mikewarot

3시간 전discuss

Sandalwood - it's nice and luxurious, just like paying full freight on tokens

5시간 전discuss

It'll be interesting to see the eventual application of this to moon bounce. If they put it on a sufficiently stiff backplane, and calibrate things carefully, the 240 channel system they hint at looks like it might be able to pull it off.

Personally, I'd like to have a coherent SDR transceiver for HF/VHF that I can expand arbitrarily, and doesn't cost multiple thousands of dollars.

mikewarot

5시간 전discuss

Be sure to check your states registry. Like many professionals, Engineering requires extensive study, testing, on the job training, and eventually, licensure. The Computer Engineering test in Indiana had a 56% initial pass rate last year.

Also, AI can't assume liability, that's the responsibility of humans.

mikewarot

어제discuss

I've learned that I'm just not cynical enough. Here's the relevant Peter Girnus post[1], he tells the underlying truth through satire.

  10 years. We report our compliance every 60 days, then once a year, which is to say we have turned handing you your own machine into a paperwork schedule that runs through 2036. Any new repair tool we build, we only have to share once more than 50% of our dealers already have it. We decide when a dealer has it. We build the clock and we wind the clock.

https://x.com/gothburz/status/2075211022198096180

mikewarot

어제discuss

The only authors who would sign such a deal would be near the end of their lives, and wishing to provide for their heirs. I don't think that it would be a widely accepted practice.

mikewarot

어제discuss

During WW2 in the United States, you had all sorts of consumer goods companies reorganized to output a prodigious amount of military supplies. There were multiple companies making the same model of things, with fairly rigorous QA to ensure quality and uniformity.

For example, the BC-348 receiver, widely used in aircraft, was produced initially by RCA, and eventually "farmed out" to 3 other manufacturers.

More than 4 million M1903 Springfield Rifle were produced by the Smith-Corona typewriter company.

Here's a really good example, look at how the production of proximity fuzes, was distributed.[1]

The key thing is to have second sources for everything. Something the US military seems to have forgotten, or decided to ignore in their pursuit of gold-plated weapons systems that give the most kick-backs.

[1] https://usautoindustryworldwartwo.com/vtproximityfuze.htm

mikewarot

어제discuss

How is it that I created $6.02 Trillion of shareholder value, but only got $15.11 million of it?

Oh no... I just sympathized with Elon and the Epstein class.

Let's walk that back... $15.11m for 14 years work, over a million per year, because I got lucky in the end. It's all a casino, and your labor/life is the table stakes.

That $6.02 Trillion?, that's an illusion, because if anyone decided to sell big, half would vanish immediately, more likely 80% or more.

What an impressive, and depressing game.

mikewarot

그저께discuss

Trying out AI is a fun and sometimes useful thing to do.

But you should never trust it completely. A lot of people have already died because they trusted it with their lives. "Full self driving" isn't.

mikewarot

그저께discuss

I've got 459,203 words posted here on HN alone, according to clickhouse

I'm currently ranked #495

https://serjaimelannister.github.io/hn-words/

I'm out of the labor pool with nothing but time until I age out. I also have strong minority opinions.

mikewarot

그저께discuss

Relevant Peter Girnus post, satire, sort of

https://x.com/gothburz/status/2074902503599059357

I just know insurance companies are going to use this to try to weasel out of doing their job

mikewarot

그저께discuss

I feel like crimes over a certain threshold should include accessories in the punishment, like the Death Penalty here in Indiana does.

It certainly helps motivate them to testify in court.

Strong minority opinion: If someone steals more than the average total lifetime earnings of 10 people, I think the death penalty should apply. That's about 25 million USD

mikewarot

4일 전discuss

Free Pascal, Assembly, STOIC, Metamine, Python 2

mikewarot

5일 전discuss

Whenever the nostalgia kicks in hard, I fire up OpenVMS 7.2 on the virtual 11/780 in my cheap Motorola phone, then SSH in.

I love OpenSimH and Termux.

mikewarot

5일 전discuss

>Personal wikis always died for the same reason.

Mine (WikidPad) died when I switched to Linux, and learned that breaking changes to WxPython rendered it worthless, as none of the dialog boxes functioned after that point. Sure, the source from 2012 was available, but my Wiki really wasn't.

Eventually this forced me back to Windows... but it was too late, now I'm back on Windows, and still don't trust WikidPad.

>Keeping them current was tedious, and humans hate tedium. But the tedium is the one thing language models are immune to. They will happily re-link, re-summarize, and reconcile contradictions across a hundred files without complaining.

Yeah, and you're going to trust the LLM to reliably maintain this? Not a wise choice.

I really wish we had a reliable way to annotate and interlink documents using hypertext. Unfortunately, HTML doesn't actually let you mark up (annotate) hypertext.

We still, 81 years later, don't have a Memex! 8(

mikewarot

7일 전discuss

I need nuclear for cheap power, I almost never actually rush to make rockets, I just want to avoid the grind for technology research. ;-)

mikewarot

8일 전discuss

Well, this is certainly an interesting point at which we all find ourselves. I'm personally going to have to learn what a KV cache is, how big they are, and the costs associated with storing this value, instead of merely caching it. The cost of the loop ends up quadratic because of reprocessing everything each time a new query was added. In theory, this means that after a year, there could be a billion input tokens, just if you say "thanks!" to the model at the end of a conversation, which is nuts. Explicitly storing the model state should make it a linear cost again.

I've been following Ed Zitron's reporting on AI and costs/profits. When Mike Pound starts talking about the costs at 4:12, he hints at the fact that none of us know the actual cost of handling a token. Ed's reporting hints that it may be quite a bit higher than we've been lead to suspect, it may, in fact, be far above current retail prices, subsidized as part of the "marketing" expenses on the AI companies balance sheet to try to gain market share.

It seems we're in for a "call to Jesus" moment, and a big pop in the markets as a result. This video is part of the structure that does the popping.

I think that actually storing the model state after each query/result will become the standard, to save reprocessing tokens. Switching between models would thus become discouraged because it would lose that state. It wouldn't be impossible, it just wouldn't be cheap. I could see storing the entire AI state, including "thinking", model version, etc. along with the code in the GitHub repository, right next to the commit comment. Gits delta compressor should make storing all that binary data tractable.

We're living in interesting times. Agentic coding certainly seems valuable, but it may turn out to be ruinously expensive, compared to just using us human programmers. We'll have to wait and see.

mikewarot

8일 전discuss

I think it's a horrible idea. It's also impractical.

What company is going to willingly face the inevitable wrongful death and injury lawsuits that result?

mikewarot

10일 전discuss

>Reduced centrifuge stack size to 10.

Well, that'll make Rocket Rush a bit more fun to play. It was taking forever to get up to 40 U235s.

[Edit] Nope.... it doesn't change the requirements from 40 U235, just makes it so a centrifuge will fit on a rocket. 8(

Am I the only one who likes rocket rush?

mikewarot

10일 전discuss

I think the ASIC we discussed yesterday[1] from Phantafield is a hint of what is to come. Instead of constantly moving weights from HBM to MACs, they have local dynamic ram areas that hold the weights locally, using a novel architecture that is more efficient at using space than the traditional SRAM cell, especially in a mostly-read environment.

Instead of doing MAC using 16 bit wide hardware, they do it all bit serial, so that they can fit far more computing into the same chip area, trading a bit of the speed they get from all that local RAM into efficiencies in usage. It also makes it easy to do FP8 support, or any other bit size required.

I expect another few orders of magnitude improvement in the cost of compute over the next decade. Moore's law isn't dead yet.

[1] https://news.ycombinator.com/item?id=48713686

mikewarot

10일 전discuss

>The knowledge is not in danger, in fact, it has never been safer. The AI models have read every manual that no human reads. They will recite, flawlessly and forever, exactly how all machines work. If this were only about competence, it would be the most secure moment in the history of computing.

Obviously this person doesn't have enough experience with the current crop of AIs. This is the exact reason why we should be concerned in the first place. Someone has to design and build the next version of everything, and AI isn't qualified to do the job, and might never be.

>What is dying is acquaintance. The plain, unglamorous intimacy of having fought a particular machine, and lost, and gone back, and finally felt the thing give.

This is true. But then again, who among us had to get up in the middle of the night, and trudge out to an outhouse? Who had to shovel a bit more coal into the furnace? Who had to pump water, and heat it on the stove? Who had to split wood, and stack it for the winter?

There are many, many things we're no longer acquainted with, and that's sad, but still ok. What's not ok is losing the competence required to maintaining the infrastructure and supply chain supporting society and civilization.

Submissions

Ask HN: Is AI Alignment about to be solved, for profit?

Ask HN: Is Google search broken now?

comments