I’m a little late to this but Salsa exposes queries and falls under the umbrella of Query Based Compiler Architecture. I think Kotlin’s K2 and C#'s Roslyn also expose queries over compiler internals and the AST.
ChaseGilliam
sodapopcan
As per @derek-zhou, I think there are always going to be people who hand-code and who are interested in programming languages themselves. There are lots of people who experience the act of writing code as a creative outlet… as a side note, I’ve recently learned that there are a lot of, let’s call them “techno-purists” out there who seem offended by this notion, but what can ya do.
If somehow all the maintainers of current popular languages and frameworks get DHH’d, then hobbyists will just move to other languages, fork, or create their own. I don’t really see JavaScript going anywhere in the next few decades, though, unless somehow it gets decided that backward compatibility no longer matters on the web. Crazier things have happened, I guess.
This is a good point I never thought of.
1
caleb-bb
Frontier models are only a few months ahead of open models, you know.
The local models won’t be as good but they’ll serve. I strongly suspect there will be jailbroken local LLMs that will do whatever anyone wants.
And there’s nothing stopping you from spinning up a fairly heavy-duty jailbroken LLM in an AWS box or something and having it do whatever you want. Might be cheaper than the mainstream services as all this new compute gets built out.
1
plcholder
I think China could probably close that gap quite easily if it had the latest hardware, Hence the reason Chinese labs have had to be so creative with model architecture and training efficiency. IMO, K3 was one of the models that came closest to the US sota at the time of its release. The problem is that it’s still pretty expensive to run at scale, $3 input / $15 output per million tokens.
This is anecdotal, but I also feel like we haven’t seen a gpt 4 to gpt-4.5 level jump in general model intelligence for a while. It feels like a lot of the progress has been increasingly focused on things like better tool use, longer reasoning, agentic workflows, and getting models to catch and correct their own mistakes. That’s still huge progress, but I’m not sure we’re seeing the same kind of obvious qualitative jump in general intelligence that we saw in earlier generations. I think local models will catch up, hence the IPO panic and export ban appeals from US labs
1
tozz
Once market saturation hits for real and dependencies are entrenched, prices will skyrocket. I would not be surprised if we in five years have 10-100x prices, it’s classic marketing really. Chinese models are cheaper, sure, but why would they not want to make more money? Nobody is doing this from the goodness of their heart, it’s business.
0
caleb-bb
mudasobwa
Creator of Cure
Well, ragex is using dllb which treats AST as a first-class citizen.
That said, ragex literally queries AST directly. The random example from docs:
-- Find functions similar to this one, called by modules importing 'tokio'
SELECT id, name, file_path, vector::distance::knn() AS similarity
FROM ast_node
WHERE kind = 'function'
AND source_embedding <|20,100|> $my_fn_embedding
AND <-contains<-module<-imports<-module[WHERE name CONTAINS 'tokio']
ORDER BY similarity
LIMIT 10;
1