Hacker Newsnew | past | comments | ask | show | jobs | submit | murkt's commentslogin

Where did you try it, how can you use it? I have nothing in my Workspace (I am admin), Gemini app, AI Studio. Only 3.6 Flash

Ohh, horizontal wheels. They’re about as good as I expected, models have pretty bad spatial awareness. I would expect Fable to be a bit better than old models, though.


It is dry heat, if you look at it from certain point of view. I once experienced smth like 27-28°C with fog, 100% humidity. Was almost as bad as 40°C with low humidity.

And if it would be 42°C/107°F with fog and 100% humidity, lots of people would be dead already. But it’s more like saying “at least this building is not on fire right now”. Or 1km wide meteor is not crashing down. It is not, and anything near 40°C is really unpleasant, regardless of burning homes and falling meteors.


42C and 100% would be a humidex of 81~ that would be every animal at that location instantly dying.


Or as we called it growing up in Missouri: “summer”.

I exaggerate. Slightly.


Exactly! Proper wet heat. /s


Yeah, along with UPPERCASE labels, which is a staple of AI designed pages


It should not be difficult to change the CSS. LLMs have learned from us.


It triggered for me on a completely pedestrian game design prompt a couple of days ago. I’ve sent feedback and continued with Opus, but that was really unexpected


I remember similar kind of visualization from a decade ago, called paperscape. Looked cool, worked on clustering using citations and references.

Never got any idea on any use case that would be covered by such visualizations, apart from looking cool.


ResearchRabbit is free and has this feature!

https://www.researchrabbit.ai/

ConnectedPapers also has this but they started to limit unless you pay:

https://www.connectedpapers.com/

A few other ones I know of:

https://litmaps.com

https://consensus.app/home/features/citation-graph/


I also made my own variant recently. I really liked the idea of litmaps but I didn't really like the UI/UX or "graph expand" feature (it uses some internal heuristic that is not very clear).

https://paper-explorer.aziis98.com/ (also on github https://github.com/aziis98/paper-explorer)

This uses OpenAlex as a source of articles and to let you explore the citation graph of papers. This is still a prototype mostly made to test out how far I could go with vibecoding (well I still checked the code now and then) something without a js framework. Someday I think I will add more features to it, but now there is already a somewhat working version of import/export so I'm fine with it.


Today you could plug this into your agent as a second brain to give it more insight on related topics and ground it's responses when doing research or building products that can benefit from cross polinating between fields.


That's usually the case with graph visualizations or clustering for networks, imo (beyond revealing obvious statistics(


I love them! It's a really nice, fun way to explore a corpus. Cosmograph for this sort of thing is great, it supports graphs as well as 2D projections, and is blazing fast.

That said, I've never had a client or stakeholder show any interest in using one, beyond an initial "that's cool".

And UMAP etc., is just as much an art as a science. You'll go mad trying to get the perfect layout.

Great toy if you're into that sort of thing, but yeah, fiddly and overwhelming for most.


Hi, I love the genre too. Cosmograph is wonderful, I did try it, but because of its license restriction I could not use it for this project. I do agree that beyond an initial "that's cool" this map may not contribute much, "and that's why I didn't make it the main product. I already had the data as I was building other things (extension, paper page) and wanted a bit of a cool factor so people would take a look at the project. The value is what's under each dot, the enriched page (TLDR, genes/drugs/diseases, trials, 3D structures, code, datasets, full text), extension and the MCP for agents.


Hello, I agree with you, viz are just cool and might not really have a usecase. In this project map is not the product, it is 1 of 4 parts and to be honest the least important. The value is what is under each dot, the enriched page (TLDR, genes/drugs/diseases, trials, protein structures, code, datasets, full text, images, reviews, etc) and the MCP for agents. You are welcome to use whichever part of the project is most useful to you (whether that is the map, paper pages, browser extension, or MCP).


This supports my feelings of what works best when working with the models. Especially together with the directional bias in recall that’s mentioned here in the comments.

First, the model attention is actually limited, so less rules is usually better, but that’s common knowledge already. Or maybe it’s as common as common sense, and a lot of people still employ lots of rules and try to cram everything in one step.

Second, it’s often quite sufficient to just namedrop a technique and LLM will work differently. For example, when debugging, LLMs tend to try to brute force the problem and often end up in the weeds. Just add “use scientific method for debugging and keep journal file” is usually sufficient to improve their skill here.

Another example is refactoring. Just add “use Mikado method”, and it’s sufficient to wholly change the approach and produce much better results.


Some problems are very hard to solve with stupid code. This can easily be the case (computational geometry)


I am having a hard time picturing what could be the problem that you were solving.

Redis, Cassandra, RabbitMQ and Clickhouse. RabbitMQ looks like a black sheep in this lineup.


Also like redis and clickhouse are diametrically different platforms. Wild to go from one to the other.


I’m gonna guess that it’s because Ruby is dog slow and not concurrent.


I find all current LLMs to have pretty poor spatial awareness. It is becoming better, but still very poor. How are you dealing with that? Got any special tricks, any advice?


I write about this in detail here: https://adam.new/blog/bitter-lesson-ai-cad

This is improving greatly in recent model releases


Opus 4.5-4.7 was pretty bad at it, 4.8 was a bit better, and I have not tried Fable much.

So basically you have a good enough code that’s “intuitive” for a model, screenshots, and that’s it?


fable is a fair bit better, but to an extent its that it tried more things to get an understanding of whats happening than opus does


Fable is considerably better from my experience: https://x.com/LLMJunky/status/2065229625702109340?s=20

Fingers crossed it comes back!


bro, with all respect... your post says:

"Before working at Adam I worked at an AI Lab called Adept. We trained foundation models to do actions on a computer.

What does computer use now? The best general models. They just got good at it."

You were working for 4 months in Adept. What could you deliver or even learn in such a short period of time?

Sounds like an excuse tbh


My favorite spatial reasoning benchmark: https://minebench.ai/

no tricks, I'd definitely be curious to know how much screenshots help


Guidelines | FAQ | Lists | API | Security | Legal | Apply to YC | Contact

Search: