strangecasts

> 1) Get any porn site that is going to obey this to instead list themselves in some open way (TXT record in DNS, blockchain, some other form of list)

Exists in the form of the "rating"/RTA <meta> tag - https://developers.google.com/search/docs/crawling-indexing/...

> Generative AI produces work that is voluminous and difficult to check. It presents such challenges to people who apply it that they, in practice, do not adequately validate the output.

As a matter of scope I could understand leaving the social understanding of "AI makes errors" separate from technical evaluations of models, but the thing that really horrified me is that the author apparently does not think past experience should be a concern in other fields:

> AI both frustrates the producer/consumer dichotomy and intermediates access to information processing, thus reducing professional power. In response, through shaming, professionals direct their ire at those they see as pretenders. Doctors have always derided home remedies, scientists have derided lay theories, sacerdotal colleges have derided folk mythologies and cosmogonies as heresy – the ability of individuals to “produce” their own healing, their own knowledge, their own salvation. [...]

If you don't permit that scientists often experiencing crank "lay theories" is a reason for initial skepticism, can you really explain this as anything other than anti-intellectualism?

That's a much better way of phrasing what I was trying to get at :)

I didn't mean to suggest otherwise! That's why I also linked the CryptoNets paper - to show that you're transforming the inference to happen inside an FHE context, not trying to learn encrypted data

There are FHE schemes which effectively allow putting together arbitrary logical circuits, so you can make larger algorithms FHE by turning them into FHE circuits -- Jeremy Kun's 2024 overview [1] has a good summary

[1] https://www.jeremykun.com/2024/05/04/fhe-overview/ - discussed previously: https://news.ycombinator.com/item?id=40262626

Good encryption schemes are designed so that ciphertexts are effectively indistinguishable from random data -- you should not be able to see any pattern in the encrypted text without knowledge of the key and the algorithm.

If your encryption scheme satisfies this, there are no patterns for the LLM to learn: if you only know the ciphertext but not the key, every continuation of the plaintext should be equally likely, so trying to learn the encryption scheme from examples is effectively trying to predict the next lottery numbers.

This is why FHE for ML schemes [1] don't try to make ML models work directly on encrypted data, but rather try to package ML models so they can run inside an FHE context.

[1] It's not for language models, but I like Microsoft's CryptoNets - https://www.microsoft.com/en-us/research/wp-content/uploads/... - as a more straightforward example of how FHE for ML looks in practice

It's designed for 2D but there are libraries which add 3D rendering, e.g. Menori (https://github.com/rozenmad/Menori) and G3D (https://github.com/groverburger/g3d)

Alternatively, the LÖVR framework - https://lovr.org/ - is heavily inspired by LÖVE, but is aimed towards making VR games (but can also be used as a regular 3D game engine)

Cool! Going with a 3D-printed bracket to hold individually wired buttons instead of a separate input board is neat - it does mean the GPIO pins constrain how many inputs you can have, but the nice thing about keeping the USB ports is you can connect another controller/keyboard if the game calls for them :)

Huh, I didn't know SSB let you do combined first/last name searches, I assumed it would only let you search the names individually

(Is Tromsø where they keep all the spare Mads Mikkelsens??)

It would be quite a feat for the local paper [1] to interview him without realizing his name was fake!

Mads and Mikkelsen are both common Norwegian names (SSB has >4000 Mads and Mikkelsens individually), and unless his parents were big Refn enthusiasts I don't think they would have been aware of the actor...

[1] https://www.nordlys.no/mads-sin-drommereise-til-usa-spolert-...

Was lucky enough to get my Fairphone 4 on sale, but I'd happily pay full price now - even though the Fairphones are pricey for the specs, unless you absolutely need 24 cores etc. I'd say they are worth it, knowing the company is at least trying to improve the parts supply chain, and knowing you stand a chance of fixing the devices yourself (luckily I've only had to replace the USB-C port, which was trivial)

About the only thing I'd ding Fairphone on is not communicating earlier that they were having trouble getting Android 14 out to the FP4s, but the security patches have been consistent.

(Okay I'm also dinging them on getting rid of the headphone jack, yes I know it's a lost cause... )

The model itself on HF: https://huggingface.co/stabilityai/stable-audio-open-small

As an aside, a few early 90s games did experiment with spheroid sprites to approximate 3D rendering, including the DOS game Ecstatica [1] and the (unfortunately named) SNES/Genesis game Ballz 3D [2]

[1] https://www.youtube.com/watch?v=nVNxnlgYOyk

[2] https://www.youtube.com/watch?v=JfhiGHM0AoE

Obviously I can only speculate since I neither have access to their dataset nor interest in paying for API access, but crawling and dataset cleaning have gotten much better since the GPT-2 days, especially after Microsoft's PHI models [1] demonstrated how much dataset construction matters for parameter efficiency and toxicity. Having some basic content filtering is a pretty established part of data cleanup -- e.g. the fastText toxicity classifiers in the Dolma pipeline [2] -- which obviously still leaves in bad data, but certainly won't leave in the entirety of /b/

If shoddy data collection was the problem, we should expect the model to do much worse on overall leaderboards like [3], which require models to answer questions without sudden detours into Holocaust denialism. A change to the system prompt is more consistent with this, and as an added benefit, only requires one person to be completely out of their gourd.

[1] https://www.microsoft.com/en-us/research/publication/textboo...

[2] https://arxiv.org/pdf/2402.00159

[2] https://livebench.ai/#/

In the case of this and the earlier "white genocide" replies, it is way more likely someone changed the system prompts than that someone tampered with the training data, considering the conspiracy theory was brought up unprompted in wildly unrelated situations [1]

[1] As one example, https://bsky.app/profile/jdcmedlock.bsky.social/post/3lp6eal...

> We are in an unpleasant situation. It's probably going to require an unpleasant solution. You seem to be in denial of this reality.

It is an incredible coincidence that every time I hear this, the unpleasant situation happens to be fixable by just tearing away a few more individual rights, which will surely be back once the unpleasant situation is gone and not replaced by five more unpleasant situations

The problem is that it is much easier to implement such a check in a way which lets the verification service link the site to the user, with no discernable difference to the end user

e: I get the same feeling as I do reading about key escrow schemes in the Clipper chip vein, where nobody claimed it was theoretically impossible to have a "spare key" only accessible by warrant, but the resulting complexity and new threat classes [1] just was not worth it

[1] https://academiccommons.columbia.edu/doi/10.7916/D8GM8F2W

That would be Microsoft Mathematics, the previous version - I think they've removed the download in favor of this, unfortunately :(

That rules, have you made a writeup of it anywhere?

They're not complex in the Big O sense, they are complex systems in the sense that they have to incorporate and reconcile organizational knowledge from very different domains - e.g. Ron Jeffries' "Why Is Payroll Hard" http://wiki.c2.com/?WhyIsPayrollHard - while working in a predictable fashion with implicit requirements on availability and consistency, which the average web app stack more often than not leaves as "someone else's problem"

Keras' code examples - https://keras.io/examples/ - tend to be structured divided into dataset-preprocessing-training, and tend to start with a somewhat underparameterized model to show off the method quickly, rather than attaining the best performance. Keras 3 is also set up to be backend-independent, so you should be able to borrow the examples' preprocessing regardless of whether you're using PyTorch, TensorFlow or Jax.

In particular, the monocular depth estimation example seems like a good start: https://keras.io/examples/vision/depth_estimation/

Apply the same reasoning with the start cutoff at 2010 then

Okay, but your objection was requiring more evidence for the less likely assertion constituted an "arbitrary bar", you cannot argue that and quote _an arbitrary estimate by another commenter_

> Depending how we count, there have been a dozen or so pandemics comparably novel or deadly to the 1977 flu in the last fifty years. Of those, at least one was research-origin. So are you really saying p ~ 1/12 is an "extraordinary claim"?

Yes, I think it is extraordinary to extrapolate one event in 1977 to a "once every fifty years" rate

Prior probability estimates are not part of science?

The argument is not "research accidents categorically do not happen" it was "are they the more common event and therefore the more probable explanation, absent anything else"

> It's so strange to demand so much more proof of lab origin than one does of non-lab origin.

It's not, zoonotic transmission has happened many times, lab leak is the extraordinary claim requiring extraordinary evidence

Like Moq last year (https://news.ycombinator.com/item?id=37051967) this strikes me as another poor solution to the legitimate problem of funding work on OSS libraries - in addition to the licensing issues pointed out here, introducing the possibility that major version upgrades can create licensing violations risks teaching people that they should never upgrade dependencies

This is not the experience of the musicians I know, and I am really sad for people who feel that way - developing the skill and your own workflow are a huge part of what makes music enjoyable!

Oh, yeah, this would fit very well inside the free tier - I haven't used R2 but I'll give it a look