Noctua Fan Swap on a Calnex Rb/GPS Frequency Reference
3 by aray07 | 0 comments on Hacker News.
Saturday, October 31, 2026
Friday, October 30, 2026
Wednesday, October 28, 2026
Tuesday, October 27, 2026
Monday, October 26, 2026
Sunday, October 25, 2026
New top story on Hacker News: Show HN: ReVault – new secure archive format
Show HN: ReVault – new secure archive format
3 by bsutto | 1 comments on Hacker News.
reVault is a new, secure-by-default archive format called a lockbox, designed to be simple to use. I've been working on reVault for about 2 years. It was originally written in Dart, but with the improvements in AI, I decided to port it to Rust using AI. It's now in beta, so I'm looking for feedback, particularly around the different language bindings. A lockbox is encrypted, signed and compressed by default. Each can be disabled selectively. - Hybrid post-quantum or passphrase encryption, including metadata. - Zstd compression. - Pure Rust libraries wherever possible. We intend to phase out remaining C code as alternatives become available. The project has been scanned using OpenAI's Daybreak cyber security tools. The manual is a work in progress: https://ift.tt/iLdvFwM reVault supports Linux, macOS and Windows, x64 and arm64, plus Wasm. The core is Rust, with bindings for C/C++, Java, Dart, Python, Ruby, JS/TypeScript, Go, C#, Kotlin, Lua, PHP and Swift. Within a lockbox you can store: - Files, directories, permissions and symlinks. - Variables. - Forms. Variables, forms and the lockbox itself can have descriptions. reVault uses a portable vault (just another lockbox with structured content) to hold your keys, contacts' keys, lockbox aliases and form definitions. It ties into your platform's secure store, so logging into your desktop unlocks your vault. You don't have to enter a passphrase every time you access a lockbox. You can disable this. There are lots of shortcuts. Creating a lockbox stores an alias in your vault, based on its filename or a name you choose: lbx create /some/path/mybox.lbox lbx mybox open --duration 1h lbx mybox var set KEY VALUE Set a default lockbox and you can leave its name out: lbx session default mybox lbx var get KEY You can mirror a directory tree into a lockbox and update it when the directory changes: lbx store.lbox mirrors project create --from ./project --to /projects/project lbx store.lbox mirrors update The API lets you read, write and seek within files without extracting them to the filesystem. Each lockbox has a table of contents, so extracting a single file from a multi-gigabyte lockbox is fast. Recovery Updates are transactional, with automatic rollback for failed updates. The format is designed for recovery: content is stored in pages, allowing unaffected content to be recovered if part of a lockbox is corrupted. Sharing An encrypted lockbox is more useful if you can share it. reVault has an open-source keyserver for exchanging public keys, with validation through a second trusted channel. It's still experimental. OnePub has donated a server to run a single keyserver. If you're interested in donating resources, I'd like to bring up a second redundant server. It's pure Rust, uses under 100 MB RSS and supports rate limits. To publish your default profile's public key: lbx vault profiles email you@example.com lbx vault profiles publish Give the publish code to your contact. They can import your key: lbx vault contacts receive
Display your profile's fingerprint and exchange it through a second trusted channel: lbx vault profiles fingerprint
Repeat the process in the opposite direction to get their key. Once you have it, grant them access to your default lockbox: lbx access grant alice
You can then email the encrypted lockbox to Alice. To install the CLI and get started: cargo install revault_cli
lbx vault init
lbx completion install
Completion installation detects your shell. To diagnose issues: lbx doctor
The language bindings are published to their respective package repositories. I'm looking for constructive criticism of the project on its own merits, particularly the APIs and language bindings. I'm not interested in a debate about using AI.
3 by bsutto | 1 comments on Hacker News.
reVault is a new, secure-by-default archive format called a lockbox, designed to be simple to use. I've been working on reVault for about 2 years. It was originally written in Dart, but with the improvements in AI, I decided to port it to Rust using AI. It's now in beta, so I'm looking for feedback, particularly around the different language bindings. A lockbox is encrypted, signed and compressed by default. Each can be disabled selectively. - Hybrid post-quantum or passphrase encryption, including metadata. - Zstd compression. - Pure Rust libraries wherever possible. We intend to phase out remaining C code as alternatives become available. The project has been scanned using OpenAI's Daybreak cyber security tools. The manual is a work in progress: https://ift.tt/iLdvFwM reVault supports Linux, macOS and Windows, x64 and arm64, plus Wasm. The core is Rust, with bindings for C/C++, Java, Dart, Python, Ruby, JS/TypeScript, Go, C#, Kotlin, Lua, PHP and Swift. Within a lockbox you can store: - Files, directories, permissions and symlinks. - Variables. - Forms. Variables, forms and the lockbox itself can have descriptions. reVault uses a portable vault (just another lockbox with structured content) to hold your keys, contacts' keys, lockbox aliases and form definitions. It ties into your platform's secure store, so logging into your desktop unlocks your vault. You don't have to enter a passphrase every time you access a lockbox. You can disable this. There are lots of shortcuts. Creating a lockbox stores an alias in your vault, based on its filename or a name you choose: lbx create /some/path/mybox.lbox lbx mybox open --duration 1h lbx mybox var set KEY VALUE Set a default lockbox and you can leave its name out: lbx session default mybox lbx var get KEY You can mirror a directory tree into a lockbox and update it when the directory changes: lbx store.lbox mirrors project create --from ./project --to /projects/project lbx store.lbox mirrors update The API lets you read, write and seek within files without extracting them to the filesystem. Each lockbox has a table of contents, so extracting a single file from a multi-gigabyte lockbox is fast. Recovery Updates are transactional, with automatic rollback for failed updates. The format is designed for recovery: content is stored in pages, allowing unaffected content to be recovered if part of a lockbox is corrupted. Sharing An encrypted lockbox is more useful if you can share it. reVault has an open-source keyserver for exchanging public keys, with validation through a second trusted channel. It's still experimental. OnePub has donated a server to run a single keyserver. If you're interested in donating resources, I'd like to bring up a second redundant server. It's pure Rust, uses under 100 MB RSS and supports rate limits. To publish your default profile's public key: lbx vault profiles email you@example.com lbx vault profiles publish Give the publish code to your contact. They can import your key: lbx vault contacts receive
Saturday, October 24, 2026
Friday, October 23, 2026
New top story on Hacker News: Show HN: Swurf – Swipe through the real web
Show HN: Swurf – Swipe through the real web
13 by webtjb096 | 3 comments on Hacker News.
Hello HN, Swurf allows you to swipe through the real web one page after another. It learns what you like and doesn’t show you the same page twice. The idea came naturally: there’s a seemingly infinite amount of high quality blog posts, random topic rabbit holes, and trending or cool sites that exist which the average person simply will never load on their browser. Swurf solves that by allowing you to rapidly traverse a constantly growing and learning feed of sites that it thinks you’d want to view. Quickly sift through endless pages and find things you like - effortlessly. One of the core drivers for making this project was the desire for real content. The project started as a simple tool for myself to find new and interesting things to learn and read about. Hours spent scrolling on my phone - while it still has its place - is increasingly losing its sense of reward. It’s genuinely hard to put into words exactly how swurfin’ feels but it’s almost refreshing in a sense. I encourage you to try it out and provide feedback. At this time there are no accounts and you can jump in and start swurfin’ immediately. There is a three dot menu button on the bottom right which allows you to submit links, save links, control adult content, and search similar topics. One of the things that’s great about Swurf is that its almost like a group browsing session (though entirely private) where everyone else who is spending time hunting down great reads or solid sites is helping massage the feed for you and you for them - all within your own set of interests. Enjoy and please ask any questions you have as theres plenty of room to grow.
13 by webtjb096 | 3 comments on Hacker News.
Hello HN, Swurf allows you to swipe through the real web one page after another. It learns what you like and doesn’t show you the same page twice. The idea came naturally: there’s a seemingly infinite amount of high quality blog posts, random topic rabbit holes, and trending or cool sites that exist which the average person simply will never load on their browser. Swurf solves that by allowing you to rapidly traverse a constantly growing and learning feed of sites that it thinks you’d want to view. Quickly sift through endless pages and find things you like - effortlessly. One of the core drivers for making this project was the desire for real content. The project started as a simple tool for myself to find new and interesting things to learn and read about. Hours spent scrolling on my phone - while it still has its place - is increasingly losing its sense of reward. It’s genuinely hard to put into words exactly how swurfin’ feels but it’s almost refreshing in a sense. I encourage you to try it out and provide feedback. At this time there are no accounts and you can jump in and start swurfin’ immediately. There is a three dot menu button on the bottom right which allows you to submit links, save links, control adult content, and search similar topics. One of the things that’s great about Swurf is that its almost like a group browsing session (though entirely private) where everyone else who is spending time hunting down great reads or solid sites is helping massage the feed for you and you for them - all within your own set of interests. Enjoy and please ask any questions you have as theres plenty of room to grow.
Thursday, October 22, 2026
Wednesday, October 21, 2026
Tuesday, October 20, 2026
Monday, October 19, 2026
New top story on Hacker News: RuneScape Bench: MMOs for Agents [video]
RuneScape Bench: MMOs for Agents [video]
40 by evakhoury | 15 comments on Hacker News.
Related: https://ift.tt/cXg0zIo
40 by evakhoury | 15 comments on Hacker News.
Related: https://ift.tt/cXg0zIo
Sunday, October 18, 2026
Saturday, October 17, 2026
Friday, October 16, 2026
Thursday, October 15, 2026
Wednesday, October 14, 2026
New top story on Hacker News: Show HN: I wrote an email app that accepts only one sender per inbox
Show HN: I wrote an email app that accepts only one sender per inbox
10 by edpin2000 | 5 comments on Hacker News.
Infinite inboxes, including prefixes where you don't have to create inboxes before you use them -- they self create when you add a suffix. Three types of inboxes: - One sender. Locks to the first domain to send email (except gmail, outlook, etc where it locks to an email account for obivous reasons. It bounces everyone else. - Temporary. Lasts for a while (24h, 7days, etc) then bounces. Received messages stay or go as per your choice. - Prefix. The most powerful. Declare a prefix like "service" and then invent any suffix, like "service-netflix" or "service$bank" and it creates a new one-sender inbox locked to whoever wrote first. Bring you own domain. Bring your own AI (MCP server ready). Forwarding. Please help me test it.
10 by edpin2000 | 5 comments on Hacker News.
Infinite inboxes, including prefixes where you don't have to create inboxes before you use them -- they self create when you add a suffix. Three types of inboxes: - One sender. Locks to the first domain to send email (except gmail, outlook, etc where it locks to an email account for obivous reasons. It bounces everyone else. - Temporary. Lasts for a while (24h, 7days, etc) then bounces. Received messages stay or go as per your choice. - Prefix. The most powerful. Declare a prefix like "service" and then invent any suffix, like "service-netflix" or "service$bank" and it creates a new one-sender inbox locked to whoever wrote first. Bring you own domain. Bring your own AI (MCP server ready). Forwarding. Please help me test it.
Tuesday, October 13, 2026
New top story on Hacker News: Show HN: Deslop – Oxlint rules for AI-generated TypeScript
Show HN: Deslop – Oxlint rules for AI-generated TypeScript
5 by maharshi365 | 0 comments on Hacker News.
Catches unsafe types, duplicate code, and low-value tests.
5 by maharshi365 | 0 comments on Hacker News.
Catches unsafe types, duplicate code, and low-value tests.
Monday, October 12, 2026
Sunday, October 11, 2026
Saturday, October 10, 2026
Friday, October 9, 2026
Thursday, October 8, 2026
Wednesday, October 7, 2026
Tuesday, October 6, 2026
Monday, October 5, 2026
New top story on Hacker News: Show HN: Nightwatch – a Mac menu-bar app that tells you when tonight is clear
Show HN: Nightwatch – a Mac menu-bar app that tells you when tonight is clear
8 by delphidolphin | 1 comments on Hacker News.
Here in the north of England clear nights are at a premium, and it's never guaranteed that you'll be able to see anything but cloud. That's why I developed Nightwatch, so that I could be notified when one of those precious few nights where I could see into the heavens was available. Finding out whether tonight was worth setting up meant reading forecast charts every evening, usually to learn that it was not. Nightwatch does that reading for me. It sits in the Mac's menu bar and refreshes the forecast every 30 minutes. An hour before sunset it sends one notification, and only if there is an unbroken run of clear sky in astronomical darkness long enough to image in. Otherwise it says nothing: no news means "clouds". I hope this is useful for other hobbyist stargazers too! When I was a child I wanted desperately to be an astronaut or a dinosaur. I'm now old enough to be the latter and was born in a country where the former was never an option. I've had a fascination with the stars since a a friend of my father gifted me a telescope, and have faded memories of sharing time with my father staring at the stars. My father passed away a few years back and the memory of those nights came back, along with the want to share the same experience with my kids and grandchild. The app uses two forecasts: Apple's WeatherKit is the main source and Open-Meteo is a second opinion. When they disagree, the alert says so instead of averaging them. Thin high cloud counts for half. Stacking many short frames works through a veil of cirrus, so treating it as full cloud threw away usable nights. It works on your own horizon: hills come from terrain data, and you set the houses and trees in each direction, so they are accounted for when it says what is up. It works out, a year ahead, when the Moon will pass in front of a planet, a bright star or the Pleiades as seen from your site. You don't need an account. What leaves the Mac is the coordinates of the place being forecast, sent to the forecast and sky-survey services listed in PRIVACY.md. Limits: Mac only, and the aurora alerts are UK only. The worldwide light-pollution data is new this week. I have checked it against Britain's finer grid and a handful of cities, so I would especially like to hear from people elsewhere whether the darkness it suggests for your site looks right.
8 by delphidolphin | 1 comments on Hacker News.
Here in the north of England clear nights are at a premium, and it's never guaranteed that you'll be able to see anything but cloud. That's why I developed Nightwatch, so that I could be notified when one of those precious few nights where I could see into the heavens was available. Finding out whether tonight was worth setting up meant reading forecast charts every evening, usually to learn that it was not. Nightwatch does that reading for me. It sits in the Mac's menu bar and refreshes the forecast every 30 minutes. An hour before sunset it sends one notification, and only if there is an unbroken run of clear sky in astronomical darkness long enough to image in. Otherwise it says nothing: no news means "clouds". I hope this is useful for other hobbyist stargazers too! When I was a child I wanted desperately to be an astronaut or a dinosaur. I'm now old enough to be the latter and was born in a country where the former was never an option. I've had a fascination with the stars since a a friend of my father gifted me a telescope, and have faded memories of sharing time with my father staring at the stars. My father passed away a few years back and the memory of those nights came back, along with the want to share the same experience with my kids and grandchild. The app uses two forecasts: Apple's WeatherKit is the main source and Open-Meteo is a second opinion. When they disagree, the alert says so instead of averaging them. Thin high cloud counts for half. Stacking many short frames works through a veil of cirrus, so treating it as full cloud threw away usable nights. It works on your own horizon: hills come from terrain data, and you set the houses and trees in each direction, so they are accounted for when it says what is up. It works out, a year ahead, when the Moon will pass in front of a planet, a bright star or the Pleiades as seen from your site. You don't need an account. What leaves the Mac is the coordinates of the place being forecast, sent to the forecast and sky-survey services listed in PRIVACY.md. Limits: Mac only, and the aurora alerts are UK only. The worldwide light-pollution data is new this week. I have checked it against Britain's finer grid and a handful of cities, so I would especially like to hear from people elsewhere whether the darkness it suggests for your site looks right.
Sunday, October 4, 2026
Saturday, October 3, 2026
Friday, October 2, 2026
New top story on Hacker News: Show HN: We built a real-time desktop companion for patient's telehealth visits
Show HN: We built a real-time desktop companion for patient's telehealth visits
6 by nikilr | 2 comments on Hacker News.
Hey HN, Some friends and I built an app called Sidekick that joins your telehealth visits in real-time. It explains medical terms and suggests useful questions, including ones you may not have thought to ask. After the visit, it provides a summary including next steps, so you can focus on the conversation instead of taking notes. The idea of Sidekick came from our team's experiences with healthcare. Our founder, Rahul, had stage three cancer at 17. His mom had to manage everything around his care: taking notes through treatment, figuring out what questions to ask, etc. A decade later and he’s now in remission, but still sees a lot of specialists and deals with the same challenges. Another teammate became a caregiver after his dad had a severe stroke. He found himself taking notes, looking up medical terms, and trying to keep track of what was said during every doctor visit. Others have dealt with situations like this when navigating pancreatitis and other serious health issues. These experiences got us interested in how we might be able to reduce the burden on the patient or caregiver. A lot of products focus on simply recording or summarizing the visit. That's useful and a step in the right direction, but we kept coming back to the fact that a lot of the challenge originates while the conversation is still happening. If you can catch something in the moment, you can address it with your clinician immediately and potentially lead to better outcomes. If you realize you missed something after the visit, then you have to try to get another appointment (and in SF I often can't get another doctor visit till 2 months later). So we built Sidekick, to help patients and caregivers during a visit in real-time to translate medical jargon and suggest useful questions, ideally without becoming a distraction. If you want to try it, you can download it on our website. Sidekick is desktop-only for now, but we’re working on a mobile app for in-person visits. We would be especially interested to hear feedback on the live part of the experience. What are your thoughts on the balance between live help vs after-visit? Would something pre-visit be more helpful?
6 by nikilr | 2 comments on Hacker News.
Hey HN, Some friends and I built an app called Sidekick that joins your telehealth visits in real-time. It explains medical terms and suggests useful questions, including ones you may not have thought to ask. After the visit, it provides a summary including next steps, so you can focus on the conversation instead of taking notes. The idea of Sidekick came from our team's experiences with healthcare. Our founder, Rahul, had stage three cancer at 17. His mom had to manage everything around his care: taking notes through treatment, figuring out what questions to ask, etc. A decade later and he’s now in remission, but still sees a lot of specialists and deals with the same challenges. Another teammate became a caregiver after his dad had a severe stroke. He found himself taking notes, looking up medical terms, and trying to keep track of what was said during every doctor visit. Others have dealt with situations like this when navigating pancreatitis and other serious health issues. These experiences got us interested in how we might be able to reduce the burden on the patient or caregiver. A lot of products focus on simply recording or summarizing the visit. That's useful and a step in the right direction, but we kept coming back to the fact that a lot of the challenge originates while the conversation is still happening. If you can catch something in the moment, you can address it with your clinician immediately and potentially lead to better outcomes. If you realize you missed something after the visit, then you have to try to get another appointment (and in SF I often can't get another doctor visit till 2 months later). So we built Sidekick, to help patients and caregivers during a visit in real-time to translate medical jargon and suggest useful questions, ideally without becoming a distraction. If you want to try it, you can download it on our website. Sidekick is desktop-only for now, but we’re working on a mobile app for in-person visits. We would be especially interested to hear feedback on the live part of the experience. What are your thoughts on the balance between live help vs after-visit? Would something pre-visit be more helpful?
Thursday, October 1, 2026
New top story on Hacker News: Show HN: Open-source model routing for coding agents at Astra-level performance
Show HN: Open-source model routing for coding agents at Astra-level performance
17 by adchurch | 1 comments on Hacker News.
A few months ago we started building a model router for coding agents because we thought we could outperform any single model with an ensemble approach. Recently we’ve achieved that milestone and I want to talk about how we did it. First of all, a quick explanation: the Weave Router ( https://ift.tt/8pon6xR ) plugs into any coding agent (e.g. Claude Code or Codex) and intelligently switches between LLMs. So, for example, Astra handles tricky debugging or complex system design tasks, and Deepseek v4 Flash handles simple frontend updates. What we’re announcing today is our new routing model, which we’re calling Weave Router 2.0. We benchmarked 2.0 against GPT-6 Astra on Terminal Bench 4.0 and SWE Atlas. On both benchmarks, the router had equivalent pass rates. On Terminal Bench, the router hit 52% of Astra’s cost, and completed tasks 2.2x faster. On SWE Atlas, the router cost 54% as much as Astra and ran 2.5x faster. (Full results on our website at https://ift.tt/ErVtglT !) It turns out training a model to route effectively - taking into consideration model capabilities, costs, cache awareness, and more - is a really hard problem! I want to talk about three ways we were able to improve so much over the last few months: 1) a new architecture, 2) larger training data set size, and 3) smarter cache-eviction impact calculation. 1) a new architecture. Our initial approach used an RL model without many priors. While RL is still an important part of the story, the cost of fully exploring the space of routing decisions is very high, so we’ve taken some shortcuts that have significantly improved performance. Consider how large the search space for the routing problem is. Take a typical coding agent session, with ~100 agent turns (i.e. 100 LLM API calls). Technically there are 100 chances to select a model. If we assume a roster of ~10 models (of course there are lots more but we can remove any that are Pareto dominated), then there are 10^100 possible paths through that session. We simply cannot explore all of them! So that's why clever tricks to shrink this space are so important. In particular: we trained a hidden Markov model to trace the session state, then a classifier maps the session to one of a few buckets of similar models. Using the HMM allows us to evaluate not just where a session is currently, but how it got there . We've gotten significantly better performance on bucket selection by incorporating that information - we believe this is because two sessions that might look quite similar to a naive classifier are much better distinguished by this HMM approach. Using this HMM + classifier to select a bucket first significantly shrinks the space to explore, by throwing out most models that could not reasonably serve the given session. This rearchitecture was the single biggest performance unlock! 2) larger training data set size (much less technically interesting but still an important part of the story). By using frontier LLMs to help us label a larger and more diverse set of coding agent sessions, we were able to bootstrap the two models discussed in 1) to a better state, while also providing even richer reward signals for RL. 3) smarter cache-eviction impact calculation. One of the hardest parts of routing well (if you care about saving money) is using the model caches intelligently. We built a subsystem that can calculate the expected value of switching models (and thus paying a high one-time cost to fill up a different cache) much more accurately, helping us avoid costly and unnecessary switches in more cases, while still switching when the benefit outweighs the cost. This is where most of our improvement on cost has come from. We still have a lot of room to continue to improve (we won’t rest until we’re consistently beating Astra/Fable, not just tying!) but matching frontier model performance was a huge milestone for our routing model, and in my opinion validates our initial hypothesis that an ensemble of models can do better than any single model ever could. Our router is open source ( https://ift.tt/8pon6xR ) so anyone can try it out. Or if you prefer you can use our hosted version ( https://ift.tt/ErVtglT ).
17 by adchurch | 1 comments on Hacker News.
A few months ago we started building a model router for coding agents because we thought we could outperform any single model with an ensemble approach. Recently we’ve achieved that milestone and I want to talk about how we did it. First of all, a quick explanation: the Weave Router ( https://ift.tt/8pon6xR ) plugs into any coding agent (e.g. Claude Code or Codex) and intelligently switches between LLMs. So, for example, Astra handles tricky debugging or complex system design tasks, and Deepseek v4 Flash handles simple frontend updates. What we’re announcing today is our new routing model, which we’re calling Weave Router 2.0. We benchmarked 2.0 against GPT-6 Astra on Terminal Bench 4.0 and SWE Atlas. On both benchmarks, the router had equivalent pass rates. On Terminal Bench, the router hit 52% of Astra’s cost, and completed tasks 2.2x faster. On SWE Atlas, the router cost 54% as much as Astra and ran 2.5x faster. (Full results on our website at https://ift.tt/ErVtglT !) It turns out training a model to route effectively - taking into consideration model capabilities, costs, cache awareness, and more - is a really hard problem! I want to talk about three ways we were able to improve so much over the last few months: 1) a new architecture, 2) larger training data set size, and 3) smarter cache-eviction impact calculation. 1) a new architecture. Our initial approach used an RL model without many priors. While RL is still an important part of the story, the cost of fully exploring the space of routing decisions is very high, so we’ve taken some shortcuts that have significantly improved performance. Consider how large the search space for the routing problem is. Take a typical coding agent session, with ~100 agent turns (i.e. 100 LLM API calls). Technically there are 100 chances to select a model. If we assume a roster of ~10 models (of course there are lots more but we can remove any that are Pareto dominated), then there are 10^100 possible paths through that session. We simply cannot explore all of them! So that's why clever tricks to shrink this space are so important. In particular: we trained a hidden Markov model to trace the session state, then a classifier maps the session to one of a few buckets of similar models. Using the HMM allows us to evaluate not just where a session is currently, but how it got there . We've gotten significantly better performance on bucket selection by incorporating that information - we believe this is because two sessions that might look quite similar to a naive classifier are much better distinguished by this HMM approach. Using this HMM + classifier to select a bucket first significantly shrinks the space to explore, by throwing out most models that could not reasonably serve the given session. This rearchitecture was the single biggest performance unlock! 2) larger training data set size (much less technically interesting but still an important part of the story). By using frontier LLMs to help us label a larger and more diverse set of coding agent sessions, we were able to bootstrap the two models discussed in 1) to a better state, while also providing even richer reward signals for RL. 3) smarter cache-eviction impact calculation. One of the hardest parts of routing well (if you care about saving money) is using the model caches intelligently. We built a subsystem that can calculate the expected value of switching models (and thus paying a high one-time cost to fill up a different cache) much more accurately, helping us avoid costly and unnecessary switches in more cases, while still switching when the benefit outweighs the cost. This is where most of our improvement on cost has come from. We still have a lot of room to continue to improve (we won’t rest until we’re consistently beating Astra/Fable, not just tying!) but matching frontier model performance was a huge milestone for our routing model, and in my opinion validates our initial hypothesis that an ensemble of models can do better than any single model ever could. Our router is open source ( https://ift.tt/8pon6xR ) so anyone can try it out. Or if you prefer you can use our hosted version ( https://ift.tt/ErVtglT ).