Hacker News: deepsquirrelnet

New comment by deepsquirrelnet in "Can we replace a $150k Looker with a bucket of HTML files?"

deepsquirrelnet — Fri, 23 Oct 2026 19:59:16 +0000

For a related use case, I have been using claude to add or change features in my own tensorboard log reader. The flexibility of being able to change anything at any time so I can see what I want to see is really nice.

I've always preferred local monitoring tools, but tensorboard is seemingly just on life support at this point.

https://github.com/dleemiller/BlernsBoard

New comment by deepsquirrelnet in "Voxtral 3"

deepsquirrelnet — Wed, 21 Oct 2026 14:08:34 +0000

The model cards list all of our data sources. They are all publicly available datasets that are commonly used in training most speech recognition models. It is good to take a wide variety of domains for robustness. Clean speech vs noisy backgrounds. Prepared speech vs spontaneous speech. Studio microphones vs cheap laptop microphones. More exposure is better.

For raw number of hours, I would set a goal of 40k hours. That number worked well for us, although if you are training bigger models you might still benefit from more.

New comment by deepsquirrelnet in "Voxtral 3"

deepsquirrelnet — Wed, 21 Oct 2026 13:39:44 +0000

Yes, because MIT does not prevent public and private mass-surveillance initiatives from using our models against people.

New comment by deepsquirrelnet in "Voxtral 3"

deepsquirrelnet — Tue, 20 Oct 2026 18:30:29 +0000

Our "plus" models have speaker segmentation and speaker verification (embedding) that you can use for speech to text and diarization from one model. There's an example script in the model repo that shows how to do it.

https://huggingface.co/collections/futo-org/asr4all

If you're looking for a single stage "end-to-end" model just for diarization only, Nvidia released one recently: https://huggingface.co/nvidia/Nemotron-3-Diarization

New comment by deepsquirrelnet in "Voxtral 3"

deepsquirrelnet — Tue, 20 Oct 2026 17:45:52 +0000

Yep exactly! Streaming models have to run faster, because they're more compute heavy. Some compute from every batch is wasted when the next set of frames come in. Streaming models also have higher WER than fully offline models.

Our models are more comparable to nemotron-3.5-asr-streaming in that regard and I think the medium/large sizes do perform quite favorably. For offline models, I don't think it's worthwhile for us to try and compete with parakeet v2/3. Those are already really good, and we're in the process of adding those to FUTO Voice.

These models are filling some gaps we see for the kinds of things we tend to work on. But broadly they can slot into a lot of places where there isn't accelerated compute (Raspberry pi 4, laptops, local client side transcription, etc).

New comment by deepsquirrelnet in "Voxtral 3"

deepsquirrelnet — Tue, 20 Oct 2026 16:46:51 +0000

You could try ours. They're very small and english only right now (training Spanish now). But they're a CPU optimized architecture and even the largest size will work on a single thread with realtime streaming.

https://huggingface.co/collections/futo-org/asr4all

You can run them client-side using wasm-simd. I setup a space for the small sized model. There's a big step up in accuracy going to the medium size though:

https://huggingface.co/spaces/futo-org/asr4all-client-side

New comment by deepsquirrelnet in "Ousted Savannah police officer gave ICE agent husband access to Flock system"

deepsquirrelnet — Mon, 19 Oct 2026 14:41:17 +0000

That's a truly horrible story. It strikes me as one that a reputable investigative journalism outlet might take up. Have you thought about contacting somewhere like ProPublica to see if they would be interested in writing about it?

New comment by deepsquirrelnet in "A new model for open source: closed to pull requests, open to remix"

deepsquirrelnet — Thu, 15 Oct 2026 03:00:52 +0000

> All right, I'm going to open up a history lesson because it's been a couple of generations of software engineers since this essay

> It created really weird social-structure expectations: that someone who releases code has an obligation to maintain it and get paid nothing for the work.

Fundamentally, yeah that is why it took off. Big tech doesn't sell software anymore, they made it a commodity through open source. Now they sell cloud services or consulting, or ads or your data.

Just read what Eric Raymond said about it: https://www.oreilly.com/openbook/opensources/book/raymond2.h...

> It seemed clear to us in retrospect that the term "free software" had done our movement tremendous damage over the years. Part of this stemmed from the well-known "free-speech/free-beer" ambiguity. Most of it came from something worse--the strong association of the term "free software" with hostility to intellectual property rights, communism, and other ideas hardly likely to endear themselves to an MIS manager.

The intention was to find a business model and market it. That's why they distinguished themselves from FOSS in the first place. It's what you see big tech doing today (that's the business model).

The outcomes being observed and complained about are literally just the expected outcomes of making software a commodity, which is that businesses need to make money and they don't do it by selling software.

New comment by deepsquirrelnet in "ICE Wants to Use AI to Monitor Detainee Calls and Movements"

deepsquirrelnet — Thu, 15 Oct 2026 00:45:45 +0000

When we recently released our ASR models, we decided to use a custom license that forbids using it for monitoring people in custody. I know it probably won't make a difference in the grand scheme of things, but I think with the capabilities that we're developing it's worth considering how it might be used.

https://huggingface.co/futo-org/asr4all-l/blob/main/LICENSE

New comment by deepsquirrelnet in "Strands Decider: Why Not an Encoder?"

deepsquirrelnet — Wed, 14 Oct 2026 19:07:25 +0000

That's interesting results, and a bit surprising. I tried training an NLI model on the T5 architecture, which I think is similar to a lot of Jev-style use cases. But I found the T5 encoder half was better than enc+dec at roughly equal parameters and still worse than ModernBERT/Ettin.

I also found Ettin to be better than ModernBERT, which is not that surprising if you look back at the published benchmarks in their paper.

Qwen3.5 0.8B was looking significantly worse than all of those, but I added LoRA adapters to 27B and that easily beat all of my encoders. So parameter for parameter, encoders were still the best for me, but with a ceiling at 1B for size. I'm planning to train more Qwen sizes and I guess I'll see at what point they can overtake the encoders I have.

New comment by deepsquirrelnet in "Pareto Frontier Speech Recognition for CPU, Mobile and Edge"

deepsquirrelnet — Wed, 14 Oct 2026 16:22:52 +0000

New open weight streaming ASR models you can run on a Raspberry Pi 4 or a phone. These models include all kinds of nifty features, like voice activity detection (VAD), and features that support a diarization pipeline.

https://huggingface.co/collections/futo-org/asr4all

Pareto Frontier Speech Recognition for CPU, Mobile and Edge

deepsquirrelnet — Wed, 14 Oct 2026 16:22:52 +0000

Article URL: https://futo.tech/blog/asr4all

Comments URL: https://news.ycombinator.com/item?id=50090211

Points: 5

# Comments: 1

New comment by deepsquirrelnet in "CFTC Staff Were Told to Retract Analysis of Suspicious Oil Trades"

deepsquirrelnet — Tue, 13 Oct 2026 15:10:54 +0000

> The trading in question was eye-popping. In a two-minute window starting at 6:49 a.m. in New York on March 23, traders sold contracts equivalent to at least 6 million barrels of Brent and West Texas Intermediate, more than 8.5 times the volume that had changed hands that time in the morning over the prior week. A similar increase in activity was observed in futures tied to the S&P 500, with contracts worth more than $2 billion in notional value traded — far more than would be expected in an otherwise quiet premarket session. Unusual volumes were also visible in natural gas and other futures markets.

> Those trades paid off 15 minutes or so later, when Trump wrote on Truth Social that he was reversing his threat to attack Iranian infrastructure unless the country re-opened the Strait of Hormuz.

Likely the entire purpose of the war is to create opportunities for corruption in my opinion.

New comment by deepsquirrelnet in "Bag of Decisions Reranker"

deepsquirrelnet — Mon, 12 Oct 2026 15:54:04 +0000

> 1 - Is the document about desks designed for children?

> 2 - Does the document offer products or information on kids’ desks?

> ...

> 9 - Are brands or models of children’s desks listed?

> 10 - Does it cover where to buy or how to choose a desk for kids?

> Each of these can be turned into a noul question. P(yes) of each can be summed to score overall relevance

It seems to me that this is a rediscovery of natural language inference (NLI), which was very active in NLP research pre-LLMs. I have a side hobby of training NLI cross encoders, and I'd bet that many of these use cases could be serviced by a good quality small NLI model.

See: https://huggingface.co/dleemiller/finecat-nli-v2-l

And also: "Less Annotating, More Classifying: Addressing the Data Scarcity Issue of Supervised Machine Learning with Deep Transfer Learning and BERT-NLI" https://www.semanticscholar.org/paper/Less-Annotating%2C-Mor...

New comment by deepsquirrelnet in "Typesafe AI raises $870M at $7.5B"

deepsquirrelnet — Fri, 09 Oct 2026 19:21:33 +0000

Here's another leaderboard: https://huggingface.co/spaces/multimodalart/jev-decision-ind...

New comment by deepsquirrelnet in "Iranian campaign planted fake articles in real U.S. publications using ChatGPT"

deepsquirrelnet — Fri, 09 Oct 2026 14:04:19 +0000

And what do you call it when a country acts in adversarial ways to its own citizens? Call it cynical, but it sure seems like there's a lot of that in the world right now.

New comment by deepsquirrelnet in "Open-sourcing AstaBrief, the fast report-generation model in Asta"

deepsquirrelnet — Sat, 03 Oct 2026 00:06:10 +0000

This is exciting. I use asta all the time for finding papers. It's one of my favorite resources for exploring new topics.

New comment by deepsquirrelnet in "Decision models like Jev don't beat LLM-as-a-judge or traditional classifiers"

deepsquirrelnet — Fri, 02 Oct 2026 21:30:55 +0000

BART is quite an old model for this kind of test, and probably not a good very good choice for much these days. I'm working on replicating their benchmark on my own NLI model that targets zero-shot guardrail applications. I don't think it'll beat much larger models, but should give a better baseline for what a crossencoder can do.

https://huggingface.co/dleemiller/crossingguard-nli-l

New comment by deepsquirrelnet in "OpenAI is well positioned to fast-follow Jev"

deepsquirrelnet — Tue, 22 Sep 2026 19:16:10 +0000

It's hard to say without knowing their architecture, but I'd guess something like block attention. You can process the prompt separately from the classifications into a latent space and then do some kind of late interaction with the encodings from the classifications.

There are plenty of other ways to do zero shot classification that would result in more "token usage" (really just having to reprocess everything for each class), but the pricing and the way they describe it narrows it down somewhat.

New comment by deepsquirrelnet in "OpenAI is well positioned to fast-follow Jev"

deepsquirrelnet — Tue, 22 Sep 2026 19:09:17 +0000

I'm having a really hard time wrapping my head around why Jev is getting so much hype. It feels manufactured to me. I don't think they've proven a significant market for their product, and there's no independent benchmarks that prove anything. To me that's doesn't pass the smell test.

But if they had to show how well their product worked they might give away the whole game... because they'd have to compare their "noul" class against an NLI benchmark for instance, and possibly show they're losing to cross encoders and give away the fact that they are just rebranding NLI. Or rerankers (choice) or zero-shot classifiers.