bryantwolf

This piece is great. And it’s totally true that we’re surrounded by walls of slop. I stop reading things all the time when I notice the tinge of AI.

That said, I don’t think that just because this is the way things are, this is this is the way things have to be. Even with our interactions with AI. And I have hope.

Ted Chiang had a great piece in the New Yorker years ago called “Why AI isn’t going to make Art”[1] where he describes something similar. There are 10,000 choices to make in every good pieces of art. AI makes the average choice for every decision and the average choice is derivative and uninteresting. There in I think lies both the problem and the solution. Rather than autocompleting our work, we’d all be so lucky to have a captivated and motivated reader who asks all of the questions this author suggests, so we can get better and better at asking them ourselves and making great art.

[1] https://dmd-program.github.io/dmd-300-master/assets/Chiang-A...

If you think all of these models aren’t finding important bugs everywhere I think your not being honest with yourself.

In fact I think the opposite is true. The Zcash bug was found with opus 4.6 or something like that. Many worse models currently in the wild might be very capable but not yet industrialized for bug finding.

Another example from today. Cryptography system that’s been around for decades: https://eprint.iacr.org/2026/1630

It’s going to be a bumpy few months

Sure, I think it’s reasonable to tell Anthropic the barn door is already open.

Though, like, I guess I expect that when this comes out, all the opus traffics will move over. It does appear to be much more capable, just jury is out about how much more capable

All of this discourse seems very bizarre.

If smaller models can find these things, that doesn’t mean mythos is worse than we thought. It means all models are more capable.

Also if pointing models at files and giving them hints is all it takes to make them find all kinds of stuff, well, we can also spray and pray that pretty well with llms can’t we.

It just points to us finding a lot more stuff with only a little bit more sophistication.

Hopefully the growing pains are short and defense wins

The advantage is that the password never leave the device. It has a public key and signs challenges with the private key but nothing sensitive goes over the wire on every login

Shoot a way to contact you to tecoholic at bawolf.com and I'll invite you to a copy of the board you can duplicate and mess around with.

I appreciate the sentiment and I agree. While I think there are countless humans who could do way better, I was never going to hire someone to illustrate this. Furthermore, I don't think it reads very interestingly without the images. I doubt it would have even gotten published.

But now thousands of people have seen it, it's shown that it can strike a chord. Maybe it is worth polishing a little more. It would be adorable as a small book

I bought the starter membership which is 20k credits for $18 and I have 5.7k credits left. They charge 50 credits a generation for nano banana. There was some LLM usage to plan the images too so the math is a little blurry but that should give you a rough sense

I actually thought a lot about the eye whites! He didn’t look nervous enough without them

Thank you for saying so. I’m glad it connected with you

You’re entitled to your opinion, but I don’t think that’s what I wrote.

I too am skeptical we’ll really be able to catch everyone. By making public tools we just create evals to beat the tools etc.

Still, right now I think we can tell, so I focused on making sure they were my words, but I let an llm help edit and I think it honestly made it much more readable

What parts felt too stretched? Or just as a composition would you have preferred I narrow it down? I thought it was fun that it goes a little overboard. To me it felt like doing so captured just how much these journeys can change over time.

I tried to focus on 3 main bits:

Early exploration, problems between people, and then how much is ultimately in or out of your control

I’d love some thoughts here on the consistency. I thougt I actually did pretty well! But I can see how I might be blind to mind own work here.

What ruined it for you?

Hey hacker news! I wrote this and I’m glad it connected with folks.

To answer a few of your comments:

Writing is all mine but I had Claude proofread it, in addition to some close friends. Honestly it pointed out some great weaknesses in the original draft.

The art is all nano-banana through a tool called flora ai. I’d love to work with a human illustrator for something like this. I can draw, but I can’t paint and there’s an aesthetic here I think it handles better than I would have.

Man, it’s amazing that I can get something out there that expresses a vision all by myself. If this were a revenue generating project like an actual children’s book or something I’d love to work with someone that could bring it to life a bit more.

You might find this fun. There this sonic the hedgehog special stage from sonic three that takes place on a sphere. Recently I made a remake of it and they project a 2d grid onto a sphere, but the projection adjusts as you move, so no matter where you go, you never end up at a pole. The poles always stay on your sides.

The other interesting bit is that you can have an arbitrary map size and just repeat it. The game is 32x32 but it could be whatever.

https://blue-sphere.fly.dev/play?map=s3-01

Anyway, great stuff you have here!

I prefer ‘clankerism’

People will like them when they’re good content. Right now we’re stuck in the in between where it’s kind of all or nothing ai, but it will get grayer when the feedback loops are tighter and building ai movies is more interactive. Same thing with any special effects really

I'm happy you enjoyed it!

The article talks about what I did and how. If you want to try it out you can send a video through the model at https://breakingvision.bawolf.com/

If you need a video to try it out, maybe grab a clip from one of these: https://www.youtube.com/results?search_query=halo+windmill+s...

Hey, this looks great! I'm a huge fan of vectors in Postgres or wherever your data lives, and this seems like a great abstraction.

When I write a sql query that includes a vector search and some piece of logic, like: ``` select name from users where age > 21 order by <vector_similarity(users.bio, "I like long walks on the beach")> limit 10; ``` Does it filter by age first or second? I've liked the DX of pg_vector, but they do vector search, followed by filtering. It seems like that slows down what should be the superpower of a setup like this.

Here's a bit more of a complicated example of what I'm talking about: https://blog.bawolf.com/p/embeddings-are-a-good-starting-poi...

Elixir has had typespecs for a long time which can handle static type checking which is a lot like using typescript with vite. The compiler doesn’t type check, it relies on your IDE to do the dev time analysis.

https://hexdocs.pm/elixir/1.12/typespecs.html

They’ve recently started gradual typing which will feel a lot like typescript in the early days where you can type some things but you’ll frequently need the escape hatch ‘dynamic’ which is their ‘any’ to make it compatible with existing code

https://elixir-lang.org/blog/2023/09/20/strong-arrows-gradua...

I've been using Destucture [1] for a few years. It's an interesting compromise that works for atom and string keys because it's explicit, kind of like the sigil approach. Though this seems way more streamlined, especially as someone who uses JS for front-end apps.

[1]https://github.com/danielberkompas/destructure/

Hybrid searches are great, though I'm not sure they would help here. Neither 'crown' nor 'ruler' would come back from a text search for 'king,' right?

I bet if we put a better description into the embedding for 'ruler,' we'd avoid this. Something like "a straight strip or cylinder of plastic, wood, metal, or other rigid material, typically marked at regular intervals, to draw straight lines or measure distances." (stolen from a Google search). We might be able to ask a language model to look at the icon and give a good description we can put into the embedding.

You'd have to update the embedding every time the data used to generate it changes. For example, if you had an embedding for user profiles and they updated their bio, you would want to make a new embedding.

I don't expect to have to change the embeddings for each icon all that often, so storing them seemed like a good choice. However, you probably don't need to cache the embedding for each search query since there will be long-tail ones that don't change that much.

The reason to use pgvector over blobs is if you want to use the distance functions in your queries.

I love it!

Yeah, these can be cute, but they're not ideal. I think the user feedback mechanism could help naturally align this over time, but it would also be gameable. It's all interesting stuff

As an individual, I love the idea of pushing to simplify even further to understand these core concepts. For the ecosystem, I like that vector stores make these features accessible to environments outside of Python.

This is a good call out. OpenAI embeddings were simple to stand up, pretty good, cheap at this scale, and accessible to everyone. I think that makes them a good starting point for many people. That said, they're closed-source, and there are open-source embeddings you can run on your infrastructure to reduce external dependencies.