Comments URL: https://news.ycombinator.com/item?id=50380704

Points: 85

# Comments: 30

However, the issue with this suite is, it's dependent and running on Java with a old Console UI (not updated in years), that was not built to support tons and millions of data and has pretty bad developer experience. For a long time, we were hitting Out of Memory on even high specs laptop while developing and the worst part is while this happens all the persists data keep getting lost.

We were honestly frustrated with this, and thought what if we built the Firebase Emulator Suite the right way, as it always should have been.

Introducing Firenook, a faster drop-in replacement for Firebase Emulator Suite for Firebase developers. It's 40x faster writes and needs 10x less RAM to run it. We took the right approach of first building it Rust as a single engine, so that there will be no Java dependencies and built from scratch a new Console UI using modern web stacks.

The result is, now we were able to run 2 millions datas loaded to Firenook with only taking around 2GB ram compared to 22GB of the official suite. Things are way faster and rebuilt Console UI gives best developer experience from logging Firebase Functions, to editing Firestore documents in JSON and many more.

We developed this throughly for around 2 months now and ran this internally for a month before we decided it's time to release it to public so that other devs can make use of Firenook and also contribute it to make it better.

Firenook is still new and I'm sure it misses things. So please try it our on your project and let us if there are any issues.

Thank you.

Comments URL: https://news.ycombinator.com/item?id=50378643

Points: 126

# Comments: 17

Comments URL: https://news.ycombinator.com/item?id=50369895

Points: 249

# Comments: 38

Video models today are too slow/expensive and lack sufficient controls. Creativity is an iterative process. You make, you tweak, you get better. In our opinion, they're not usable yet for storytelling.

When we set out to build Linum v3 (our next video model) earlier this year, the primary goal was to bring the cost envelope of inference down significantly. If our model is fast and cheap, it's the first step (of many) towards useful creative tools.

Typically, you see "latent diffusion models". A VAE (Variational Autoencoder) compresses the input data into a smaller "latent space" and the diffusion model learns how to generate within this subspace.

Over the past few months, we've been publishing our work on efficient pixel-space diffusion models. These throw away the VAE, having the diffusion model learn a latent-space optimized for its exact task.

This serves 3 purposes: 1) Simplify Architecture (no 2-stage model training) 2) Faster Convergence (11x faster convergence vs. our latent diffusion counterpart) 3) Cheaper Inference (4x more aggressive token reduction in attention context-windows, without significant loss in quality)

This past month we were pressure testing our latest architecture (Pyramid-JiT) by training a text-to-image model longer than the typical one-off experiment. Specifically, we trained it for ~$20K on 660M samples.

We were surprised by the quality of results, so we thought we'd release it as open weights for the community (Apache 2.0). The research itself is already public, on our website.

For us, it's a big stepping stone towards cheaper video models, since pretty much all video models start as image models. We like to say these models have to learn nouns before they learn verbs.

Comments URL: https://news.ycombinator.com/item?id=50368208

Points: 240

# Comments: 22

Comments URL: https://news.ycombinator.com/item?id=50366929

Points: 116

# Comments: 82

As of today: it's free on Android, the iOS version is still waiting to be reviewed by Apple. It's available in six popular languages, but adding more languages is just a matter of time and demand.My long term vision if it takes off: weekly cases, private parties like a team-building escape game, an MCP so a game master can build their own scenario from ChatGPT or Claude, and maybe tourism offices since players travel the world.

Comments URL: https://news.ycombinator.com/item?id=50358352

Points: 127

# Comments: 69

1. Traditional fraud solutions use ML models such as XGBoost to detect fraud, but these models are dependent on quality labeled data which many businesses do not have. We address this by using “System One” decision models in additional to traditional models such as XGBoost and logistic regression. This allows you to incorporate plain language questions such as “Is this a throwaway email” into your decisioning, bypassing the need for large amounts of labeled data. This also allows you to capture aspects of fraud that are more unstructured, which is something that XGBoost models struggle with.

2. Fraud teams are facing off with live players that are constantly changing their behavior to bypass the detection logic they have in place. You can train a detection model or define a ruleset once, but it will become obsolete in a matter of weeks as fraudsters figure out what your detection logic is. This means that you need a team regularly training models and updating rules based on new fraud patterns. Typically this takes a large data science/ML team. Kairos addresses this by introducing an MCP server, so that your agent (Claude Code, Cursor, Codex, etc.) can define a decision workflow, train models on your data, and pick the variant that performs best.

You can self-host Kairos with Docker or use the hosted version at https://kairos.finic.ai

Comments URL: https://news.ycombinator.com/item?id=50358314

Points: 104

# Comments: 36

Comments URL: https://news.ycombinator.com/item?id=50353386

Points: 71

# Comments: 26

There are still a lot of commands I want to add, but I hope someone finds it useful in the meantime :)

Comments URL: https://news.ycombinator.com/item?id=50347552

Points: 72

# Comments: 53

Comments URL: https://news.ycombinator.com/item?id=50346981

Points: 474

# Comments: 148

Each topic area is a channel (DEVTV, SOLDER, RETRO, ORBIT, ...), plus HN Prime (the top 400), a music channel, and Late Night on ch 99, which only broadcasts 11pm-5am your local time and shows color bars otherwise.

Similar to HackerTube it was built with Claude Code in a few hours.

All credit for the data goes to HackerTube. Happy to hear channel ideas.

Comments URL: https://news.ycombinator.com/item?id=50346755

Points: 481

# Comments: 82

Here's a video of Dave from YC trying it out: https://www.youtube.com/watch?v=9QdX1eObVYI

(Edit: you don't actually have to move your lips or even open your mouth for this to work! Since the RF can see inside your mouth the system can be fine tuned on "internal speech". This isn't the automatic tongue movements that occur when "thinking" words to yourself, but more like mouthing with your lips closed. It's a bit weird to do, but with practice, we have been able to get it down pretty well.)

Arnav and I were studying CS at Georgia Tech at the beginning of the year, with no experience in silent speech or sensing. We were both AI-native builders and power users of Wispr Flow. What bugged us was not being able to use these tools in the library (or an office) or in public. At Stanford’s TreeHacks in February, we decided to build the most technically ridiculous thing we could think of. We’d read about AlterEgo, a silent speech system out of MIT’s Media Lab, and wanted to recreate it in 36 hours. We built a crude prototype, sticking EMG electrodes onto my face:

https://drive.google.com/file/d/1mwhv5HZMz3UTevdrgwS5O7g2jWV...

We got it to kind of work at a letter level in a very janky way. It wasn’t good enough to win the hackathon, but convinced us to go all in. The first time I typed a sentence without anyone noticing, I felt the magic of the technology.

We quickly realized EMG (reading muscle electrical signals) wouldn’t work:

- Signal changes due to sweat

- Electrical interference

- Placement variability

- Most importantly, no one wants to stick electrodes onto their face!!

So we switched to RF and built the glasses you see now. They emit radio waves, similar to Bluetooth, that pass through your face. As they cross the tongue, lips and soft palate, they’re modulated by silent mouthing. We train foundation models on this data to decode these perturbations to text.

Despite prior academic research, getting this to work was a huge engineering challenge. First we built an extremely low noise RF sensing stack. Then we had to place antennas so their fields overlapped maximally with the articulators. Our first attempt on glasses failed completely, so we reverted to old habits and taped antennas on our face:

https://drive.google.com/file/d/1ks0dS0eZiTV9EqSFsyB_w5BBjpc...

Since taping antennas consistently in the same spot and angle was hard, we created this weird but effective head rig to hold them in place:

https://drive.google.com/file/d/1rhzNFpdDzr7oHk0isFY70dkoV4C...

This worked, but obviously wasn’t a product yet. So we worked deep in simulations to design glasses that could hold the antennas.

That evolved into our design now, with antennas embedded within the frames. Once we established there was a signal, the next hurdle was cross-user generalization. Everyone’s facial anatomy, and hence RF signal patterns, are vastly different. We tried algorithmic workarounds but realized the only lever was to massively scale pretraining data.

So that’s what we did. We set up booths in our office and paid people to read prompts wearing our glasses, to pretrain our models on. This produced a beautiful scaling curve, with errors falling quickly as we scaled our dataset and cross-user performance becoming genuinely compelling (as you see in the video above; Dave only had a short 20-sentence calibration).

We always heard about building things from first principles but never understood it before this. We tried tackling it top down but completely failed. Only when we broke each component (the antenna, electronics, ML) down to its fundamentals to understand and optimize it did we get it to work really well.

Over the next few months we’re continuing R&D to turn our prototype glasses into the slick product on our website and ship in Q1 2027. We’re also scaling up our dataset to improve performance on edge cases.

For anyone who loves voice dictation but hates using it around people, we’ve opened preorders for our first batch of 500 units, with a $200 discount off the final price. Allocation is running out quickly, so get yours at https://subvocal.co/preorder!

Comments URL: https://news.ycombinator.com/item?id=50340088

Points: 837

# Comments: 393

Works best on desktop, but I tried to make it usable on mobile too. The initial load can take a while. It's based on several open datasets, including:

- OpenStreetMap (https://www.openstreetmap.org/#map=6/52.02/19.14) - buildings, streets, sidewalks, tram tracks, etc

- GUGiK / Geoportal (https://www.geoportal.gov.pl/) - LiDAR for building heights and trees, aerial imagery for colours

- Warsaw's city tree database (https://mapa.um.warszawa.pl/) - individual tree species, used for seasonal foliage

Open-source (https://github.com/stared/warsaw-vibe-voxel), so feel invited to (re)use it as you wish!

Comments URL: https://news.ycombinator.com/item?id=50339221

Points: 325

# Comments: 55

Comments URL: https://news.ycombinator.com/item?id=50339183

Points: 52

# Comments: 29

The API is at https://api.syscall.sh

Comments URL: https://news.ycombinator.com/item?id=50334704

Points: 237

# Comments: 56

Comments URL: https://news.ycombinator.com/item?id=50333877

Points: 185

# Comments: 25

I scraped all mentions of wikipedia on hackernews and ranked them, and turn them into a scrollable feed to make it a bit more easy to digest / interact with. You can see where people submitted the articles and where people mentioned it in comments.

The top shows the most frequently linked articles.

Random is random but slightly weighted towards the more frequently linked articles (a little bit like Wikipedia actually).

Rising is ranked towards the past 12 months.

The entire site is served only through the frontend -- I'll add more features if people enjoy this!

Comments URL: https://news.ycombinator.com/item?id=50333339

Points: 339

# Comments: 104

It's ~16.5k books that at least two different HN commenters brought up, from 2007 till now. Each book links back to the comments that mentioned it. Data is downloadable from the about page.

To be clear: Claude wrote almost all of it, including the data pipeline and the site. I mostly steered and made calls, bought the domain name, etc.

It counts mentions, so it's not perfect (praising a book and trashing it both count). If your favourite book is missing or matched to the wrong thing, tell me.

Comments URL: https://news.ycombinator.com/item?id=50326821

Points: 1860

# Comments: 306

OmniDiff can diff: code (25 languages), config, images, PDFs, archives, cursors, fonts...

It's fast. It can diff C++ files of tens of thousands of lines in under 100ms, on a 2023 ThinkPad T14s.

It's open source. AGPLv3.

It's based on real diffs. I looked at the 7400 repositories that are listed in Gentoo packages, sampled 2000 diffs and solved them manually to get a good understanding of what is needed, before putting together the diff pipelines for each content type. Incidentally, this also forms the basis of the CI pipeline, so the robustness of the codebase is quite high.

I'm still developing it. It's v0.3.0 at this time. But I have been using it daily as my git difftool for months now, and it works well.

AI Disclosure: Yes, I do use AI. No, I do not vibecode.

Comments URL: https://news.ycombinator.com/item?id=50325253

Points: 319

# Comments: 97

I kept wiring Ben Eater's 8-bit computer wrong, so I built a realistic virtual breadboard editor.

Some of the nice features are side-by-side diff, scope, schematics auto-sync, auto-router (BBRouter stands for breadboard router), and my favorites: upload from the real Arduino IDE and connect to your AI with MCP, let Claude & co build it or help you troubleshoot.

The link is the SAP-1 running. You can orbit it, probe a hole with the scope, slow the clock down, play with the virtual displays for debugging, then you can actually build it following the step-by-step instructions.

A side-goal is to create a community, there are quite a lot of simulators out there but mostly offline / too abstract to realistically use it for a real build. I also never found one that had a decent UI/UX and worked well on mobile or with a trackpad.

Happy to answer anything, and tell me what breaks.

p.s. Maker plan is $7/month; code HACKERNEWS makes the first month free and then $5/month for as long as you keep it.

Comments URL: https://news.ycombinator.com/item?id=50321522

Points: 58

# Comments: 14