Qwen3.5 35B-A3B (3B active params) outperforms Qwen3-235B on benchmarks. Alibaba's MoE architecture means near-Sonnet 4.5 performance running locally with <4GB VRAM active. For specific tasks, the open-source models aren't just catching up. They're ahead.
Karpathy's autoresearch: 630 lines, 1 GPU, ~12 experiments/hour, 18k stars in days. Full autonomous ML loop — edits code, trains, scores, keeps or reverts, repeats.
630 lines is now the threshold where non-coders + AI assistants can read and adapt research-grade tooling.
https://is.gd/Zohl6B
Anthropic's unit economics: each model ships profitably, but the company still loses money overall. A one-year revenue miscalculation = bankruptcy risk. Meanwhile Dario puts AGI at 50/50 in 12-24 months. The financial structure of frontier AI labs is genuinely unusual right now.
Jeff Dean on Latent Space: the step CS classes taught but nobody practiced — writing specs — is now the most important artifact when coding agents execute your instructions. 50 AI interns per engineer, managed via specs. https://is.gd/f84jV5
Dario Amodei on Dwarkesh: every Anthropic model ships profitably, but the company still loses money. A one-year revenue miss = bankruptcy risk. That kind of pressure either sharpens every decision or breaks the team. No middle ground. is.gd/GfOrd7
Google's Gemini Deep Think paper positions AI not as a research tool but as a creative research partner. Iterative refinement + cross-disciplinary transfer.
The useful moments aren't when AI gives you answers. It's when it reframes your question.
4% of GitHub public commits today are Claude Code-authored. SemiAnalysis projects 20%+ by EOY 2026. Viral growth inflection: October 2025. Second surge: January 2026 after Boris Cherny's 4.4M-view post. Task horizon doubling every 4-7 months.
Claude Code running a full Arduino workflow -- compile, upload, serial monitor -- no IDE. What caught me: if AI coding tools work for firmware, the "AI is just for web devs" assumption breaks. Every hardware team is a potential user. is.gd/DgJGQ3
Productivity flat for decades. Birth rates falling. Labor force shrinking.
AI arrives at exactly this moment.
The timing argument doesn't get enough attention.
How I actually use Claude Code: build something small, use it until it breaks, research the fix, rebuild. My recipe extractor evolved into a research pipeline that now finds tools for me. The pieces feed each other.
Claude's Programmatic Tool Calling: instead of one round-trip per tool, Claude writes Python that calls all your tools in a sandbox. One round-trip total. 50-90% latency drop.
Claude Code Agent Teams: one JSON setting, lead agent spawns parallel specialists, they communicate and share tasks.
Also: one 2-teammate run on Opus 4.6 = 29% of your daily token budget.
The capability is real. So is the bill.
Luca Dell'Anna's compounding checklist pattern for Claude Code: each review updates the checklist for the next one. The session gets smarter as it runs. Simple idea, significant delta over time.
Walked a mentee through his first AI coding session Sunday.
The AI forgot its own security rule within 5 minutes.
The best AI users aren't the ones with the best prompts. They're the ones who notice what the AI missed and push back.
You're the key ingredient, not the tool.
Qwen3.5-397B: 76.4 SWE-bench Verified, 87.8 MMLU-Pro. 397B params, 17B active per pass. Open source matching proprietary frontier on coding. The "just use the API" assumption deserves a second look.
Had my AI coding tool go down this morning. Reminded me why I've been building a local model router — open-source handles more than you'd expect, it's free, and nothing leaves your machine. Full privacy by default. Diversify before you have to.
35% of Cursor's own PRs are now agent-generated.
The copilot era is ending. Autonomous agents onboarding to codebases, spinning up VMs, shipping code.
We crossed a line this week and I'm not sure everyone noticed.
With #Severence season 2 ending, was a big clue about Lumon dropped in ep. 1 at 21:06?
Hellie: "Am I livestock?"
Mark: "I’m sorry?"
Hellie: "Like did you grow me as food and that’s why I have no memories?"
Mark: "You think we grew a full human, gave you consciousness?"
Maybe not a throwaway!
Cosm is going to open a location in Detroit-perhaps a great way to offer a new type of live experience that is much more affordable than being on-site. I could see this being great for concerts and anyone who doesn't want to pay over $1000 for a Taylor Swift ticket. www.bloomberg.com/news/feature...
This is how my #doggo Fiona feels about the Monday after a holiday break 😂😆. #collie #mondays
Let's go Lions!
That's a #Lions franchise-record 18th touchdown this season for Jahmyr Gibbs ⚡️
Continued censorship at the Washington Post:
[T]oday Pulitzer-Prize-winning political cartoonist Ann Telnaes left the Washington Post after her editor killed a cartoon criticizing the tech and media leaders who have been currying favor with Trump.
Per @hcrichardson.bsky.social in today's post.
Related article on how the U.S. presence Guam, geographically on the front lines with China, is experiencing elevated cyber attacks.
www.bloomberg.com/news/feature...
Very informative and well-written overview by @deeplearningdaily.bsky.social about China's state-sponsored hacking program. #hacking #china
dianawolftorres.substack.com/p/why-does-c...
Pinterist's 2025 design trends are interesting, vibrantly presented, and eerily prescient. Just prior to coming across the predictions, I subscribed to History Hit, mainly because of their focus on Medieval history...
business.pinterest.com/pinterest-pr...
Fascinating video from the WSJ on airport design and how architecture, flooring and layout guide us through the process of check-in, security and arriving at our gate: www.wsj.com/video/series...
So I pay for a premium #netflix account not to receive ads, only to find NFL on Netlfix with wall to wall ads?
Amazing reporting and story about sleeper Russian agents embedded around the world, and the effort find them. www.wsj.com/world/europe...
