The part where the LLMs (actually, VLMs) outperformed the CNN was that the LLMs were better at avoiding false negatives (see the quote below). That kind of Type 2 error is the most dangerous one since it persuades people not to get advice. Since the LLMs are designed to be general-purpose and public-facing, the headline could be flipped to: "A specialised 10-year old CNN is a worse public facing tool than a general frontier LLM". Which would surprise nobody.
| GPT |Claude | CNN |
Unsafe lesions kept (of 100) | 97 | 96 | 77 at its preset [sic] threshold; 98 or 99 when matched |
If I had an opinion on an issue but no dog in the fight; and if I were to explain my opinion to politicians in the hope they might learn from or even share it -- I may well be "lobbying", but I don't believe I'd be corrupt.
Well the problem is that access to politicians (who are supposed to represent you) is almost entirely through lobbyism. If you want anything, you can't afford it - and you definitely can't afford to go against existing lobbyists. It becomes corruption when only certain people have access to a democratic mechanism. It doesn't help that in the US, corporations are people and cash is their speech and you can't do or say anything about it because that would be treasonous.
That’s not what lobbying is, though. If have “no dog in the fight”, you are unlikely to try to influence a politician any which way. Lobbyists appear because someone with money is willing to pay for the lobbyist to apply their talents, and why would they if they had no “dog in the fight”?
I think the last sentence of your comment is the most exciting thing about this whole article!
There are at least three enormous gulfs here:
First, just understanding music and even jazz as a prerequisite to being able to do this.
Second, being able to do this using machines instead of a piano itself, which is unaffordable for a lot of people.
Third, the capturing of the je ne sais quoi of musical style, which is almost impossible to articulate.
So for me bringing this all together, it allows you to capture an improved understanding of different styles without needing to invest a huge amount of time because you can feed the same input into multiple style models and see how it diverges. That's a control which can't usually be obtained!
As a technical feat, what AI research allows across domains is a remarkable thing.
> First, just understanding music and even jazz as a prerequisite to being able to do this.
Do you have to understand music to train an ai to do this? You can be a fan without understanding music. You can be motivated and have an ear for what Cedar Walton sounds like without knowing anything about what harmonic and rhythmic vocabulary he had or didn’t have.
> Second, being able to do this using machines instead of a piano itself, which is unaffordable for a lot of people.
What? You can get a piano or a keyboard for a couple hundred dollars.
Unless you mean actually training and being good at the piano being unaffordable, no, that's just lack of skills and/or will to train to become good at it. Anyone can afford to buy an instrument and become good at it if they cared.
I'm sure no actual musician is actually impressed by this slop, and as the GP comment says the actual book written by a human to whose experience I can relate to and be inspired by is really impressive.
The cut-and-stick nets aren't but the rest most certainly are! Look up "The Complete Book of Origami Polyhedra". I've made some of them myself: they use only square paper without tearing, cutting or sticking.
> ... particularly concerned about is social media. Most images and videos uploaded to most social media is re-compressed by the target platform. This is a door to tracking that is far too easy for social media platforms to open.
This has been going on for a while with Facebook. They seem to embed custom metadata tags so that images shared outside the platform can be traced back:
reply