nwiswell

Here's my thinking:

Cloud:

- You cannot directly execute a remotely-hosted program.

- You cannot run inference on an API-served model.

---

Closed-source:

- You can execute a program with the binary. You cannot generate a new binary, but you could try to reverse-engineer it or (painfully) modify its execution.

- You can run inference on a model with the weights. You cannot re-produce a new set of weights from scratch, but you can fine-tune.

---

Truly open:

- You can freely modify the source and produce new binaries.

- You can use the original training data and model architecture to independently re-produce the weights (assuming you've got the compute). You can modify the model architecture to get the weights that would've resulted from training the model that way.

---

To me these are pretty clear parallels... I don't think the weights provided in a vacuum are in the spirit of open source, historically speaking.

The policy argument is totally separate, of course, and I fully understand why none of the frontier labs are truly open.

You can make parallel arguments for binary executables.

I associate that phrase with Dr. Henry Killinger

https://youtu.be/9FUik3QotBQ?si=27AcY47wLNov0Vzw

I believe this is due to a surge in imports in response to looming tariffs - supply chains are trying to stockpile before they hit. I am skeptical that this stockpiling is significantly displacing real investment.

The way that the figures are calculated views imports as a negative factor to GDP (because NET exports is an input to the model). Please correct me if I am wrong.

In any event, view the headline with suspicion.

> Anyone who can pay this, can run private tests and compare scores.

Yes, by uploading the tests to a server controlled by OpenAI/Anthropic/etc

Companies with actual profits can, and do, simply buy back their own stock. Inflating market expectation of future cashflows is not the only way to realize asset appreciation.

> i'm the one who set the password

You, personally, set the password to a public internet-facing database to 'abc123'?

And if you really did, how much do you estimate that increased your 'velocity'?

> the comment I was replying to only mentioned as a benefit that Mars could be a second home

The first and second sentences of that comment literally say

> A Mars mission would benefit humanity, but less directly. The past lunar missions and space program benefited humanity in many ways.

And then it goes on to acknowledge the "second home" element, but only as a small consideration.

No, he clearly said that a "second home for humanity" is of dubious (but potentially nonzero) value.

Rather, the main benefit would lie in the technological advances made in order to enable such a Mars mission in the first place (similar to advances during Apollo).

As of 1/1/2025 the residential average rate paid for the municipal utility of Santa Clara, CA is $0.175/kWh vs a residential average of $0.425/kWh for PG&E

https://www.siliconvalleypower.com/residents/rates-and-fees

PG&E has an 11% margin on those rates because they keep burning the state down and having to pay for it. Municipal utilities don't have to worry about that.

The only thing that could be seriously considered a downside is that Santa Clara, CA is now absolutely jam-packed with data centers that have low employment per sqft.

If you come into contact with a harvester, it should be clear that it's a harvester, correct.

If you come into contact with a dough mixer, it should be clear that it's a dough mixer.

But if you come into a contact with a book, I imagine you might have trouble recognizing it.

No, in other words, states and local governments are also bound by the Constitution in many of the the same ways that the federal government is.

The major difference is the Tenth Amendment, which sets the states apart by specifying that any powers not "delegated to" the federal government are reserved exclusively for the states. (In practice courts have found many "implied powers" that are not explicitly enumerated).

Federal laws are distinct from the Constitution.

> * Having written all that, I will add that "government" above means the US Federal government, not all the other ones. State, local, have a lot of latitude to make whatever laws they want, unless a federal law specifically prohibits it.

This is not entirely correct. In general many elements of the Constitution are incorporated and apply at all levels of government. It even outranks state constitutions where the two conflict.

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

> A single house is exposed on five sides.

Six. The most heat escapes through the roof, but thermal loss through the floor is generally about 10-15% of the total.

> It’s too late. It was too late 18 months ago. Nvidia is over 30x bigger. Qualcomm is 2x, AMD is 2x.

AMD came back from odds much longer than this. At its nadir, AMD's market cap was about a billion. A billion! Jensen's jacket collection is worth more than that.

Anyway, it's not looking good for Intel but it's certainly not "too late".

Because it drives inequality. We're proposing a policy that explicitly benefits large owners of capital over small ones.

Inequality is a natural outcome of capitalism, and critically it will get worse without limit unless there is significant policy intervention (ref. Capital in the Twenty-First Century). Existing progressive income taxes are not sufficient, even in Europe where they are far more aggressive. So from a policy perspective, this is exactly the opposite of what is desired.

My argument is

1) that something that takes in 500m in its useful lifetime should not be subject to a lower tax rate than something that takes in only 100k in its useful lifetime (in fact it should be subject to a higher tax rate, in line with income tax policy);

2) something that takes in 500m should not automatically be entitled to a longer useful lifetime than something that takes in 100k.

Unfortunately both things are true for this proposal. #1 can be ameliorated by offering an income tax credit, but #2 is fundamental.

Small owners of IP get to enjoy less value from their creations due to weaker IP protections. That's backward.

Moreover the amount of tax paid as a fraction of total value realized is actually lower for the large owner of IP because the total tax payment is dominated by the final years, but the total revenue is determined by the number of years. In the example above, we had:

$5,100 / $40,000 = 12.75% tax for the small author, and

$671,088,620 / $12,500,000,000 = 5.37% tax for Dreamworks.

The ratios would be even worse if the small author could've just barely justified the 9th year. Pretty much unconscionable.

The fact that we're collecting tax from IP is not interesting. We have progressive income tax for this purpose.

This clearly benefits wealthy owners of IP (disney, movie studios, game publishers, etc) over small-time artists (self-published authors, small bands, etc) since the period of time that IP protection remains an economic choice is strongly tied to the value of the IP.

E.g., if you write a book and realize $5,000 in sales per year, then 10 * 2^x=5000 where x is 8.97, so you only enjoy 8 years of revenues ($40,000) and you've paid Σ(1->8) 10 * 2^x = $5,100 for the privilege, for a net $34,900 or 6.98x the yearly royalty value.

If Dreamworks sees $500M a year in Minions merchandising, then 10 * 2^x=500,000,000 where x = 25.56 and so Dreamworks realizes 25 years of revenues ($12,500,000,000) and pays Σ(1->25) 10 * 2^x = $671,088,620 for the privilege, for a net $11,828,911,380 or 23.66x the yearly royalty value.

This is backward, in my opinion.

> It realistically affects practically nobody alive

Do people in the Global South not use GPS or consume weather forecasts?

> A coding standard that assumes a 10 point font size would violate the Americans with Disabilities Act's 'reasonable accommodation' mandate.

Would it? You can still set the size to 18, you just might have to scroll or line wrap. That's a mild inconvenience, not "inaccessible".

Once physical hardware is involved, the problem gets a lot more complicated.

If you're talking about $2M worth of 4090s you have to concern yourself not only with putting them in hosts, but also with adequate space, power distribution, cooling, and networking. You have to figure out how to gracefully handle hardware failures. You have to install operating systems, or at least come up with some mechanism to deploy automatically. And you just straight up have to supply labor to get 1000+ of these things plugged in.

All that before you can get even get to the "fun part". If you're a multi-millionaire you might be willing to just pay the spot price premium.

While I'm sympathetic to democratizing access to simple fabrication technology, I have serious misgivings about hobbyists getting involved.

There's the obvious stuff: you can't avoid HF, and it's nasty stuff. You can die. But that's not what I'm the most worried about; people can make smart decisions to reduce risk, and ultimately people can make their own decisions about their risk tolerance.

What I'm worried about is the SF6 for the RIE. Kg for kg, that stuff has a global warming potential of more than 24,000 TIMES the warming potential of CO2. If it's all broken down in the plasma chamber, or there's exhaust scrubbers involved like you'd have at an industrial fab, then it's no issue.

But hobbyists are going to be spilling and purging a bunch of unmodified SF6. It's kind of an ecological catastrophe. Some things are better not done at home.

Glass has a tensile strength more than an order of magnitude below mono-Si. The elastic modulus for soda-lime is within a factor of 2, but that's not usually what is meant by "strength", rather "stiffness".

You're right that fracture toughness is comparable, but that's not usually as relevant in a MEMS context: for one thing there's less potential for unanticipated collisions, but also crack propagation is significantly inhibited at microscopic length scales.

> out of something that normally doesn't move (polycrystalline silicon)

More commonly, bulk micromachined MEMS devices are monocrystalline.

> and is very fragile at that

Normalizing for density, monocrystalline silicon is stronger than steel!

> if they don't make the numbers make sense soon then I don't see things ending well for OpenAI.

This is pretty much obvious just from the valuations.

The wild bull case where they invent a revolutionary superintelligence would clearly value them in the trillions, so the fact that they're presently valued an order of magnitude less implies that it is viewed as an unlikely scenario (and reasonably so, in my opinion).

> Inside the battery is perhaps the best hidden, but you'd need to own a bunch of battery manufacturing facilities (expensive).

Do you?

What stops you from just taking a smaller battery and packing it with some plastic explosive into the typical "battery foil"? I'm sure the IDF is capable of doing that at scale.

I don't doubt you meant well, but that's a lot to infer from a couple of sentences. I think what the parent meant to imply is that this seems like the classic case of "when all you have is a hammer, everything looks like a nail."

Oh, that's super. I've been really impressed recently with the wisdom of our collective selections.

This didn't sound right, so I did the math.

The volume of all water is 1,386,000,000 km^3, which is then 1.386e+21 liters, or right about the same number of kilograms.

The mass of Earth is about 5.972e+24 kg. So the percent fraction by mass is 0.0232%.

A "drop" is typically estimated at 1/20th of one mL, which is then 0.05 grams. We can estimate the mass of a small-ish bowling ball at 5kg, or 5000 grams. 0.05 / 5000 * 100 = 0.001%.

So it's an order of magnitude shy, but that's still closer than I expected! It's about 1 ml of beer on a bowling ball - a small splash. Or maybe a very large drop.