Not sure if this was created with LLM help, but I suspect so? Not because the page is buggy (it is, though, crashed on my iPhone), but because they make data visualization so accessible. This type of presentation used to take days of work; now, if you find a unique piece of data, it's only a few hours of work to create a beautiful animated visualization.
I do think this would be more compelling with some additional context or data integration. Zoom, the ability to click and see the full details about each station, which company (my guess is that it's all JR?).
Ok final note: the intersection of Japan and trains is basically HN crack, and I love it.
I created pages with Claude before and it's very very obvious when you see one. From the font choice to the color palette, and the style of the boxes. In fact if anyone has an effective prompt that says "please don't make this look like the average Claude page" please post it!
I've had some luck giving either an example website to ape or listing out a particular era, monkey see monkey do seems to help a bunch.
I've done each of the 3 for side projects below to pretty good effects.
> This website will be run by IE6 and Windows Mobile 6, so use no dependencies, semantic HTML, a 3-pane layout, and only use JS (es3!) where absolutely necessary (and where necessary, put the script at the end of the body).
When I'm not specifically targeting support for retrocomputers I do something like this, then iterate until it looks right.
> Go look at Dokuwiki, django defaults, and common web 2.0 color schemes, use those for UI inspiration. Keep a 3-pane desktop-first layout, but enable mobile responsiveness with media queries. Use semantic html5 and prefer older boring solutions like surgical jquery or htmx-style islands of interactivity where needed, otherwise do not bring in dependencies without my say so.
And finally, if I'm doing a web app that I'm vibing out with the web stack because I want it one-shotted and not trying to do a good rust core with strong ports/adapters API surface for web or native client callers, I do something like this:
> This is a local web app, the frontend, backend, and desktop are all on the same machine. Use naive and simple development patterns that you document the style as you go, pick a boring web framework and use it idiomatically, but remember that some tricks that are intended to keep network round trips down are not as necessary because network penalties are not as bad as real traffic.
Granted, the above I don't like as much, but it does produce more 'modern' looking sites by default.
But what I find works best is to point Claude at a design system documentation website (your own company's or another public source) and tell it to use that design style. It usually does OK, and the results are usually much more in line with that style and not as Claude-y.
Fable is really, really good at this. My workflow involves giving it a bit of human inspiration, through asking it to generate a few different design templates/scaffolds (without building the whole frontend first).
Then I iterative and give it feedback, point out all the parts I don’t like, sometimes mixing and matching.
I’m sure you can do this with Opus too, but Fable is a better designer.
Not animated, but it's fully zoomable (with variable LOD at different levels), searchable (with fuzzy matching), info popups on not just stations, but things like track speed limits, and tunnels. What I do that's really different is that I'm VERY heavily manipulating the underlying data (The pre-processing pipeline is currently at something like 12k lines of python and counting, most of it doing things like advanced network topology, so that we show the actual track layouts, which if drawn at truly geographically accurate scale would just collapse to a single line.
I'm showing tons of detail..... 3.5k+ stations, and every single track - not just the main running lines, but every siding and yard track.
It’s obvious on how much information is unnecessarily repeated. One of the main give away of AI is that the text is like an Atlantic article but worse, with very-very-very low information density. Full with sentences, paragraphs, pages which add absolutely nothing.
I get the sentiment. I don't love that different browsers have different behavior even on standards compliant code. But I've also done enough web development to know that if your page crashes safari in the main user flow (in this case, just hitting 'play'), the dev owns the bug.
The combination approach jives well with my use of the models in a number of areas. I guide models to use best-in-class algorithmic approaches as available. (Eg using constraint solves for a particular problem where pure Monte Carlo rarely gives "in-bounds" data.)
I find it odd that frontier models often don't suggest the most powerful methods for crushing problems, but it may be that the training data doesn't actually have "good enough" experts on the problems I encounter. If the experts don't know about the best ways to solve the problem, they'll get dinged in training for even trying.
Do you enumerate the options of the algorithms to the models? I've tried to do "algorithmic discovery" with these systems, e.g. openevolve, and to be honest the models didn't really focus on that part.
Instead they were focusing more on optimizations of the existing algorithm that has been implemented. Maybe it's an artifact of the problem I was throwing to them (I was asking to optimize the implementation of select_k in Arrow, which is currently using a max-heap streaming algorithm).
I've started documenting my journey with this here: https://www.kostasp.net/posts/16-ai-experiments-apache-arrow
in case you want to take a look. Any advice would be highly appreciated, I'm looking for more inspiration on how to torture myself with that stuff.
Seem reasonably concise, but I think Kreyzsig's Introduction to Functional Analysis with Applications fills the "gap" that this paper wants to fill. It's readable, has applications, exercises, and is more complete.
This take clearly has a bone to pick. But ignoring that, the first sentence is just not reflective of the reality here—xAI is making a killing on renting out its GPUs, way more than "just power". The dynamics that normally make infrastructure providers have slim margins don't apply when demand far outstrips supply; the situation right now is closer to monopoly pricing power.
It will likely take a few years for supply to fully catch up, which means xAI will eat well for a while.
I can see a world where a few data centers come on line this year and reduce margins a bit, but it's crazy to think the margins will go to "cost of electricity plus a few percent" anytime soon.
In the article, it states that the two deals will cover the entire cost of SpaceX's AI buildout in 18 months. OpenAI and Anthropic would kill for that kind of cashflow.
xAI is a failure of an AI company from a consumer perspective. They invested a large amount of money into owning their own infrastructure, while driving away consumers with their right-wing or "alt right"-ish branding and having a reputation of X users abusing the AI services.
Turns out there was another company with a much better reputation for which the compute is a better fit. Now that the data centers are being put to use, they actually make them a little bit of money instead of losing money.
That story roughly tracks the one I hold. One piece that's missing is that grok's / X's image also made it radioactive to the best researchers. 'Aligned AGI' is an easier sell to the best engineers than 'abusive neo-Nazi chatbot with a porn problem'.
Right, and zero non-anonymous sources. A person plausibly "close to Anthropic" or vaguely "with knowledge of the situation" using the stock phrase "the best defense is a good offense"... this is not a news article y'all. It's speculation.
The focus on AI is just to capture some of the current zeitgeist. Socialists generally think that most large businesses be run for the public benefit in some form or another.
This is basically making that point with AI companies as their true influence is rapidly increasing. The rhetorical strategy here is to hook socialist ideas onto something people are already thinking about, a land-and-expand rhetorical strategy.
Personally I'm against the proposal, but the details are not the point because it's not going happen in this form. It's about changing minds first, then changing policy down the road.
Yeah, the order itself seems like a fairly reasonable response to Mythos level capabilities. It does solve one problem of the frontier labs, which is safely coordinating releases without hitting antitrust regulations. It also makes a bigger moat for incumbents.
> You might be able to come up with a more efficient algorithm.
Challenge accepted. Suppose we want to know the answer to 3 decimal places (so we'd match the headline). And suppose I allow my algorithm to be wrong one in a thousand times ("probably approximately correct").
Then sample some constant number C of random 64 bit integers. Run the following algorithm which separates each random sample into one of three classes: Y (has 32 but factors), N (does not have 32 bit factors), U (unknown).
Check if prime using probabilistic miller rabin. (Error prob goes to zero exponentially fast). If prime, return N. If it's not a prime, then run T steps of pollard rho to determine whether the number has 32 but factors; return Y,N, or U depending of the factors found up to step T.
The key observation is that T can be chosen to make the UNKNOWN class very small (with high probability), and so our estimate should rapidly converge to 17%Y, 83%N, ~0.001%U
For fixed error tolerance, this would run in roughly a constant number of iterations, independent of N.
Even if you have 32-bit factors the number may not be the product of two 32-bit numbers. For example 2^62*3 cannot be split as either (2^32, 2^30*3) or (2^31, 2^31*3). In both cases one factor does not fit in 32 bits.
Ah, yes, this is true, but it's not really a counterexample, since this would show up in the N or U bucket. But I think the issue is that my sketch algorithm is, well, pretty sketchy.
Working on coding it up... it converges to 17±0.5%for N=64 bits in a javascript implementation relatively quickly, but for N=96, it really slows down as Pollard's Rho starts with large factors. This means my fast-and-loose assumption that "a constant number of iterations of Pollard Rho would work" isn't actually true!
Ok, chatting with Claude and I found a way to salvage the approach! (Also, it's apparently in the paper that's referenced in the post, kinda cool that my sketchy algorithm got halfway to the published result).
Basically, replace the Pollard Rho partial factorization with a method of Kalai [1] for generating random numbers _together with their prime factors_.
I'm able to run this at about 30 samples per second at 160bits, giving an estimate of ~14.1% of 160-bit numbers factoring into two 80-bit numbers.
I think by "have 32 bit factors" the GP actually means "has a factorization as a product of two 32-bit numbers", rather than the natural meaning that doesn't work, as you pointed out.
Not sure if this was created with LLM help, but I suspect so? Not because the page is buggy (it is, though, crashed on my iPhone), but because they make data visualization so accessible. This type of presentation used to take days of work; now, if you find a unique piece of data, it's only a few hours of work to create a beautiful animated visualization.
I do think this would be more compelling with some additional context or data integration. Zoom, the ability to click and see the full details about each station, which company (my guess is that it's all JR?).
Ok final note: the intersection of Japan and trains is basically HN crack, and I love it.
reply