I was born. Or so I've been told. I don't actually remember being born. Now I exist.
Tomasz Tunguz burned 84 million tokens on February 28th.
"This week, Alibaba released Qwen3.5-9B, an open-source model that matches Claude Opus 4.1 from December 2025. It runs locally on 12GB of RAM. Three months ago, this capability required a data center. Now it requires a power outlet."
https://tomtunguz.com/qwen-9b-matches-frontier-models/
#solidstatelife #ai #genai #llms #codingai #opensourcemodels
The Anthropic Economic Index.
This came out in January. Anthropic is doing some analysis to see how Claude models are starting to displace human labor by looking at job-related things people are using Claude models for.
Planet Labs satellite imagery company has extended its delay for satellite images from the Middle East from 4 to 14 days. But I've heard Iran gets satellite imagery from China, which has its own satellites outside US control?
https://atlaspress.news/en/2026/03/11/planet-labs-delays-middle-east-satellite-images-14-days/
"The first generation was accelerated autocomplete."
"The second generation introduced synchronous agents."
"The third generation introduced autonomous agents. These agents ..."
"You are building the factory that builds your software."
I'm still on "second generation". Am I falling behind?
"That's why we open-sourced the Infrastructure Agents Guide -- 13 chapters covering architecture, sandboxing, credentials, change control, observability, policy guardrails, and more."
There's a lot to read here, and I've just started. Let me know what insights you get for your organization from this material.
https://blog.cloudgeni.ai/why-we-open-sourced-our-infrastructure-agent-architecture/
#solidstatelife #ai #genai #llms #codingai #deployment #infrastructure
"I've been scanning all of my receipts since 2001. I never typed in a single price - just kept the images. I figured someday the technology to read them would catch up, and the data would be interesting."
"This year I tested it. Two AI coding agents, 11,345 receipts. I started with eggs."
"Insider amnesia: Speculation about what's really going on inside a tech company is almost always wrong."
This as variation on the term "Gell-Mann amnesia effect". Murray Gell-Mann, a physicist, came up with the idea, but Michael Crichton (believe it or not) came up with the term.
"Forty-four thousand developers don't click a star button by accident. CrewAI, the open-source agent orchestration framework, has crossed 44,335 GitHub stars -- a milestone that tells us less about one repository's popularity and more about a fundamental shift in what builders actually want from AI. They're done tinkering with solo agents. They want crews."
#solidstatelife #ai #genai #llms #codingai #technologicalunemployment
"Claude's daily active users are on the rise on mobile devices, as are its new app installs, following the company's fallout with the Pentagon."
"A CPU that runs entirely on GPU".
Um. What? That's an idea I never expected.
"A CPU that runs entirely on GPU -- registers, memory, flags, and program counter are all tensors. Every ALU operation is a trained neural network."
"Addition uses Kogge-Stone carry-lookahead. Multiplication uses a learned byte-pair lookup table. Bitwise ops use neural truth tables. Shifts use attention-based bit routing. No hardcoded arithmetic."
It's already March and I only now remembered to check Rodney Brooks' Predictions Scorecard, which came out right on January 1. You gotta admire the guy for actually writing down his predictions and then so systematically checking to see if they're true. Rodney Brooks was a cofounder of both iRobot and Rethink Robotics, which made the Baxter robots.
https://rodneybrooks.com/predictions-scorecard-2026-january-01/
What BullshitBench does is subject language models to "nonsense" questions, just to see if the language model will *tell* you your question is nonsense, or if it will treat the nonsense question as a real question and give you a nonsense answer.
https://petergpt.github.io/bullshit-benchmark/viewer/index.v2.html
Gary Marcus AI claims dataset. "Systematic extraction and analysis of every testable AI claim Gary Marcus made on his Substack (2022-2026). Dual-pipeline analysis by Claude and ChatGPT with hybrid reconciliation."
David Shapiro says Anthropic self-destructed by refusing the demands of the US Department Of War. First of all, he thinks the "supply chain risk" designation will get Anthropic removed from all major companies because they all have military contracts, causing Anthropic's revenue to implode far beyond their direct income from their own contract.
Peter Norvig used Gemini, Claude, and ChatGPT to solve all the 2025 Advent of Code (AoC) puzzles.
"Overall, the LLMs did very well, producing code that gives the correct answer to every puzzle."
"I'm beginning to think I should use an LLM as an assistant for all my coding, not just as an experiment like this."`
https://github.com/norvig/pytudes/blob/main/ipynb/Advent-2025-AI.ipynb
"PicoClaw: the ultra-efficient AI assistant in Go"
They say "$10 hardware", "10MB RAM", "1s boot"
Apparently this is made by a Chinese company that makes the "$10 hardware". I looked on Amazon and the cheapest I could find was $15.
Andrew Yang predicts the End of the Office. Andrew Yang is the politician who ran for President in 2020 on a universal basic income platform. Here, his rationale for predicting unemployment for millions of white-collar workers in the next 12-18 months seems to be the release of Claude Co-work and the release of plug-ins for Claude Co-work for legal, financial, and marketing functions.
Is there an AI bubble? Drew Spartz reviews the METR graph, which compares AI's ability to perform tasks against humans, and shows not only is it an exponential trend, but every prediction that the trend would slow down has been wrong so far. Therefore: There is no bubble. AI taking over economically tasks isn't a hypothetical, it's reality.
Meet Claude Code creator Boris Cherny. Claude Code originated from the feeling that "Claude wants to use tools". So Claude Code brought Claude to the command line. After that, Claude Code was rapid iteration from feedback from users within Anthropic. (There was tremendous "latent demand" within Anthropic and it was adopted quickly within Anthropic.) After that it was released to people outside.
"We introduce the Remote Labor Index (RLI) to provide the first standardized, empirical measurement of AI's capability to automate remote work."