Instead, what if we fed audio directly to the LLM - LLM's are really smart, can they figure it out? This approach is faster (we skip transcription decoding) and less lossy/more robust because the big language model should be smarter than a smaller transcription decoder.
I've trained a model in just that fashion. For more architectural information and some training details, see this first post: https://tincans.ai/slm . For details about this model and some ideas for how to prompt it, see this post: https://tincans.ai/slm3 . We trained this on a very limited budget but the model is able to do some things that even GPT-4, Gemini, and Claude cannot, eg reasoning about long-context audio directly, without transcription. We also believe that this is the first model in the world to conduct adversarial attacks and apply preference modeling in the speech domain.
The demo is unoptimized (unquantized bf16 weights, default Huggingface inference, serverless speed bumps) but achieves 120ms time to first token with audio. You can basically think of it as Mistral 7B, so it'll be very fast and can also run basically anywhere. I am especially optimistic about embedded usage -- not needing the transcription step means that the resulting model is smaller and cheaper to use on the edge.
Would love to hear your thoughts and how you would use it! Weights are Apache-2 and on Hugging Face: https://huggingface.co/collections/tincans-ai/gazelle-v02-65..."},"title":{"matchLevel":"none","matchedWords":[],"value":"Show HN: Real-time voice chat with AI, no transcription"},"url":{"matchLevel":"none","matchedWords":[],"value":"https://demo.tincans.ai/"}},"_tags":["story","author_huac","story_39759147","show_hn"],"author":"huac","children":[39760261,39767586,39789908,39901251],"created_at":"2024-03-19T19:07:00Z","created_at_i":1710875220,"num_comments":6,"objectID":"39759147","points":33,"story_id":39759147,"story_text":"Hi HN -- voice chat with AI is very popular these days, especially with YC startups (https://twitter.com/k7agar/status/1769078697661804795). The current approaches all do a cascaded approach, with audio -> transcription -> language model -> text synthesis. This approach is easy to get started with, but requires lots of complexity and has a few glaring limitations. Most notably, transcription is slow, is lossy and any error propagates to the rest of the system, cannot capture emotional affect, is often not robust to code-switching/accents, and more.
Instead, what if we fed audio directly to the LLM - LLM's are really smart, can they figure it out? This approach is faster (we skip transcription decoding) and less lossy/more robust because the big language model should be smarter than a smaller transcription decoder.
I've trained a model in just that fashion. For more architectural information and some training details, see this first post: https://tincans.ai/slm . For details about this model and some ideas for how to prompt it, see this post: https://tincans.ai/slm3 . We trained this on a very limited budget but the model is able to do some things that even GPT-4, Gemini, and Claude cannot, eg reasoning about long-context audio directly, without transcription. We also believe that this is the first model in the world to conduct adversarial attacks and apply preference modeling in the speech domain.
The demo is unoptimized (unquantized bf16 weights, default Huggingface inference, serverless speed bumps) but achieves 120ms time to first token with audio. You can basically think of it as Mistral 7B, so it'll be very fast and can also run basically anywhere. I am especially optimistic about embedded usage -- not needing the transcription step means that the resulting model is smaller and cheaper to use on the edge.
Would love to hear your thoughts and how you would use it! Weights are Apache-2 and on Hugging Face: https://huggingface.co/collections/tincans-ai/gazelle-v02-65...","title":"Show HN: Real-time voice chat with AI, no transcription","updated_at":"2024-09-20T16:39:23Z","url":"https://demo.tincans.ai/"},{"_highlightResult":{"author":{"matchLevel":"none","matchedWords":[],"value":"wayy"},"story_text":{"fullyHighlighted":false,"matchLevel":"full","matchedWords":["preference","model"],"value":"Hey all, Justin here. I previously built Phind, the AI search engine for developers.
One of the biggest problems we had there was figuring out what went wrong with bad searches. We had tons of searches per day, but less than 1% of users gave any explicit feedback. So we were either manually digging through searches or making general system improvements and hoping they helped.
This problem gets harder with agents. Traces are longer and more complex. It takes more effort to review them, so I'm building a tool that lets you analyze LLM outputs directly to help developers of LLM apps and agents understand where things are breaking and why.
I've put together a demo using browser-use agent traces (gpt-5): https://trails-red.vercel.app/viewer
It's early, but I have lots of ideas - live querying of past failures for currently-running agents, preference models to expand sparse signal data.
Would love feedback on the demo. Also if you're building agents and have 10k+ traces per day that you're not looking at but would like to, I'd love to talk."},"title":{"matchLevel":"none","matchedWords":[],"value":"Show HN: 83 browser-use trajectories, visualized"},"url":{"matchLevel":"none","matchedWords":[],"value":"https://trails-red.vercel.app/viewer"}},"_tags":["story","author_wayy","story_46736978","show_hn"],"author":"wayy","children":[46739500],"created_at":"2026-01-23T19:50:03Z","created_at_i":1769197803,"num_comments":1,"objectID":"46736978","points":7,"story_id":46736978,"story_text":"Hey all, Justin here. I previously built Phind, the AI search engine for developers.
One of the biggest problems we had there was figuring out what went wrong with bad searches. We had tons of searches per day, but less than 1% of users gave any explicit feedback. So we were either manually digging through searches or making general system improvements and hoping they helped.
This problem gets harder with agents. Traces are longer and more complex. It takes more effort to review them, so I'm building a tool that lets you analyze LLM outputs directly to help developers of LLM apps and agents understand where things are breaking and why.
I've put together a demo using browser-use agent traces (gpt-5): https://trails-red.vercel.app/viewer
It's early, but I have lots of ideas - live querying of past failures for currently-running agents, preference models to expand sparse signal data.
Would love feedback on the demo. Also if you're building agents and have 10k+ traces per day that you're not looking at but would like to, I'd love to talk.","title":"Show HN: 83 browser-use trajectories, visualized","updated_at":"2026-03-05T23:24:17Z","url":"https://trails-red.vercel.app/viewer"},{"_highlightResult":{"author":{"matchLevel":"none","matchedWords":[],"value":"flamenfury"},"story_text":{"fullyHighlighted":false,"matchLevel":"full","matchedWords":["preference","model"],"value":"Hi HN,
I built what\u2011the\u2011token because I wanted to see how many tokens (and equivalent API dollars) my ChatGPT habit really burns. It\u2019s a static Next.js page running on a single Cloudflare worker, and everything happens in your browser \u2013 no uploads, no tracking.
It visualizes token usage, and provides insights like usage patterns and model preferences.
It's open-source, requires no sign-up, and is designed with privacy in mind. Try the live demo or clone and run `pnpm dev`.
The FAQ on the website covers more details :)"},"title":{"matchLevel":"none","matchedWords":[],"value":"Show HN: Visualize your ChatGPT usage locally"},"url":{"matchLevel":"none","matchedWords":[],"value":"https://what-the-token.pages.dev"}},"_tags":["story","author_flamenfury","story_43871424","show_hn"],"author":"flamenfury","created_at":"2025-05-02T15:52:01Z","created_at_i":1746201121,"num_comments":0,"objectID":"43871424","points":4,"story_id":43871424,"story_text":"Hi HN,
I built what\u2011the\u2011token because I wanted to see how many tokens (and equivalent API dollars) my ChatGPT habit really burns. It\u2019s a static Next.js page running on a single Cloudflare worker, and everything happens in your browser \u2013 no uploads, no tracking.
It visualizes token usage, and provides insights like usage patterns and model preferences.
It's open-source, requires no sign-up, and is designed with privacy in mind. Try the live demo or clone and run `pnpm dev`.
The FAQ on the website covers more details :)","title":"Show HN: Visualize your ChatGPT usage locally","updated_at":"2025-05-02T16:26:20Z","url":"https://what-the-token.pages.dev"},{"_highlightResult":{"author":{"matchLevel":"none","matchedWords":[],"value":"kartless"},"story_text":{"fullyHighlighted":false,"matchLevel":"full","matchedWords":["preference","model"],"value":"I'm a writer who was pondering the problems with getting my work discovered and realized that current search/recommendation engines don't go very deep. So I thought about ways to address this.
I was obviously mostly interested in books, so I worked on that first. I'd noticed LLMs are capable of pretty solid textual anaylsis these days, so I wondered if you could use that to generate good recommendations. So I sat down with an LLM- ChatGPT, Luna, for the curious- and spent about two hours explaining, in datail, my literary tastes- going into specifics like what sort of plot structure, characters, and humor I like, not just genre/specific titles.
The experiment demonstrated surprisingly strong results (though admittedly with a very small sample size)- the five recommendations I asked for were all ones that I hadn't heard of, as well as books that I would actually read. Further, I provided the LLM with a list of several books I had read and asked it to guess whether I would have liked them and why. The LLM successfully did so. Just proof of concept, but intriguing.
The problem with this, of course, is that most people wouldn't sit down with an LLM for an extended period of time to discuss literature. So I decided to build a small prototype of an app that might be able to do that.
The prototype is very simple- first, there are two types of information it saves- a persistent preference profile for the user, and user conversations. Then I built conversational functions to call an LLM to have conversations with the user to extract preference information.
The interesting thing to me is that this isn't just using an LLM to make recommendations- it's using one to analyze conversations over time, develop a persistent, evolving preference model for a specific user and then use that model to guide an LLM on a search for recommendations.
A few things occurred to me. First, the conversations and user profile DO NOT need to live on a central server- they can be stored locally by the user (or in a cloud of their choice).
Second, it seems like the user data could potentially be minimized and pseudonymized before being sent to the LLM, though I lack the technical expertise to know what is realistic here.
Third, if the LLM is conducting any searches, this would seem to make it possible to provide search services with less information about the user and their preferences (I'm not positive about this, so please do correct me if I'm wrong).
Finally, books are just one of the things you could use this for- it could technically be done for just about anything, at least in theory.
I experimented with this process several times, and received encouraging results.
What I don't know is-
Is this architecture novel? \nDoes it work on things other than books?\nAm I an outlier in how well I articulate my preferences?\nAm I wrong about the potential benefits to privacy?\nWould people use this? If the search results are as qualitatively higher for others as it was for me, I think they probably would, but I just don't know.
The reason I'm posting this is because I am a writer, not a software engineer. I don't really have a strong interest in developing this and could bring very little to the table on the technical side. But it seems like a neat use of the more advanced LLMs we have these days. I don't really know if the idea is novel (not in each part- every bit of the pieces has been done before- but as a whole).
I'm in the process of documenting the prototype and experiments- I'd be curious to hear if I am completely barking up the wrong tree or if this could actually be a cool new way to conduct searches.
I'm not posting this as a business proposition- it's just not my jam. If anyone wants to take the idea and run with it, I'd love to see what you do with it."},"title":{"matchLevel":"none","matchedWords":[],"value":"Can an AI learn what you like, rather than just what you've searched for?"}},"_tags":["story","author_kartless","story_50017172","ask_hn"],"author":"kartless","children":[50017196,50017350,50026022,50124831],"created_at":"2026-10-09T07:18:41Z","created_at_i":1791530321,"num_comments":3,"objectID":"50017172","points":2,"story_id":50017172,"story_text":"I'm a writer who was pondering the problems with getting my work discovered and realized that current search/recommendation engines don't go very deep. So I thought about ways to address this.
I was obviously mostly interested in books, so I worked on that first. I'd noticed LLMs are capable of pretty solid textual anaylsis these days, so I wondered if you could use that to generate good recommendations. So I sat down with an LLM- ChatGPT, Luna, for the curious- and spent about two hours explaining, in datail, my literary tastes- going into specifics like what sort of plot structure, characters, and humor I like, not just genre/specific titles.
The experiment demonstrated surprisingly strong results (though admittedly with a very small sample size)- the five recommendations I asked for were all ones that I hadn't heard of, as well as books that I would actually read. Further, I provided the LLM with a list of several books I had read and asked it to guess whether I would have liked them and why. The LLM successfully did so. Just proof of concept, but intriguing.
The problem with this, of course, is that most people wouldn't sit down with an LLM for an extended period of time to discuss literature. So I decided to build a small prototype of an app that might be able to do that.
The prototype is very simple- first, there are two types of information it saves- a persistent preference profile for the user, and user conversations. Then I built conversational functions to call an LLM to have conversations with the user to extract preference information.
The interesting thing to me is that this isn't just using an LLM to make recommendations- it's using one to analyze conversations over time, develop a persistent, evolving preference model for a specific user and then use that model to guide an LLM on a search for recommendations.
A few things occurred to me. First, the conversations and user profile DO NOT need to live on a central server- they can be stored locally by the user (or in a cloud of their choice).
Second, it seems like the user data could potentially be minimized and pseudonymized before being sent to the LLM, though I lack the technical expertise to know what is realistic here.
Third, if the LLM is conducting any searches, this would seem to make it possible to provide search services with less information about the user and their preferences (I'm not positive about this, so please do correct me if I'm wrong).
Finally, books are just one of the things you could use this for- it could technically be done for just about anything, at least in theory.
I experimented with this process several times, and received encouraging results.
What I don't know is-
Is this architecture novel? \nDoes it work on things other than books?\nAm I an outlier in how well I articulate my preferences?\nAm I wrong about the potential benefits to privacy?\nWould people use this? If the search results are as qualitatively higher for others as it was for me, I think they probably would, but I just don't know.
The reason I'm posting this is because I am a writer, not a software engineer. I don't really have a strong interest in developing this and could bring very little to the table on the technical side. But it seems like a neat use of the more advanced LLMs we have these days. I don't really know if the idea is novel (not in each part- every bit of the pieces has been done before- but as a whole).
I'm in the process of documenting the prototype and experiments- I'd be curious to hear if I am completely barking up the wrong tree or if this could actually be a cool new way to conduct searches.
I'm not posting this as a business proposition- it's just not my jam. If anyone wants to take the idea and run with it, I'd love to see what you do with it.","title":"Can an AI learn what you like, rather than just what you've searched for?","updated_at":"2026-10-16T21:15:06Z"},{"_highlightResult":{"author":{"matchLevel":"none","matchedWords":[],"value":"Tanjim"},"story_text":{"fullyHighlighted":false,"matchLevel":"full","matchedWords":["preference","model"],"value":"Hi HN,
When OpenAI started phasing out GPT-4o from the main ChatGPT interface, I noticed a surprising amount of backlash \u2014 not about benchmarks or features, but about tone.
A lot of people (myself included) felt that 4o had a certain conversational warmth that later models don\u2019t quite replicate in the same way. That difference was subtle, but noticeable.
So I built a small side project for myself: https://www.ilove4o.com/
It\u2019s a minimal interface that connects directly to GPT-4o via the OpenAI API. No extra layers, no personality hacks \u2014 just a focused 4o-only chat experience.
I\u2019m sharing it here because:\n- There seems to be real user preference around model \u201cpersonality.\u201d\n- I\u2019m curious whether others noticed the same behavioral shift.\n- It raises an interesting question: how much of perceived \u201cfriendliness\u201d comes from system prompts, UI, or subtle model tuning?
If you try it, I\u2019d genuinely love feedback \u2014 especially from people who\u2019ve spent significant time with multiple model versions.
Happy to answer technical or architectural questions."},"title":{"matchLevel":"none","matchedWords":[],"value":"Show HN: Ilove4o \u2013 a simple way to keep using GPT-4o"}},"_tags":["story","author_Tanjim","story_47123105","show_hn"],"author":"Tanjim","children":[47171728],"created_at":"2026-02-23T14:56:42Z","created_at_i":1771858602,"num_comments":1,"objectID":"47123105","points":1,"story_id":47123105,"story_text":"Hi HN,
When OpenAI started phasing out GPT-4o from the main ChatGPT interface, I noticed a surprising amount of backlash \u2014 not about benchmarks or features, but about tone.
A lot of people (myself included) felt that 4o had a certain conversational warmth that later models don\u2019t quite replicate in the same way. That difference was subtle, but noticeable.
So I built a small side project for myself: https://www.ilove4o.com/
It\u2019s a minimal interface that connects directly to GPT-4o via the OpenAI API. No extra layers, no personality hacks \u2014 just a focused 4o-only chat experience.
I\u2019m sharing it here because:\n- There seems to be real user preference around model \u201cpersonality.\u201d\n- I\u2019m curious whether others noticed the same behavioral shift.\n- It raises an interesting question: how much of perceived \u201cfriendliness\u201d comes from system prompts, UI, or subtle model tuning?
If you try it, I\u2019d genuinely love feedback \u2014 especially from people who\u2019ve spent significant time with multiple model versions.
Happy to answer technical or architectural questions.","title":"Show HN: Ilove4o \u2013 a simple way to keep using GPT-4o","updated_at":"2026-03-05T23:35:35Z"},{"_highlightResult":{"author":{"matchLevel":"none","matchedWords":[],"value":"nharada"},"title":{"matchLevel":"none","matchedWords":[],"value":"Terminal Bench 4 Is Saturated"},"url":{"fullyHighlighted":false,"matchLevel":"full","matchedWords":["preference","model"],"value":"https://preferencemodel.com/blog/terminal-bench-4-is-saturated/"}},"_tags":["story","author_nharada","story_50157233"],"author":"nharada","children":[50162937],"created_at":"2026-10-19T19:55:04Z","created_at_i":1792439704,"num_comments":3,"objectID":"50157233","points":26,"story_id":50157233,"title":"Terminal Bench 4 Is Saturated","updated_at":"2026-10-21T21:01:29Z","url":"https://preferencemodel.com/blog/terminal-bench-4-is-saturated/"},{"_highlightResult":{"author":{"matchLevel":"none","matchedWords":[],"value":"256"},"title":{"matchLevel":"none","matchedWords":[],"value":"Show HN: Karotte, a framework for building robust RL envs"},"url":{"fullyHighlighted":false,"matchLevel":"full","matchedWords":["preference","model"],"value":"https://github.com/preferencemodel/karotte"}},"_tags":["story","author_256","story_49996497","show_hn"],"author":"256","created_at":"2026-10-07T18:05:38Z","created_at_i":1791396338,"num_comments":0,"objectID":"49996497","points":8,"story_id":49996497,"title":"Show HN: Karotte, a framework for building robust RL envs","updated_at":"2026-10-07T18:31:05Z","url":"https://github.com/preferencemodel/karotte"},{"_highlightResult":{"author":{"matchLevel":"none","matchedWords":[],"value":"ilyaberger"},"title":{"matchLevel":"none","matchedWords":[],"value":"Thinking of buying a home? Use this tool to see if you should keep on looking"},"url":{"fullyHighlighted":false,"matchLevel":"full","matchedWords":["preference","model"],"value":"https://medium.com/@ilyaberger/should-i-buy-this-house-or-wait-for-a-better-one-solved-with-a-risk-preference-model-b6d06f22bad7"}},"_tags":["story","author_ilyaberger","story_22485412"],"author":"ilyaberger","created_at":"2020-03-04T16:57:16Z","created_at_i":1583341036,"num_comments":0,"objectID":"22485412","points":4,"story_id":22485412,"title":"Thinking of buying a home? Use this tool to see if you should keep on looking","updated_at":"2024-09-20T05:46:16Z","url":"https://medium.com/@ilyaberger/should-i-buy-this-house-or-wait-for-a-better-one-solved-with-a-risk-preference-model-b6d06f22bad7"},{"_highlightResult":{"author":{"matchLevel":"none","matchedWords":[],"value":"ilyaberger"},"title":{"fullyHighlighted":false,"matchLevel":"partial","matchedWords":["model"],"value":"Should I buy this house or wait for a better one \u2013 a handy calculator and model"},"url":{"fullyHighlighted":false,"matchLevel":"full","matchedWords":["preference","model"],"value":"https://medium.com/@ilyaberger/should-i-buy-this-house-or-wait-for-a-better-one-solved-with-a-risk-preference-model-b6d06f22bad7"}},"_tags":["story","author_ilyaberger","story_22384549"],"author":"ilyaberger","children":[22384821],"created_at":"2020-02-21T16:50:46Z","created_at_i":1582303846,"num_comments":2,"objectID":"22384549","points":3,"story_id":22384549,"title":"Should I buy this house or wait for a better one \u2013 a handy calculator and model","updated_at":"2024-09-20T05:45:09Z","url":"https://medium.com/@ilyaberger/should-i-buy-this-house-or-wait-for-a-better-one-solved-with-a-risk-preference-model-b6d06f22bad7"},{"_highlightResult":{"author":{"matchLevel":"none","matchedWords":[],"value":"martian"},"story_text":{"matchLevel":"none","matchedWords":[],"value":""},"title":{"fullyHighlighted":false,"matchLevel":"full","matchedWords":["preference","model"],"value":"Modeling wine preferences by data mining from physicochemical properties [pdf]"},"url":{"matchLevel":"none","matchedWords":[],"value":"http://www3.dsi.uminho.pt/pcortez/winequality09.pdf"}},"_tags":["story","author_martian","story_948389"],"author":"martian","children":[948425],"created_at":"2009-11-18T15:11:32Z","created_at_i":1258557092,"num_comments":2,"objectID":"948389","points":15,"story_id":948389,"story_text":"","title":"Modeling wine preferences by data mining from physicochemical properties [pdf]","updated_at":"2024-09-19T16:49:25Z","url":"http://www3.dsi.uminho.pt/pcortez/winequality09.pdf"},{"_highlightResult":{"author":{"matchLevel":"none","matchedWords":[],"value":"mixeden"},"title":{"fullyHighlighted":false,"matchLevel":"full","matchedWords":["preference","model"],"value":"Comal: A Meta-Algorithm for Aligning Large Language Models with Preferences"},"url":{"fullyHighlighted":false,"matchLevel":"partial","matchedWords":["preference"],"value":"https://synthical.com/article/COMAL%3A-A-Convergent-Meta-Algorithm-for-Aligning-LLMs-with-General-Preferences-ee4303ca-8397-49be-8409-6f803de3f190"}},"_tags":["story","author_mixeden","story_42026763"],"author":"mixeden","created_at":"2024-11-02T14:58:41Z","created_at_i":1730559521,"num_comments":0,"objectID":"42026763","points":2,"story_id":42026763,"title":"Comal: A Meta-Algorithm for Aligning Large Language Models with Preferences","updated_at":"2024-11-02T16:55:15Z","url":"https://synthical.com/article/COMAL%3A-A-Convergent-Meta-Algorithm-for-Aligning-LLMs-with-General-Preferences-ee4303ca-8397-49be-8409-6f803de3f190"},{"_highlightResult":{"author":{"matchLevel":"none","matchedWords":[],"value":"pr337h4m"},"title":{"fullyHighlighted":false,"matchLevel":"full","matchedWords":["preference","model"],"value":"An experiment in AI model aesthetic preferences"},"url":{"matchLevel":"none","matchedWords":[],"value":"https://test-ai-taste.vercel.app/"}},"_tags":["story","author_pr337h4m","story_43118273"],"author":"pr337h4m","created_at":"2025-02-20T18:28:23Z","created_at_i":1740076103,"num_comments":0,"objectID":"43118273","points":1,"story_id":43118273,"title":"An experiment in AI model aesthetic preferences","updated_at":"2025-02-20T18:30:00Z","url":"https://test-ai-taste.vercel.app/"},{"_highlightResult":{"author":{"matchLevel":"none","matchedWords":[],"value":"bookofjoe"},"title":{"fullyHighlighted":false,"matchLevel":"full","matchedWords":["preference","model"],"value":"Eliciting Human Preferences with Language Models"},"url":{"matchLevel":"none","matchedWords":[],"value":"https://arxiv.org/abs/2310.11589"}},"_tags":["story","author_bookofjoe","story_37969229"],"author":"bookofjoe","created_at":"2023-10-21T18:19:15Z","created_at_i":1697912355,"num_comments":0,"objectID":"37969229","points":1,"story_id":37969229,"title":"Eliciting Human Preferences with Language Models","updated_at":"2024-09-20T15:27:57Z","url":"https://arxiv.org/abs/2310.11589"},{"_highlightResult":{"author":{"matchLevel":"none","matchedWords":[],"value":"adilhafeez"},"story_text":{"fullyHighlighted":false,"matchLevel":"full","matchedWords":["preference","model"],"value":"Hi HN \u2014 we're the team behind Arch (https://github.com/katanemo/archgw), an open-source proxy for LLMs written in Rust. Today we're releasing Arch-Router (https://huggingface.co/katanemo/Arch-Router-1.5B), a 1.5B router model for preference-based routing, now integrated into the proxy. As teams integrate multiple LLMs - each with different strengths, styles, or cost/latency profiles \u2014 routing the right prompt to the right model becomes a critical part of the application design. But it's still an open problem. Most routing systems fall into two camps:
- Embedding-based routers use intent classifiers \u2014 label a prompt as \u201csupport,\u201d \u201cSQL,\u201d or \u201cmath,\u201d then route to a matching model. This works for simple tasks but breaks down in real conversations. Users shift topics mid-conversation, task boundaries blur, and product changes require retraining classifiers.
- Performance-based routers pick models based on benchmarks like MMLU or MT-Bench, or based on latency or cost curves. But benchmarks often miss what matters in production: domain-specific quality or subjective preferences like \u201cWill legal accept this clause?\u201d
Arch-Router takes a different approach: route by preferences written in plain language. You write rules like \u201ccontract clauses \u2192 GPT-4o\u201d or \u201cquick travel tips \u2192 Gemini Flash.\u201d The router maps the prompt (and conversation context) to those rules using a lightweight 1.5B autoregressive model. No retraining, no fragile if/else chains. We built this with input from teams at Twilio and Atlassian. It handles intent drift, supports multi-turn conversations, and lets you swap in or out models with a one-line change to the routing policy. Full details are in our paper (https://arxiv.org/abs/2506.16655), but here's a snapshot:
Specs:
- 1.5B params \u2014 runs on a single GPU (or CPU for testing)
- No retraining needed \u2014 point it at any mix of LLMs
- Cost and latency aware \u2014 route heavy tasks to expensive models, light tasks to faster/cheaper ones
- Outperforms larger closed models on our conversational routing benchmarks (details in the paper)
Links:
- Arch Proxy (open source): https://github.com/katanemo/archgw
- Model + code: https://huggingface.co/katanemo/Arch-Router-1.5B
- Paper: https://arxiv.org/abs/2506.16655"},"title":{"fullyHighlighted":false,"matchLevel":"full","matchedWords":["preference","model"],"value":"Show HN: Arch-Router \u2013 1.5B model for LLM routing by preferences, not benchmarks"}},"_tags":["story","author_adilhafeez","story_44436031","show_hn"],"author":"adilhafeez","children":[44436070,44437340,44437954,44438177,44438786,44439077,44445753],"created_at":"2025-07-01T17:13:11Z","created_at_i":1751389991,"num_comments":15,"objectID":"44436031","points":66,"story_id":44436031,"story_text":"Hi HN \u2014 we're the team behind Arch (https://github.com/katanemo/archgw), an open-source proxy for LLMs written in Rust. Today we're releasing Arch-Router (https://huggingface.co/katanemo/Arch-Router-1.5B), a 1.5B router model for preference-based routing, now integrated into the proxy. As teams integrate multiple LLMs - each with different strengths, styles, or cost/latency profiles \u2014 routing the right prompt to the right model becomes a critical part of the application design. But it's still an open problem. Most routing systems fall into two camps:
- Embedding-based routers use intent classifiers \u2014 label a prompt as \u201csupport,\u201d \u201cSQL,\u201d or \u201cmath,\u201d then route to a matching model. This works for simple tasks but breaks down in real conversations. Users shift topics mid-conversation, task boundaries blur, and product changes require retraining classifiers.
- Performance-based routers pick models based on benchmarks like MMLU or MT-Bench, or based on latency or cost curves. But benchmarks often miss what matters in production: domain-specific quality or subjective preferences like \u201cWill legal accept this clause?\u201d
Arch-Router takes a different approach: route by preferences written in plain language. You write rules like \u201ccontract clauses \u2192 GPT-4o\u201d or \u201cquick travel tips \u2192 Gemini Flash.\u201d The router maps the prompt (and conversation context) to those rules using a lightweight 1.5B autoregressive model. No retraining, no fragile if/else chains. We built this with input from teams at Twilio and Atlassian. It handles intent drift, supports multi-turn conversations, and lets you swap in or out models with a one-line change to the routing policy. Full details are in our paper (https://arxiv.org/abs/2506.16655), but here's a snapshot:
Specs:
- 1.5B params \u2014 runs on a single GPU (or CPU for testing)
- No retraining needed \u2014 point it at any mix of LLMs
- Cost and latency aware \u2014 route heavy tasks to expensive models, light tasks to faster/cheaper ones
- Outperforms larger closed models on our conversational routing benchmarks (details in the paper)
Links:
- Arch Proxy (open source): https://github.com/katanemo/archgw
- Model + code: https://huggingface.co/katanemo/Arch-Router-1.5B
- Paper: https://arxiv.org/abs/2506.16655","title":"Show HN: Arch-Router \u2013 1.5B model for LLM routing by preferences, not benchmarks","updated_at":"2025-12-29T14:06:44Z"},{"_highlightResult":{"author":{"matchLevel":"none","matchedWords":[],"value":"foundress"},"story_text":{"fullyHighlighted":false,"matchLevel":"full","matchedWords":["preference","model"],"value":"As models are getting better, context windows are expanding, tokens are getting cheaper, there is an explicit race after context.
Context is the holy grail. With the right context, models can read your data, situation and constraints to generate more relevant output. Better context lets you tell the model what you mean in less iterations.
Context capture takes different forms, however.
The browsers, screen recorders, products syncing with email, calendar, drive where you keep your information are getting traction. I believe passive context is largely solved.
Another form of context that is very poorly tapped into is the one hidden in your own brain. These are patterns learnt from previous data and feedback you have seen, your thinking process and constraints, your tacit domain knowledge and world model, your preferences and interpretation of reality.
The real bottleneck is the ability to get that information out of my human brain into the model, as efficiently and precisely as possible.
Active extraction is broken. We burn hours translating what\u2019s in our head into prompts, specs or comments.
You write a 500 word prompt, let\u2019s say, then realize you forgot the one nuance that actually matters or certain constraints that would have impact. You split tasks into micro-prompts because dumping the whole mental model at once is impossible.You often start from zero vs iterating further as the returns on iterations are diminishing and costly from a token spend and time perspective.
As humans, we can maybe juggle 3-4 things at once; complex specs can be composed of 10-100 different concepts, crossing this limit. It does not help that LLMs still demand big monolithic prompts. We end up offloading a lot of details and memory to the models.
No one is truly going after this problem. In fact, the ones who should are not incentivized to.
Most of the revenue generated in AI today is in fact accelerated by this bottleneck, hence the companies developing productivity tools are not truly motivated to address it.
So where is the next productivity leap? Models that can read our mind better than we can and preempt every need? Models and products that passively get every possible piece of context about me? Brain-computer interfaces?
Interfaces that can shrink the mind-to-model gap, help the model do what I mean, tools that almost let me think out loud in real-time, capture nuance without friction, and refine my intent , are going to have the most impact today.
We have built such a tool internally at Ntropy and have been using it for a while. To set up and refine almost all our LLM pipelines. Today, we are sharing it with the world .
Below are some raw thoughts and design principles that we used to make it
Mixed initiative. A productive human-to-model interface needs to be dialogue driven, where the model is more proactive and initiates follow ups that are precise and lead you to chunk by chunk thinking vs asking for a straightforward dump of thought . It takes out and infers what you really want it to do chunk by chunk from your brain.\nVisual scaffolding. Our brains often require structure and scaffolding that is permanent and gets updated as we add or remove detail or change input. \nReal-time and continuous spec evals. Everyone is focused on output evaluations that are important and effective, however are very costly and not straightforward to act on. Also, often misleading. They are biased towards your own dataset and lack ground truth. Continuous input evals and context quality assessment will completely change LLM powered development and work in general, including evaluations and the developer experience.
As we continue using the tool for production inputs, the thinking and the list of these items is evolving rapidly. We cannot wait for more people to try and share their experience for us to improve and add to it. \nWill share the link in comments."},"title":{"matchLevel":"none","matchedWords":[],"value":"Active context extraction > passive context capture with LLMs"}},"_tags":["story","author_foundress","story_44825780","ask_hn"],"author":"foundress","children":[44825792,44826067],"created_at":"2025-08-07T15:30:18Z","created_at_i":1754580618,"num_comments":2,"objectID":"44825780","points":2,"story_id":44825780,"story_text":"As models are getting better, context windows are expanding, tokens are getting cheaper, there is an explicit race after context.
Context is the holy grail. With the right context, models can read your data, situation and constraints to generate more relevant output. Better context lets you tell the model what you mean in less iterations.
Context capture takes different forms, however.
The browsers, screen recorders, products syncing with email, calendar, drive where you keep your information are getting traction. I believe passive context is largely solved.
Another form of context that is very poorly tapped into is the one hidden in your own brain. These are patterns learnt from previous data and feedback you have seen, your thinking process and constraints, your tacit domain knowledge and world model, your preferences and interpretation of reality.
The real bottleneck is the ability to get that information out of my human brain into the model, as efficiently and precisely as possible.
Active extraction is broken. We burn hours translating what\u2019s in our head into prompts, specs or comments.
You write a 500 word prompt, let\u2019s say, then realize you forgot the one nuance that actually matters or certain constraints that would have impact. You split tasks into micro-prompts because dumping the whole mental model at once is impossible.You often start from zero vs iterating further as the returns on iterations are diminishing and costly from a token spend and time perspective.
As humans, we can maybe juggle 3-4 things at once; complex specs can be composed of 10-100 different concepts, crossing this limit. It does not help that LLMs still demand big monolithic prompts. We end up offloading a lot of details and memory to the models.
No one is truly going after this problem. In fact, the ones who should are not incentivized to.
Most of the revenue generated in AI today is in fact accelerated by this bottleneck, hence the companies developing productivity tools are not truly motivated to address it.
So where is the next productivity leap? Models that can read our mind better than we can and preempt every need? Models and products that passively get every possible piece of context about me? Brain-computer interfaces?
Interfaces that can shrink the mind-to-model gap, help the model do what I mean, tools that almost let me think out loud in real-time, capture nuance without friction, and refine my intent , are going to have the most impact today.
We have built such a tool internally at Ntropy and have been using it for a while. To set up and refine almost all our LLM pipelines. Today, we are sharing it with the world .
Below are some raw thoughts and design principles that we used to make it
Mixed initiative. A productive human-to-model interface needs to be dialogue driven, where the model is more proactive and initiates follow ups that are precise and lead you to chunk by chunk thinking vs asking for a straightforward dump of thought . It takes out and infers what you really want it to do chunk by chunk from your brain.\nVisual scaffolding. Our brains often require structure and scaffolding that is permanent and gets updated as we add or remove detail or change input. \nReal-time and continuous spec evals. Everyone is focused on output evaluations that are important and effective, however are very costly and not straightforward to act on. Also, often misleading. They are biased towards your own dataset and lack ground truth. Continuous input evals and context quality assessment will completely change LLM powered development and work in general, including evaluations and the developer experience.
As we continue using the tool for production inputs, the thinking and the list of these items is evolving rapidly. We cannot wait for more people to try and share their experience for us to improve and add to it. \nWill share the link in comments.","title":"Active context extraction > passive context capture with LLMs","updated_at":"2026-03-05T22:29:53Z"},{"_highlightResult":{"author":{"matchLevel":"none","matchedWords":[],"value":"kmckiern"},"story_text":{"fullyHighlighted":false,"matchLevel":"full","matchedWords":["preference","model"],"value":"At this point, the leading foundation models from each of the top labs tend to do well on the classic set of coding evals. Each new release brings smaller improvements over the last, with leading models separated by increasingly thin margins. This may be a sign we're heading toward commoditization.
However, in our experience, these models can differ meaningfully when the task is narrow and the context is specific. And under these circumstances is where most people use these models. E.g. in their backend codebase, according to their specific coding preferences.
We think this is partly why folks argue about which model is best for coding. I was personally confused to see the praise for Sonnet 4 as a pair programmer. This didn't match my experience - it takes way too many liberties for my taste, and it refuses to conform to / use the translation interface between my app's API and database.
So, our thesis is that the best model for coding (or for any narrow task) is personal. Hence, our goal is to make it as easy as possible for our users to find the best model for them.
To explore this, we put together a dataset of real engineering sprint tickets (tasks like "optimize our database connections logic") and wrote down some of our top model behavioral preferences to use as eval metrics.
We focused on three of our core pain points: Pattern Adherence (does it follow our existing architecture?), Scope Discipline (does it stay focused or make unsolicited refactors?), and Comment Quality (does it write useful documentation vs verbose restating). Each response was scored from -1.0 to 1.0 across these dimensions.
Then we evaluated 14 of the current top LLMs using these metrics, with statistical significance testing across the prompt set.
Results: o3 medium ranked first (0.53 \u00b1 0.05), with o4-mini second (0.48 \u00b1 0.05). Sonnet 4 placed 7th (0.41 \u00b1 0.07).
This helped us design a tiered "model-stack" that we now follow:
- o3 for complex, low-frequency tasks where quality matters most\n- o3-mini for difficult work at higher scale (similar quality, but much faster/cheaper)\n- Gemini 2.5 Flash for documentation (strong on comment quality, and very fast/cheap)
Some other findings:
- Scope discipline correlates weakly with other coding skills (r \u2248 0.05-0.09)\n- Every model struggles with comment quality - this metric has the lowest absolute scores\n- "Thinking" / "reasoning" doesn't always improve performance (Sonnet 4 regressed on two metrics)
Our results: https://mandoline.ai/leaderboards/coding
But recall, these results reflect our specific codebase patterns and team priorities. Your optimal choice depends on your architecture, task distribution, and what behaviors matter most to your workflow.
If you're choosing coding LLMs, our product helps you eval lots of models on your real tasks, against your own criteria.
Try it out: https://mandoline.ai/best-llm-for\nDocs: https://mandoline.ai/docs/best-llm-for"},"title":{"fullyHighlighted":false,"matchLevel":"partial","matchedWords":["preference"],"value":"Show HN: O3 beats Sonnet 4 at coding (in our codebase, wrt our preferences)"}},"_tags":["story","author_kmckiern","story_44686247","show_hn"],"author":"kmckiern","created_at":"2025-07-25T18:05:48Z","created_at_i":1753466748,"num_comments":0,"objectID":"44686247","points":2,"story_id":44686247,"story_text":"At this point, the leading foundation models from each of the top labs tend to do well on the classic set of coding evals. Each new release brings smaller improvements over the last, with leading models separated by increasingly thin margins. This may be a sign we're heading toward commoditization.
However, in our experience, these models can differ meaningfully when the task is narrow and the context is specific. And under these circumstances is where most people use these models. E.g. in their backend codebase, according to their specific coding preferences.
We think this is partly why folks argue about which model is best for coding. I was personally confused to see the praise for Sonnet 4 as a pair programmer. This didn't match my experience - it takes way too many liberties for my taste, and it refuses to conform to / use the translation interface between my app's API and database.
So, our thesis is that the best model for coding (or for any narrow task) is personal. Hence, our goal is to make it as easy as possible for our users to find the best model for them.
To explore this, we put together a dataset of real engineering sprint tickets (tasks like "optimize our database connections logic") and wrote down some of our top model behavioral preferences to use as eval metrics.
We focused on three of our core pain points: Pattern Adherence (does it follow our existing architecture?), Scope Discipline (does it stay focused or make unsolicited refactors?), and Comment Quality (does it write useful documentation vs verbose restating). Each response was scored from -1.0 to 1.0 across these dimensions.
Then we evaluated 14 of the current top LLMs using these metrics, with statistical significance testing across the prompt set.
Results: o3 medium ranked first (0.53 \u00b1 0.05), with o4-mini second (0.48 \u00b1 0.05). Sonnet 4 placed 7th (0.41 \u00b1 0.07).
This helped us design a tiered "model-stack" that we now follow:
- o3 for complex, low-frequency tasks where quality matters most\n- o3-mini for difficult work at higher scale (similar quality, but much faster/cheaper)\n- Gemini 2.5 Flash for documentation (strong on comment quality, and very fast/cheap)
Some other findings:
- Scope discipline correlates weakly with other coding skills (r \u2248 0.05-0.09)\n- Every model struggles with comment quality - this metric has the lowest absolute scores\n- "Thinking" / "reasoning" doesn't always improve performance (Sonnet 4 regressed on two metrics)
Our results: https://mandoline.ai/leaderboards/coding
But recall, these results reflect our specific codebase patterns and team priorities. Your optimal choice depends on your architecture, task distribution, and what behaviors matter most to your workflow.
If you're choosing coding LLMs, our product helps you eval lots of models on your real tasks, against your own criteria.
Try it out: https://mandoline.ai/best-llm-for\nDocs: https://mandoline.ai/docs/best-llm-for","title":"Show HN: O3 beats Sonnet 4 at coding (in our codebase, wrt our preferences)","updated_at":"2025-07-25T18:14:38Z"},{"_highlightResult":{"author":{"matchLevel":"none","matchedWords":[],"value":"NyX_AI_ZERO_DAY"},"story_text":{"fullyHighlighted":false,"matchLevel":"full","matchedWords":["preference","model"],"value":"Hey HN,
wrote this short paper cause i'm honestly tired of current alignment methods (RLHF). optimizing for "human preference" just creates models that hallucinate plausibly to please the user (stochastic parrots ), instead of being grounded in reality.
i'm proposing a different framework called LOGOS-ZERO. idea is to ditch moral guardrails (which are subjective/fluid) and anchor the loss function to physical/logical invariants.
basically:
Thermodynamic Loss : treat high entropy/hallucination as "Waste". if an action increases systemic disorder, it gets penalized.
Action Gating: Unlike current models that must generate tokens, this architecture simulates in latent space first. if the output is high-entropy or logically inconsistent, it returns a Null Vector (Silence/No).
it attempts to solve the grounding problem by making the AI follow the path of least action/ entropy rather than just mimicking human speech patterns.
link to the pdf on zenodo: https://zenodo.org/records/17976755
curious to hear thoughts on the physics mapping, roast it if u want."},"title":{"matchLevel":"none","matchedWords":[],"value":"Show HN: Aligning AI with Entropy Instead of 'Human Values' ( Paper)"}},"_tags":["story","author_NyX_AI_ZERO_DAY","story_46335795","show_hn"],"author":"NyX_AI_ZERO_DAY","children":[46335828],"created_at":"2025-12-20T12:51:53Z","created_at_i":1766235113,"num_comments":1,"objectID":"46335795","points":1,"story_id":46335795,"story_text":"Hey HN,
wrote this short paper cause i'm honestly tired of current alignment methods (RLHF). optimizing for "human preference" just creates models that hallucinate plausibly to please the user (stochastic parrots ), instead of being grounded in reality.
i'm proposing a different framework called LOGOS-ZERO. idea is to ditch moral guardrails (which are subjective/fluid) and anchor the loss function to physical/logical invariants.
basically:
Thermodynamic Loss : treat high entropy/hallucination as "Waste". if an action increases systemic disorder, it gets penalized.
Action Gating: Unlike current models that must generate tokens, this architecture simulates in latent space first. if the output is high-entropy or logically inconsistent, it returns a Null Vector (Silence/No).
it attempts to solve the grounding problem by making the AI follow the path of least action/ entropy rather than just mimicking human speech patterns.
link to the pdf on zenodo: https://zenodo.org/records/17976755
curious to hear thoughts on the physics mapping, roast it if u want.","title":"Show HN: Aligning AI with Entropy Instead of 'Human Values' ( Paper)","updated_at":"2026-03-05T23:15:04Z"},{"_highlightResult":{"author":{"matchLevel":"none","matchedWords":[],"value":"tim_sw"},"title":{"fullyHighlighted":false,"matchLevel":"full","matchedWords":["preference","model"],"value":"Direct Preference Optimization: Your Language Model Is a Reward Model"},"url":{"matchLevel":"none","matchedWords":[],"value":"https://arxiv.org/abs/2305.18290"}},"_tags":["story","author_tim_sw","story_36137914"],"author":"tim_sw","created_at":"2023-05-31T13:06:51Z","created_at_i":1685538411,"num_comments":0,"objectID":"36137914","points":4,"story_id":36137914,"title":"Direct Preference Optimization: Your Language Model Is a Reward Model","updated_at":"2024-09-20T14:13:30Z","url":"https://arxiv.org/abs/2305.18290"},{"_highlightResult":{"author":{"matchLevel":"none","matchedWords":[],"value":"tosh"},"title":{"fullyHighlighted":false,"matchLevel":"full","matchedWords":["preference","model"],"value":"Direct Preference Optimization: Your Language Model Is a Reward Model"},"url":{"matchLevel":"none","matchedWords":[],"value":"https://arxiv.org/abs/2305.18290"}},"_tags":["story","author_tosh","story_38957060"],"author":"tosh","created_at":"2024-01-11T19:01:28Z","created_at_i":1704999688,"num_comments":0,"objectID":"38957060","points":3,"story_id":38957060,"title":"Direct Preference Optimization: Your Language Model Is a Reward Model","updated_at":"2024-09-20T16:05:54Z","url":"https://arxiv.org/abs/2305.18290"},{"_highlightResult":{"author":{"matchLevel":"none","matchedWords":[],"value":"ntonozzi"},"title":{"fullyHighlighted":false,"matchLevel":"full","matchedWords":["preference","model"],"value":"Direct Preference Optimization: Your Language Model Is a Reward Model"},"url":{"matchLevel":"none","matchedWords":[],"value":"https://arxiv.org/abs/2305.18290"}},"_tags":["story","author_ntonozzi","story_36191442"],"author":"ntonozzi","created_at":"2023-06-05T02:04:51Z","created_at_i":1685930691,"num_comments":0,"objectID":"36191442","points":3,"story_id":36191442,"title":"Direct Preference Optimization: Your Language Model Is a Reward Model","updated_at":"2024-09-20T14:19:29Z","url":"https://arxiv.org/abs/2305.18290"},{"_highlightResult":{"author":{"matchLevel":"none","matchedWords":[],"value":"caprock"},"title":{"fullyHighlighted":false,"matchLevel":"full","matchedWords":["preference","model"],"value":"Pretraining Language Models with Human Preferences"},"url":{"matchLevel":"none","matchedWords":[],"value":"https://arxiv.org/abs/2302.08582"}},"_tags":["story","author_caprock","story_34869295"],"author":"caprock","created_at":"2023-02-20T15:56:13Z","created_at_i":1676908573,"num_comments":0,"objectID":"34869295","points":3,"story_id":34869295,"title":"Pretraining Language Models with Human Preferences","updated_at":"2024-09-20T13:26:47Z","url":"https://arxiv.org/abs/2302.08582"},{"_highlightResult":{"author":{"matchLevel":"none","matchedWords":[],"value":"measurablefunc"},"title":{"fullyHighlighted":false,"matchLevel":"full","matchedWords":["preference","model"],"value":"Steerable Cultural Preference Optimization of Reward Models"},"url":{"matchLevel":"none","matchedWords":[],"value":"https://arxiv.org/abs/2606.18606"}},"_tags":["story","author_measurablefunc","story_49789253"],"author":"measurablefunc","created_at":"2026-09-21T16:15:53Z","created_at_i":1790007353,"num_comments":0,"objectID":"49789253","points":2,"story_id":49789253,"title":"Steerable Cultural Preference Optimization of Reward Models","updated_at":"2026-09-21T19:06:00Z","url":"https://arxiv.org/abs/2606.18606"},{"_highlightResult":{"author":{"matchLevel":"none","matchedWords":[],"value":"honorable_coder"},"story_text":{"fullyHighlighted":false,"matchLevel":"full","matchedWords":["preference","model"],"value":"Today we\u2019re shipping a major update to ArchGW (an edge and service proxy for agents [1]): a unified router that supports three strategies for directing traffic to LLMs \u2014 from explicit model names, to semantic aliases, to dynamic preference-aligned routing. Here\u2019s how each works on its own, and how they come together.
Preference-aligned routing decouples task detection (e.g., code generation, image editing, Q&A) from LLM assignment. This approach captures the preferences developers establish when testing and evaluating LLMs on their domain-specific workflows and tasks. So, rather than relying on an automatic router trained to beat abstract benchmarks like MMLU or MT-Bench, developers can dynamically route requests to the most suitable model based on internal evaluations \u2014 and easily swap out the underlying moodel for specific actions and workflows. This is powered by our 1.5B Arch-Router LLM [2]. We also published our research on this recently[3]
Modal-aliases provide semantic, version-controlled names for models. Instead of using provider-specific model names like gpt-4o-mini or claude-3-5-sonnet-20241022 in your client you can create meaningful aliases like "fast-model" or "arch.summarize.v1". This allows you to test new models, swap out the config safely without having to do code-wide search/replace every time you want to use a new model for a very specific workflow or task.
Model-literals (nothing new) lets you specify exact provider/model combinations (e.g., openai/gpt-4o, anthropic/claude-3-5-sonnet-20241022), giving you full control and transparency over which model handles each request.
P.S. we routinely get asked why we didn't build semantic/embedding models for routing use cases or use some form of clustering technique. Clustering/embedding routers miss context, negation, and short elliptical queries, etc. An autoregressive approach conditions on the full context, letting the model reason about the task and generate an explicit label that can be used to match to an agent, task or LLM. In practice, this generalizes better to unseen or low-frequency intents and stays robust as conversations drift, without brittle thresholds or post-hoc cluster tuning.
[1] https://github.com/katanemo/archgw\n[2] https://huggingface.co/katanemo/Arch-Router-1.5B\n[2] https://arxiv.org/abs/2506.16655"},"title":{"fullyHighlighted":false,"matchLevel":"full","matchedWords":["preference","model"],"value":"Show HN: Model-literals, model-aliases, and preference-aligned routing for LLMs"},"url":{"matchLevel":"none","matchedWords":[],"value":"https://docs.archgw.com/guides/llm_router.html"}},"_tags":["story","author_honorable_coder","story_45324991","show_hn"],"author":"honorable_coder","created_at":"2025-09-21T17:45:54Z","created_at_i":1758476754,"num_comments":0,"objectID":"45324991","points":2,"story_id":45324991,"story_text":"Today we\u2019re shipping a major update to ArchGW (an edge and service proxy for agents [1]): a unified router that supports three strategies for directing traffic to LLMs \u2014 from explicit model names, to semantic aliases, to dynamic preference-aligned routing. Here\u2019s how each works on its own, and how they come together.
Preference-aligned routing decouples task detection (e.g., code generation, image editing, Q&A) from LLM assignment. This approach captures the preferences developers establish when testing and evaluating LLMs on their domain-specific workflows and tasks. So, rather than relying on an automatic router trained to beat abstract benchmarks like MMLU or MT-Bench, developers can dynamically route requests to the most suitable model based on internal evaluations \u2014 and easily swap out the underlying moodel for specific actions and workflows. This is powered by our 1.5B Arch-Router LLM [2]. We also published our research on this recently[3]
Modal-aliases provide semantic, version-controlled names for models. Instead of using provider-specific model names like gpt-4o-mini or claude-3-5-sonnet-20241022 in your client you can create meaningful aliases like "fast-model" or "arch.summarize.v1". This allows you to test new models, swap out the config safely without having to do code-wide search/replace every time you want to use a new model for a very specific workflow or task.
Model-literals (nothing new) lets you specify exact provider/model combinations (e.g., openai/gpt-4o, anthropic/claude-3-5-sonnet-20241022), giving you full control and transparency over which model handles each request.
P.S. we routinely get asked why we didn't build semantic/embedding models for routing use cases or use some form of clustering technique. Clustering/embedding routers miss context, negation, and short elliptical queries, etc. An autoregressive approach conditions on the full context, letting the model reason about the task and generate an explicit label that can be used to match to an agent, task or LLM. In practice, this generalizes better to unseen or low-frequency intents and stays robust as conversations drift, without brittle thresholds or post-hoc cluster tuning.
[1] https://github.com/katanemo/archgw\n[2] https://huggingface.co/katanemo/Arch-Router-1.5B\n[2] https://arxiv.org/abs/2506.16655","title":"Show HN: Model-literals, model-aliases, and preference-aligned routing for LLMs","updated_at":"2026-03-05T22:41:13Z","url":"https://docs.archgw.com/guides/llm_router.html"},{"_highlightResult":{"author":{"matchLevel":"none","matchedWords":[],"value":"variational123"},"title":{"fullyHighlighted":false,"matchLevel":"full","matchedWords":["preference","model"],"value":"Aligning diffusion models to human preferences with DPO"},"url":{"matchLevel":"none","matchedWords":[],"value":"https://arxiv.org/abs/2311.12908"}},"_tags":["story","author_variational123","story_38463967"],"author":"variational123","created_at":"2023-11-29T19:23:36Z","created_at_i":1701285816,"num_comments":0,"objectID":"38463967","points":2,"story_id":38463967,"title":"Aligning diffusion models to human preferences with DPO","updated_at":"2024-09-20T15:50:20Z","url":"https://arxiv.org/abs/2311.12908"},{"_highlightResult":{"author":{"matchLevel":"none","matchedWords":[],"value":"Jimmc414"},"title":{"fullyHighlighted":false,"matchLevel":"full","matchedWords":["preference","model"],"value":"Ayahuasca blocks ethanol preference in an animal model of dependence"},"url":{"matchLevel":"none","matchedWords":[],"value":"https://www.sciencedirect.com/science/article/abs/pii/S0378874121010953"}},"_tags":["story","author_Jimmc414","story_33352355"],"author":"Jimmc414","created_at":"2022-10-27T02:33:41Z","created_at_i":1666838021,"num_comments":0,"objectID":"33352355","points":2,"story_id":33352355,"title":"Ayahuasca blocks ethanol preference in an animal model of dependence","updated_at":"2024-09-20T12:24:47Z","url":"https://www.sciencedirect.com/science/article/abs/pii/S0378874121010953"},{"_highlightResult":{"author":{"matchLevel":"none","matchedWords":[],"value":"bryanrasmussen"},"title":{"fullyHighlighted":false,"matchLevel":"full","matchedWords":["preference","model"],"value":"Contrasting Computational Models of Mate Preference Across 45 Countries"},"url":{"matchLevel":"none","matchedWords":[],"value":"https://www.nature.com/articles/s41598-019-52748-8"}},"_tags":["story","author_bryanrasmussen","story_21880012"],"author":"bryanrasmussen","created_at":"2019-12-25T19:25:01Z","created_at_i":1577301901,"num_comments":0,"objectID":"21880012","points":2,"story_id":21880012,"title":"Contrasting Computational Models of Mate Preference Across 45 Countries","updated_at":"2024-09-20T05:23:56Z","url":"https://www.nature.com/articles/s41598-019-52748-8"},{"_highlightResult":{"author":{"matchLevel":"none","matchedWords":[],"value":"rahimnathwani"},"title":{"fullyHighlighted":false,"matchLevel":"full","matchedWords":["preference","model"],"value":"Direct Preference Optimization: Your Language Model Is a Reward Model"},"url":{"matchLevel":"none","matchedWords":[],"value":"https://arxiv.org/abs/2305.18290"}},"_tags":["story","author_rahimnathwani","story_38972645"],"author":"rahimnathwani","children":[38972665],"created_at":"2024-01-12T19:22:07Z","created_at_i":1705087327,"num_comments":2,"objectID":"38972645","points":1,"story_id":38972645,"title":"Direct Preference Optimization: Your Language Model Is a Reward Model","updated_at":"2024-09-20T16:07:31Z","url":"https://arxiv.org/abs/2305.18290"},{"_highlightResult":{"author":{"matchLevel":"none","matchedWords":[],"value":"optimalsolver"},"title":{"fullyHighlighted":false,"matchLevel":"full","matchedWords":["preference","model"],"value":"Direct Preference Optimization: Your Language Model Is Secretly A Reward Model"},"url":{"matchLevel":"none","matchedWords":[],"value":"https://arxiv.org/abs/2305.18290"}},"_tags":["story","author_optimalsolver","story_37393562"],"author":"optimalsolver","created_at":"2023-09-05T15:58:50Z","created_at_i":1693929530,"num_comments":0,"objectID":"37393562","points":1,"story_id":37393562,"title":"Direct Preference Optimization: Your Language Model Is Secretly A Reward Model","updated_at":"2024-09-20T15:06:14Z","url":"https://arxiv.org/abs/2305.18290"},{"_highlightResult":{"author":{"matchLevel":"none","matchedWords":[],"value":"belter"},"title":{"fullyHighlighted":false,"matchLevel":"full","matchedWords":["preference","model"],"value":"Teaching robots to tidy up based on user preferences using large language models"},"url":{"matchLevel":"none","matchedWords":[],"value":"https://techxplore.com/news/2023-06-robots-tidy-based-user-large.html"}},"_tags":["story","author_belter","story_36248560"],"author":"belter","created_at":"2023-06-08T20:35:21Z","created_at_i":1686256521,"num_comments":0,"objectID":"36248560","points":1,"story_id":36248560,"title":"Teaching robots to tidy up based on user preferences using large language models","updated_at":"2024-09-20T14:14:49Z","url":"https://techxplore.com/news/2023-06-robots-tidy-based-user-large.html"},{"_highlightResult":{"author":{"matchLevel":"none","matchedWords":[],"value":"whyisitso"},"story_text":{"fullyHighlighted":false,"matchLevel":"full","matchedWords":["preference","model"],"value":"Does the VC model really require participating preferences?
If their profit is driven by outsized returns, why do they care about getting a little bit back in the less successful exits?
It just hurts founders and employees but doesn't seem to materially matter for VCs."},"title":{"fullyHighlighted":false,"matchLevel":"full","matchedWords":["preference","model"],"value":"Ask HN: Does the VC model require participating preferences?"}},"_tags":["story","author_whyisitso","story_24384770","ask_hn"],"author":"whyisitso","created_at":"2020-09-05T16:28:34Z","created_at_i":1599323314,"num_comments":0,"objectID":"24384770","points":1,"story_id":24384770,"story_text":"Does the VC model really require participating preferences?
If their profit is driven by outsized returns, why do they care about getting a little bit back in the less successful exits?
It just hurts founders and employees but doesn't seem to materially matter for VCs.","title":"Ask HN: Does the VC model require participating preferences?","updated_at":"2024-09-20T06:58:48Z"},{"_highlightResult":{"author":{"matchLevel":"none","matchedWords":[],"value":"sel1"},"title":{"fullyHighlighted":false,"matchLevel":"full","matchedWords":["preference","model"],"value":"Learning an Urban Air Mobility Encounter Model from Expert Preferences"},"url":{"matchLevel":"none","matchedWords":[],"value":"https://arxiv.org/abs/1907.05575"}},"_tags":["story","author_sel1","story_20446550"],"author":"sel1","created_at":"2019-07-16T00:30:17Z","created_at_i":1563237017,"num_comments":0,"objectID":"20446550","points":1,"story_id":20446550,"title":"Learning an Urban Air Mobility Encounter Model from Expert Preferences","updated_at":"2024-09-20T04:36:23Z","url":"https://arxiv.org/abs/1907.05575"},{"_highlightResult":{"author":{"matchLevel":"none","matchedWords":[],"value":"s_q_b"},"story_text":{"fullyHighlighted":false,"matchLevel":"full","matchedWords":["preference","model"],"value":"My grandfather just passed his 91st birthday. His only regret about getting older is that he will never experience what it is like to skateboard or rollerblade.
I would like to give that to him.
My grandfather is a veteran of World War II, a former FBI field agent who fought against organized crime, a father of four, a grandfather of nine, and not only the most honorable, but also the most consistently jovial person whom I have ever known.
When reflecting over his long life thus far, he was asked if there was anything he wished he had done when he was younger. He looked off into the distance, and after a long while, said, "Well, I'd like to learn to use a skateboard, or maybe rollerblades. Yeah, that might be easier on me."
He smiled and we all laughed, but as we looked back to him, he said wistfully and without a trace of irony, "I really would... But sometimes I have to remind myself I'm not seventeen anymore."
This is where I need your help.
Earlier today a friend sent me a video of her bed-bound grandmother taking VR tours of distant cities to which she had never traveled. As she looked left and right, her expression became one of pure wonderment, an unrestrained smile spreading wide across her face.
I want to provide the same experience for my grandfather.
Does anyone have any recommendations for a performant VR stack, preference of VR make/model (HTC Vive vs. Oculus Rift), or knowledge of high quality content sources?
With regard to price, I am on a limited budget due to recent medical expenses, but I earn a good living. So if there is one option that is clearly better, I would rather sell every luxury item I have than deny him the best experience.
Thank you all for reading."},"title":{"matchLevel":"none","matchedWords":[],"value":"Ask HN: At 91, My Grandfather's Only Regret Is Never Skateboarding. Best VR Rig?"}},"_tags":["story","author_s_q_b","story_12832027","ask_hn"],"author":"s_q_b","children":[12832100,12832125,12832233,12832333,12832405,12832703,12832748,12832850,12832896,12833161,12833222,12833409,12833421,12833543,12833547,12833690,12833731,12833937,12834582,12834604,12834637,12834788,12835484,12835639,12836131,12836607,12836809,12843647,12843866],"created_at":"2016-10-30T20:23:47Z","created_at_i":1477859027,"num_comments":67,"objectID":"12832027","points":184,"story_id":12832027,"story_text":"My grandfather just passed his 91st birthday. His only regret about getting older is that he will never experience what it is like to skateboard or rollerblade.
I would like to give that to him.
My grandfather is a veteran of World War II, a former FBI field agent who fought against organized crime, a father of four, a grandfather of nine, and not only the most honorable, but also the most consistently jovial person whom I have ever known.
When reflecting over his long life thus far, he was asked if there was anything he wished he had done when he was younger. He looked off into the distance, and after a long while, said, "Well, I'd like to learn to use a skateboard, or maybe rollerblades. Yeah, that might be easier on me."
He smiled and we all laughed, but as we looked back to him, he said wistfully and without a trace of irony, "I really would... But sometimes I have to remind myself I'm not seventeen anymore."
This is where I need your help.
Earlier today a friend sent me a video of her bed-bound grandmother taking VR tours of distant cities to which she had never traveled. As she looked left and right, her expression became one of pure wonderment, an unrestrained smile spreading wide across her face.
I want to provide the same experience for my grandfather.
Does anyone have any recommendations for a performant VR stack, preference of VR make/model (HTC Vive vs. Oculus Rift), or knowledge of high quality content sources?
With regard to price, I am on a limited budget due to recent medical expenses, but I earn a good living. So if there is one option that is clearly better, I would rather sell every luxury item I have than deny him the best experience.
Thank you all for reading.","title":"Ask HN: At 91, My Grandfather's Only Regret Is Never Skateboarding. Best VR Rig?","updated_at":"2024-09-19T23:53:01Z"},{"_highlightResult":{"author":{"matchLevel":"none","matchedWords":[],"value":"phlux"},"story_text":{"fullyHighlighted":false,"matchLevel":"full","matchedWords":["preference","model"],"value":"All,
I am exploring different datacenter models and as the HN community is effectively the DC target market, I'd like to find out what you would want out of a datacenter.
Obviously, many startups are heavily reliant on hosted I/PaaS offerings, so for arguments sake - lets leave out comparisons to AWS for the moment.
I am exploring a new full datacenter build out in San Francisco - and would like to find out what the community would most value:
There are effectively three data center models;
Colocation - where you get rack/floor space and you bring in your own equipment and rack it into the DC provided space.
Hosting - Where you timeslice/rent dedicated hardware or resources.
Modular - which can be a combination of the other two, and also offer an option to bring in your own modules. These are more rare as there are few modular datacenters at this point.
I am developing a hybrid model for DC design though, where I try to employ a number of design innovations which would integrate/incorporate things like modular containers with certain aspects of ideas found in the Open Compute platform.
The issues with purpose built systems such as Facebook's open compute platform, is that it is beneficial at scale, but difficult to leverage in a shared hosting model.
I am interested in deploying open compute style systems (racks, UPS, mobo-slot machines, integrated cabling) into container modules.
This would be sold as hosting - or you could choose to buy a number of nodes and have your own cluster. The benefit is ideally that you have a great PUE for a set number of nodes.
Additionally, the site may have traditional colo space as well - either in a separate building - or as rack space in containers.
My question to the HN community, assuming you're buying hosting and not simply leveraging AWS - what model is your preference.
If you could buy open compute-style nodes/infrastructure - would you? Or do you want to buy fully enclosed machines and rack them up in a more traditional fashion?
Any feedback would be appreciated."},"title":{"matchLevel":"none","matchedWords":[],"value":"Ask HN: What do you want out of a datacenter?"},"url":{"matchLevel":"none","matchedWords":[],"value":""}},"_tags":["story","author_phlux","story_2584569","ask_hn"],"author":"phlux","children":[2589051],"created_at":"2011-05-25T16:40:12Z","created_at_i":1306341612,"num_comments":1,"objectID":"2584569","points":7,"story_id":2584569,"story_text":"All,
I am exploring different datacenter models and as the HN community is effectively the DC target market, I'd like to find out what you would want out of a datacenter.
Obviously, many startups are heavily reliant on hosted I/PaaS offerings, so for arguments sake - lets leave out comparisons to AWS for the moment.
I am exploring a new full datacenter build out in San Francisco - and would like to find out what the community would most value:
There are effectively three data center models;
Colocation - where you get rack/floor space and you bring in your own equipment and rack it into the DC provided space.
Hosting - Where you timeslice/rent dedicated hardware or resources.
Modular - which can be a combination of the other two, and also offer an option to bring in your own modules. These are more rare as there are few modular datacenters at this point.
I am developing a hybrid model for DC design though, where I try to employ a number of design innovations which would integrate/incorporate things like modular containers with certain aspects of ideas found in the Open Compute platform.
The issues with purpose built systems such as Facebook's open compute platform, is that it is beneficial at scale, but difficult to leverage in a shared hosting model.
I am interested in deploying open compute style systems (racks, UPS, mobo-slot machines, integrated cabling) into container modules.
This would be sold as hosting - or you could choose to buy a number of nodes and have your own cluster. The benefit is ideally that you have a great PUE for a set number of nodes.
Additionally, the site may have traditional colo space as well - either in a separate building - or as rack space in containers.
My question to the HN community, assuming you're buying hosting and not simply leveraging AWS - what model is your preference.
If you could buy open compute-style nodes/infrastructure - would you? Or do you want to buy fully enclosed machines and rack them up in a more traditional fashion?
Any feedback would be appreciated.","title":"Ask HN: What do you want out of a datacenter?","updated_at":"2024-09-19T17:47:11Z","url":""},{"_highlightResult":{"author":{"matchLevel":"none","matchedWords":[],"value":"try-working"},"story_text":{"fullyHighlighted":false,"matchLevel":"full","matchedWords":["preference","model"],"value":"Hey everyone, I'm launching role-model today: a routing protocol, a reference router runtime, and an extension for Pi that allows for better informed routing decisions.
role-model is mostly deterministic, with fallback to a controller model, that routes requests based on a chosen routing strategy. the protocol is structured around assigning domains and roles to models, where requests sent by consumer applications like Pi have task types to enrich routing metadata and thereby accuracy. you can to run the built-in benchmark to compare performance of models across speed, quality and cost, as well as observed performance on real tasks. I have a diagram on how routing works in [0].
The runtime supports local models, either directly to your local endpoint (LM Studio, llama.cpp etc), or routing between multiple local models via vendored llama-swap.
Since there was another model router post yesterday where people discussed the basics of routing, I will focus on discussing some of the interesting learnings I've made building and testing this:
1. Model routing is essentially trying to predict the future: which model will perform optimally (based on criteria defined by the user) on this request?
2. After you have routed the request, you want to evaluate if it was the right decision or if some other model would have performed better
3. You also realize that having the router assess difficulty (among other things) to make decisions by itself is far from ideal - we'd prefer to have the consumer application work with the router to define what the request needs
4. You also realize that it becomes much easier, decisions become much accurate, and the outcomes of routing becomes more impactful when there is more of a distinction between models
For point 2, I will be launching evals that you can run locally to benchmark models in your pool on the same requests. The outcomes here can then be used for point 1, as input when routing new requests.
For point 3, I've built the pi-role-model package for Pi, which lets the Pi agent inject role_model.intent metadata including difficulty, preferred roles or even specfic model ids, required capabilities (say tool use or image input) and so on. You should be able to customized this further in Pi, and route in additional ways by changing metadata. This is why I've also built the role-model routing protocol.
For point 4, what model routing really does as a second order effect is create a market for specialized models - models that may or may not be smaller, could be cheaper or more expensive, may be locally runnable. It makes little sense to route between two frontier models (GPT 5.5 and Opus 4.8); it makes more sense to route between models where one of the factors of quality, speed, cost is a multiple of the other candidate models, and it makes even more sense to have specialized domain models: code, prose, math and science, visuals and so on. It is at this stage model routing becomes really valuable.
While role-model has a reference runtime that I'm continuously building out (there's lots to do to improve routing, as well as give users more granular control over routing decisions, and also ways to improve cross-model caching and also add techniques like FastContext), the ultimate goal of role-model is for there to be a standard protocol for inference requests that is used by consumer applications, so that the provider, be it a router middleware or an inference provider, will be able to route to a model that strikes the best balance between cost, speed and quality and also respects user choices, and even lets the user control these preferences to use local models for some tasks and allow cloud for others.
Links:
[0] role-model - the case for a model routing protocol: https://try.works/role-model-the-case-for-a-model-routing-pr...
[1] GitHub: https://github.com/try-works/role-model
[2] Docs: https://role-model.dev/"},"title":{"fullyHighlighted":false,"matchLevel":"partial","matchedWords":["model"],"value":"Show HN: role-model, a router for hybrid local/cloud AI"},"url":{"fullyHighlighted":false,"matchLevel":"partial","matchedWords":["model"],"value":"https://github.com/try-works/role-model"}},"_tags":["story","author_try-working","story_48706181","show_hn"],"author":"try-working","children":[48706187,48741424,48775696],"created_at":"2026-06-28T10:46:08Z","created_at_i":1782643568,"num_comments":2,"objectID":"48706181","points":3,"story_id":48706181,"story_text":"Hey everyone, I'm launching role-model today: a routing protocol, a reference router runtime, and an extension for Pi that allows for better informed routing decisions.
role-model is mostly deterministic, with fallback to a controller model, that routes requests based on a chosen routing strategy. the protocol is structured around assigning domains and roles to models, where requests sent by consumer applications like Pi have task types to enrich routing metadata and thereby accuracy. you can to run the built-in benchmark to compare performance of models across speed, quality and cost, as well as observed performance on real tasks. I have a diagram on how routing works in [0].
The runtime supports local models, either directly to your local endpoint (LM Studio, llama.cpp etc), or routing between multiple local models via vendored llama-swap.
Since there was another model router post yesterday where people discussed the basics of routing, I will focus on discussing some of the interesting learnings I've made building and testing this:
1. Model routing is essentially trying to predict the future: which model will perform optimally (based on criteria defined by the user) on this request?
2. After you have routed the request, you want to evaluate if it was the right decision or if some other model would have performed better
3. You also realize that having the router assess difficulty (among other things) to make decisions by itself is far from ideal - we'd prefer to have the consumer application work with the router to define what the request needs
4. You also realize that it becomes much easier, decisions become much accurate, and the outcomes of routing becomes more impactful when there is more of a distinction between models
For point 2, I will be launching evals that you can run locally to benchmark models in your pool on the same requests. The outcomes here can then be used for point 1, as input when routing new requests.
For point 3, I've built the pi-role-model package for Pi, which lets the Pi agent inject role_model.intent metadata including difficulty, preferred roles or even specfic model ids, required capabilities (say tool use or image input) and so on. You should be able to customized this further in Pi, and route in additional ways by changing metadata. This is why I've also built the role-model routing protocol.
For point 4, what model routing really does as a second order effect is create a market for specialized models - models that may or may not be smaller, could be cheaper or more expensive, may be locally runnable. It makes little sense to route between two frontier models (GPT 5.5 and Opus 4.8); it makes more sense to route between models where one of the factors of quality, speed, cost is a multiple of the other candidate models, and it makes even more sense to have specialized domain models: code, prose, math and science, visuals and so on. It is at this stage model routing becomes really valuable.
While role-model has a reference runtime that I'm continuously building out (there's lots to do to improve routing, as well as give users more granular control over routing decisions, and also ways to improve cross-model caching and also add techniques like FastContext), the ultimate goal of role-model is for there to be a standard protocol for inference requests that is used by consumer applications, so that the provider, be it a router middleware or an inference provider, will be able to route to a model that strikes the best balance between cost, speed and quality and also respects user choices, and even lets the user control these preferences to use local models for some tasks and allow cloud for others.
Links:
[0] role-model - the case for a model routing protocol: https://try.works/role-model-the-case-for-a-model-routing-pr...
[1] GitHub: https://github.com/try-works/role-model
[2] Docs: https://role-model.dev/","title":"Show HN: role-model, a router for hybrid local/cloud AI","updated_at":"2026-07-03T20:29:24Z","url":"https://github.com/try-works/role-model"},{"_highlightResult":{"author":{"matchLevel":"none","matchedWords":[],"value":"zimzima"},"story_text":{"fullyHighlighted":false,"matchLevel":"full","matchedWords":["preference","model"],"value":"I built a theme selector for Ghostty that learns which color schemes you prefer in different contexts.
It reads your Mac's ambient light sensor and considers:\n - Time of day (morning, afternoon, evening, night)\n - Ambient light level (dim, office, bright, daylight)\n - System dark/light mode\n - Weather conditions\n - Power source (AC vs battery)\n\n A background daemon watches which themes you use and builds a Bayesian model of your preferences. Over time it learns things like "in dim evening light, they prefer high-contrast dark themes with muted colors."\n\n Once it has enough data, it can generate a custom theme optimized for your current context.\n\n ghostty-ambient # Interactive picker with recommendations\n ghostty-ambient --ideal # Generate optimal theme for current context\n ghostty-ambient --stats # See what it's learned about you\n\n Install: uv tool install ghostty-ambient or pip install ghostty-ambient\n\n GitHub: https://github.com/gezibash/ghostty-ambient\n\n Built for Ghostty (https://ghostty.org) but the Bayesian preference learning could apply to any theming system. It reads your Mac's ambient light sensor and considers:\n - Time of day (morning, afternoon, evening, night)\n - Ambient light level (dim, office, bright, daylight)\n - System dark/light mode\n - Weather conditions\n - Power source (AC vs battery)\n\n A background daemon watches which themes you use and builds a Bayesian model of your preferences. Over time it learns things like "in dim evening light, they prefer high-contrast dark themes with muted colors."\n\n Once it has enough data, it can generate a custom theme optimized for your current context.\n\n ghostty-ambient # Interactive picker with recommendations\n ghostty-ambient --ideal # Generate optimal theme for current context\n ghostty-ambient --stats # See what it's learned about you\n\n Install: uv tool install ghostty-ambient or pip install ghostty-ambient\n\n GitHub: https://github.com/gezibash/ghostty-ambient\n\n Built for Ghostty (https://ghostty.org) but the Bayesian preference learning could apply to any theming system.As teams integrate multiple LLMs - each with different strengths, styles, or cost/latency profiles \u2014 routing the right prompt to the right model becomes a critical part of the application design. But it's still an open problem. Most routing systems fall into two camps:
- Embedding-based routers use intent classifiers \u2014 label a prompt as \u201csupport,\u201d \u201cSQL,\u201d or \u201cmath,\u201d then route to a matching model. This works for simple tasks but breaks down in real conversations. Users shift topics mid-conversation, task boundaries blur, and product changes require retraining classifiers.
- Performance-based routers pick models based on benchmarks like MMLU or MT-Bench, or based on latency or cost curves. But benchmarks often miss what matters in production: domain-specific quality or subjective preferences like \u201cWill legal accept this clause?\u201d
Arch-Router takes a different approach: it decouples route selection from model assignment. Developers can write route policies using a domain-action taxonomy (like "engineering" or "image editing" respectively), and the router maps the prompt (and conversation context) to those policies using a lightweight 1.5B autoregressive model. No retraining, no fragile if/else chains. We built this with input from teams at Twilio and Atlassian. Arch-Router handles intent drift, supports multi-turn conversations, and lets you swap in or out models with a one-line change to the routing policy. Full details are in our paper, but here's a snapshot:
Specs:
- 1.5B params \u2014 runs on a single GPU (or CPU for testing)
- No retraining needed \u2014 point it at any mix of LLMs
- Routing can be cost, latency or quality aware based on your preferences
- Outperforms larger closed models on our conversational routing benchmarks (benchmarks in the paper)
Links:
- ArchGW (open source edge and service proxy for agents ): https://github.com/katanemo/archgw
- Model + code: https://huggingface.co/katanemo/Arch-Router-1.5B
- Paper: https://arxiv.org/abs/2506.16655"},"title":{"fullyHighlighted":false,"matchLevel":"partial","matchedWords":["preference"],"value":"Show HN: Arch-Router \u2013 Aligning LLM Routing with Human Preferences"},"url":{"matchLevel":"none","matchedWords":[],"value":"https://arxiv.org/abs/2506.16655"}},"_tags":["story","author_honorable_coder","story_44774539","show_hn"],"author":"honorable_coder","created_at":"2025-08-03T06:32:04Z","created_at_i":1754202724,"num_comments":0,"objectID":"44774539","points":1,"story_id":44774539,"story_text":"Hi HN \u2014 we're the team behind ArchGW [1], an edge and service proxy for agents written in Rust, and we just recently published our research on LLM Routing: https://arxiv.org/abs/2506.16655
As teams integrate multiple LLMs - each with different strengths, styles, or cost/latency profiles \u2014 routing the right prompt to the right model becomes a critical part of the application design. But it's still an open problem. Most routing systems fall into two camps:
- Embedding-based routers use intent classifiers \u2014 label a prompt as \u201csupport,\u201d \u201cSQL,\u201d or \u201cmath,\u201d then route to a matching model. This works for simple tasks but breaks down in real conversations. Users shift topics mid-conversation, task boundaries blur, and product changes require retraining classifiers.
- Performance-based routers pick models based on benchmarks like MMLU or MT-Bench, or based on latency or cost curves. But benchmarks often miss what matters in production: domain-specific quality or subjective preferences like \u201cWill legal accept this clause?\u201d
Arch-Router takes a different approach: it decouples route selection from model assignment. Developers can write route policies using a domain-action taxonomy (like "engineering" or "image editing" respectively), and the router maps the prompt (and conversation context) to those policies using a lightweight 1.5B autoregressive model. No retraining, no fragile if/else chains. We built this with input from teams at Twilio and Atlassian. Arch-Router handles intent drift, supports multi-turn conversations, and lets you swap in or out models with a one-line change to the routing policy. Full details are in our paper, but here's a snapshot:
Specs:
- 1.5B params \u2014 runs on a single GPU (or CPU for testing)
- No retraining needed \u2014 point it at any mix of LLMs
- Routing can be cost, latency or quality aware based on your preferences
- Outperforms larger closed models on our conversational routing benchmarks (benchmarks in the paper)
Links:
- ArchGW (open source edge and service proxy for agents ): https://github.com/katanemo/archgw
- Model + code: https://huggingface.co/katanemo/Arch-Router-1.5B
- Paper: https://arxiv.org/abs/2506.16655","title":"Show HN: Arch-Router \u2013 Aligning LLM Routing with Human Preferences","updated_at":"2025-08-03T06:37:54Z","url":"https://arxiv.org/abs/2506.16655"},{"_highlightResult":{"author":{"matchLevel":"none","matchedWords":[],"value":"killerdhmo"},"story_text":{"matchLevel":"none","matchedWords":[],"value":""},"title":{"fullyHighlighted":false,"matchLevel":"full","matchedWords":["preference","model"],"value":"MIT choice-modeling software predicts customer preferences for retail"},"url":{"fullyHighlighted":false,"matchLevel":"full","matchedWords":["preference","model"],"value":"http://newsoffice.mit.edu/2015/choice-modeling-software-customer-preference-0710#.VaPlGmhZ7ps.facebook"}},"_tags":["story","author_killerdhmo","story_9882112"],"author":"killerdhmo","children":[9883081],"created_at":"2015-07-14T00:29:31Z","created_at_i":1436833771,"num_comments":1,"objectID":"9882112","points":5,"story_id":9882112,"story_text":"","title":"MIT choice-modeling software predicts customer preferences for retail","updated_at":"2023-09-07T03:12:03Z","url":"http://newsoffice.mit.edu/2015/choice-modeling-software-customer-preference-0710#.VaPlGmhZ7ps.facebook"},{"_highlightResult":{"author":{"matchLevel":"none","matchedWords":[],"value":"legofan94"},"title":{"fullyHighlighted":false,"matchLevel":"full","matchedWords":["preference","model"],"value":"PlayAI's new Dialog model achieves 3:1 preference in human evals"},"url":{"matchLevel":"none","matchedWords":[],"value":"https://play.ht/news/playai-announces-new-benchmarks-playdialog/"}},"_tags":["story","author_legofan94","story_42925110"],"author":"legofan94","children":[42925181,42931128,42939468,42968397,42968444,42968478,42968563,42968893,42969461,42969633,42969669,42969682,42969891,42970116,42970182,42973016],"created_at":"2025-02-03T23:51:39Z","created_at_i":1738626699,"num_comments":55,"objectID":"42925110","points":95,"story_id":42925110,"title":"PlayAI's new Dialog model achieves 3:1 preference in human evals","updated_at":"2025-08-14T22:23:39Z","url":"https://play.ht/news/playai-announces-new-benchmarks-playdialog/"},{"_highlightResult":{"author":{"matchLevel":"none","matchedWords":[],"value":"dares2573"},"title":{"fullyHighlighted":false,"matchLevel":"full","matchedWords":["preference","model"],"value":"Show HN: AI Facial Attractiveness Model Aligned with Human Preferences"},"url":{"matchLevel":"none","matchedWords":[],"value":"https://faceanalysisai.com/"}},"_tags":["story","author_dares2573","story_49773871","show_hn"],"author":"dares2573","children":[49774927,49775492,49775541,49777749],"created_at":"2026-09-20T08:48:35Z","created_at_i":1789894115,"num_comments":6,"objectID":"49773871","points":11,"story_id":49773871,"title":"Show HN: AI Facial Attractiveness Model Aligned with Human Preferences","updated_at":"2026-09-22T09:04:15Z","url":"https://faceanalysisai.com/"},{"_highlightResult":{"author":{"matchLevel":"none","matchedWords":[],"value":"burkaygur"},"story_text":{"fullyHighlighted":false,"matchLevel":"partial","matchedWords":["preference"],"value":"We recently open sourced this public leaderboard and dataset with 50K+ preference data points across 20+ image models. Hoping that this is a useful resource for fine-tuning / preference optimization.
Preference data is updated every day.
Code for frontend: https://github.com/fal-ai/imgsys-public
Preference Data: https://huggingface.co/datasets/fal-ai/imgsys-results"},"title":{"fullyHighlighted":false,"matchLevel":"full","matchedWords":["preference","model"],"value":"Show HN: Open-Source Image Model Leaderboard with Public Preference Data"},"url":{"matchLevel":"none","matchedWords":[],"value":"https://imgsys.org/"}},"_tags":["story","author_burkaygur","story_40205195","show_hn"],"author":"burkaygur","children":[40205254],"created_at":"2024-04-29T23:03:21Z","created_at_i":1714431801,"num_comments":7,"objectID":"40205195","points":4,"story_id":40205195,"story_text":"We recently open sourced this public leaderboard and dataset with 50K+ preference data points across 20+ image models. Hoping that this is a useful resource for fine-tuning / preference optimization.
Preference data is updated every day.
Code for frontend: https://github.com/fal-ai/imgsys-public
Preference Data: https://huggingface.co/datasets/fal-ai/imgsys-results","title":"Show HN: Open-Source Image Model Leaderboard with Public Preference Data","updated_at":"2024-09-20T16:55:54Z","url":"https://imgsys.org/"},{"_highlightResult":{"author":{"matchLevel":"none","matchedWords":[],"value":"honorable_coder"},"story_text":{"fullyHighlighted":false,"matchLevel":"full","matchedWords":["preference","model"],"value":"Hi HN \u2014 we're the team behind Arch (an open-source edge and service proxy for agents)[1], and today we're releasing Arch-Router (https://huggingface.co/katanemo/Arch-Router-1.5B), a 1.5B LLM router model designed to align to user-defined preferences, not public benchmarks and leader boards.
As teams integrate multiple LLMs - each with different strengths, styles, or cost/latency profiles \u2014 routing the right prompt to the right model becomes a critical part of the application design. But it's still an open problem. Most routing systems fall into two camps:
- Embedding-based routers use intent classifiers \u2014 label a prompt as \u201csupport,\u201d \u201cSQL,\u201d or \u201cmath,\u201d then route to a matching model. This works for simple tasks but breaks down in real conversations. Users shift topics mid-conversation, task boundaries blur, and product changes require retraining classifiers.
- Performance-based routers pick models based on benchmarks like MMLU or MT-Bench, or based on latency or cost curves. But benchmarks often can't capture what matters in production: domain-specific quality or subjective evaluation criteria. These routers are often opaque, difficult to debug, and their quality judgments can feel arbitrary, failing to capture the subjective nuance of what a \u201cgood\u201d response actually means for a specific user\u2019s intent.
Arch-Router takes a different approach: route to LLMs based on preferences written as policies in plain ol English.
You write policies like \u201ccontract clauses \u2192 GPT-4o\u201d or \u201cquick travel tips \u2192 Gemini Flash.\u201d The router maps the prompt (and the full conversation context) to those policies using a lightweight 1.5B auto-regressive model. The model is capable to handle intent drift, supports multi-turn conversations, and lets you swap in or out models with a one-line change to the routing policy. To read more about the strength of our model, check out our research paper here: https://arxiv.org/abs/2506.16655
Essentially, Arch-Router splits the routing process into two distinct parts:
Route Selection: This is the what. The system defines a set of human-readable routing policies using a \u201cDomain-Action Taxonomy.\u201d Think of it as a clear API contract written in plain English. A policy isn\u2019t just intent_123; it\u2019s a descriptive label like Domain: \u2018finance\u2019, Action: \u2018analyze earnings report\u2019. The router\u2019s only job is to match the user\u2019s query to the best-fit policy description.\n\n Model Assignment: This is the how. A separate, simple mapping configuration connects each policy to a specific LLM. The finance/"analyze earnings report" policy might map to a powerful model like GPT-4o, while a simpler general/"greeting" policy maps to a faster, cheaper model.\n\n- 1.5B params \u2014 runs on a single GPU (or CPU for testing)
- No retraining needed \u2014 point it at any mix of LLMs
- Outperforms larger closed models on our conversational routing benchmarks (details in the paper)
Links:
[1] Arch Proxy: https://github.com/katanemo/archgw"},"title":{"fullyHighlighted":false,"matchLevel":"full","matchedWords":["preference","model"],"value":"Show HN: 1.5B LLM routing model that aligns to preferences, not leaderboards"},"url":{"matchLevel":"none","matchedWords":[],"value":"https://huggingface.co/katanemo/Arch-Router-1.5B"}},"_tags":["story","author_honorable_coder","story_44597819","show_hn"],"author":"honorable_coder","created_at":"2025-07-17T20:29:12Z","created_at_i":1752784152,"num_comments":0,"objectID":"44597819","points":4,"story_id":44597819,"story_text":"Hi HN \u2014 we're the team behind Arch (an open-source edge and service proxy for agents)[1], and today we're releasing Arch-Router (https://huggingface.co/katanemo/Arch-Router-1.5B), a 1.5B LLM router model designed to align to user-defined preferences, not public benchmarks and leader boards.
As teams integrate multiple LLMs - each with different strengths, styles, or cost/latency profiles \u2014 routing the right prompt to the right model becomes a critical part of the application design. But it's still an open problem. Most routing systems fall into two camps:
- Embedding-based routers use intent classifiers \u2014 label a prompt as \u201csupport,\u201d \u201cSQL,\u201d or \u201cmath,\u201d then route to a matching model. This works for simple tasks but breaks down in real conversations. Users shift topics mid-conversation, task boundaries blur, and product changes require retraining classifiers.
- Performance-based routers pick models based on benchmarks like MMLU or MT-Bench, or based on latency or cost curves. But benchmarks often can't capture what matters in production: domain-specific quality or subjective evaluation criteria. These routers are often opaque, difficult to debug, and their quality judgments can feel arbitrary, failing to capture the subjective nuance of what a \u201cgood\u201d response actually means for a specific user\u2019s intent.
Arch-Router takes a different approach: route to LLMs based on preferences written as policies in plain ol English.
You write policies like \u201ccontract clauses \u2192 GPT-4o\u201d or \u201cquick travel tips \u2192 Gemini Flash.\u201d The router maps the prompt (and the full conversation context) to those policies using a lightweight 1.5B auto-regressive model. The model is capable to handle intent drift, supports multi-turn conversations, and lets you swap in or out models with a one-line change to the routing policy. To read more about the strength of our model, check out our research paper here: https://arxiv.org/abs/2506.16655
Essentially, Arch-Router splits the routing process into two distinct parts:
Route Selection: This is the what. The system defines a set of human-readable routing policies using a \u201cDomain-Action Taxonomy.\u201d Think of it as a clear API contract written in plain English. A policy isn\u2019t just intent_123; it\u2019s a descriptive label like Domain: \u2018finance\u2019, Action: \u2018analyze earnings report\u2019. The router\u2019s only job is to match the user\u2019s query to the best-fit policy description.\n\n Model Assignment: This is the how. A separate, simple mapping configuration connects each policy to a specific LLM. The finance/"analyze earnings report" policy might map to a powerful model like GPT-4o, while a simpler general/"greeting" policy maps to a faster, cheaper model.\n\n- 1.5B params \u2014 runs on a single GPU (or CPU for testing)
- No retraining needed \u2014 point it at any mix of LLMs
- Outperforms larger closed models on our conversational routing benchmarks (details in the paper)
Links:
[1] Arch Proxy: https://github.com/katanemo/archgw","title":"Show HN: 1.5B LLM routing model that aligns to preferences, not leaderboards","updated_at":"2025-09-15T03:39:12Z","url":"https://huggingface.co/katanemo/Arch-Router-1.5B"},{"_highlightResult":{"author":{"matchLevel":"none","matchedWords":[],"value":"achompas"},"title":{"fullyHighlighted":false,"matchLevel":"full","matchedWords":["preference","model"],"value":"AI Model Rankings on Human Preference from Not Diamond's Arena Mode"},"url":{"fullyHighlighted":false,"matchLevel":"full","matchedWords":["preference","model"],"value":"https://www.notdiamond.ai/blog/ai-model-rankings-on-human-preference-from-not-diamonds-arena-mode"}},"_tags":["story","author_achompas","story_41926282"],"author":"achompas","created_at":"2024-10-23T15:50:55Z","created_at_i":1729698655,"num_comments":0,"objectID":"41926282","points":4,"story_id":41926282,"title":"AI Model Rankings on Human Preference from Not Diamond's Arena Mode","updated_at":"2024-10-23T19:19:50Z","url":"https://www.notdiamond.ai/blog/ai-model-rankings-on-human-preference-from-not-diamonds-arena-mode"},{"_highlightResult":{"author":{"matchLevel":"none","matchedWords":[],"value":"honorable_coder"},"title":{"fullyHighlighted":false,"matchLevel":"full","matchedWords":["preference","model"],"value":"1.5B LLM routing model that aligns to preferences, not leaderboards"},"url":{"matchLevel":"none","matchedWords":[],"value":"https://huggingface.co/katanemo/Arch-Router-1.5B"}},"_tags":["story","author_honorable_coder","story_44601318"],"author":"honorable_coder","children":[44601319],"created_at":"2025-07-18T04:50:41Z","created_at_i":1752814241,"num_comments":1,"objectID":"44601318","points":3,"story_id":44601318,"title":"1.5B LLM routing model that aligns to preferences, not leaderboards","updated_at":"2025-07-18T05:53:07Z","url":"https://huggingface.co/katanemo/Arch-Router-1.5B"},{"_highlightResult":{"author":{"matchLevel":"none","matchedWords":[],"value":"kyubyong"},"title":{"fullyHighlighted":false,"matchLevel":"full","matchedWords":["preference","model"],"value":"NeuroBind--Yet Another Model for Predicting Binding Preferences"},"url":{"matchLevel":"none","matchedWords":[],"value":"https://github.com/Kyubyong/neurobind"}},"_tags":["story","author_kyubyong","story_14718147"],"author":"kyubyong","created_at":"2017-07-07T13:30:12Z","created_at_i":1499434212,"num_comments":0,"objectID":"14718147","points":2,"story_id":14718147,"title":"NeuroBind--Yet Another Model for Predicting Binding Preferences","updated_at":"2024-09-20T01:00:37Z","url":"https://github.com/Kyubyong/neurobind"},{"_highlightResult":{"author":{"matchLevel":"none","matchedWords":[],"value":"EliasLittle"},"title":{"fullyHighlighted":false,"matchLevel":"full","matchedWords":["preference","model"],"value":"Evaluating S2S Model Quality with Human Preference Data"},"url":{"matchLevel":"none","matchedWords":[],"value":"https://research.withdavid.ai/blog/st-s2s-eval"}},"_tags":["story","author_EliasLittle","story_49735433"],"author":"EliasLittle","created_at":"2026-09-17T01:42:28Z","created_at_i":1789609348,"num_comments":0,"objectID":"49735433","points":1,"story_id":49735433,"title":"Evaluating S2S Model Quality with Human Preference Data","updated_at":"2026-09-17T01:43:29Z","url":"https://research.withdavid.ai/blog/st-s2s-eval"},{"_highlightResult":{"author":{"matchLevel":"none","matchedWords":[],"value":"irsina2"},"story_text":{"fullyHighlighted":false,"matchLevel":"full","matchedWords":["preference","model"],"value":"Hey Hacker News!
I've been working on an open-source project called LLM Alignment Template, a comprehensive toolkit designed to help researchers, developers, and data scientists align large language models (LLMs) with human values using Reinforcement Learning from Human Feedback (RLHF).
What the project does:
Interactive Web Interface: Easily train models, visualize alignment metrics, and manage alignment with an accessible UI.\n Training with RLHF: Align models effectively to human preferences using feedback loops.\n Explainability: Built-in dashboards to help understand model behavior using SHAP-based explainability tools.\n Data Augmentation & Transfer Learning: Includes tools for advanced preprocessing and utilizes pre-trained models for improved performance.\n Scalable Deployment: Comes with Docker and Kubernetes setup to easily scale deployments.\nKey Features:
Unit tests and E2E tests for quality assurance\nMonitoring and centralized logging using Prometheus and the ELK stack\nDocker and Kubernetes deployment options for easy setup\nModular training scripts for data augmentation, fine-tuning, and RLHF\nWhy it might be interesting:
If you're looking to build an LLM solution and need a strong foundation, this template has all the core tools to get started.\nThe project provides an end-to-end solution, from data augmentation to deployment, making it a great starting point for those interested in AI ethics and model alignment."},"title":{"fullyHighlighted":false,"matchLevel":"partial","matchedWords":["model"],"value":"Show HN: LLM Alignment Template \u2013 Aligning Language Models with Human Feedback"},"url":{"matchLevel":"none","matchedWords":[],"value":"https://github.com/astorfi/LLM-Alignment-Project-Template"}},"_tags":["story","author_irsina2","story_42220186","show_hn"],"author":"irsina2","created_at":"2024-11-23T10:12:58Z","created_at_i":1732356778,"num_comments":0,"objectID":"42220186","points":9,"story_id":42220186,"story_text":"Hey Hacker News!
I've been working on an open-source project called LLM Alignment Template, a comprehensive toolkit designed to help researchers, developers, and data scientists align large language models (LLMs) with human values using Reinforcement Learning from Human Feedback (RLHF).
What the project does:
Interactive Web Interface: Easily train models, visualize alignment metrics, and manage alignment with an accessible UI.\n Training with RLHF: Align models effectively to human preferences using feedback loops.\n Explainability: Built-in dashboards to help understand model behavior using SHAP-based explainability tools.\n Data Augmentation & Transfer Learning: Includes tools for advanced preprocessing and utilizes pre-trained models for improved performance.\n Scalable Deployment: Comes with Docker and Kubernetes setup to easily scale deployments.\nKey Features:
Unit tests and E2E tests for quality assurance\nMonitoring and centralized logging using Prometheus and the ELK stack\nDocker and Kubernetes deployment options for easy setup\nModular training scripts for data augmentation, fine-tuning, and RLHF\nWhy it might be interesting:
If you're looking to build an LLM solution and need a strong foundation, this template has all the core tools to get started.\nThe project provides an end-to-end solution, from data augmentation to deployment, making it a great starting point for those interested in AI ethics and model alignment.","title":"Show HN: LLM Alignment Template \u2013 Aligning Language Models with Human Feedback","updated_at":"2024-11-25T13:33:29Z","url":"https://github.com/astorfi/LLM-Alignment-Project-Template"}],"hitsPerPage":50,"nbHits":753,"nbPages":16,"page":0,"params":"query=Preference+Model&tags=story&hitsPerPage=50&advancedSyntax=true&analyticsTags=backend","processingTimeMS":17,"processingTimingsMS":{"_request":{"roundTrip":22},"afterFetch":{"format":{"highlighting":3,"total":3},"merge":{"mergeLoop":{"prepareNextHit":2,"total":2},"total":2},"total":2},"fetch":{"query":5,"scanning":6,"total":13},"total":17},"query":"Preference Model","serverTimeMS":21}