MontyCarloHall

>What are you basing this off of? Do you actually have any experience making art? Or is this just learned helplessness talking?

I've done ≥weekly life drawing classes for 20+ years, and have observed the distribution of progress people make over time. Based on my observations and conversations with my teachers, I agree with you that a nontrivial fraction of adults starting with zero artistic ability can be trained to an advanced degree. But I disagree that this holds true for "anyone"; a larger fraction cannot be trained beyond a basic level.

>Also please stop implying I said any 12 year old can do this. I didn't.

You literally said: "It's impressive that he did it at 12, but he had years of focused practice under his belt before he did this one. Anyone can do this level of work - they just need to actually learn it." To me, that heavily implies any 12 year old with sufficient training is encompassed by that "anyone."

I very much doubt the majority of adults with sufficient practice could do this level of work. I can say with 100% certainty that an infinitesimal minority of 12 year olds with sufficient practice could do this level of work.

>Anyone can do this level of work - they just need to actually learn it.

Sorry, that's like saying with enough math practice, any kid could perform at the level of young Terry Tao (e.g. teaching himself calculus at 8, winning a gold medal at the International Math Olympiad at 12). Some people are just intrinsically talented at certain things, and no amount of hard work in people lacking those intrinsic talents will get them to that level. This is indisputable when it comes to athletic talent; everyone would agree that no matter how much an average tall person practices basketball, they will never play at the level of Michael Jordan, LeBron James, or even the lowest ranked NBA player [0], for that matter. Artistic and intellectual talent is no different.

[0] https://old.reddit.com/r/todayilearned/comments/1oxpng5/til_...

Of course it's much better, Schongauer was ~25 when he did the engraving. Michelangelo was 12 when he copied it. Likewise, it goes without saying that Haydn's symphonies circa 1765 were much better than Mozart's from the same time, since Haydn was ~30 years old and Mozart was ~10 years old.

The remarkable thing about the early painting/symphonies isn't the absolute quality of the work, it's that they showcase the artists' intrinsic baseline talents, which they would then leverage as their skills improved with maturity to become some of the greatest artists of all time.

Agreed! I knew about the second chance pool but didn't realize it reset all the timestamps.

I'd call it "Stochastic Newton's Method" then. :-)

>If the Hessian-vector product is Hv for some fixed vector v, we're interested in solving Hx=v for x. The hope is to soon use this as a preconditioner to speed up stochastic gradient descent.

Silly question, but if you have some clever way to compute the inverse Hessian, why not go all the way and use it for Newton's method, rather than as a preconditioner for SGD?

Bizarre; this was originally posted 5 days ago (as evidenced by clicking on the poster's submission history [0]), but timestamps on the submission and on comments throughout the post show 11 hours ago. @dang, is this a bug in HN software? I was feeling some serious déjà vu scrolling seeing this and scrolling through the comments.

[0] https://news.ycombinator.com/submitted?id=jszymborski

It is indeed a builtin, but `/bin/[` still exists for compatibility reasons!

   $ which [
   /bin/[
   $ type [
   [ is a shell builtin

The same is true for the `test` command:

   $ which test
   /bin/test
   $ type test
   test is a shell builtin

If it {tastes,feels,looks} like eating shrimp, I don't see why people wouldn't go for it. After all, much weirder looking bugs of the sea (e.g. lobsters, crabs) are considered luxuries.

All user data is E2E encrypted, so the government literally cannot force this. This has been the source of numerous disputes [0, 1] that either result in the device itself being cracked [0] (due to weak passwords or vulnerabilities in device-level protection) or governments attempting to ban E2E encryption altogether [1].

[0] https://en.wikipedia.org/wiki/Apple%E2%80%93FBI_encryption_d...

[1] https://en.wikipedia.org/wiki/Crypto_Wars

Because their focus on user privacy makes it difficult for them to train at scale on users' data in the way that their competitors can. Ironically, this focus on privacy initially stemmed from fumbling the ball on Siri: recall that Apple never made privacy a core selling point until it was clear that Siri was years behind Google's equivalent, which Apple then retroactively tried to justify by claiming "we keep your data private so we can't train on it the way Google can." The result was a vicious cycle: initially botch AI rollout -> justify that failure with a novel marketing strategy around privacy that only makes it harder to improve their AI capabilities -> botch subsequent AI rollouts as a result -> ...

To be clear, I'd much rather have my personal cloud data private than have good AI integration on my devices. But strictly from an AI-centric perspective, Apple painted themselves into a corner.

A neural-GP could probably be trained with the same parallelization efficiency via consistent discretization of the input space. I think their absence owes more to the fact that discrete data (namely, text) has dominated AI applications. I imagine that neural-GPs could be extremely useful for scale-free interpolation of continuous data (e.g. images), or other non-autoregressive generative models (scale-free diffusion?)

It's utterly baffling to me that there hasn't been more SOTA machine learning research on Gaussian processes with the kernels inferred via deep learning. It seems a lot more flexible than the primitive, rigid dot product attention that has come to dominate every aspect of modern AI.

100% agreed. Attention finally clicked for me when I realized "wait, it's just a transformed, weighted dot product and has nothing to do with key/value lookups." I would have gotten this a lot faster had they called the key matrix \Sigma.

The confusing thing about attention in this article (and the famous "Attention is all you need" paper it's derived from) is the heavy focus on self-attention. In self-attention, Q/K/V are all derived from the same input tokens, so it's confusing to distinguish their respective purposes.

I find attention much easier to understand in the original attention paper [0], which focuses on cross-attention for machine translation. In translation, the input sentence to be translated is tokenized into vectors {x_1...x_n}. The translated sentence is autoregressively generated into tokens {y_1...y_m}. To generate y_j, the model computes a similarity score of the previously generated token y_{j-1} against every x_i via the dot product s_{i,j} = x_i*K*y_{j-1}, transformed by the Key matrix. These are then softmaxed to create a weight vector a_j = softmax_i(s_{i,j}). The weighted average of X = [x_1|...|x_n] is taken with respect to a_j and transformed by the Value matrix, i.e. c_j = V*X*a_j. c_j is then passed to additional network layers to generate the output token y_j.

tl;dr: given the previous output token, compute its similarity to each input token (via K). Use those similarity scores to compute a weighted average across all input tokens, and use that weighted average to generate the next output token (via V).

Note that in this paper, the Query matrix is not explicitly used. It can be thought of as a token preprocessor: rather than computing s_{i,j} = x_i*K*y_{j-1}, each x_i is first linearly transformed by some matrix Q. Because this paper used an RNN (specifically, an LSTM) to encode the tokens, such transformations on the input tokens are implicit in each LSTM module.

[0] https://arxiv.org/pdf/1508.04025 (predates "Attention is all you need" by 3 years)

>They all are on foliage, which changes depending on season and wind. Geolocating that picture accurately _automatically_ using features is next to impossible.

Seems plausible enough to me. The trees are evergreens in a place that doesn't get snow, and the keypoints are mostly grounded on stable parts of the trees (trunks or thick branches), which barring gale-force winds probably don't fluctuate all that much.

The part that gives me pause are the keypoints that map the hood of the car to the pavement, and the point on the far right that maps the ledge to the pavement. How can a system robust enough to map foliage also return such blatant false matches?

They both fit Gaussians, just different ones! OLS fits a 1D Gaussian to the set of errors in the y coordinates only, whereas TLS (PCA) fits a 2D Gaussian to the set of all (x,y) pairs.

>But at a more fundamental level, our material quality of life is that of literal kings.

This cannot be overstated. To wit, a Honda Accord (or equivalent mid-range car of today) is objectively superior to a Rolls Royce from the 90s in terms of amenities, engine power/efficiency, quietness, build quality, safety, etc. The same is true for quality-of-life improvements across a vast swath of consumer goods, and therefore consumer lifestyles.

Without growth, it's unlikely we'd see those improvements manifest. Carefully consider the lifestyle of someone living several decades ago. Would you honestly want to live such a lifestyle yourself? That's where degrowth likely leads. As the article says, "I feel it’s impossible to convince Europeans to act in their self interest. You can’t even convince them to adopt air conditioning in the summer."

Exactly. If you're working on an application where these numbers matter, Python is far too high-level a language to actually be able to optimize them.

I agree it's incredibly stupid. Just saying that our current society's valuation of scarcity makes access to an elite education very much a zero-sum game.

Very interesting observation. Multicore systems have been fairly standard for the last 10+ years, and while you occasionally notice a misbehaving process hog an entire core, it never visibly impacts system performance because there are still several other idle cores, so you don't notice said "hogs."

It's much rarer to see misbehaving multithreaded processes hog all of the cores. Perhaps most processes are not robustly multithreaded, even in 2025. Or perhaps multithreading is a sufficiently complex engineering barrier that highly parallelized processes rarely misbehave, since they are developed to a higher standard.

>Celeron CPUs were usually CPUs that shared the same core architecture as the current Pentium standard, but often had a lower core clock speed, lower core memory speed, and/or had smaller L2 caches.

This had an interesting side effect: Celerons of that era overclocked extremely well (stable 300 -> 500MHz+), due to the smaller and simpler on-die L2 cache relative to the Pentiums of the era, whose L2 cache was much larger but had to be off-die (and less amenable to overclocking) as a result.

An overclocked dual Celeron could easily outperform the highest-end Pentiums of the era on clock-sensitive, cache-insensitive applications, especially those designed to take advantage of parallelism.

>The fact of resource extraction from society and externalities like pollution not being counted by capitalist because they "can’t count them" and just bundle them as externalities demonstrably destroy any concept of non-zero sum game

The article explicitly addresses this:

   The fact that Capitalism is non-zero-sum doesn't mean it is necessarily positive-sum. An economy that gets out of balance can produce very negative results (which are still non-zero). Cons of capitalism: — Can not be relied on to provide adequate social services, including healthcare and education. — Can be expected to run at a cost to externalities like the environment. — Can produce products that are detrimental to well-being.

Based on your other comment [0], it seems you have a bad-faith axe to grind against this site.

[0] https://news.ycombinator.com/item?id=46434065

>The difference between these two sides is which value they prioritize.

Yup. Though there is a third option: completely ignore the meritocracy vs. equity zero-sum game and simply argue that demographic-based weighting of applicants is an ineffective way to rectify those historical injustices. It is treating a symptom, not the underlying disease.

>Indeed, I'm sure the author would agree that part of the comprehensive solution is to increase the amount of university admission slots.

A large part of the value of elite education is its scarcity, and adding more slots dilutes that value.

>The conflict is not a disagreement over effectiveness. It's a misalignment between meritocracy and equity.

A lot of proponents of affirmative action will agree with this. They'll explicitly acknowledge that people admitted under AA will be underqualified, due to factors mentioned in the article:

   [Minorities] may lack foundational skills (taken for granted in more affluent households and schools) and therefore might require breaks from study, which can lead to dropping out. They might have developed unhelpful habits or attitudes formed in teen years, or a sense of identity tied up with being part of a historically maligned group, affecting confidence and performance. [Affirmative action] does nothing to address these factors.

Said proponents would agree that AA is a failure if assessed strictly by these criteria. However, they would then go on to say that the benefits conferred by an elite education to the current crop of AA beneficiaries lead to future generations of minorities being less likely to experience the aforementioned issues, so after accounting for all future externalities, AA is a net good. As Justice O'Connor famously wrote in Grutter v. Bollinger (2003) [0],

   It would be a sad day indeed, were America to become a quota-ridden society, with each identifiable minority assigned proportional representation in every desirable walk of life. But that is not the rationale for programs of preferential treatment; the acid test of their justification will be their efficacy in eliminating the need for any racial or ethnic preferences at all. […] It has been 25 years since Justice Powell first approved the use of race to further an interest in student body diversity in the context of public higher education [California v. Bakke (1978)]. Since that time, the number of minority applicants with high grades and test scores has indeed increased. We expect that 25 years from now, the use of racial preferences will no longer be necessary to further the interest approved today.

That said, it's been almost 25 years since she wrote that (and 50 years since California v. Bakke), and it's debatable whether those future externalities have manifested.

[0] https://supreme.justia.com/cases/federal/us/539/306/

Relevant XKCD: https://xkcd.com/808/

Might the creator of Claude Code have some … incentives … to develop like that, or at least claim that he does?

As someone who frequently uses Claude Code, I cannot say that a year's worth of features/improvements have been added in the last month. It bears repeating: if AI is truly a 10x force multiplier, you should expect to see a ~year's worth of progress in a month.

Very true! I liken AI to having an endless supply of newly hired interns with near-infinite knowledge but intern-level skills.