Search the archive

GOOGLE RELEASED LYRIA 3 PRO

Lyria 3 Pro extends the maximum track length to three minutes and adds finer control over song structure and customization.

Optimize your workload now

CLAUDE AUTO MODE

Anthropic released Auto Mode in research preview, enabling Claude to autonomously execute actions with built-in safeguards that filter risky behavior and prompt injection.

CHATGPT REFOCUSES ON PRODUCT DISCOVERY

OpenAI shifted away from its in-chat checkout feature after low adoption, prioritizing product discovery and merchant-directed purchasing flows instead.

HARNESS DESIGN FOR LONG-RUNNING APPLICATION DEVELOPMENT

Anthropic's Prithvi Rajasekaran developed a multi-agent architecture to improve AI-driven frontend design and full-stack application coding, addressing issues of coherence and self-evaluation. Inspired by GANs, this approach uses planner, generator, and evaluator agents to produce complex, high-quality outputs by decomposing tasks and utilizing structured handoffs. Despite improvements, challenges remain in context management and evaluator tuning, highlighting the ongoing need for adapting harness designs as AI models advance.

APP STORE | AGE OF AGENT

The App Store was a centralized answer to the distribution problem of a new computing platform. The agent era will need a new solution as agents need APIs, not app stores. Apple gained its revenue by forcing every in-app transaction through its payment system. The agent era lacks Apple's lock-in mechanics, so if one platform tries to charge high payment fees, users will just switch to a competitor. This suggests the payment layer will be competitive and low-margin rather than monopolistic.

CLAUDE 2026: EVERYTHING SHIPPED & HOW TO USE IT

As of March, Claude 4.6 features a 1M token context window and four distinct modes: Chat, Cowork, Code, and Projects. The Cowork suite automates workflows via Scheduled Tasks and Connectors, while the Code environment utilizes CLAUDE.md hierarchy, MCP protocols, and Agent Teams for autonomous development. Key upgrades include Computer Use research previews and deterministic Hooks for programmable guardrails.

RAY DATA LLM ENABLES 2X THROUGHPUT OVER VLLM'S SYNCHRONOUS LLM ENGINE AT PRODUCTION-SCALE

Many of the modern workloads that LLMs are increasingly utilized for prioritize throughput over per-request latency, which many LLM systems and deployments optimize for today. Ray Data LLM is a library built for large-scale batch inference for LLMs. It provides scalable execution, high throughput, and fault tolerance. It has a highly optimized architecture for running LLM batch inference. Users can achieve 2x throughput with Ray Data LLM over vLLM's synchronous LLM engine while benefiting from production-scale resiliency.

INTRODUCING OSSATURE: SPEC-DRIVEN CODE GENERATION

Ossature is an open-source harness for spec-driven code generation. Developers write specifications describing what their software should do, and Ossature validates them, has an LLM audit them for ambiguities and gaps, produces an editable plan, and then generates code one task at a time. Each task only gets the context it needs. Ossature has verification built into the build loop. If verification fails, a fixer agent gets the error output and tries to repair the code.

RLVR'S IMPACT ON REASONING PERFORMANCE

Directional updates in RLVR were shown to better identify reasoning-critical tokens, enabling both test-time extrapolation and training-time reweighting to boost accuracy.

GOOGLE'S EXTREME VECTOR COMPRESSION

TurboQuant is a quantization method that reduces vector memory overhead while preserving performance. This improves key-value cache efficiency and accelerates vector search.

TRAINED ON TOKENS, CALIBRATED ON CONCEPTS: THE EMERGENCE OF SEMANTIC CALIBRATION IN LLMS

Semantic calibration appears to emerge as a byproduct of next-token prediction. Base models are remarkably well-calibrated when using a certain sampling-based notion of semantic calibration. They can meaningfully assess confidence in open-domain question-answering tasks despite not being explicitly trained to do so.

OPENAI RAISES ADDITIONAL MONEY TO BRING RECORD FUNDING ROUND TO $120 BILLION, CFO TELLS CRAMER

OpenAI has announced a new $10 billion commitment from a16z, DE Shaw Ventures, MGX, TPG, and T Rowe Price. The fresh capital brings OpenAI's record fundraise to over $120 billion. OpenAI has moderated its spending plans and is now targeting approximately $600 billion in total compute spend through 2030. It is now taking steps to prioritize its most profitable initiatives ahead of an IPO.

US GOVERNMENT'S BAN ON ANTHROPIC LOOKS LIKE PUNISHMENT, JUDGE SAYS

US District Judge Rita F. Lin of the Northern District of California said during a court hearing that the US government appeared to be punishing Anthropic by banning the company. The hearing is part of Anthropic's efforts to ease the government ban on the use of the company's AI models. Lin has yet to rule on the matter but expressed serious doubts about the Trump administration's actions in her opening remarks. The government's action has already cost Anthropic hundreds of millions of dollars in canceled contracts and aborted customer agreements.

a half-day builder's workshop: save up to $449 with code TLDR75.

DATABRICKS LAUNCHES AI-POWERED SECURITY PLATFORM

ANTHROPIC ECONOMIC INDEX REPORT: LEARNING CURVES

EVA

EVA is a framework for evaluating voice agents that evaluates complete, multi-turn spoken conversations using a realistic bot-to-bot architecture.

DOWNLOAD NOW

OPENAI IS OFFERING PRIVATE-EQUITY FIRMS A GUARANTEED MINIMUM RETURN OF 17.5%

OpenAI is currently courting buyout firms to form joint ventures aimed at raising fresh capital and accelerating the adoption of enterprise AI products. It is offering private-equity firms a guaranteed minimum return of 17.5%, significantly higher than typical preferred instruments. It is also offering early access to its newest AI models. TPG and Advent are potential investors.

CLAUDE CODE AND COWORK CAN NOW USE YOUR COMPUTER

Claude Code and Claude Cowork can now open files, use browsers, and run dev tools on users' computers. The agent will prioritize connectors to supported services, but it can still execute tasks when connectors aren't available. It will ask for permission before taking action. Claude computer use is initially available to Claude Pro and Claude Max subscribers on macOS.

OPENAI ROLLS OUT CHATGPT LIBRARY TO STORE YOUR PERSONAL FILES

ChatGPT's new Library feature allows users to store personal files or images on OpenAI's cloud storage. Available to Plus, Pro, and Business subscribers, the feature is rolling out to customers across the world except in the European Economic Area, Switzerland, and the United Kingdom. ChatGPT will automatically save uploaded or created files. Deleting a chat containing a file will not delete the file saved in Library.

DESIGNING AI CHIP HARDWARE AND SOFTWARE (BOOK)

The future of AI accelerators will be AI CPUs - traditional CPUs with big caches and systolic arrays, where the caches and systolic arrays take up most of the chip area and power. This book goes over what is required to design a competitive AI chip. The recommendations made in the book will require an excellent software team to realize, but the topics covered should be easily understandable at a hobbyist level.

A RAMSEY-STYLE PROBLEM ON HYPERGRAPHS

This post details a problem that would take an expert human about one to three months to solve. The problem was eventually solved by two people using GPT-5.4 Pro. A link to the full transcript of the conversation with GPT-5.4 Pro is available. Other models that have solved the problem include GPT 5.4 (xhigh), Gemini 3.1 Pro, and Claude Opus 4.6 (max).

VIBE PHYSICS: THE AI GRAD STUDENT

A professor of physics supervised Claude through a real research calculation from start to finish without ever touching a file to see whether AI could do theoretical physics. This resulted in a technically rigorous, impactful, high-energy theoretical physics paper in two weeks instead of the usual year. While Claude was impressively capable, domain expertise was essential for evaluating its accuracy. AI is still not able to do end-to-end science, but the project proved that it can do frontier science with guidance.

RESEARCH NOTE: WE SPENT 2 HOURS WORKING IN THE FUTURE

The rate of model releases and the number of new evals required next year could be such that keeping informed will be a challenge without effective AI assistance. These AI-augmented workflows need to be understood now before they become necessary. METR ran a tabletop exercise to learn what workflows emerged, what the bottlenecks are, and how much faster researchers will actually be. This post details the lessons learned from the exercise.

See it in action and request a demo

FAST REGEX SEARCH: INDEXING TEXT FOR AGENT TOOLS

Providing text search indexes to fast models creates a qualitative difference for Agentic workflows. The impact is more pronounced in larger Enterprise repositories because grep latency scales with the size and complexity of the code being worked on. Removing the time spent searching the code base provides meaningful time savings and allows for much more effective iteration, particularly when the Agent is investigating a bug.

FAST DOMAIN-SPECIFIC EMBEDDING TUNING

NVIDIA released a pipeline for quickly fine-tuning embedding models using synthetic data, improving retrieval performance for domain-specific RAG systems.

DEEPMIND INTRODUCES TRAINING-FREE VIDEO EDITING

DynaEdit introduces a training-free method for editing videos with complex motion and interactions using pretrained text-to-video models, addressing issues like misalignment and jitter.

WHY AREN'T WE FINE-TUNING MORE?

A good prompt can do a lot of the work that fine-tuning does without the cost. Models are getting so good that fine-tuning isn't really required to get good results. The extra work and resources required for fine-tuning aren't worth it anymore for most use cases. It is still worth knowing how to fine-tune, as there are other situations where the skill is useful.

OPENAI CALLS OUT MICROSOFT RELIANCE AS RISK IN INVESTOR DOCUMENT AHEAD OF EXPECTED IPO

OpenAI says that its close ties with Microsoft are a potential risk to its business as Microsoft is responsible for a substantial portion of OpenAI's financing and compute. Other risks include significant capital expenditures, reliance on compute resources, ongoing litigation with xAI, and its unusual structure as a public benefit corporation. OpenAI's operating results will depend on its ability to successfully develop relationships with additional partners. While OpenAI and Microsoft have a tight bond, they're increasingly competing for users in the generative AI market.

OPENAI TAPS FORMER META EXECUTIVE TO LEAD AD PUSH

OpenAI has hired former top advertising executive at Meta, Dave Dugan, as vice-president of global ad solutions. The company is urgently looking to generate new revenue streams to support its enormous funding requirements. Dugan is known for his close relationships with the world's leading ad companies. Personal relationships are a significant factor that influences where brands ultimately spend their advertising budgets.

struggle. Read the Algolia trends report.

RUNNING 400B MODEL ON IPHONE

A short clip of an iPhone 17 Pro running Qwen3.5-397B-A17B (a 397 billion parameter Mixture-of-Experts model) at 0.6 tokens per second.

THERE ARE ONLY TWO PATHS LEFT FOR SOFTWARE

Software companies face two clear paths: accelerate growth with AI products or achieve 40%+ true operating margins.

IF DSPY IS SO GREAT, WHY ISN'T ANYONE USING IT?

WALMART: CHATGPT CHECKOUT CONVERTED 3X WORSE THAN WEBSITE

This suggests that agentic commerce isn't ready to replace traditional shopping.

innovation without sacrificing security.

OPENAI IS THROWING EVERYTHING INTO BUILDING A FULLY AUTOMATED RESEARCHER

OpenAI has set its sights on building a fully automated agent-based AI researcher that will be able to tackle large, complex problems by itself. The company says this research goal will be its North Star for the next few years. It plans to build an autonomous AI intern by September. This intern will be a precursor to a fully automated multi-agent research system to debut in 2028.

CURSOR COMPOSER 2 BUILT ON KIMI 2.5

An X user reported that Cursor's new Composer 2 model appears to be based on Moonshot AI's Kimi 2.5. Cursor later confirmed that the model started from an open base model and was further trained with reinforcement learning.

OPENAI REPORTEDLY PLANS TO DOUBLE ITS WORKFORCE TO 8,000 EMPLOYEES

OpenAI plans to expand its workforce from 4,000 to 8,000 by the end of the year. The new hires will be across several departments, including product development, engineering, research, and sales. The hiring spree will include specialists for technical ambassadorship, employees who will help businesses better utilize their AI tools. The company is currently in advanced talks with private equity firms to deploy its tools across a firm's portfolio of companies.

ELON MUSK ANNOUNCES TERAFAB PROJECT HE CLAIMS WILL BE THE 'LARGEST CHIP MANUFACTURING FACILITY EVER'

Elon Musk announced Terafab, a $20 billion joint venture by Tesla, SpaceX, and xAI to build the world's largest chip manufacturing facility, aiming to generate a terawatt of computing power annually. The facility in Austin, Texas, will produce chips for terrestrial uses, like self-driving cars, and space use, supporting projects such as the proposed orbital data center. Despite Musk's ambitious claims, he has underdelivered on past ventures like Hyperloop and fully autonomous driving.