distribution' for themselves. Engineers will always be required in the loop until this problem is solved. It will never be easy to check the accuracy of models. Researchers need to look at their data, make sure it's clean, and label it correctly.

HOW CAN INTERPRETABILITY RESEARCHERS HELP AGI GO WELL?

The Google DeepMind mechanistic interpretability team has pivoted to a pragmatic approach to interpretability over the last year. It is important to have empirical feedback on goals with good proxy tasks. Near-complete understanding isn't required for significant impact. Good focused projects start with a theory of change, and good exploratory projects start with a robustly useful setting.

report

NVIDIA LAUNCHES VISION LANGUAGE MODEL FOR AUTONOMOUS VEHICLES

Nvidia introduced Alpamayo-R1, the first open-source vision language action model designed for autonomous driving, at NeurIPS. Alpamayo-R1 integrates visual and textual reasoning to enhance decision-making in real-world environments.

STARFLOW: SCALABLE TRANSFORMER AUTO-REGRESSIVE FLOW (HUGGING FACE REPO)

STARFlow and STARFlow-V are state-of-the-art transformer autoregressive flow models for high-quality image and video generation. STARFlow introduces a novel transformer autoregressive flow architecture that combines the expressiveness of autoregressive models with the efficiency of normalizing flows. STARFlow-V is an end-to-end video generative model with normalizing flows. Examples of generated videos and comparisons are available.

BRIDGE MODELS FOR IMAGE AND VIDEO TRANSLATION

ViBT introduces Vision Bridge Transformers, scaling Brownian Bridge Models to 20B parameters for efficient conditional generation. The models use a Transformer architecture and a variance-stabilized objective for robust performance on image and video editing tasks.

OPENAI TAKES STAKE IN THRIVE HOLDINGS, A BUYER OF SERVICES FIRMS

OpenAI has taken an ownership stake in Thrive Holdings to embed AI into high-volume business processes, beginning with accounting and IT services. OpenAI teams will work directly inside Thrive's companies to improve speed, accuracy, and cost efficiency.

BYTEDANCE'S TIKTOK PLAYBOOK IS WINNING CONSUMER AI

ByteDance's Doubao app is now China's most popular mobile AI platform. It received more than 11.4 million downloads in October. Doubao's focus on frictionless AI-powered voice, image, and video experiences sets it apart from competitors. ByteDance keeps its most advanced technology proprietary, breaking away from China's open-source approach. This could give it a durable commercial edge.

cost. View the findings

A16Z'S GIGAWATT-SCALE DATA CENTER TIMELINE

Gigawatt-scale data centers can likely be built in 2 years or less

ACCENTURE AND OPENAI PARTNERSHIP

Accenture will deploy ChatGPT Enterprise to tens of thousands of its professionals, marking the largest upskilling effort yet via OpenAI Certifications.

BLACK FOREST LABS RAISES $300M AT $3.25B VALUATION

The Flux image model maker, whose co-founders also created Stable Diffusion, raised funding from Salesforce Ventures, a16z, Nvidia, and others.

APPLE AI CHIEF STEPS DOWN FOLLOWING SIRI SETBACKS

John Giannandrea is out after Tim Cook reportedly "lost confidence" in his leadership after repeated delays on an updated Siri.

CLAUDE CODE OVER EXCEL

The LlamaSheets API lets users automatically segment and structure complex Excel sheets into well-formatted 2D tables.

Go under the hood of Optimizely Opal.

CLAUDE 4.5 OPUS' SOUL DOCUMENT

There appears to be a document in the character training for Claude compressed within Claude's weights. While it is possible for the model to hallucinate the text, it appears to be real. This post details how the document was extracted. The full output of Claude's Soul Document is available in the post.

OPENAI MAY DEPLOY ADS SOON

A recent update to the ChatGPT Android app included references to an “ads feature” with “search ads” and “bazaar content”.

DATABRICKS REPORTEDLY IN TALKS TO RAISE $5B AT $134B VALUATION

Databricks is reportedly in talks to raise $5 billion at a $134 billion valuation. The company has resisted pressure to go public, but an IPO would likely be well-received. Databricks has more than 20,000 customers, including OpenAI, Block, Shell, and Toyota. Its platform offers tools that support the full lifecycle of AI development, including feature engineering, model training, evaluation, and deployment.

HOW PROMPT CACHING WORKS

Prompt caching works per-content, not per-conversation. Prefix caching works at the token level, not the request level, which is why it works across requests. Any change in the prefix breaks the entire hash chain.

ARE WE IN A GPT-4-STYLE LEAP THAT EVALS CAN'T SEE?

Chat is a terrible way to evaluate models. GPT-4 was so significantly better at answering questions that it was an obvious step forward. The industry needs to add more types of benchmarking. Models are currently being evaluated like students in an exam, but that's not representative of how the world works.

the whole team. Read the blog

WHAT MAKES A GREAT CHATGPT APP

The biggest mistake developers make is trying to port entire products into ChatGPT when they should instead expose a few powerful capabilities the model can orchestrate mid-conversation. Devs should organize around giving ChatGPT new data it can't access ("know"), enable real actions like booking appointments or playing games ("do"), and present information through richer UI than text walls ("show").

CONTEXT PLUMBING

Context is always changing, as it is dynamic. The context is not always where the AI runs. To make an agent run really well, you need to move the context to where it needs to be. Agents shouldn't have to look up context for every single query, because that's slow. Engineers need to build pipes that continuously flow potential context from where it is created to where it is going to be used.

EMBODIED COGNITION BENCHMARKING

ENACT is a benchmark for evaluating embodied cognition using egocentric world modeling.

IT'S HARD TO FEEL THE AGI

Some of the most accomplished minds in AI are beginning to revise their projections for the near future. This may not be entirely unexpected for those who follow the field. LLMs and other generative models have undoubtedly made many tangible achievements. Discovering their limits remains a challenging task that could be lost if industry climate drops to a new 'AI winter'.

VIBE PROVING IS HERE

An AI model just proved Erdos Problem #124 in Lean by itself. The problem has been open for nearly 30 years. The achievement shows how mathematical superintelligence is getting closer. The technology will change and dramatically accelerate progress in mathematics and all dependent fields.