a16z speedrun is one of the biggest accelerators in the world. We submitted our application a bit back on the recommendation of one of our friends. Our EdubaWare software has been in closed testing for a bit (get ready, next week thousands of you will get access to it and see our changes) We received and Email yesterday saying they want to hear more from us and to send them a video in under three minutes that explains: - Your backgrounds and greatest accomplishments, professional or personal - What makes you exceptional, and why you're the team to win this market - What you're building and what problem it solves - Why this is a big, underserved opportunity - What progress you've made so far  Now fitting all that into one video is challenging, but we did our best. I hope this gives you a view into the other side of my life which is my company Eduba!

Found this cool YouTube shorts that seems to Make any website into APi. Let me know your thoughts. https://youtube.com/shorts/RYqPb9e0RXg?is=m1JTei4kJ6ZjF9b5

I kept reanswering the same question at midnight: which model should this conversation run on? So I built the thing that answers it for me. https://github.com/dvogeldev/model-picker A skill for any agent that reads skills. Ask it "which model for X?" and it pulls the live Nous Portal and OpenRouter catalogs, ranks every model on three axes, and gives you one pick plus one cheaper alternate, with current prices and promotions. A promoted flagship can beat an unpromoted one at equal fit. Free tiers count as real picks. Every model gets classified: capability tier (frontier, flagship, value, small), deployment (API, or open weights you can self-host), and use case (long context, hard reasoning, agentic, coding). In pi you get /set-model coding. One command. The standing coding model for the session, thinking level included. A daily timer re-ingests both catalogs at 6:30 AM, so the standing map never goes stale. You stop defaulting to whatever you used last week and start choosing on purpose. Install is two lines. MIT. Your API keys stay in your own password manager. No key lands in a config file. It works in pi, opencode, hermes, and grok. Same skill in every harness. The standing map is a JSON file in your home directory: edit it, override it. Your call which model does which work. Try it, then ask your agent "which model should handle this?" Tell me below what it picked and whether you agreed. That's the part I read the comments for. ========================================================================= Why Nous Portal as the provider Fair question, since the skill also reads OpenRouter. portal.nousresearch.com is Nous Research, the same lab behind Hermes Agent. Three reasons it's my default surface. 1. The catalog is OpenRouter underneath. Same schema, same model IDs, same naming. That's why one picker script parses both catalogs with one parser. 2. It's built for the Hermes environment. Hermes is Nous's own model line, and the same lab runs the inference. One account, one key, one bill. hermes setup --portal and your agent is on it.

everyone's chasing tok/s with 4-bit quants right now. i wanted real numbers on what that costs, so i measured it instead of eyeballing a few prompts. the setup: Qwen3.6-35B-A3B (the MoE). i kept everything the same and only changed the precision of the expert weights, which are most of the model. i compared each version against the 8-bit model on three things: - perplexity on normal text and on code - how often the model picks a different next token than the 8-bit version - the same thing on real agentic tool-calling transcripts from my own coding agent, not benchmark prompts the results (vs 8-bit): - 4-bit: perplexity +1.73%. same next token as 8-bit: 93.9% on text, 90.4% on agentic work - 4-bit mix (gate/up 4-bit, down 8-bit): perplexity +1.04%. same next token: 95.2% text, 92.0% agentic - 6-bit: perplexity +0.05%. same next token: 97.6% text, 95.8% agentic - 8-bit: perplexity +0.00%. same next token: 98.7% text, 96.4% agentic (the 8-bit row isn't 100% because it's a re-encoding of the original 8-bit weights into the card's native 8-bit format, so it's the noise floor.) what this actually means: - a perplexity hit of +1.7% sounds tiny. but on agentic work the 4-bit model disagrees with 8-bit on about 1 in 10 tokens. in a tool call, one wrong token can be a broken argument or a malformed call, and the agent goes off the rails from there. that's the "works great for a few prompts, falls apart on real work" thing people keep running into. - 6-bit is close to 8-bit. it's a solid fallback when memory forces your hand. - 8-bit is basically lossless. on a smaller model i measured 8-bit at only +0.175% perplexity vs full 16-bit. - in practice i saw the same thing on my bank: the 4-bit version of a 9B coding model was broken for tool calling, so i run 6-bit there. the part people miss: on my hardware, 4-bit isn't even faster. on the Tenstorrent Blackhole card i measured the 4-bit and 8-bit matrix math at the same speed (532.6 vs 534.1 GFLOP/s). you give up quality and get nothing back. and with exact speculative decoding i'm getting ~50 tok/s on a 9B model at full 8-bit, with output identical to normal decoding.

High Tea 17 (10 Oct) and High Tea 18 (24 Oct) are both up. Each one comes with the full recording and a zip package built for the session. Drop the folder on your machine, open it in Claude or whatever you use, and say what you’re working on. It routes you to the right file, and every strategy file ends with a prompt to run on your own work. ━━━━━━━━━━━━━━━━━━━━ 📼 HIGH TEA 17: Outcome First (1:40) Get the outcome, then refine. • Why you run out of tokens, and how to fix it: say where your data lives, and push repeat jobs to a script or the cheapest model • Using Claude to replace Claude • A mounted Google Drive as the simplest shared folder for a team • HTML dashboards that live in your folder, and GitHub Pages in two clicks • Cutting big tender and RFP files down to the paragraphs that matter, with Jev as a filter • Pricing against software a company already pays for 📦 In the zip: 8 strategy files, plus 4 kits you can use today: a token diet worksheet, a shared Drive starter, an HTML dashboards folder, and a tender intake system with working scripts. ━━━━━━━━━━━━━━━━━━━━ 📼 HIGH TEA 18: The Next Yes (1:21) Get paid before the product exists. • Audit, pilot, contract, product: each stage pays for the next • Grants that take no equity, and using one to pay the client you want • How clients find you • Selling what you already know as the audit • Proving AI’s return when it seems to vanish across a team • One basket, and a plan B that gets better every year • Setting Claude up on a client’s machine 📦 In the zip: 9 strategy files, plus a pricing ladder, a grant-hunt kit and an ROI worksheet. ━━━━━━━━━━━━━━━━━━━━ Both zips also hold the full transcript, the member Q&A and a term sheet. Grab them in the vault. Questions go in the Discord. 💬