akoboldfrying

I think that's overstating it. Carbon promises a reliable transition; getting an LLM to rewrite in Rust depends on the LLM being smart enough to never make a mistake that it can't catch with (existing or its own freshly created) tests.

LLMs are very good now, but they are still stochastic (when temp > 0), and Google has a lot of code -- i.e., many rolls of the die.

It's more nuanced than a lie.

Quality of life during the Industrial Revolution dropped considerably for most of the working population, but afterwards saw big improvements. You seem to be implying that life now is worse than before the IR. Are you? If not, in what way was the loss of jobs to mechanisation then more acceptable than it is now?

I've now read the GMP terms on their website, and I've moved closer to your position.

> They don't get a battery backup.

The customer in effect gets a "probabilistic" battery backup -- whatever is in the battery at outage time. As you say, the terms (para. 15) explicitly allow that GMP can use the entire battery at any time; it seems I was confusing those terms with different ones ("at most 36 times per year") mentioned by another commenter, which were actually for a different company, in NC. This is not as good as a battery that you fully control, and it would have been nice to see GMP place limits on its usage level in the terms, but it's much better than nothing, as made clear by commenter xoa and the customers interviewed in TFA. GMP wants high total charge levels in customer batteries, so their incentives are aligned with customers; the only potential misalignment I see is that it may not matter to GMP how that charge is distributed across customer batteries (to them, everyone at 50% charge may look identical to 50% of customers at full charge and 50% empty, but the former is much better from customers' perspective), but I also can't think of any reason why GMP would benefit from an unbalanced charge distribution, and maintaining rough balance seems straightforward, so I don't think that would be a conflict in practice.

> maybe get to draw power from it if there's a service outage (the company isn't obligated to let them use any of it).

This is simply wrong. Para. 15 explicitly obligates GMP to permit this:

> As Lessee, Customer’s control over the Energy Storage System is limited to its usage as a backup power source in the event of a power outage up to the point that the battery is completely depleted.

Outages are obviously "bidirectional", so there's no conceivable way for GMP to draw on the battery in the event of one anyway!

I didn't see any mention of net metering or payment for exporting electricity back out to the grid, which surprised me. Without such an agreement in writing, it seems unlikely that GMP would ever pay for this access, so I agree with you there.

The internet usage clause is reasonable and common sense, IMO. I suppose it would be better to include an upper limit to make sure that it isn't hogging the connection, but it would be my very last concern.

The only clause that I really disliked was para. 20, clearing them of any liability for damages caused by the battery.

I don't see the terms as exploitative.

IIUC, the benefit for the customer is a battery backup that is much cheaper than buying the batteries themselves would be, plus the ability to make money when their battery is taken over for a few hours during occasional peak usage events. Provided the former override the latter, as seems to be the case, I don't see what the issue is.

I think you're claiming that battery backup shouldn't be needed in the first place if the power company was doing their job properly, since then there wouldn't be frequent outages in the first place. But I gather that this just isn't practical in rural Vermont. To put it another way: In that environment, a power company guaranteeing 99.999% uptime could not offer this at a price that is acceptable to most customers. So the equilibrium naturally shifts towards citizens tolerating more frequent outages than would be tolerated in, say, central NYC, and/or paying for mitigations like generators or batteries.

With any luck he'll change the company name again, this time to "Surveillance Glasses", months before shutting down the whole division due to everyone hating the idea.

> Instead of backtracking, Claude just imported an industrial-strength library made to solve these sorts of problems. OR-Tools CP-SAT is put out by Google and is made for solving constrained optimization problems, as well as satisfiability problems like this one.

I'm not familiar with CP-SAT, but TTBOMK all SAT solvers use a type of backtracking search underneath called DPLL. Modern ones are highly tuned in terms of which variable they choose to branch on next, and in what order to try its possible values; this can have an enormous impact on runtime. They probably use several tricks on top of that; the big one that I'm aware is conflict-driven clause learning, where the solver adds new constraints that it discovers as it goes along (e.g., it might be able to determine that x and y always have the same value in every solution), which can shrink the search space a lot.

You're right, you should subtract off the compressed sizes of the respective reference files before comparing. (This suffices if we assume that later input data does not influence the compression of earlier input data, which is true except for certain unusual conditions like a repeated substring at the end of the reference data that also appears at the beginning of the test data.)

Basically I agree with your second point. I'd been hung up on the notion that people were maligning specific technical choices (and indeed there are sub-threads where this argument is being advanced), but that's not the case here. Y Combinator's goal with HN is to make money indirectly, by discreetly, almost implicitly advertising itself as an outfit at the centre of things of interest to the "higher end" of the tech field, who might then (a) absorb awareness of it by osmosis and later (b) think of it first when planning their next startup. As such its goals are much more closely aligned with what that crowd enjoys, than with maximising engagement and thus eyeballs on ads.

Everyone being served the same algorithmic results is a genuine technical difference (I believe; do we actually know that HN doesn't customise the front page for logged-in users?). I think it's quite a fine distinction, though -- both sites are trying to hold the user's attention, one just omits to use an available level of customisation. I have no moral objection to either approach.

I don't see any technical difference? On HN, you scroll through an algorithmically generated list of submissions, vote them up and down, and post and vote on comments. This is pretty similar to FB. Just replace "cat videos" with "pelicans on bikes".

I think HN is good, BTW.

Absence of money does not at all imply perfect information. The problem you describe would occur equally with an ordinary debt between two parties, or barter, or indeed any system that doesn't entail the (laborious, impractical, necessarily incomplete) gathering of perfect information.

You mean like he already emphasises in TFA at various points, including the conclusion?

Ah, that makes sense, thanks! I was wrong.

I think I had the same issue. An L-seam looks to be any (horizontal line segment, vertical line segment) pair having a line segment endpoint in common, at least one of which could not be cut by any sequence of guillotine (end-to-end) cuts. So the (minimum possible) number of L-seams is well defined for a given problem instance, but not (necessarily) their specific locations.

I think you buried the lede a bit -- the "L-seams" in your diagram look like ordinary T-junctions, and I didn't figure out what you meant by the term until your later mention of maximising guillotine cuts (a term you don't explain but which I'm familiar with).

At the linked squaring.net site, the definition of a "Mrs. Perkins quilt" is a bit unclear. It says (a little offhandedly) that "An additional constraint is that the side lengths cannot have a common factor", however the example solution they provide violates this. Even if we interpret this constraint as narrowly as possible by pretending that 1 and the full side length of a square are not "factors" of its side length, there is still in their solution a 6x6 square and a 4x4 square, which share common factor 2. I guess they are looking for a way to prevent trivial solutions (e.g., any square with even side length can be partitioned into 4 equal-size squares), but either I'm misunderstanding something about their current definition or they are.

Oh, I think we could crank that headline even harder while still keeping one toe on the truth. For example:

"Elon Musk's Brain is Two Separate Organs"

That, and Every Sentence Is Punchy.

Every sentence sounds like it's trying to be in the trailer for a film.

Good point, I was wrong. Groups of weights having more zeros will be more likely, so should be favoured. Due to independence there won't be any meaningful difference in frequencies between two groups of weights that have the same number of zeros, but that doesn't invalidate the above.

Thank you, don't know how I missed that!

I totally agree. Characterising Google as "flailing" because Gemini isn't the #1 model for coding at the moment is... not realistic, to put it politely. In addition to the factors you mention, Google has enormous cash flow from its ad business, which translates to a long runway when compared to other frontier AI outfits.

It's possible that they will nevertheless snatch defeat from the jaws of victory, of course, but I personally think they're in the strongest position of them all.

I laughed :)

I'm not betting against Google at this stage, though. I just don't think Gemini is targeting the same "coding savant" niche as OpenAI and Anthropic. Gemini is fast with good general knowledge, and the TPUs behind it give Google a degree of freedom that Nvidia-dependent outfits lack.

For now, I'd say the biggest challenge Google has is overcoming the well-earned fear developers have that they will drop support or introduce backwards-incompatible changes at a moment's notice.

I was only objecting to this claim in the parent post:

> It won't matter what hard problem solving moat you think you have

If that is "a new dimension that is irrelevant to the original argument", you'll have to take that up with them. (I think it is a load-bearing part of their core argument.)

Yes. But (as a side effect) this has entailed resolving a number of conjectures that have been open for decades, only one of which (the Navier-Stokes blow-up) is attached to any controversy.

Impressive work!

I have a side question. I looked into the linked Raspberry Pi hacking challenge, and there's something very basic I couldn't figure out: It looks like the relevant script in the repo just writes 0xc0ff 0xffee a few times to the OTP as the "secret" to unlock. But given that $20000 was up for grabs, this can't possibly be the genuine secret being sought to claim the prize. (Indeed, I can't think of a secure way to install a secret from a public GitHub repo unless it involves running on-device code that encrypts something using some other, factory-installed secret key, which is just kicking the can down the road.) And given that the OTP on a brand new RP23550 is initialised to all zeros, it can't be that the genuine secret is programmed in at the factory either.

What am I missing? How does the genuine secret get installed on a person's RP2350?

> Lol no one cares. Most of that is smoke and mirrors anyways.

Lol tell me you don't know any practicing mathematicians without telling me you don't know any practicing mathematicians lol.

Are you aware of what frontier AI has done to mathematics in the last 6 months?

> It won't matter what hard problem solving moat you think you have

This is false on its face. There's a reason why a company, even today, would hire John Carmack over a person with no reputation, or offer him a higher salary if employing both people.

Unionising is definitely something worth considering. But, as with the question of whether it's better to compete or cooperate with other people in your field, is a nuanced question that depends on details you aren't acknowledging.

I disagree.

> AI finding proofs to open problems does not solve at all the question of how to produce new problems, and there is no indication imo that there is way to go with that with AI.

This has not yet been explored with AI only because solving hard problems is where everyone, practicing mathematician or layperson, understands 99.99% of the prestige to be.

AI's attention will not be directed towards generating interesting new conjectures until all the low-hanging prestige-rich fruit of famous decades-old conjectures have been mined, because it makes no economic sense for frontier AI companies to do so.

Agreed. One possible objection might be that they need fast random access to weights, but I skimmed parts of the paper and it looks like they process 128 entries at a time, which to me sounds like it should be amenable to better compression: short enough that better compression results could still be efficiently cached in faster local RAM, long enough that better compression would save useful amounts of memory per block.

I think the weights are iid distributed, so all 729 patterns will be roughly equally likely. That doesn't make this a bad idea though -- it just means there's no point trying to select the most common 256 to keep, since any 256 will be roughly as good.

With the exception of navigation waypoints, none of the things you listed would actually change AFAICT.

If the industry as a whole refuses to take measures whose only impact is to drastically improve its efficiency, that would be a net negative for everyone except jet fuel sellers. In a competitive environment, such bloat would soon die. But given the level of international regulation that (I assume) exists around air travel, it might linger indefinitely.