Mizza

Even at $10k/yr/employee, you'd need 30 million people on the 10k/yr plan to hit 300B ARR. I think that's a hell of a big swing. 3 million, recoup over ten years? Maybe, but I still don't think so. And then competition between 4 or 5 vendors, larger customers figuring out it's cheaper to train their own models for one thing that gives them 90% of the productivity gains, etc.

But rather than speculating, I'm generally curious what the companies are saying to their investors about the matter.

What's the path to recouping that money?

Even if every major company in the US spends $100,000 a year on subscriptions and every household spends $20/month, it still doesn't seem like enough return on investment when you factor in inference costs and all the other overhead.

New medical discoveries, maybe? I saw OpenAI's announcement about gpt-bio and iPSCs which was pretty amazing, but there's a very long gap between that and commercialization.

I'm just wondering what the plan is.

Until there's a paradigm shift and we get data and instructions in different bands, I don't see how it can get better over time.

It's like we've decided to build the foundation of the next ten years of technology in unescaped PHP. There are ways to make it work, but it's not the easiest path, and since the whole purpose of the AI initiative seems to be to promote developer laziness, I think there are bigger fuck-ups yet to come.

That's great, I thought about building a similar thing data from PuzzGrid, an onlyconnect fan site, but some of the questions there are a bit iffy compared to the ones on the show. How did you build the dataset - just binge watching with a notepad?

It's amphetamine. All of these people are constantly tweaking. They're annoying people to begin with, but they're all constantly yakked up and won't stop babbling. It's really obvious, I don't know why it isn't highlighted more in all these post Ziz articles.

> It's like writing an entire diatribe about how it sucks that people can insult you online, and how that should change ... and not even mentioning the benefits of freedom of speech.

They'll make you a minister in the UK for that.

Related to this is Loeb's proposal to nudge the Juno spacecraft, currently orbiting Jupiter and soon facing EOL, into the path of 3I/Atlas to try to scan it and snap some pictures. I doubt it has enough fuel left, but I hope they're looking into it.

https://avi-loeb.medium.com/how-close-can-the-juno-spacecraf...

That first one looks like a parody of 'View of the World from 9th Avenue' but I don't know what Acius was!

https://en.wikipedia.org/wiki/View_of_the_World_from_9th_Ave...

Demos here: https://resemble-ai.github.io/chatterbox_demopage/ (not mine)

This is a good release if they're not too cherry picked!

I say this every time it comes up, and it's not as sexy to work on, but in my experiments voice AI is really held back by transcription, not TTS. Unless that's changed recently.

What are you doing with it?

I can see this being addictive! Would like a harder AI, though, it's not enough of a challenge.

What's the optimal strategy? I tried a few just for fun and found I got the highest score by taking all the corners to start, but I don't think it's a strategy which would work well against a better AI.

Looks like this is the AI's strategy: https://github.com/schaoss/wall-go/blob/6d5af58ffe9f4ce8e648... which is the opposite of mine. But I get more points!

She didn't even break the top 10 in this content: https://www.societyforscience.org/regeneron-sts/2025-student...

I'm impressed beyond words by these kids, though I think I'd give her the top prize. Watching my grandfather's final days taken away from him by the effects of morphine has always made me wish so much that we had much more effective non-narcotic painkillers

What does the foot pedal do? On/off for the recorder?

Yeah, you can try it out here: https://llm.hunyuan.tencent.com/#/chat/hy-t1

It takes a bit of nudging to get it to switch to English, but it's fast and accurate after that.

That's pretty cool, I didn't know this type of stuff was available to consumers. The reviews on their site seem really enthusiastic, lots of older folks and mobility-restricted folks seem extremely pleased with the product and post action videos. Can't wait to see where this technology goes by the time I need it.

Acetates are still used in the manufacturing process - even pressed records need to be cut once - but they aren't great as products to send to home-users because they degrade from usage very quickly.

We are also running a soundsystem and mainly cutting for other DJs.

Records are cool!

One of the many secrets of lathe cutting is that these aren't actually PVC, but rather PETG. They actually play nicer and are more durable than pressed PVC records and can have a lower noise floor. On a good day, we can make a cut record that sounds nicer than a pressed one.

A buddy of mine just acquired one of these machines (actually a beta version of a new product that will hopefully enter the market soon), and I've been helping him along the way as we try to get some good records out of it.

It's not easy! It's not like plugging in a printer. There are so many variables to consider (materials, temperatures, pressures, angles, sharpnesses, volumes, phase), magical anti-static solution formulas, airflow, isolation, line control. The margin of error can be as low as 100 microns, and if you screw up, you might break a $200 diamond shank which have to be specially manufactured and imported.

And that's before you even get to the actual audio quality!

Still, it's fun to use and great to have 1-of-a-kind dubplates.

You can even tell that's ChatGPT rather than, say, Claude. Redundant obvious comments are a stylistic signature of their models.

You can 'distill' with data from a smaller, better model into a larger, shittier one. It doesn't matter. This is what they said they did on the livestream.

There are plenty of distills of reasoning models now, and they said in they livestream they used training data from "smaller models" - which is probably every model ever considering how expensive this one is.

Sounds like it's a distill of O1? After R1, I don't care that much about non-reasoning models anymore. They don't even seem excited about it on the livestream.

I want tiny, fast and cheap non-reasoning models I can use in APIs and I want ultra smart reasoning models that I can query a few times a day as an end user (I don't mind if it takes a few minutes while I refill a coffee).

Oh, and I want that advanced voice mode that's good enough at transcription to serve as a babelfish!

After that, I guess it's pretty much all solved until the robots start appearing in public.

Have any multimodal models been reasoning-trained yet?

Marcan is a legend, it's a shame it ended like this. The communities around certain software technologies can act incredibly hostile and entitled towards developers, paradoxically even more so for free and open source projects. I've seen this many times happen in audio production and game emulation software.

But Marcan clearly has true hacker spirit, I'd wager we'll see him again in the future with an equally cool project. It's often best if the visionaries just spend their efforts to get the ball rolling and then let the project evolve on it's own as they move onto their next challenge.

This is Adderall-brained LessWrong gibberish. Bayesian epistemology is not "a fairly accurate description of how we acquire knowledge" compared to critical rationalism. There's only a small, specific community of extremely online people who think otherwise.

Then it's not a very useful framework, is it.

The first part is is precisely correct, and then the second part is precisely wrong:

> We have no way to operate on hypotheses nobody has thought of, so we just operate on the set of hypotheses we have thought of. Since the sum of this set of probabilities must be 1, decreasing the probability of one hypothesis does increase the probability of all the other hypotheses in the set.

We've been able to go to the moon because we understood general relativity, which absolutely could not have been created from a purely Bayesian approach for the reasons you made clear - it required new ideas.

But those two rows are vacuous truths, which Popper obviously doesn't care about in context as they're not falsifiable.

Thank you for beating me to it. Astounding how those LessWrong dorks were able to revive the corpse of this dead end of epistemology after it was so thoroughly destroyed by Popper. And to what benefit.. sex crimes, massive financial fraud, murder cults and a far right government?

Here's a fun one https://sci-hub.3800808.com/10.1038/302687a0

If this is the startup that finally unleashes AI spam bot articles and comments to the top of HackerNews, I'm gonna quit the internet and move into a log cabin.