- yfontanaIf I were to do this (and I might give it a try, this is quite an interesting case), I would try to run a detection model on the image, to find bounding boxes for the planets and their associated text. Even a small model running on CPU should be able to do this relatively quickly.
- On the professional side, they also often let you interact with their experts and architects directly, as part of your support contract. With most other companies, you either have to go through front-office support exclusively, or pay extra for Professional Services.
- > I’m downplaying because I have honestly been burned by these tools when I’ve put trust in their ability to understand anything, provide a novel suggestion or even solve some basic bugs without causing other issues.?
I've had that experience plenty of times with actual people... LLMs don't "think" like people do, that much is pretty obvious. But I'm not at all sure whether what they do can be called "thinking" or not.
- > In the examples given, it’s much faster, but is that mostly due to the missing indexes? I’d have thought that an optimal approach in the colour example would be to look at the product.color_id index, get the counts directly from there and you’re pretty much done.
So I tried to test this (my intuition being that indexes wouldn't change much, at best you could just do an index scan instead of a seq scan), and I couldn't understand the plans I was getting, until I realized that the query in the blog post has a small error:
> AND c1.category_id = c1.category_id
should really be
> AND p.category_id = c1.category_id
otherwise we're doing a cross-product on the category. Probably doesn't really change much, but still a bit of an oopsie. Anyway, even with the right join condition an index only reduces execution time by about 20% in my tests, through an index scan.
- Interestingly, "aggregate first, join later" has been the standard way of joining fact tables in BI tools for a long time. Since fact tables are typically big and also share common dimensions, multi-fact joins for drill-across are best done by first aggregating on those common dimensions, then joining on them.
Makes you wonder how many cases there are out there of optimizations that feel almost second nature in one domain, but have never been applied to other domains because no one thought of it.
- > - It says it's done when its code does not even work, sometimes when it does not even compile.
> - When asked to fix a bug, it confidently declares victory without actually having fixed the bug.
You need to give it ways to validate its work. A junior dev will also give you code that doesn't compile or should have fixed a bug but doesn't if they don't actually compile the code and test that the bug is truly fixed.
- I think I was shadow-banned because my very first comment on the site was slightly snarky, and have now been unbanned.
- Properly measuring "GPU load" is something I've been wondering about, as an architect who's had to deploy ML/DL models but is still relatively new at it. With CPU workloads you can generally tell from %CPU, %Mem and IOs how much load your system is under. But with GPU I'm not sure how you can tell, other than by just measuring your model execution times. I find it makes it hard to get an idea whether upgrading to a stronger GPU would help and by how much. Are there established ways of doing this?
- Open source models like Flux Kontext or Qwen image edit wouldn't refuse, but you need to either have a sufficiently strong GPU or get one in the cloud (not difficult nor expensive with services like runpod), then set up your own processing pipeline (again, not too difficult if you use ComfyUI). Results won't be SOTA, but they shouldn't be too far off.
- I pay for chatgpt because, in my experience, o3 and o4 are currently the best at combining reasoning with information retrieval from web searches. They're the best models I've tried at emulating the way I search for information (evaluating source quality, combining and contrasting information from several sources, refining searches, etc.), and using the results as part of a reasoning process. It's not necessarily significant for coding, but it is for designing.
- From the article:
> Besides protein folding, the canonical example of a scientific breakthrough from AI, a few examples of scientific progress from AI include:1
> Weather forecasting, where AI forecasts have had up to 20% higher accuracy (though still lower resolution) compared to traditional physics-based forecasts.
> Drug discovery, where preliminary data suggests that AI-discovered drugs have been more successful in Phase I (but not Phase II) clinical trials. If the trend holds, this would imply a nearly twofold increase in end-to-end drug approval rates.
- > It’s insane to me that maybe every bank I use requires SMS 2FA, but random services I use support apps.
It never ceases to surprise me how much American banks always seem to lag behind with regards to payment tech. My (european) bank started sending hardware TOTP tokens to whoever requested one like a decade ago. They've since switched to phone app MFA.
- I've been working on extracting text from some 20 million PDFs, with just about every type of layout you can imagine. We're using a similar approach (segmentation / OCR), but with PyMuPDF.
The full extract is projected to run for several days on a GPU cluster, at a cost of like 20-30k (can't remember the exact number but it's in that ballpark). When you can afford this kind of compute, text extraction from PDFs isn't quite a fully solved problem, but we're most of the way there.
What the article in the OP tries to do is, as far as I understand, somewhat different. It's trying to use much simpler heuristics to get acceptable results cheaper and faster, and this is definitely an open issue.
- We don't know for sure that the universe is a closed system.
- Primordial black holes are black holes that formed right after the big bang. Basically areas where gravity caused the extremely dense matter of the universe's first instants to collapse into black holes before expansion could pull it apart. Their existence has been hypothesized but not confirmed (or definitively rejected) so far.
- It's absolutely not crazy, and I don't know how a car could only cost you 1k/year. This is just an example but has the average cost of motoring in Ireland in 2019 at almost 11k€/year : https://www.theaa.ie/motoring-advice/cost-of-motoring/.
One of the issues with trains is that people often severely underestimate the true total cost of car trips.
- A while ago I worked on a system handling call records for a large telco. Call records were considered sensitive information at that company, and distributed only where definitely needed. I'm sure security wasn't bulletproof, but there were regular audits to check that employees and contractors didn't store records in places they weren't supposed to.
One of the main functions of the system that I worked on was to create various anonymous and/or aggregated versions of the data, which could be distributed and used more widely (for stuff like fraud detection, network provisioning, marketing...).
- This is inaccurate. Reserve requirements are not the only cost that banks have to pay for the money they create. If bank A lends $20B, and the borrower spends it to buy something from someone with an account at bank B, bank B isn't going to accept $20B of "made-up" bank A money, it'll want $20B of central bank money from bank A. If bank A doesn't have that that $20B of central bank money, it'll have to borrow it from the central bank, at the cost of the reserve interest rate. Being able to transfer central bank money to bank B is essentially what backs the money that bank A created.
(in reality some of that $20B would probably make it back to bank A as money circulates, so the actual amount of reserve money that banks have to borrow depends on more complicated factors. This also doesn't talk about what happens if the borrower defaults. But the fundamental principle remains the same, and is what ultimately limits how much money a bank can lend).
- I think the hardest bug I've had to work on was just plain irreproducible. Once we exhausted all other ideas, we just attributed it to some sort of bit flip. Not a very satisfying resolution, but kinda cool to have encountered such an issue at least once.
- I'm working on a projet that uses PaddleOCR to get bounding boxes. It's far from perfect, but it's open source and good enough for our requirements. And it can mostly handle a 150 MB single-page PDF (don't ask) without completely keeling over.