New top story on Hacker News: Show HN: Open-source model routing for coding agents at Astra-level performance

Show HN: Open-source model routing for coding agents at Astra-level performance
6 by adchurch | 0 comments on Hacker News.
A few months ago we started building a model router for coding agents because we thought we could outperform any single model with an ensemble approach. Recently we’ve achieved that milestone and I want to talk about how we did it. First of all, a quick explanation: the Weave Router ( https://ift.tt/HfW7y2I ) plugs into any coding agent (e.g. Claude Code or Codex) and intelligently switches between LLMs. So, for example, Astra handles tricky debugging or complex system design tasks, and Deepseek v4 Flash handles simple frontend updates. What we’re announcing today is our new routing model, which we’re calling Weave Router 2.0. We benchmarked 2.0 against GPT-6 Astra on Terminal Bench 4.0 and SWE Atlas. On both benchmarks, the router had equivalent pass rates. On Terminal Bench, the router hit 52% of Astra’s cost, and completed tasks 2.2x faster. On SWE Atlas, the router cost 54% as much as Astra and ran 2.5x faster. (Full results on our website at https://ift.tt/WNLDiIJ !) It turns out training a model to route effectively - taking into consideration model capabilities, costs, cache awareness, and more - is a really hard problem! I want to talk about three ways we were able to improve so much over the last few months: 1) a new architecture, 2) larger training data set size, and 3) smarter cache-eviction impact calculation. 1) a new architecture. Our initial approach used an RL model without many priors. While RL is still an important part of the story, the cost of fully exploring the space of routing decisions is very high, so we’ve taken some shortcuts that have significantly improved performance. Consider how large the search space for the routing problem is. Take a typical coding agent session, with ~100 agent turns (i.e. 100 LLM API calls). Technically there are 100 chances to select a model. If we assume a roster of ~10 models (of course there are lots more but we can remove any that are Pareto dominated), then there are 10^100 possible paths through that session. We simply cannot explore all of them! So that's why clever tricks to shrink this space are so important. In particular: we trained a hidden Markov model to trace the session state, then a classifier maps the session to one of a few buckets of similar models. Using the HMM allows us to evaluate not just where a session is currently, but how it got there . We've gotten significantly better performance on bucket selection by incorporating that information - we believe this is because two sessions that might look quite similar to a naive classifier are much better distinguished by this HMM approach. Using this HMM + classifier to select a bucket first significantly shrinks the space to explore, by throwing out most models that could not reasonably serve the given session. This rearchitecture was the single biggest performance unlock! 2) larger training data set size (much less technically interesting but still an important part of the story). By using frontier LLMs to help us label a larger and more diverse set of coding agent sessions, we were able to bootstrap the two models discussed in 1) to a better state, while also providing even richer reward signals for RL. 3) smarter cache-eviction impact calculation. One of the hardest parts of routing well (if you care about saving money) is using the model caches intelligently. We built a subsystem that can calculate the expected value of switching models (and thus paying a high one-time cost to fill up a different cache) much more accurately, helping us avoid costly and unnecessary switches in more cases, while still switching when the benefit outweighs the cost. This is where most of our improvement on cost has come from. We still have a lot of room to continue to improve (we won’t rest until we’re consistently beating Astra/Fable, not just tying!) but matching frontier model performance was a huge milestone for our routing model, and in my opinion validates our initial hypothesis that an ensemble of models can do better than any single model ever could. Our router is open source ( https://ift.tt/HfW7y2I ) so anyone can try it out. Or if you prefer you can use our hosted version ( https://ift.tt/WNLDiIJ ).

NIT, WBIT cutting down to 24 teams in wake of expanded NCAA tournaments

In the aftermath of the NCAA tournaments expanding to 76 teams starting this season, the NIT and Women's Basketball Invitation Tournament will drop from 32 teams to 24 beginning in 2027, the NCAA announced.

from www.espn.com - TOP https://ift.tt/um5YNfs

New top story on Hacker News: Show HN: Ledge.sh – Runnable Markdown Notes

Show HN: Ledge.sh – Runnable Markdown Notes
28 by dancablam | 18 comments on Hacker News.
Hi HN, Ledge is a Markdown notebook that runs shell commands, code, SQL, etc from inside your own notes. I built Ledge because I spend much of my day copy/pasting commands from my notes into the terminal. I was inspired by how much cmux helped me organize my terminals - but there was still a split brain between my notes and frequently run commands. I've been daily-driving it for the past few weeks and use it for deploys, API calls, smoke tests, etc. Ledge runs your real shell just like a terminal app and can be hosted locally or remotely over SSH using ledge-server. I've been building it since July and have recently added support for all the major platforms: Mac (Silicon), Windows (WSL required), Linux, iOS, Android. Mobile devices require SSH access to a ledge-server and Android is still in beta and looking for beta testers (see link on website)! It's built on Bun and Electrobun and is free and open-source. Feel free to review the code and contribute at https://ift.tt/Nyb7uQC Feedback is very much welcome. Any must-have features that are missing?

Texas AD says Longhorns' vulgar celebrations 'embarrassing'

Texas athletic director Chris Del Conte apologized for the top-ranked Longhorns' vulgar on-field celebrations during their win over Tennessee and called them "embarrassing" for the school.

from www.espn.com - TOP https://ift.tt/zqXaMPL

New top story on Hacker News: New PlayStation 5 Console Jailbreak Released

New PlayStation 5 Console Jailbreak Released
5 by therepanic | 0 comments on Hacker News.

Crochet, out since April, available out of bullpen for Red Sox

Red Sox left-hander Garrett Crochet, who has not pitched since April because of shoulder inflammation, will be available out of the bullpen vs. the Yankees.

from www.espn.com - TOP https://ift.tt/zymOIi7

New top story on Hacker News: Show HN: HN.watch – Videos of all Hacker News posts

Show HN: HN.watch – Videos of all Hacker News posts
30 by mrborgen | 8 comments on Hacker News.
Hi HN, I’m Per, founder of Scrimba (YC S20). We’ve spent the last decade teaching people how to code with an HTML-based video format. We’ve now plugged an LLM into it, so that people can create explainer videos about anything. It’s called “Scrimba Explain”. To demo this technology for Hacker News, we built HN.watch. It’s like HN, but with explainer videos instead of articles. We create them on-the-fly the first time someone clicks on a link. While there are obvious visual drawbacks of using HTML instead of diffusion models, there are three big benefits: - Speed: Much faster to generate than pixel-based videos (just a few seconds from click to playback) - Cost: Our cost per video is ~$0.04. (Excluding image generation, which some videos utilize. Quickly blows up the cost) - Easy editing: the above benefits also make AI-assisted editing cheap & fast Our hypothesis is that if video creation goes from “dollars and minutes” to “cents and seconds”, a bunch of new use cases will be unlocked. Here are some we see already: - A video explanation of every single Pull Request (we do this internally) - Give every page in your internal/extrernal docs a video - Turn a complex article into a video in ~4 seconds (via our Chrome extension) - Course creators can quickly draft lessons before recording the real thing - People also create a lot of personal stuff stories for their kids, wedding invitations, birthdays, etc The stack is based on an open-source programming language (Imba) created by our CTO, Sindre Aarsæther. It compiles to JavaScript, so it interoperates fully with the npm + node ecosystem. You can learn more here: https://imba.io/ We’ve also built our own sync engine (OP), and a context management system for agents (Q). We feared this would make the LLMs struggle when writing code for us, as neither is in their training data (there’s very little Imba in there too). However, we’ve been pleasantly surprised to see that LLMs actually are really good at our stack. This is probably because the stack is extremely dense. Imba is compact, and so is OP, where a single declaration sets storage, sync, permissions, UI, and what the AI sees. This means there’s no translations between frontend, API, db and JSON where the model can get confused and get things wrong. Simply said, instead of using React.js, Express, Supabase, and LangChain, we built it all from scratch. Definitely suffering from the “not invented here” syndrome, lol! As for the models, we use Gemini, GPTs, Inworld, ElevenLabs, and a few others. If you want to try it out, just take your pick: - The Web UI (scrimba.com/explain) - MCP (add it to your coding agent) - ChatGPT Plugin - Chrome Extension You can find a link to all of the above in our docs: https://ift.tt/e7gantS And finally, a real pixel-based video of the tool: https://www.youtube.com/watch?v=k6rbHmBxSEs Would love to hear your feedback and if anyone has ideas for other use cases. PS: I expect quite a bit of pushback from HN for this launch, given how fan of text the HN crowd is. This kind of tool is not for everyone. But there are a lot of people today who prefer videos over text, especially in the younger generations.