wongarsu

12 ms·

331.
by wongarsu 4mo ago
English around 1800-1900 reads almost like modern English. At least if the author didn't write in an overly pretentious style, which was more common back then
332.
by wongarsu 4mo ago
Seemingly any decent way to tune in a steady-state supplementation involves multiple iterations of the "get blood test, adjust dosage, repeat" loop On the other hand, when our body synthesizes Vitamin D itself it seems to regulate
333.
by wongarsu 4mo ago
The bugs I can trivially find, the bugs Claude can trivially find and the bugs Codex can trivially find are not necessarily the same. The most obvious bugs would be obvious to any of us three, but beyond that we would't agree on a defi
334.
by wongarsu 4mo ago
The manifesto is somehow devoid of coherent arguments. You are still personally listed as the author of your commits, adding a co-author is not a mechanism for moving all accountability to that co-author. That's not what a co-author is
335.
by wongarsu 4mo ago
In the EU, that list would be regulated under GDPR. Just as any other list with people on it. You would be allowed to have it, but storing it insecurely or giving people who don't need it access to it would be a violation
336.
by wongarsu 4mo ago
not if you put it in a google doc so all your bouncers have easy access to it and you can update it anytime
337.
by wongarsu 4mo ago
Never had fraud on my credit or debit cards, and with 3D Secure it's difficult to pull off (basically 2FA for all online credit card purchases). But I did have someone fraudulently making direct debit transfers from my bank account. My
338.
by wongarsu 4mo ago
Which is just a flashy way to say "we have low margins and lots of overhead". Considering how much they spend on sales, marketing and R&D that doesn't sound that absurd
339.
by wongarsu 4mo ago
An interesting procedural detail is how an admin decided to just close the discussion and ban the user before the mandatory discussion period was over, and got a lot of pushback for the sloppy decision making process. This was overturned, o
340.
by wongarsu 4mo ago
Ok, saying things like "The left marched through this institution. There’s no reason we can’t march right back" is pretty bad. To be clear: it would be equally bad if you swapped left and right in that sentence. I don't know
341.
by wongarsu 4mo ago
So to summarize: he resuggested "WikiProject Intellectual Diversity", a group with the goal to make "Wikipedia more intellectually diverse" and "ensure fair and open decision-making and governance, broaden the range
342.
by wongarsu 4mo ago
Exactly my experience. If you try to OCR hand-filled forms with a fixed structure, traditional OCR models are great. Vision-llms can improve a bit on character recognition, but at the cost of harder to detect failure modes. But if you are t
343.
by wongarsu 4mo ago
Most of the world follows the Vienna Convention on Road Signs and Signals, where all the important road signs are understandable without reading. This is how no entry signs look around the world [1] Especially important in places like Europ
344.
by wongarsu 4mo ago
It's DHL. If you want your package or pallet shipped to some island it might take a long while until a big container ship makes the next stop there. If it can stop there at all. For a package/freight company, the capability to run
345.
by wongarsu 4mo ago
economies of scale are enough to explain the entire price difference. Running 8 concurrent requests at 100 token/s on $100k hardware is a lot cheaper than running one concurrent request at 20 token/s on $20k hardware
346.
by wongarsu 4mo ago
If you invest the minimum to run the model, obviously that's more expensive per-token than investing the optimum to get the best price/performance tradeoff (which for GLM 5.2 is at least five times that figure) If you can bring th
347.
by wongarsu 4mo ago
Anything below one billion parameters you can run on the CPU at acceptable speed For larger sizes you still can , it just becomes slower and slower. For a simple classification task (small input, tiny output, and you can constrain output t
348.
by wongarsu 4mo ago
JSON-LD is 12 years old. Just four years after Facebook introduced Open Graph to make their links prettier. Maybe an appeal to implement it today is 25 years too late. But there were plenty of appeals 10 years ago, or to implement open grap
349.
by wongarsu 4mo ago
Tbf, most of the "real benchmarks" have issues that are just as bad. Assessing LLM performance is just hard
350.
by wongarsu 4mo ago
Because nearly all benchmarks measure "accuracy" by giving you a point for a correct answer, and 0 points for everything else. If you have 100 questions you are 10% certain on, answering "I don't know" to all of tho
351.
by wongarsu 4mo ago
I learned that exact style of writing in a marketing workshop, pre-AI. It's effective, satisfying, and a random third thing I can't be bothered to come up with right now. As a proportion of all easily crawled text on the internet,
352.
by wongarsu 4mo ago
English has this weird dichotomy where most of the words in a typical sentence are Germanic, while most of the words in the dictionary are French. Fun fact: according to a quick count by AI using web search, the previous sentence contains 2
353.
by wongarsu 4mo ago
I suspect quickly slapped together 3d renders photoshopped into actual landscape images. With very limited attention to detail when it comes to matching perspective or lighting between render and photo, or when it comes to blending them tog
354.
by wongarsu 4mo ago
Are you even legally signing anything if they can't show you the document you are signing? I am not familiar with the nitty gritty of US law, but under German law that signature would be worthless. Even signing a document you have but
355.
by wongarsu 4mo ago
Any statement about "Europe" as a whole will always be an oversimplification. But in Germany all of this rings true. We have lots of on-street parking, business are frequently required to provide parking spots (depends on the muni
356.
by wongarsu 4mo ago
My recollection of HN 15 years ago includes a lot of annoyance with apps that could have been a website and how these walled gardens harm our freedom
357.
by wongarsu 4mo ago
Git does great with text files, but game development contains a lot of binary assets (textures, videos, 3d models) and correspondingly huge repos. git-lfs tries to patch around that, but that makes a complex tool that creatives struggle to
358.
by wongarsu 4mo ago
Yes. Most benchmarks just measure how many answers are correct. The best way to optimize that is to confidently state something, in hopes it's correct. Which is exactly how most LLMs behave, despite plenty of evidence that they do know
359.
by wongarsu 4mo ago
Imagine a forum where comment ids are client-generated UUIDs, and comments are inserted with "ON CONFLICT IGNORE". Submitting the same comment twice would simply be a noop But what the Query method really targets are things like a
360.
by wongarsu 4mo ago
For an 8-bit quant (what people call "near lossless") you are looking at something like 4xMI350X, which comes out to about $150k after adding the rest of the server. More if you go with Nvidia instead of AMD. More if you want more

More ›