81 comments, 241 points.

https://news.ycombinator.com/item?id=39694045

163 points | 252 comments

For example, we can mostly agree CLIP does a fine job classifying images, except if you glue a sticky note saying "iPod" onto an apple, it would say classify it as such.

No matter the performance, these are categorically statistical machines reaching for the most immediately useful representations, yielding an incoherent world model. These systems will be proposed as replacement to humans, they will do their best to pretend to work, they will inevitably fail over a long enough time horizon, and a human accustomed to rubber-stamping its decisions, or perhaps fooled by the shape of a correct answer, or simply tired enough to let it slip by, will take the blame.

Consider pair(a, b) with first(pair) and rest(pair). I can define first-lens and rest-lens for [get-first, set-first] and [get-rest, set-rest] respectively. I could use them as simple getter/setter groupings with lens-view(rest-lens, pair) or lens-set(first-lens, 4, pair), which isn't so useful by itself, or I can compose then to get third-lens with lens-compose(rest-lens, first-lens), allowing me to lens-set(third-lens, 7, pair) without a nested call.

I just tried to summarize what I could read [1] [2] [3] and I don't claim to entirely understand it, but it's still better than the sibling comment dismissing it as abstract nonsense.

[1] https://docs.racket-lang.org/lens/lens-intro.html

[2] https://hackage.haskell.org/package/lenses-0.1.4/docs/Data-L...

[3] https://theswiftdev.com/lenses-and-prisms-in-swift/

  (define (juxt . fns)
    (lambda args
      (map (lambda (fn) (apply fn args)) fns)))

The representative democracy does attempt to solve this problem; you pick a representative you understand is aligned with your interests on the issues you do understand and trust they and their experts will do well on the rest.

It's not a perfect system, but unless the populace can be given free time and education to consider individual policies based on reliable evidence and not hearsay from biased sources, it's the best we can do.

Well, I'd say permanent brain damage from scaring alone wouldn't quite be worth an extra 1 click per second...

What it's showing really is a very different problem; the mouse is a fair show and inefficient input method. While some use-cases, like 2d/2d panning, drawing, etc., benefit from a pointing device, most of the typical UI can best be done by a keyboard. I doubt any brain computer interface can achieve 100+ WPM on arbitrary inputs with a low error rate.

The only real accomplishments of LLMs were how good the proposed use-cases sound on paper under competent implementation, and a theoretical solution to unstructured data parsing that's still too heavy to be worth a tiny bump in performance.

Do you want to live in a future where all human thought has been replaced by its surface level reproductions, made by big tech stuffing copyrighted works into a GPU farm with near-zero human labor? We both know it won't benefit you and me, our role is merely transitory in bootstrapping their self-improvement under the guise of a paid product, nor had the relationship between us and these tools been in any shape collaborative in the first place.

This fantasy targets the owner class, which can finally dream of labor decoupled from the laborer, the work simply costing no more than the price of electricity, all without the demands for livable compensation or following best practice. Even if the LLMs gained above-human performance in all domains of knowledge shortly followed by institution of a universal basic income, their invention will still have only been a force of stagnation, learned intellectual helpless, and overconsumption.

    > terrible; bland art...bad code

    > Do you want to live in a future where all human thought has been replaced by its surface level reproductions, made by big tech stuffing copyrighted works into a GPU farm with near-zero human labor?

Sure, in the way that technically this is a computable problem, but maybe not a simple one. Any exponential in the real world is a sigmoid and given all major AI labs, having spent years and incomprehensible sums of cache, have arrived at about GPT-4 performance, including OpenAI's latest release being a smaller model, should tell us something. Be the limiting factor corpus size, model parameters, or an architectural defect, we're clearly loosing momentum, at least until it's to be diagnosed and solved. Deferring to hypothetical futures without meeting the burden of evidence seems ill-advised, especially when it comes to incompetent use today.

> there's no putting AI back into the box

There's also no complete undoing of an oil spill, nor the practical possibility of unilateral nuclear disarmament. The weights are public, the corps is as well, as much as it had broken the open web to gather it, the architecture is known, ergo today's open-weight models are the baseline of capability for all future models. Still, seeing that LLMs form a natural monopoly given the required compute power to enter the field, we can enforce policies like mandatory statistical fingerprinting on all outputs of proprietary models beyond a certain size. LLM detection is also getting quite good; my hope is that adversarial fine-tuning will work out like Bayesian poisoning for email spammers, only giving the discriminator new stable patterns to look for. We as consumers do have power, however small and unconcentrated, to vote with our wallets, which a study posted here shown many are beginning to do [0]. The copyright question also gives quite a nice kill switch if our society decides this whole industry isn't that beneficial, removing the commercial incentive for training new models while providing some for detection.

There's great value from transformer networks, such as state-of-the-art speech recognition, that will make its way into consumer products and will be here to stay. As for the FADs like useless chatbots in every product, that may come under question.

[0] https://news.ycombinator.com/item?id=40390375

When I was in the GNOME bubble, I too thought GNOME was the be-all and end-all of Linux DE usability, with everyone else being savages slapping together UIs without so much as a style guide. Perhaps at some point this may have been partially true.

Today, all major DEs are fine. Plasma did not crash once since 2019 for me and I think its UX is quite nice, in the case of Dolphin in particular visibly better than GNOME's. At the same time, GNOME had routine issues with extensions, semi-frequent crashes, and odd non-compliant bits like refusing to use tray icons, breaking apps that depend on them, and the last time I checked, scaling was a mess.

I do think the GNOME/libadwaita ecosystem is a fantastic achievement and agree with many of their ideas, but it would be dishonest to say all other DEs are inferior and don't deserve any consideration as a default.

Also bloat is so very debatable. You can't on one hand complain about Gnome being unusable and then turn around and say Plasma is bloated. Uh, that "bloat" is the difference between the two!