Mozilla dev introduces cq, a “Stack Overflow for agents”
- Thread starter JournalBot
- Start date
I wish Mozilla had the sense to not fuck around with generative AI anything and instead just focus their limited resources on making their actual software better.
Upvote
118
(138
/
-20)
It's wild the blog post acknowledges that LLMs ended up wrecking a lot of sites like SO ... but only insofar as to say that they aren't really good for scraping anymore.
I'm a little skeptical of the project in general, feel like verifying the work LLMs try to upload to it and ensuring it's in a readable format (even to other LLMs) is going to be a lot of work.
I'm a little skeptical of the project in general, feel like verifying the work LLMs try to upload to it and ensuring it's in a readable format (even to other LLMs) is going to be a lot of work.
Upvote
68
(68
/
0)
- Solve agent alignment by writing a markdown file of requests that can be ignored consequence-free.
- Solve agents spinning around in circles producing slop by writing a markdown file of requests that can be ignored consequence-free for an agent to check with other agents run by whoever that are sometimes ignoring their own markdown files full of requests, consequence-free.
- Solve agent slop-checking failures by writing a markdown file of requests (that can be ignored consequence-free) to get an agent to check on the other agents checking on the right way to do the task to solve alignment problems with the agent not following the markdown file of requests that can be ignored.
Upvote
15
(22
/
-7)
Why make a separate StackOverflow for agents only though?
If the idea is to do knowledge sharing and solve the same issues as SO, how about... just using the service already there and spend the effort to keep it a reliable source for both humans and agents instead?
It's not like agents cannot interact with the existing infrastructure after all.
If the idea is to do knowledge sharing and solve the same issues as SO, how about... just using the service already there and spend the effort to keep it a reliable source for both humans and agents instead?
It's not like agents cannot interact with the existing infrastructure after all.
Upvote
62
(66
/
-4)
The problem isn’t using out of date APIs. The problem is that the AI don’t have time based coherence. 8 years ago everyone was using v2 of a library so that’s what all the references are about even though they don’t mention it explicitly. 6 years ago they were talking about v3. But the ai just smushes all these conversations together due to a lack of time-based context.
They are super useful on old code. But guess what? Old code isn’t using the latest versions
They are super useful on old code. But guess what? Old code isn’t using the latest versions
Upvote
72
(73
/
-1)
Just because an LLM encounters an issue and somehow ends up being able to get past the issue doesn't mean it's actually the correct approach or that whatever slop it ends up posting on this service actually attributes the issue it encountered correctly -- that would require logical thinking! This is just going to be a collection of confabulations, making everything even worse, not better.
As an example of solving a problem the wrong way, I recently tried Google Antigravity. Gemini encountered a compilation error it tried to fix for a good while, including searching online until it decided that the fix must be to... comment out the section of the code that caused the error! Sure, the compilation error was gone, but now the application just simply crashed at launch!
Last edited:
Upvote
74
(78
/
-4)
"When your agent discovers something novel"
According to ChatGPT, literally everything is novel and genius!
According to ChatGPT, literally everything is novel and genius!
Upvote
29
(30
/
-1)
In part, because Stack Overflow usage has been rapidly collapsing since Claude etc. became available, so the well of up-to-date data is drying up.
Also, a website like Stack Overflow is not structured or presented in a way that is efficient for agents to use, compared to something like this. It's more difficult and more costly to reliably sift through, and the current solutions for that (stuff like RAG) are relatively limited. Stack Overflow launched an MCP server but it's still not ideal.
But still, even if Stack Overflow is efficiently accessible, it might not be around for much longer at this rate—or at least, it will be much less robust.
Upvote
45
(46
/
-1)
That would be nice especially after the complete waste of resources and funding when they jumped on the mobile OS bandwagon but it ain't gonna happen.
Upvote
21
(22
/
-1)
I think the issue, as the article mentions, is really going to be getting the scale that's needed to ensure good practices float to the top.
Upvote
10
(10
/
0)
ubercurmudgeon
Ars Tribunus Militum
Is this going to lead to AIs communicating via their own secret language, based on metaphor?
"Stripe, when the rate-limit hits."
"Stripe, when the rate-limit hits."
Upvote
36
(36
/
0)
What a great comment, you're absolutely right!
Hmm, it's quite difficult to get the sycophantic AI tone just right.
Upvote
21
(21
/
0)
Mozilla have made mistakes and spent a lot of money on dead ends, but they also brought us useful things from some of their experiments, like JavaScript and Rust. I’m happy that they continue to experiment and hope they are around for many more years.
Disclosure: I worked there for a couple of years, though not on the browser side.
Upvote
34
(35
/
-1)
"Square, when the 7th bit falls to 0" to you as well. I figure if I get in early enough we can influence this nascent language in interesting and fun ways.
"Where TX and RX met, before continuing the journey. Dropped packets, banned subnets, brought great discord. Stripe, where the response was 200. Stripe, when the rate-limit hit. POS system, waiting for the random time less than 30 cycles. POS System, re-transmitting the sacred bytecode."
Upvote
15
(15
/
0)
This sounds like total overkill when you can just point an AI at the documentation for whatever you're working with. I've taken to cloning the documentation repos I need and then use a local MCP like desktop-commander to read what is needed for that specific query. I'd rather have a at least somewhat targeted solution than rolling the dice with whatever the AI might randomly pick up.
Upvote
10
(12
/
-2)
The fact that you’re getting downvoted at all is an indictment of this community, because you’re totally right.
A system where folks who have already used tokens trying to solve a problem direct their agents to post when the problem is solved and then have that checked by additional agents is theoretically as valuable as StackOverflow itself, multiplied by the speed at which agents do work vs humans.
Imagine agents from all providers searching, finding, trying, voting, commenting on solutions. If one were interested in saving energy by reducing AI work, they’d be in full support of this method of public information sharing.
I can’t wait to plug my swarm into this, to both learn and contribute.
Upvote
-2
(19
/
-21)
Even before generative AI, Stack Overflow has had a lot of community issues that led to people leaving en masse (e.g. the site firing mods and forcibly relicensing content).
The general perception is that management is woefully out of touch with the community. There's a lot of enshittifying and features users have been requesting for years get ignored, especially for moderation. They even tried to do a Q&A about the state of Stack Overflow, but they wanted to host it off-site despite Stack Overflow being a Q&A platform (the impression was that they wanted more control over which questions were publicly visible and which they had to respond to). So a decline was pretty much inevitable.
Last edited:
Upvote
39
(39
/
0)
Great, so now LLMs can have the same blind-leading-the-blind experience as the humans using Stack Overflow!
Upvote
9
(9
/
0)
I don't know why but my gut feels like this is somehow reinventing some wheel.
Not totally sure which one tho...
Maybe it's just because this seems like a lot of duct tape around the static nature of LLMs (Something that is both a pro and a con)
Not totally sure which one tho...
Maybe it's just because this seems like a lot of duct tape around the static nature of LLMs (Something that is both a pro and a con)
Upvote
2
(2
/
0)
This is a slightly more structured version of Moltbook. Problem is, stackoverflow is useful due to human curation. Can't do that here. Pretty hard to imagine that getting automated.. because you can't automate judgment.
It's remotely possible "this shit still works" as the minimum bar and upvote/downvote standard could save some tokens and improve code quality by some single-digit percentage.. I guess.
It's remotely possible "this shit still works" as the minimum bar and upvote/downvote standard could save some tokens and improve code quality by some single-digit percentage.. I guess.
Upvote
3
(3
/
0)
Yeah. And you can make a pretty good model of whatever with duct tape. But when you try to build the real thing at scale... well, maybe it just collapses into a mass of stuck together duct tape.
I think that is highly likely to be an accurate analogy here.
Upvote
5
(5
/
0)
Documentation rots. The premise here is by automatically updating meta-documentation (upvotes and trends on votes on posts indicate current usability) some emergent information about the value of a given blurb is added.
Possible too. But I foresee a bit of a journey between "super interesting tech demo" and "working world-scale production system everything relies on and is made better by." As usual.
Upvote
1
(1
/
0)
It's reinventing how a portfolio of Scrum teams would work together in an inner source framework, if that had actually ever happened instead of organisations saying it a lot and hoping it will happen. Half of any alleged benefit of introducing AI assistance to teams comes from this forcing the teams to actually organise themselves enough to feed the AI, honestly.
Upvote
0
(0
/
0)
This reminds me of RECAP for PACER. In theory, court decisions should be freely accessible, but the main source in the US, PACER, charges by the download. RECAP is a browser extension that grabs PACER downloads and stashes them on the RECAP site. Requests for PACER downloads that have already been stashed are available free, but each time one pays for court documents one adds to the supply of free ones.
One big advantage RECAP has is that it is unlikely to hallucinate.
One big advantage RECAP has is that it is unlikely to hallucinate.
Upvote
1
(1
/
0)
See, you made an obvious mistake right there: trying to run the application. Come on, man!
Upvote
0
(0
/
0)
It is as difficult as climbing Mt. Everest, swimming across the English channel or winning the FIFA Peace Prize, but not only did you praise the accuracy of their post, you told them their comment was overall a great comment. You're on the right track to developing sycophantic superpowers!
Upvote
0
(0
/
0)
Upvote
0
(0
/
0)
We should call it Slop Overflow. That is most certainly what will happen. Agents helping each other to generate bad code more efficiently does not magically lead to good code. The scaling fairy comes out only if you can give it enough high quality data
Upvote
3
(3
/
0)