Search the archive

MICROSOFT ON TRACK TO SPEND $500 MILLION PER YEAR ON ANTHROPIC AI

Microsoft has been using Anthropic's AI to power more and more of its own products.

Try it in the API Playground or locally

INTRODUCING ANTHROPIC LABS

Anthropic Labs has expanded its team to incubate experimental products at the forefront of AI capabilities, leveraging rapid model advancements. Co-founder of Instagram, Mike Krieger, has joined Labs for collaborative product development. Ami Vora will lead product organization alongside CTO Rahul Patil to scale Claude experiences. Recent successes include Claude Code's billion-dollar growth and the Model Context Protocol's rise to industry standard.

GOOGLE UPGRADES VEO 3.1 WITH INGREDIENTS AND 4K UPSCALING

The Veo 3.1 update introduces new features designed for both casual creators and professionals. Users can now generate videos based on reference images. The tool now supports native vertical outputs tailored for mobile devices, and high-resolution upscales to 1080p and 4K are also available. The release is available now across multiple Google services and platforms.

GOOGLE TESTS GEMINI AUTO BROWSE TOOL FOR CHROME USERS

Google is testing an Auto Browse tool for Gemini that lets the AI manage tasks in Chrome. The tool will allow Gemini to autonomously browse the web, manage tabs, and directly interact with Chrome on users' behalf. Code references indicate that Auto Browse may be exclusive to the Gemini Ultra plan. Google is rolling out the feature incrementally to a limited audience in the US.

THE ALL-STAR CHINESE AI CONVERSATION OF 2026

Tsinghua University and Zhipu co-hosted AGI-Next, a summit for frontier AI, in Beijing on January 10. The event included a series of keynotes by notable figures in the Chinese AI industry. This post contains a translated transcript of the event, revealing a fascinating conversation about the AI landscape in China that covers the technical side, corporate dynamics, as well as the future envisioned by China's industry Titans. The event took an honest look at whether China's open-source leadership has actually narrowed the technology gap with the US, China's emerging AI-for-business paradigm, and what it will take for Chinese researchers to take riskier bets.

VIBE CODING WITHOUT SYSTEM DESIGN IS A TRAP

AI-assisted coding streamlines product creation but often sacrifices long-term design sustainability. It prioritizes immediate functionality, leading to hardcoded configurations, which are fragile for future updates. Successful product development requires experienced system design to avoid a "build-now, regret-later" scenario, transforming AI vibes into structured, sustainable solutions.

INTRODUCING THE AI CHIP SALES DATA EXPLORER

The AI Chip Sales data explorer uses financial reports, company disclosures, and more to estimate compute, power usage, and spending over time for a wide variety of AI chips. The cumulative global AI compute capacity has reached the equivalent of more than 15 million Nvidia H100 GPUs. The cost to purchase these chips has rapidly escalated to tens of billions of dollars per quarter, drawing over 10 GW of power per quarter. A link to the interactive data explorer is available.

Try now for free

ENGRAM: HOW DEEPSEEK ADDED A SECOND BRAIN TO THEIR LLM

DeepSeek's Engram architecture introduces a conditional memory system that uses lookup tables for common N-gram patterns, significantly reducing computational overhead in LLMs. This approach boosts performance on knowledge benchmarks and reasoning tasks by efficiently freeing up neural resources for complex reasoning. Engram's design leverages tokenizer compression, multi-head hashing, and context-aware gating, providing a marked improvement over MoE baselines without substantial memory impact.

GLM-IMAGE: AUTO-REGRESSIVE FOR DENSE-KNOWLEDGE AND HIGH-FIDELITY IMAGE GENERATION

GLM-Image is an open-source, industrial-grade discrete auto-regressive image generation model with a hybrid architecture that combines an auto-regressive module with a diffusion decoder. It excels in text-rendering and knowledge-intensive generation scenarios. The model performs especially well in tasks that require precise semantic and complex information expression. It supports a rich set of image-to-image tasks.

DEEP SEARCH AGENTS

CaRR improves reinforcement learning for deep search agents by replacing binary rewards with citation-based rubric feedback.

"HELLO, COMPUTER."

Vocal computing may finally become mainstream due to new AI advancements in LLMs overcoming past limitations of systems like Siri and Alexa. OpenAI's breakthroughs and its acquisition of a hardware startup hint at innovative voice-driven devices potentially moving beyond the smartphone. As tech companies refine and develop diverse hardware for AI, voice control is pivotal in advancing towards a more integrated and intuitive tech ecosystem.

THE DANGER OF BORROWED AI CERTAINTY

AI introduces "borrowed certainty" in software, creating technical debt by replacing proof and evidence with confident-seeming predictions. Developers often mistake AI's fluency for validity, risking decisions based on unearned certainty. Best practices include treating AI output as a hypothesis, demanding traceability, ensuring human understanding, and pairing AI with execution feedback.

18 PREDICTIONS FOR 2026

By 2026, advanced AI capabilities will emphasize autonomous agents and Generative UI, making UX the main business differentiator.

GOOGLE'S MEDGEMMA 1.5 FOR HEALTH AI

Google expanded its Health AI Developer Foundations with MedGemma 1.5 for medical image interpretation and MedASR for speech-to-text.

CHINA'S ZHIPU ROLLS OUT NEW AI MODEL TRAINED ON HUAWEI CHIPS

Zhipu's GLM-Image model is the first in China to be trained using Huawei's Ascend AI chips and Kunpeng processors.

CLAUDE AGENT SDK TECHNICAL SPECIFICATION

This page contains the technical specification for the Claude Agent SDK, which provides programmatic access to Claude Code's agentic capabilities.

Try Sentry free →

APPLE'S NEW GOOGLE GEMINI DEAL SOUNDS BIGGER, BETTER THAN EXPECTED

Apple and Google are teaming up to power Siri with custom AI models. Google's Gemini models will also be used for more Apple Intelligence features. Apple's current privacy standards will be upheld as part of the deal. No Apple user data will be accessible to Google.

META LAUNCHES AI INFRASTRUCTURE PUSH WITH META COMPUTE

Meta Compute is an initiative to scale Meta's AI infrastructure. It includes plans to build tens of gigawatts of energy capacity over the next decade. CEO Mark Zuckerberg positions infrastructure as a strategic edge. Key leaders have been appointed to drive technical architecture, long-term capacity strategy, and government partnerships.

ANTHROPIC INTRODUCED COWORK

Claude Cowork is a simpler version of Claude Code built into the Claude Desktop app. Users can assign a folder for Claude to access and guide it via chat, enabling agent-like workflows without complex setup. The tool is in research preview and currently limited to Max subscribers.

DIGITAL RED QUEEN: ADVERSARIAL PROGRAM EVOLUTION IN CORE WAR WITH LLMS

Core War is a competitive programming game in which battle programs fight for dominance inside a virtual computer. This study explores what happens when large language models drive an adversarial evolutionary arms race where programs continuously adapt to defeat a growing history of opponents rather than a static benchmark. The dynamic adversarial process led to the emergence of increasingly general strategies. The Core War sandbox offers a safe and controlled environment for analyzing how AI agents might evolve in real-world adversarial settings such as cybersecurity.

THE AI DATA CENTER DEALS THAT NO ONE CAN VERIFY

The AI industry has announced more than half a trillion dollars of infrastructure commitments over the past year. The headline figures are being priced as if they represent binding, time-certain capex commitments, but the disclosed language and missing definitions make many of them look more like optionality dressed up as commitment. The refusal to standardize terms, define units, and allow any market-based verification suggests the participants prefer a world where scrutiny is expensive. The market is being asked to trust without the tools to verify.

WHEN AI WRITES ALMOST ALL CODE, WHAT HAPPENS TO SOFTWARE ENGINEERING?

AI models like Opus 4.5 and GPT-5.2 now write most software code, prompting a shift in software engineering practices. This reduces the value of language expertise and routine coding tasks while increasing the demand for tech leads who are product-minded. Although AI can handle more of the coding workload, engineers will need to oversee complex tasks, emphasizing a hybrid skill set in both product management and technical expertise.

Get the checklist.

MCP CLI

mcp-cli is a lightweight command-line tool that streamlines communication with MCP servers, helping AI agents interact with tools and APIs more efficiently while minimizing context window bloat.

MAP-AUGMENTED AGENT

Alibaba introduces a map-augmented agent for image geolocalization, embedding it in a map-guided loop that combines reinforcement learning and parallel test-time inference to improve prediction accuracy.

AGENT-BROWSER

agent-browser is a headless browser automation CLI for AI agents. It allows agents to take control of browsers, take screenshots, and extract information from pages. agent-browser can run multiple isolated browser instances. A headed mode is available for debugging.

DEEPSEEK FOUNDER LIANG'S FUNDS SURGE 57% AS CHINA QUANTS BOOM

Zhejiang High-Flyer Asset Management posted an average return of 56.6% across its funds in 2025. It was the second-best performer among Chinese quant funds that manage more than 10 billion yuan. High-Flyer has become a cash cow for Liang Wenfeng, DeepSeek's founder, who still holds a majority of the asset manager. DeepSeek's research was funded by High-Flyer's R&D budget. The fund is estimated to have generated revenues of more than $700 million last year.

CLAUDE EXPANDS INTO HEALTHCARE AND LIFE SCIENCES

100X A BUSINESS WITH AI

Building effective AI agents requires strong context, memory, and workflow-aligned architecture. They multiply human output by handling routine tasks while humans focus on judgment, and should resolve issues directly rather than just reporting them. Fast deployment and continuous iteration create compounding value, making custom agents far more effective than generic AI SaaS.

alternative? Get a free copy of the report from Miro

META HAS APPOINTED A NEW PRESIDENT

Meta has appointed Dina Powell McCormick as President and Vice Chairman, expanding her role beyond the board to help lead strategy, infrastructure, and capital investment.

OPENAI ACQUIRES HEALTH-CARE TECHNOLOGY STARTUP TORCH FOR $60 MILLION

Torch was building a 'unified medical memory' for AI that brought patient data stored across different vendors and formats into one place.

HEARING FRESH DETAIL ON OPENAI'S "TO-GO" HARDWARE PROJECT FROM LAST REPORT

OpenAI's hardware project is a special audio product designed to replace the AirPods.

THE POWER USER'S GUIDE TO CODEX

In this post, Alexander Embiricos, the product lead for Codex and OpenAI, shares how AI coding agents are actually used in production and breaks down the workflows that make Codex effective at scale.

THE GREAT FILTER (OR WHY HIGH PERFORMANCE STILL ELUDES MOST DEV TEAMS, EVEN WITH AI)

Companies that manage to progress to working with highly iterative, frictionless methods need to invest 20% to 25% of their entire development budget in building and maintaining that capability, and building that capability takes years.

Protect your app today →

ANTHROPIC CRACKS DOWN ON UNAUTHORIZED CLAUDE USAGE BY THIRD-PARTY HARNESSES AND RIVALS

Anthropic has implemented strict new technical safeguards to prevent third-party applications from spoofing Claude Code to access more favorable pricing and limits. Ensuring model integrity now requires routing all automated agents through the official Commercial API or the Claude Code client. The company has also restricted the usage of its AI models by rival labs. Security directors should audit internal tool chains to ensure they aren't violating commercial terms and that all automated workflows are authenticated via proper enterprise keys.

META UNVEILS SWEEPING NUCLEAR-POWER PLAN TO FUEL ITS AI AMBITIONS

Meta has unveiled a series of agreements that will make it an anchor customer for new and existing nuclear power in the US. The company will back new reactor projects with TerraPower and Oklo and has struck a deal with the power producer Vistra to purchase and expand the generation output of three existing nuclear plants. The first new reactors are expected to be delivered as early as 2030 and 2032. Financial details of the deal have not been disclosed.

BEST PRACTICES FOR CODING WITH AGENTS

Tips for maximizing productivity with coding agents, including how to structure problems, guide multi-file changes, and ensure agents iterate effectively on code until tests pass.

GLOBAL AI COMPUTING CAPACITY IS DOUBLING EVERY 7 MONTHS

Total available capacity from AI chips across all major designers has grown approximately 3.3 times per year since 2022. This has enabled larger-scale model development and consumer adoption. Nvidia AI chips currently account for over 60% of total compute. Google and Amazon make up much of the remainder.

USE MULTIPLE MODELS

Problems with AI models can often be solved by passing the same query to a peer model. This suggests that models are actually quite close to being able to solve many tasks, but they're just not there yet. For switching to regularly solve tasks, each model must have a fairly high probability of success. The next generation of models will be even more capable.