sudb

There's already open-source stuff [1][2] out there that can lipread video to transcribe text, but the word error rate is pretty high (and it may not be physically possible to get this down without doing something like subvocal and having other sensors near your face).

1. https://github.com/mpc001/auto_avsr 2. https://github.com/amywork777/lipflow

One reason I think a distinction is made is that native Docker-in-Docker can be a real pain, but Docker in a Firecracker microVM "just works".

It is now also reasonably straightforward if you have access to frontier LLMs, a recent-ish mac and a recent-ish iPhone to point them at the job of porting a given model to run on the ANE - it's a reasonably easy task to hill-climb at this point!

The underlying model (DFNet3) is not particularly great but it is very small and fast - imo the best commercially usable denoising model is MossFormer2 (no affiliation - it's just excellent) with one drawback in that it can't remove reverb.

Nvidia's RE-USE model can do what MossFormer2 does _and_ can remove reverb, but it is non-commercial licensed.

This was literally in YC's Request For Startups for Fall 2026: https://www.ycombinator.com/rfs#multiplayer-ai

To add: I doubt frontier labs will build routers - they are not financially incentivized to optimize token usage (though in the short term they may be incentivized by constrained GPU capacity to reduce load).

Obviously not, no need to patronize. Did you read it? It seems to be honest about the numbers but it's pretty skillfully written to reassure.

I found a tweet purporting to show the letter that Leopold sent to his LPs - it looks pretty thoughtful and doesn't sound as bad as some news sources seem to be portraying.

https://x.com/tbpn/status/2083226453509030285

Reminds of me of the "most expensive peanut butter" at ~$2.44/g - sold by NIST, I assume for some sort of peanut butter factory process calibration: https://shop.nist.gov/ccrz__ProductDetails?sku=2387

Absolutely agreed - though the usual claim for coding agents in particular is that they are already very capable and fast, replacing and speeding up software development hugely (anecdotally reasonably true for me and my colleagues, though again depends on domain and complexity).

I was going to say that this surely depends on the type of work being done with AI but then a small dive into your HN profile shows you're working on SWE agents at Google?

I am struggling to reconcile your viewpoint with this but may have found the wrong profile!

Of course, would be happy to!

I don't know that you can trust Google's AI Overview to have accurate knowledge about Google AI Studio internals

very cool! I've been having a fun time porting non-LLM models to iOS/Android (recently had some success with making Bonsai's image generation model run on Android) - have you attempted to run any of the workflows on a mobile device at all? Mind if I try?

I think some of the trouble here is that the application of the law cannot be guaranteed to be free of discrimination or bias, regardless of the letter of the law itself?

I am a native speaker, I agree with this assessment.

You only need to find an email platform run by people you agree with more, no? I'd vouch for proton mail here.

I mentioned this specific model because it also has Tough Solar which should mean much longer between battery replacements!

I doubly enjoyed reading this because I feel like I'm reasonably sure the author never ran this through an LLM because of the "cirriculum" typo.

This reminded of this other clock project: https://mitxela.com/projects/precision_clock_mk_iv

Also: I reckon the closest best-value watch alternative that ticks most(?) of the author's boxes is the Casio GW-M5610-1ER, so long as you are inside a radio-controlled time area.

this is extremely encouraging for individuals/small companies being able to train pareto-frontier TTS models (specifically compute required to run vs quality of model output)

I don't think Gas Town was actually all that much ahead of the curve - see Cursor's swarm post from January this year, mentioned at the very top of this blog post.

I also don't know that beads is a very strong example that Yegge knows what he's doing - it's famously been described as (pseudo)malware.

The OnePlus One was exciting because I think it genuinely was an Android-flagship-competitor at a much lower price. Prices crept up though, and the last OnePlus I owned was the 5 which was still pretty excellent!

After a brief, very annoying stint using the Fairphone 4 (underpowered & expensive, though I did actually replace both the battery and the usb c port myself and it was exactly as easy as promised), I'm now finally on a Samsung S25+, though I did really really consider the newest OnePlus.

Sad to know that it won't even be an alternative for my next phone, though hopefully by then, memory/silicon prices will have settled and Nothing will have their own flagship alternatives.

For extra security against these text-to-speech model zero-shot clones, you might also want to use made-up gibberish words for which the pronunciation can't be reliably inferred from the spelling

This sounds plausible, but don't Samsung also sell tens of millions of units of their flagship galaxy range phones, in addition to their budget offering?

Perhaps the margin isn't quite as large as for Apple phones but the flagships are still not cheap by any means.

It can't come fast enough! I have a friend who works on battery storage in Europe and it sounds like an extremely busy time for them, which I'm glad for.

I was sure this couldn't be true but I couldn't find anything about high temperatures shutting down solar plants.

But I did find something about a predicted grid overload during a sunny period requiring a solar plant to go offline:

https://www.theguardian.com/environment/2026/jul/12/solar-fa...

One of the tradeoffs of Rust is its verbosity I think (in return for which Rustaceans would say you gain explicitness).

For my current purposes, I need a speech-to-text model/API to also emit word-level timestamps - for now, that makes ElevenLabs's Scribe v2 the best multiplatform, multi-language choice though it does look like this SpeechAnalyzer API provides them (although only for English).

I'd imagine that mouse movement is just one signal among many that's weighted appropriately, but I hope we get feedback from these users