Interesting story about Google publishing someone's phone number on searches for them when they gave the number to Google for account verification/security:
https://danq.me/2025/05/21/google-shared-my-phone-number/
Reminds me of the time a company I worked for (AFAIK) accidentally used phone numbers obtained the same way for ad targeting and got fined $150M
Exercises in benchmarking and experimental design, part 4:
![19. You have a Sandy Bridge processor. You measure the memory latency and find that a read from memory takes 60ns. We have the following loop, where ITERS is some large number
for (size_t i = 0; i < ITERS; i += CACHE_LINE_SIZE) {
cnt += buf[i];
}
A. What parameters do we need to know to guess how long this loop will take to run (assuming everything is warmed up and/or we're observing some kind of steady state behavior) and what are the parameters? [answer at the end to avoid spoilers, feel free to skip to the end if you don't want to search up these parameters]
B. How fast does the loop run?
C. Without trying it out, how accurate do you think ChatGPT, Gemini, etc., are at answering this question?](https://files.mastodon.social/media_attachments/files/114/414/397/986/082/365/original/1280841fcfce4e60.png)
undefined behavior is pretty well understood at this point, but a piece of the puzzle that has always been missing is "how well could a compiler like LLVM optimize, without leaning on UB"
here's a very cool new paper that takes a crack at answering this, for LLVM:
@ericswpark It's already the case that, without trying to do any kind of proof of work check, a lot of sites take already 5s+ to load on a slow Android device.
I'm not saying companies are going to get serious about this, but if they were to, most companies don't seem to care about people not on high-end devices and I don't see why they would suddenly start caring now.
Is one of the big drivers of consumer demand for compute going to be solving expensive proof of work proxies?
I'm still semi-banned from reddit (https://mastodon.social/@danluu/113426851280402188) and, ironically, had to start browsing it with a scraper because the scraper has good anti-blocking techniques and I don't, much like how way SEO spammers are better at ranking than people who make legit content.
With the resources AI companies have, proof of work can't "win", but it seems like the least ineffective technique?
@timrice Interesting idea. I don't have a good intuition for how effective that would be. I think rotten tomatoes / metacritic style aggregation is fundamentally broken for things that can be objectively evaluated because most reviews are wrong, so you need very unequal review weighting.
In principle, what you're describing could work, but I would guess that, today, LLMs wouldn't be great at evaluating credibility? But it could work fairly well with a human in the loop?
@rwv37 IIRC, this is a side effect of an issue with Hugo. If I add a title, it breaks something else.
If I update Hugo, it breaks many different things. I've been avoiding putting time into Hugo fixes since I should just migrate off, but I'm not sure it's really worth the effort when the whole thing kinda sorta works, so I have issues like this. Sorry!
And, sure, the LLM generated code isn't great, but if I compare the LLM-generated game and "AI" in https://danluu.com/codenames/ to commercially successful online board game implementations that were created by human programmers, the LLM generated version is faster and less buggy.
Per the argument in https://danluu.com/customer-service/, AI doesn't have to be very good to replace humans in a lot of roles because, in practice, humans often aren't all that good (see also, https://danluu.com/p95-skill/).
This exchange between some programmers and a non-programmer typifies what I was getting at in https://danluu.com/codenames/#appendix-writing-the-code-for-the-post
Programmer 1: GPT-4o and Claude Haiku are useless for programming
Programmer 2: Claude Sonnet is useless for programming
Non-programmer: What do you mean GPT-4o is useless? I don't know how to program and created an app that makes $10k/mo with GPT-4o
LLMs have allowed non-programmers to produce apps for years and programmers are calling these things useless for programming.
![antirez 3 hours ago | next [–]
About "people still thinking LLMs are quite useless", I still believe that the problem is that most people are exposed to ChatGPT 4o that at this point for my use case (programming / design partner) is basically a useless toy. And I guess that in tech many folks try LLMs for the same use cases. Try Claude Sonnet 3.5 (not Haiku!) and tell me if, while still flawed, is not helpful.
But there is more: a key thing with LLMs is that their ability to help, as a tool, changes vastly based on your communication ability. The prompt is the king to make those models 10x better than they are with the lazy one-liner question. Drop your files in the context window; ask very precise questions explaining the background. They work great to explore what is at the borders of your knowledge. They are also great at doing boring tasks for which you can provide perfect guidance (but that still would take you hours). The best LLMs (in my case just Claude Sonnet 3.5, I must admit) out there are able to accelerate you.
duped 2 hours ago | parent | prev | next [–]
> Try Claude Sonnet 3.5 (not Haiku!) and tell me if, while still flawed, is not helpful.
It's not as helpful as Google was ten years ago. It's more helpful than Google today, because Google search has slowly been corrupted by garbage SEO and other LLM spam, including their own suggestions.
mvkel 1 hour ago | parent | next [–]
I'm surprised at the description that it's "useless" as a programming /...](https://files.mastodon.social/media_attachments/files/113/795/447/771/216/740/original/442392981fe24725.png)
@eighty0n3 Sorry about that.
IIRC, I changed that when I saw that someone posted an old archive.org version of mine somewhere because they didn't like my "new" formatting. I often add fixes / corrections, and in my testing the current formatting performs better than the old formatting (at the time, I only had the before/after data, but I've also done some A/B testing since then).
Exercises in benchmarking and experimental design, part 3:
A version of Missile Command for the Commodore 64 where the bottom of your screen is the game state in memory and missiles cause memory corruption, which eventually causes you to lose: https://csdb.dk/release/?id=135463.
In the video below, a missile broke my controls and caused my cursor to move down and to the left so I couldn't stop other missiles.
@jefftk It's possible, but then I'd wonder why the only two EVs on the most fatal models list are Teslas. If a base Tesla is fast, a base E-tron or I-Pace is fast, the Bolt EV and some of the cheap Hyundai/Kia EVs for sale then aren't slow, and the base Taycan is very fast (3.4s measured 0-60)
It seems somewhat weird that, if it's something general about EV acceleration, only Teslas make the list of most fatal models, esp. when there are much faster EVs (because they don't sell a slow version)
@jefftk I haven't seen a more detailed analysis, but I don't think acceleration should be the primary driver because:
1. Only two fast car models are in the top 20 list.
2. If we look at how fast a typical model is for a luxury brand, a base BMW 3-series has a measured 0-60 of 5.2s (vs. 5.6s for a base Tesla Model 3) and a base BMW X5 has a 0-60 is at 3.9s (vs. 3.9s for a Tesla Model X LR, can't find base numbers). These luxury brands that sell fast cars do not have high fatality rates.
I find it really interesting/surprising that Tesla topped the fatalities per mile ranking from 2018-2022.
Fatality rate is strongly negatively correlated with price and weight and Teslas are much more expensive and heavier than average.
The commentary I've seen says that the cars are safe so it must be about Tesla drivers, but per https://danluu.com/car-safety/, maybe it's also about the cars. The most fatal manufacturers (Kia/Hyundai, Dodge, Tesla) that were rated all rank poorly there as well.
@againsthimself @lmk @HalvarFlake Yeah, I'd like to see some different questions. For a while, I thought, in terms of answer accuracy on my polls, Mastodon > Twitter > Threads, but then I realized it was mainly that, on pessimism, Mastodon > Twitter > Threads and I could get the opposite accuracy rating by asking questions where the correct answer is optimistic.
Here, if you ask, E[median lifetime income percentile] for student loan forgiveness beneficiaries, maybe you see the opposite result.
Would the world be better or worse if car horns didn't exist?
@lmk @HalvarFlake Not sure why I don't get the exact same thing, but I get a fairly analogous result.
BTW, I find graphs 4 and 5 of https://www.ipsos.com/en-us/link-between-media-consumption-and-public-opinion even funnier.
