gettingoverit

If the last story on HN was at December 26, that is.

It isn't competitive. They are paid by Google.

Worse, it's ridden with spyware, and is merely a honeypot for security-aware people that are not sufficiently paranoid to check any of the claims. Like, those VPNs from YT ads that use your IP to give AI companies residential proxies, the same kind of scam.

Spin up Wireshark and take a look at activity of Firefox. Try to shut the browser up. It won't work.

Even if they weren't a Google's proxy company, they would lose to standards commitees being infested by Google, and would have to play the "best luck catching up" game by constantly supporting new versions of JS, APIs and CSS features that nobody needs (except Google's YouTube will use them to stop you from using an adblocker).

FF is governed by ex-Oracle managers at the moment, singing the Google's song. Don't anthropomorphize your lawnmower.

> you’re testing for antibiotic resistance

...and also develop one.

By the look on the issues there, it seems the rest of the post is not that true either

Edit. Call me a hater, but... I know the guy! That's the guy from Google whose code never works in the most hilarious ways! See issues on the rest of his pinned repos.

You've made one of the best solutions, that matched what I thought of implementing myself, and at the time it was most needed. I think a couple of "thank you" are sorely missing in this comment section.

Thank you!

Dear Noam,

I remember your nickname from the Infernu project of yours. It was one of the best attempts at typechecking JavaScript I've ever seen, and it's as needed in modern days of TS team going insane as it never has been.

The things we don't really need more of are editors, CLI software, or Rust software.

Please reconsider your goals.

Okay, this will be a bit on a conspiracy theory side, but there was a paper recently describing how to do matrix computations on a RAM stick connected to FPGA, and they've shown it's possible to do it cheaper per flop that GPUs. Except, of course, there's a variety of RAM producers.

It might either be artificial to keep GPU prices where they are, or someone already started building their RAM-based AI datacenter.

Oh, that smell of molten keyboard plastic, those yellow spots burned into a display with its own heat exhaust, those laser-machined loudspeaker holes next to keyboard, all filled with grime! How I miss that time on a Macbook, with all the chords you have to press whenever you need a Home or End button to edit the line! Not to mention the power button right next to backspace.

It's so rewarding when its charger dies in a month, and you feel superior to your colleague, whose vintage 6 months old charging cable with none of that extraneous rubber next to the connector catches fire along with your office. What a time to be alive!

The best part is the motherboard produced in a way to fail due to moisture in a couple of years, with all the uncoated copper, with 0.1mm pitch debugging ports that short-circuit due to a single hair, and the whole Louis Rossmann's youtube worth of other hardware features meant to remind you to buy a new Apple laptop every couple of years. How would you otherwise be able to change the whole laptop without all the walls around repair manuals and parts? You just absolutely have to love the fact even transplanting chips from other laptops won't help due to all the overlapping hardware DRMs.

I'll go plug the cable into the bottom of my wireless Apple mouse, and remind myself of all the best times I had with Apple's hardware. It really rocks.

Oh, it's not a problem, until you're a Kuznetsov-father trying to get on a plane with your Kuznetsova-daughter, and you become a suspect of kidnapping until you don't prove it's actually your daughter. After you get your mortally scared daughter back from authorities, you probably start thinking that having the same surname is something you actually need.

And no, US authorities won't make it easy for you or her.

I think keeping it unsupported for a couple of years, and reluctantly pushing it off to volunteers who barely have enough technical experience to support it is quite close to "was killed".

Until recently Jest had a bug that made it crash due to sl (yes, the famous steam locomotive) running under the hood. This gives a hint at, ahem, the sophistication of its architecture.

The project is long in its EOL, and the only reason for its use is inertia, the jQuery kind of it.

Until they dump your startup for no reason, and you'll have to migrate off their undocumented proprietary infrastructure. Please read stories on HN. You're risking your business if you work with them.

I should remind that in a similarly cheerful mood FB dumped support of Jest and a bunch of other libraries. They have a long history of killing successful projects.

Worse, Vercel is involved, and I literally don't remember anything good about that company.

I'd recommend to be very cautious with such news, and use older versions of React for the next couple of years.

So there are - sloth on a tree - UFO in a forest - playground - alien on a beach

What's the 5th thing you have to find?

On top of other criticism here, I'd like to add that the article optimistically assumes that actors are completely honest with their benchmarks when billions of dollars and national security are at stake.

I'm only an "expert" in computer science and software engineering, and can say that - neither of widely available LLMs can produce answers at the level of first year CS student; - students using LLMs can easily be distingished by being wrong in all the ways a human would otherwise never be.

So to me it's not really the question of whether CS-related benchmarks are false, it's a question of how exactly did this BS even fly.

Obviously in other disciplines LLMs show similar lack of performance, but I can't call myself an "expert" there, and someone might argue I tend to use wrong prompts.

Until we see a website where we can put an intermediate problem and get a working solution, "benchmarks show that our AI solves problems on gold medalist level" will still be an obvious BS.

Loved the artstyle, but controls... uhh... let's say they were on par with a bad PC port of Sonic Adventure. I can barely control where I'm going, because the camera just rotates randomly. Something I'll never miss from the old days.

That's not idealistic, that's how arithmetics work. If you use the same generic thing more times, you have the higher chance of discovering it broken. The fact that you've run into cases means that chance is never zero, and is irrelevant to the discussion.

As was already mentioned in the article, PB solve a problem that likely only Google has, even if that. State of the art nowadays is JSON/JSONL. If it grows too large, gzip it.

When someone is using third-party closed proprietary technologies to be "not like the rest", it usually doesn't work that well for their business.

The technology is "alive" until it didn't follow the path of Closure, GWT, and the rest of "we use it on the most loaded page of the world" technology. PB will be on the same graveyard soon.

> would imply <1h runtime, or a >50W consumption at idle

That's the case.

If there are errors in implementation of general constructs, they tend to be visible at their every use, and get rapidly fixed.

Some general constructs are better than the others, because they have an algebraic theory behind them, and sometimes that theory was already researched for a few hundred years.

For example, product/coproduct types mentioned in the article are quite close to addition and multiplication that we've all learned in school, and obey the same laws.

So there are several levels where the choice of ad-hoc constructs is wrong, and in the end the only valid reason to choose them is time constraints.

If they had 24 years to figure out how to do it properly, but they didn't, the technology is just dead.

My CPU is at over 5GHz, 1% load and 70C at the moment. That's in a "power-saving mode".

If nothing would be wrong, it'd be at something like 1.5GHz with most of the cores unpowered.

> might be my Linux setup being inefficient

Given that videos spin up those coolers, there is actually a problem with your GPU setup on Linux, and I expect there'd be an improvement if you managed to fix it.

Another thing is that Chrome on Linux tends to consume exorbitant amount of power with all the background processes, inefficient rendering and disk IO, so updating it to one of the latest versions and enabling "memory saving" might help a lot.

Switching to another scheduler, reducing interrupt rate etc. probably help too.

Linux on my current laptop reduced battery time x12 compared to Windows, and a bunch of optimizations like that managed to improve the situation to something like x6, i.e. it's still very bad.

> Is x86 just not able to keep up with the ARM architecture?

Yes and no. x86 is inherently inefficient, and most of the progress over last two decades was about offloading computations to some more advanced and efficient coprocessors. That's how we got GPUs, DMA on M.2 and Ethernet controllers.

That said, it's unlikely that x86 specifically is what wastes your battery. I would rather blame Linux, suspect its CPU frequency/power drivers are misbehaving on some CPUs, and unfortunately have no idea how to fix it.

Even though this should be the way to go, it still leaves a bitter taste.

Remember ESM fiasco? Now we have a few years of that all over again, this time with different versions of TypeScript, settings, and tsgo. Good news!

Probably this wouldn't be a problem if Web was somewhat anonymous, so that merely stumbling upon a security issue, or using website in a regular way would not constitute a crime for the lack of the person to put that crime onto.

Also if things stored in those databases weren't plain strings, but tokens (in asymmetric cryptography sense) so that only the service owns it, and in case of a leak user can use it to get a payout from the service, this problem would be solved.

But no business is interested in provably making their users secure, it would be a self-sabotage. It's always just a security theater.

Nice to know!

I thought about it in terms of cardinalities of types. If A and B are types with |A| and |B| values correspondingly, there are |B|^|A| possible functions A -> B.

Another funny thing is that if you consider

forall a. (a -> a) -> (a -> a)

the type of natural numbers (weird, I know, but basically we encode numbers in unary with number of times we compose (a -> a) to itself), then exponentiation on such numbers will be

a ^ b = b a

Oh wow. I've seen so many "AI" that apparently I started reading around them.

F is a left adjoint of G, and tautology below is a tensor-hom adjunction :)

That's because you can consider categories of preorders and formulas where these operators will be morphisms.

(b >= c) && (a >= b) -> (a >= c) is a composition.

The more interesting consequence is that function types and implications are different names for the same thing. This is a Curry-Howard-Lambek correspondence.

This means that in order to prove

(b -> c) -> (a -> b) -> (a -> c)

it's enough to implement a function

f g h x = g (h x)

Another consequence is that exponentiation a^b can be considered the same thing as b -> a.

a^(bc) = (a^b)^c

(b && c) -> a = c -> b -> a

> something to do with the physics of absorbing light to drive a chemical reaction

Exactly that. Blue does two steps of the process, while red does only one. There's a cost for synthesizing all that machinery, so absorbing green would just be not worth it.

> 450nm and 680nm are not harmonics

In fact they're in 3:2 ratio with 1% margin. But they don't have to be. Take a look at fluorescence: it converts one wavelength to another, and they don't have to be multiples of each other. Once photon gets absorbed onto a chemical, the electronic structure of the molecule decides what will happen to it.

Actually AI has a lot to do with material design. I can't do a proper explanation, so please bear with this information dump:

- We don't have a strong physical theory for solid state physics. Quantum stuff doesn't scale well from 1 atom to a mole of atoms, because 10^23 goes into the exponent of number of energy levels in the system, and then we also have to model interaction of those levels.

- Physical properties of materials depend on their crystal structure, unevenness of that structure, spectrum of size of crystals, temperature, pressure, fields they're exposed to and current position of stars in the sky. Even the "wrong" solid state physics equations we have are highly non-linear.

- State-of-the-art effects are usually achieved with combinations of such materials.

- Exact parameters of the process used to put those materials together radically change the behavior of the system. Put that nanolayer with a different sort of vapor deposition or at different pressure, and the thing will stop working. Ever wondered why we don't produce all the neodymium magnets at N55 grade? Because even precise description of the process is not enough.

- AI doesn't care about the exact physics, but is sometimes very good at navigating in the large parameter space.

- Google have recently made AI that predicted thousands of novel material structures. They found more materials than were found by all human research over the whole humankind history.

I wouldn't expect AI to explain what's going on in solid-state physics anytime soon, but exploring crystal structures, doping, and process parameters automatically might actually get us new materials a couple hundred years faster.

That said, I missed the slight semantic difference between "being funded by" and "owning", even though I don't see how that would be different in practice.

Regarding the second point, I don't see how "hav[ing] a verbal agreement that these materials will not be used in model training" would actually discourage someone from not doing it, because breaking that kind of verbal agreement wouldn't cause any harm.

I have not been aware of those other claims on Twitter, but IMO they do not create sufficient basis for an investor fraud case either, because Twitter is not an official way of communcating to investors, which means they can claim whatever they want there. IANAL though.

I'm really looking for FrontierMath-level problems to be solvable by OpenAI models, and being able to validate it myself, yet I don't have much hope it will happen during my lifetime.

One piece of evidence I have is that most powerful reasoning models can answer at most 10% of my questions on PhD-student level computer science, and are unable to produce correct implementations for basic algorithms, being provided direct references to their implementations and materials that describe them. Damn, o3 can't draw an SVG arrow. Recent advancement in counting "r" in "strawberry" is basically as far as it goes.

I don't know what exactly is at play here, and how exactly OpenAI's models can produce those "exceptionally good" results in benchmarks and at the same time be utterly unable to do even a quarter of that in private evaluation of pretty much everyone I knew. I'd expect them to use some kind of RAG techniques that make the question "what was in the training set at model checkpoint" irrelevant.

If you consider that several billion dollars of investment and national security are at stake, "weird conspiracy" becomes a regular Tuesday.

Unfortunately I can't see beyond the first message of that primary source.