Search the archive

THE GENTLE OBSOLESCENCE

AI is becoming more and more capable, and this may make humans obsolete.

See the difference for yourself.

GPT-5.3-CODEX

OpenAI's GPT‑5.3‑Codex is a faster agentic coding model that combines GPT‑5.2‑Codex's coding performance with GPT‑5.2's reasoning and professional knowledge.

CLAUDE OPUS 4.6

Anthropic has released Claude Opus 4.6, an upgraded flagship model with stronger agentic coding, longer task persistence, and improved performance in large codebases. The model introduced a 1M-token context window in beta and achieved state-of-the-art results across several reasoning, coding, and economically valuable work evaluations.

MIND THE GAAP AGAIN

OpenAI desperately needs cash because it lacks economies of scale. Microsoft is shipping billions to OpenAI because it would be disastrous for the company if OpenAI went under. $10 billion of OpenAI's revenue may have been effectively a subsidy from Microsoft. This means that OpenAI was charging subscribers around 25% of what they were costing. OpenAI would have significantly fewer subscribers with an $80 per month subscription.

I SPENT $10,000 TO AUTOMATE MY RESEARCH AT OPENAI WITH CODEX

Many people drastically underestimate what Codex can do. This post shares a simple setup that is routinely allocated hundreds of millions of tokens. The author spent $10,000 on API costs this month. The post also shares some killer use-cases for Codex, and reflects on how organizations might become significantly more efficient in the near future.

move agents from POC to production. Register here to join.

VERCEL V0 ADDS PRODUCTION-GRADE FEATURES

The new v0 expands from demo tooling into a full production-ready platform with enterprise-grade security, integrations, and support for building and deploying real software with AI.

BUILDING A C COMPILER WITH A TEAM OF PARALLEL CLAUDES

Multiple Claude instances worked in parallel to build a Rust-based C compiler capable of compiling the Linux 6.9 kernel. This approach reduced human supervision and explored autonomous LLM agent capabilities via stress tests and a $20,000, 2,000-session effort. The compiler, while functional, has limitations like inefficient code and reliance on GCC for some features.

ON RECURSIVE SELF-IMPROVEMENT (PART I)

The automation of AI research and engineering will begin in earnest this year. This will create at least a step-change improvement in AI progress. It could also change the dynamics of AI competition, alter AI geopolitics, and much more. There will be more AI progress this year than in all of the years that have come before, and that's a conservative forecast.

GOOGLE'S 52X AI GROWTH

Google's first-party models now process more than 10 billion tokens per minute via direct API usage. This is a 52x increase year-over-year. Google is committing capital at an unprecedented scale to fuel this growth. The level of investment signals the company's conviction that the demand for tokens is only just beginning.

Perplexity, NVIDIA. Start with 250 free credits/month

OPENAI INTRODUCED FRONTIER

OpenAI Frontier is an enterprise platform for building, deploying, and managing AI agents across organizations.

BRAIN DUMPS AS A LITERARY FORM

Turning technology proposals and architectural initiatives into shared transcripts captures the entire reasoning process, not just conclusions.

MY AI ADOPTION JOURNEY

The article discusses lessons from adopting AI tools in software development, emphasizing efficiency.

ELEVENLABS CEO: VOICE IS THE NEXT INTERFACE FOR AI

ElevenLabs CEO Mati Staniszewski claims voice is poised to become the primary interface for AI, surpassing text and screens.

SEE WHY DEVS LOVE GREPTILE

✅ Trusted by engineering teams at NVIDIA, SCALE AI, AND BREX.

✅ Now integrated with Claude Code: install via /plugin.

GOOGLE'S GEMINI APP HAS SURPASSED 750M MONTHLY ACTIVE USERS

Google Gemini has surpassed 750 million monthly active users. The AI chatbot has quickly become a prominent player in the AI space. It had 650 million monthly active users last quarter. In comparison, MetaAI has reported nearly 500 million monthly users, and ChatGPT had an estimated 810 million monthly active users in late 2025.

META SAYS ITS NEXT-GEN AI MODEL, AVOCADO, IS ITS “MOST CAPABLE PRE-TRAINED BASE MODEL TO DATE”

Meta's Avocado model reportedly outperforms top open-source base models. Meta is seeing 10x compute efficiency wins with the model on text-related tasks compared to Maverick and over 100x efficiency gains over Behemoth, a version of Llama 4 that the company delayed last year. The model has now finished pretraining.

INSIDE THE CODEX APP SERVER

OpenAI detailed the architecture of the Codex App Server, a bidirectional JSON-RPC API layer powering Codex across platforms, offering tips for integrating coding agents into real workflows.

KIMI K2.5

Kimi K2.5 is a solid model and exceeds expectations. It is likely now the leading open weights model. The model is excellent given its price. It is an excellent choice for developers who can't afford or don't want to pay for Opus 4.5 and have to go with something cheaper to run their OpenClaw.

AI IS FINALLY EATING SOFTWARE'S TOTAL MARKET: HERE'S WHAT'S NEXT

AI's rise is shrinking the software market, causing unpredictability for SaaS companies. Key players like Salesforce and SAP are integrating AI functionality into popular platforms to adapt, while weaker firms face potential obsolescence. Success hinges on owning intent gateways and effectively leveraging AI for outcomes, positioning few companies to thrive amid industry-wide disruption.

credits/month

MISTRAL INTRODUCES VOXTRAL TRANSCRIBE 2

Mistral's Voxtral Transcribe 2 is a pair of next-gen speech-to-text models, including a real-time version with open weights that delivers sub-200ms latency and low-cost, high-accuracy transcription across 13 languages.

CLAUDE AND CODEX ARE NOW AVAILABLE IN PUBLIC PREVIEW ON GITHUB

Claude by Anthropic and OpenAI Codex are now available for Copilot Pro+ and Enterprise customers on GitHub in public preview. Users can start sessions and assign tasks to these coding agents from the web, mobile app, or VS Code without additional subscriptions. All agent actions, like drafting pull requests and task prioritization, can be managed through GitHub's existing infrastructure.

WINDSURF TAB V2: 25-75% MORE ACCEPTED CODE WITH VARIABLE AGGRESSION

Windsurf's Tab feature was very well received, but it wasn't well maintained after launch. The team has now significantly improved the underlying model and context/data engineering pipeline. This has led to direct Pareto improvements across all of Windsurf's metrics, with an average 54% increase in characters per predict. Tab now uses variable aggression to tailor the experience to each user's preferences.

OWNING A $5M DATA CENTER

comma has been running its own data center for years. All of the company's model training, metrics, and data live inside its own data center in its own office. This post describes how comma's data center works. It is aimed at inspiring other companies to build their own data centers too.

WHY NVIDIA BUILDS OPEN MODELS WITH BRYAN CATANZARO

Nvidia's focus on building open models, spearheaded by efforts like the Nemotron 3 Nano release, positions it uniquely to enhance both its product offerings and the AI ecosystem. Open model development aids Nvidia in understanding compute-heavy AI workloads, which informs their hardware and software design, while fostering greater community collaboration. Nvidia distinguishes itself by emphasizing open data and models to drive AI infrastructure, benefiting businesses without AI acting as an overarching monopolistic event.

C.AI LABS

c.ai labs is a space for rapidly testing creative AI entertainment ideas.

CLAUDE WILL REMAIN AD-FREE

Anthropic announced that Claude will not show ads or include sponsored content in conversations, aiming to preserve trust and integrity in AI interactions, especially in sensitive or complex contexts.

NEMOTRON COLEMBED V2 FOR MULTIMODAL RETRIEVAL

ColEmbed V2 is a family of late-interaction embedding models that achieve state-of-the-art results on ViDoRe benchmarks by improving multimodal document understanding and retrieval.

MANUS EVOLVE

Manus Evolve, powered by Claude Code and Codex CLI, integrates with tools like GitHub and Slack using the EVOLVE SDK.

THE OTHER LEVERAGE IN SOFTWARE & AI

Funds are showing distress - any wobble in expectations will be amplified by borrowing.

Save your spot

secure AI across your organization. Topics include

* AI AGENTS UNCOVERED: AWS explains how AI agents transform workflow and create value across enterprises. * AI AGENTS UNDER ATTACK: Which security blindspots are introduced by agents in an emerging threat landscape. * SECURING AI AGENTS NOW: Learn how real-world security teams are scaling AI security measures.

CLAUDE IN XCODE

Xcode 26.3 introduces native support for Claude Agent SDK, enabling full agentic capabilities like subagents, background tasks, and plugins within Apple's IDE. This brings Claude Code's functionality directly into developers' workflows.

INTEL IS MOVING INTO GPUS AND HAS HIRED A CHIEF ARCHITECT, CEO LIP-BU TAN SAYS

Intel is expanding into the GPU market by hiring a new chief architect, as announced by CEO Lip-Bu Tan. GPUs are crucial for AI infrastructure and are in high demand, driven by LLMs. Despite recent challenges, Intel's stock has rallied, buoyed by optimism in its foundry business and significant investments.

OPENAI POACHED ITS NEW SAFETY EXECUTIVE FROM ANTHROPIC

OpenAI hired Dylan Scandinaro from Anthropic as its new "head of preparedness." Scandinaro emphasized the rapid advancement of AI and the associated risks. He stressed the urgency of addressing these challenges.

OPEN SOURCE AI ECOSYSTEM

This blog explores open-source AI trajectory since the "DeepSeek Moment," highlighting long-term strategies by major organizations and forecasting sustained momentum through open artifact sharing and deployment-first design.

DEEP DIVE: HOW CLAUDE CODE'S /INSIGHTS COMMAND WORKS

The /insights command in Claude Code generates an HTML report that analyzes usage patterns across a user's Claude Code sessions. The tool is designed to help users understand how they interact with Claude, what's working well, where friction occurs, and how to improve workflows. This post takes a look at how the tool works. For better insights, use Claude Code regularly, give feedback, don't filter yourself, and check in monthly to see how patterns evolve.

EXPENSIVELY QUADRATIC: THE LLM AGENT COST CURVE

It doesn't take much to start spending a significant amount of tokens on context. This problem compounds as conversations get longer and only gets exponentially more important with agent and subagent workflows. Agent developers need to keep this problem in mind when developing AI tools. This post provides an interactive tool that shows how quickly context size can take up a large percentage of a cache read.