throwaway219450

In a self-supervised setting you don't need ground truth data, as long as you have a proxy task that you can train on which forces the model to learn something useful. Like cutting out bits of images and telling the model to reconstruct the gaps (i.e. MAEs).

There are also techniques to watermark model weights to confirm what training data were used: https://arxiv.org/html/2512.17075 I'm not an expert in this domain, but it seems like there are lots of pieces that could be used to (a) train a model to infer whether some data were trained on and (b) train a model to understand which weights are relevant for a particular training task in a more explainable way than backprop/gradient descent.

Have you tried going to a library or a bookstore? Or even Amazon, since the return policy is excellent. Books are still a reliable resource for anything that hasn’t changed in the last 5 years (since there will be a pre-LLM text). Fortunately that includes loads of DIY tasks.

Don’t get me started on Google’s propensity to show a full page of unwanted shopping results now without any way to show websites unless you turn verbatim on. “Brand picks” what garbage.

Oh that’s a good catch, internal environmental bias there.

If you build anything with AI that is meant to be read by anyone other than AI, assume you need to review and rewrite all the copy. No need to be a marketing person, just read what the models produce and question if it’s understandable. The bar is frankly pretty low.

“Names” is a particularly obvious one. “Pick a rock and the land keeps it in color, the rest gone grey.”. And a big bugbear of mine are sentences like this “The map lights the places it’s mapped” - why is the map the actor in the sentence? Just put: “Places are highlighted on the map”.

I think you’re better off getting a good sleeping bag and sleeping in the house in relative comfort. You can camp down to very cold temps, and your house is both better insulated than a tent wall and is a better wind break, even if it was bad enough that you got well below freezing. If we’re at the stage of “sleeping in the car”, indoor camping is reasonable to me.

For me it's the (New) 3DS. It's about the peak for handheld gaming. It can run GB, GBC, NES, SNES very well, GBA can run natively and obviously DS/3DS. You can run pretty much any Zelda game up to the GameCube on it with the official ports of Ocarina/Majora. The emulation community also has some promising demos of Wind Waker and Twilight Princess, but who knows about that. I have a Switch and it's frustrating seeing games that Nintendo haven't released for it (like WW and Pokopia), but maybe after a decade I'll get a Switch 2 OLED whenever that comes out.

I hoped it was too, with the sort of glitchy portfolio image and the words being “midnight, raven, etc”

> and it was drawn to communicate messages powerfully both in the physical world and

.. in the ethereal?

> But I can’t help but feel like they’ve created a facsimile of a cathedral - they’ve created the palace at Disneyland, not Versailles.

You could at least compare to Notre Dame. It’s a bit tone deaf to draw comparison to Versailles, which was a vanity project to glorify the French monarchy.

Usually either the ink is running out and it's no longer continuously flowing to the jets; or the print head has dried ink stuck to it and needs cleaning.

Chocolate is a product made with cacao and other ingredients. But the best manufacturers declare the origin of their beans, which might be Ghana.

You included machine learning in with LLMs in your root post. Try training a supervised model for some task, like an image classifier or a detector. LLMs have not replaced convolutional models, and even transformers can be used for non-NLP tasks. The idea this is somehow not professional is not true. Despite what people would have you believe, ResNet50 fine-tuned on a good dataset is still competent. In many environments there isn’t enough compute for bigger models, and large language models aren’t always the best choice.

You also don’t have to go straight into LLMs. You can learn most of the training mechanics via classic deep learning. It isn’t much different with LLMs except the training data and losses are specific and the models are much larger.

It’s funny that Claude picks the basically the same tagline for both projects.

> Every mountain stands at its real height

> The Matterhorn is exactly as tall and as steep as the real one.

The models are RLd so hard to declare their own legitimacy. Everything has to be correct, genuine or real, no x, y or z.

I will say from a couple days: the only issues so far are that the image carousel in the Amazon app no longer works, but clicking on the (blank) photo does. I haven’t found issues like that with any other sites or apps. You don’t seem to be able to view PDFs in Safari, but can download them and view with Preview. Not bad, it’s what I do on OS X anyway as you otherwise have to rely on browser caching to avoid re-downloading - except it’s surprisingly hard to find the download list if you didn’t recently download a file in the current tab? And it insists on showing “lockdown enabled” everywhere, which is probably a good reminder for those who need it, but I wish I could hide it. Maybe you can pick these options à la carte?

You can use structure from motion for this. Walk around and capture some video. It’s the same way that drone photogrammetry works, ie bundle adjustment. It’ll give you the intrinsics at least.

But with a high res sensor, even flex in the PCB will invalidate the calibration. Only takes a few microns for a pixel to be in the wrong place.

SAD runs in better than real time on a typical desktop CPU, even on a Pi. The author glossed over basically all of the calibration involved with stereo. It’s not quite as simple as taking two cameras and running a matcher.

Because stereo disparity (parallax) scales inversely with distance, conveniently you don’t need to search huge ranges. A point at infinity should have zero disparity in a calibrated stereo system. So normally you pick a depth range and search the corresponding pixels. Also if the cameras are calibrated, you only need to search horizontally so it works out to say 64 operations per pixel, which can be vectorized if the CPU supports it. Other approaches use more stable comparisons like census (is each pixel in the match window brighter than the central one) and hamming distance which is also very efficient.

And yes, stereo is sensitive to repeating texture but absence of texture like a blank wall is usually worse.

From what I can see in Apple docs, aside from web limitations it sounds great. Is there any real reason to not enable this? Like I would be deeply suspicious of a PDF sent via messages and no auto-connect to insecure WiFi is also common sense.

> Over the last few decades I’ve read many technical docs that were obviously mechanically translated from some other language. This is much better.

I don’t think it’s better. Any frontier LLM is capable of competent translation without significantly re-wording the text.

I use Whisper to transcribe foreign language videos and then run them through a couple of models to translate into English. It works really well, and much better than mechanical translation because they can figure out likely mistakes based on how whisper might have misheard (or surrounding context from the script).

I was talking about a function with 100 lines and no line breaks. Very difficult to read.

I use black/ruff on a lot of projects and it will only touch standard-y things like docstring spacing, or spacing between methods and classes, import sorting etc. It doesn’t touch intentional/aesthetic user spacing. With agents it is helpful to make them run ruff and mypy to keep them honest, but as long as the code adheres to the spec, it doesn’t get changed. Astra also creates things like “report” functions with 10 arguments, so it’s useful to turn on complexity checkers as well.

I don’t see how Prettier is any better here unless you really want everything to be deterministic; how could it possibly know (always) when to insert a line? Aside from some more obvious cases like adding a gap before a loop declaration. Sometimes I group lines because they’re part of a mathematical formula or other operation and it makes sense to me, but Prettier doesn’t know that.

I removed it, but it's frustrating when the options are someone's (reasonable) post on X, or slop-spam aggregators. I don't really want to drive visits to either, but at least the X poster can get the views. I don't really like xcancel either as it still encourages indirect engagement with X.

Astra writes well (which is why I often like to work with it over Claude), but it tends to produce code golf that's borderline unreadable, like completely omitting line breaks. I thought I had something funky with my config, but it seems to be a known issue.

https://lucumr.pocoo.org/2026/9/7/astra-why/

[Sorry for X] https://x.com/tenobrus/status/2096636223414808719

The generated code certainly does what's asked of it, but it's as sloppy as anything else I've seen. I don't think this benchmark can control for that. Maybe if it had a metric for minimum diff to satisfy the new requirement with the existing constraints on dead code etc?

This is the way. I can write pseudo code (ie Python) and I leave a breadcrumb trail of comments where I want it to implement. Hell the code doesn’t even need to have valid syntax if you forgot some function signature, but that small amount of guiding can produce significantly better results than a blind implementation from the model.

I’ve also tried spending an hour or two first before handing over control. That way there’s something pre-existing it can work from.

I found the coffee shop one telling. Both suffer from being a stream of bland statements. The original version is actually reasonable if you just excise the Claude-isms. The rewrite has a whiff of ASD-STE100, and reads like the sort of sentences one would write for homework while learning a new language. There is also almost no difference between the de-slopped opinion piece and the blog post.

The article misses an important bit of context (imo) that low thermal expansion glasses like CERAN have been around for ages and anyone who’s used an induction or a glass top stove would be familiar. However you can now get completely hidden stoves that look like stone countertops. If you’re not in the market for a kitchen, or know anyone people who recently remodeled it’s reasonable to not know they exist. I don’t know anyone with one, and everyone in our town still has gas. The author doesn’t explain which one they were buying, but it seems like the latter?

Normal induction glass top stoves are a unit that drops into the countertop so of course you get the glass included (because at the very least, it has the decal for the cook zones and controller built in). The invisible ones only come with the burners so you can custom fit the top. Looks like you can get them off AliExpress even. It seems like this is the point of confusion - the author thought he could go down to a glazing shop to buy some tough glass..?

I suggest go and watch Huygens Optics series about CERAN:

https://youtu.be/qi8jmEbWsxU

This website is absolutely terrible in iOS Safari, super glitchy.

The Borges reference is neat. I’m a little lost in the sloppy documentation (“funes makes four deliberate, load-bearing choices”). As every project that Claude documents, figuring out what’s actually important is hard. I want to trust this one a little more than the average vibe code because it’s from HF.

My scepticism is that agent memory must be a non trivial problem to solve, or the labs would ship a solution in the harness. The models are happy to check against previous conversations if prompted, but they rarely do it on their own by default.

One advantage of plain text “diaries” is that you can read them as well, and the agent can add commit hashes, conversation IDs, links to experiment runs and so on. It works between models without any external tool.

Again, depends on the country. Switzerland (which I appreciate is not in the EU) does not require visas for many "friendly" 3rd countries even if the person is coming to work.

https://www.sem.admin.ch/sem/en/home/publiservice/weisungen-...

A nitpick here, but "visa" gets tossed around a lot. A visa is a document which allows you to request permission to enter a country for some purpose. To work in a country you need a work and/or residency permit. Even in the US, there are situations where you can work with a an expired visa, provided your other immigration documents are current. But if you leave, you need a new one. Similarly UK citizens can still work in some European countries without a visa, because there is no requirement for them to enter with one (but the officer at the border might still refuse entry).

This ruling is not targeted towards Europeans. I doubt this will significantly affect European workers in the US. The reality is there aren't that many European immigrants in the US and not many at all on H1-Bs. The vast majority of immigrants are from the Americas (with over 22% from Mexico), India and China. European migrants as of ~2023 total around 10% and have been slowly decreasing over the past few years. Bear in mind a lot of people are still here from the post-WW2 period when immigration was much more liberal.

Anecdotally from the people I've met, the majority of Europeans are here for academia on F/J/O, entertainment, are spouses of US citizens or get temporary transfer visas like the L1. Those who go to the US because of high salaries can afford to figure out employment based green cards and they don't have multiple year waiting periods. Otherwise it's a huge pain to get permanent residency or even a normal work visa compared to the EU, but qualified Europeans generally don't have much trouble.

I’m reasonably sure this person is either lying or at least misrepresenting their success. I did some vibe-coded searching through DBLP and there are very few people who have published in all of those conferences without a publication prior to ‘23 (an arbitrary cutoff to rule out big name researchers with labs), let alone with only 2 authors. ICLR only just closed entries, but I included 2025.

But people are clearly attempting to do this, and some are going to get successful. Putting your paper into ChatGPT prior to submission to get free feedback is really table stakes these days, and you can be sure reviewers are breaking the rules by using LLMs.

Digital provides a back up for me. I’ve had negs go missing in the mail, sometimes they get damaged or scratched. So I pay for the basic scan from the lab and do it again at home. Sometimes friends will ask for a digital copy, or I need a copy for the web. I also can’t print enormous enlargements at home, so sometimes I will do a scan and (sacrilege) pay a printer to make a copy.

I don’t see how any of this detracts from the joy and presence of shooting analog, nor is it mutually exclusive with home developing and printing. Film certainly isn’t getting any cheaper either.