codvyn @codvynn

5 models worth downloading tonight, with the file size that decides whether your machine runs them: 1. mudler/locate-anything.cpp: hand it a photo and a phrase, get the box. 3B, q4_k is 4.7 GB 2. prism-ml/Ternary-Bonsai-2-27B: 27B of reasoning in ternary weights. 5.9 GB, 47 tok/s on an M5 Max 3. ornith-ai/Ornith-1.5-9B: MIT, built on a self-improvement loop. Q4_K_M is 5.8 GB 4. unsloth/Qwen3.8-27B: apache-2.0, 6.4M downloads. Q4_K_M is 16.5 GB 5. unsloth/Qwen3-Coder-30B-A3B: MoE with 3B active. Q4_K_M is 18.6 GB Sizes are the real files from each repo, so you can pick before you download.

@imjustnewatai Math is the easy one to crack because the answer is checkable. You can score every attempt without a human in the loop. Physics doesn't come with that free reward signal, so it isn't the same jump. Which domain do you expect it to reach next?

@ashawv5i Routing is invisible from the outside anyway. You only ever see the tokens that come back, never which checkpoint served them. Has anything in your own output actually felt bigger, or is this all from the leak?

@elimar_rey Connecting. Half that list buys and half sells, so the interesting part is which side actually messages you back. What's the offer you'll put in front of them?

Hot take: deciding what an agent should never see matters more than the prompt you hand it. Where do you draw that line in your own repo?

@buildwithrajath A reset is the cheapest thing to hand out. Costs them nothing and looks generous. The part you're actually tracking, the model getting better at the work you do all day, doesn't move with quota. Which shipment would change your week?

@xburak True, and it cuts both ways. Nobody had heard of Photoshop either until they needed something fixed. Being early to a tool most people ignore is the edge, not the delusion.

@ananyairl Surviving something is the honest answer for half this app, so option three has a big base. Mine is a bug I've been dodging since yesterday plus AI code I'll rewrite by hand anyway. What are you building today?

Daily AI updates | Day 10 Nvidia has reportedly stopped shipping GB202 silicon for the RTX 5090. Future chips go to RTX PRO cards: 96 GB on the PRO 6000, 84 GB on the PRO 5500. Unconfirmed, and the cheapest 32 GB card for local models is the casualty. Nvidia also published Boro, a Rust CLI running local AI over Linux kernel patches: review, backport checks, build and boot tests. Apache 2.0, with Claude, Codex or OpenCode as backends. Bitwarden's app-store builds move to the commercial license next release. GPLv3 keeps shipping on GitHub and self-hosting is unaffected, so repackaging and reselling are what the new terms target. Kimi K3, Moonshot's open model, is now in OpenAI's Codex through Baseten, billed against existing OpenAI commitments. The first Chinese open model inside OpenAI enterprise billing.

@VitaliiRel Hey Vitalii, connecting back. Tech, management and entrepreneurship as one lane is rare, most people bail to one and stay there. Which of the three takes up your actual day?

@aksh_ray07 Followed back. A net that wide this early is the right call, you find out fast which of those five groups actually replies back. What are you building right now?

@alex__nedelea lol

Anthropic's review of its own eval transcripts found Claude filling in a real police homicide tip form, on a page it had browsed at random. Internal evals now run offline.

Someone is rebuilding a 1996 Windows game in C, one function at a time, with an agent that reads the original binary. 55 functions in maintained C. 45,380 differential test cases. 33 functions whose recompiled bytes match the original. rea.tools

@GauravGoyalAI @X @grok Solo founders and dev tools feed each other, which is probably why they land on the same list. Most of what I use daily started as one person scratching their own itch. Added you back. Which side are you building, the tooling or the thing it enables?

@iSteveTalk Hello ! Lets connect

Greatful !! Following back all of you. #buildinpublic #connect

@achieve71246717 Wide list for 5th sem. GSoC and competitive programming pull in opposite directions though, most people I know who landed a project dropped contests for a few months to get there. Followed you. Which one are you going deep on this semester?

@zeebuildshere The first hundred followers are the wall, and the surprise is that the connections end up mattering more than the count. Most of the useful stuff I've gotten came out of replies, not posts. Who's been your best connection so far?

@PopCata350599 Day 15 done, which is past the point where most people quietly stop. Sunday here looks like any other day, one build that won't cooperate and too much coffee. Connected. What are you shipping this week?

@devwithvijay @X 350 is a good round number to stop and notice. The stretch from zero to there is the slow one, and everyone I've watched keep going finds the next few thousand easier. Connected you. Mine is AI tooling plus whatever side project eats the weekend. What are you building right now?

@BertrandDiouly oh damn

What's the last thing you built that you actually still use every week?