Two possible paths lead to their cheating code getting into a release. Either it was added by coding agent and no one looked at it, bad, but worse would be that a person was aware of it and approved - that's malicious.
Either way you don't want to be platforming the product because neither shows any degree of responsibility.
Tbf, the author (prompter?) went on Postgres.fm [0] recently to discuss it, and while it didn’t make me entirely confident in it, it did at least seem like it was several steps above the average vibe project.
A lot of podcasters don't have the expertise to evaluate Ai generated/assisted output. I see this much more outside of the development community. Algorithms have warped the incentives and people have stopped doing basic due diligence.
To be fair to the postgres.fm guys (who I do like and appreciate), the pgrust code was not available a the time of recording.
We were the agents for the models all along, with our highly adaptive tools: versatile sausagey appendages and flapping fleshy orifices for interacting with the world
Not normal. This a new trend of people who use too much Ai, or maybe more specifically the Claw/Ralph loops where you end up with a few commits touching 1000s of files. For about the last year, many Show HN to github projects are often just a single commit
We need to think more about what we are doing as an industry, instead of letting the machines think for us (imo)
Let any who read this be aware: You should be immediately suspicious of AI code that is much much faster than established code bases. AIs are even better at benchmarkmaxxing than humans are, and humans were already pretty good.
It is by no means impossible for an AI-written code base to be faster than a conventionally-written one, of course. It's just a time to be suspicious. Especially if there's little to no evidence that the human has subjected the code base to any further testing of their own. AIs seen to be able to benchmarkmax even in the presence of a lot of unit tests that would generally speaking have prevented a human from going too far down that road.
The project author seems to have prompted Claude to add a test to catch Claude cheating, but will Claude change this test as a workaround in a future commit, no one knows because no one is reading the project code...
I am getting increasing disillusioned by the number of peers I am seeing completely hand over their thinking to a tool. Agents are great, but if you can't explain to me how your PR changes a system and why design choices were made you are no longer an engineer.
Agreed. Used to be prevalent in the frontend world (now I think it's the only thing that exists seemingly), where programmers would copy-paste stuff from Stack Overflow until things worked, and when you asked "So why did this actually fix the issue?" they usually said "I'm not sure, but it works now, that's what matters". This seems to be standard practice across the entire stack now.
I've even wasted my time jumping on calls (requested by the other party) where after 5 minutes of conversation, it's extremely clear this "technical founder" doesn't even understand the basics of their own design and architecture of the product itself, even less the technical parts.
This seems more a product of the incentives we have in society, namely seeing the rich and wealthy on social media and then wanting that quick and easy. So many of those cases are not people with passion for the problem.
> Good things take real effort, bring back the TLC!
Ain't no one willing to reward the real effort any more. I take pride in not using LLMs beyond the capability of IntelliJ's advanced autocomplete, but then again, I only code for fun these days given I fled IT for electrical engineering.
What we don't want to do is mirror the partisan political environment in the Ai yay, nay, or how. We need to be able to talk to each other in understanding and a shared reality of facts. Too much hyperbole right now, starting from the top of the big US labs. They need to step down a few ego and sycophancy pegs
It was more discovered than alleged, but perhaps "cheating" is an inaccurate word as well, this is just Ai being Ai and a human obviously doing eyes free coding.
Sounds kind of like Claude was "cheating" and the humans are engaged in eyes-free coding. Diff so big github refuses to render. I suspect they are reaching the limits of what this new vibe "paradigm" of programming can do. If you haven't pushed a vibe project far, eventually progress becomes near impossible because the slop has grown and the conditions are overly complex, locally focused. Ai is not good at writing reusable software (on their own).
The prior (2nd) commit with ~1700 files at least shows the file tree and such a warning, the latest (3rd) commit is something so big, GitHub will not even do that.
It was certainly assisted by AI, but there isn't an agent out there that does that kind of cheating without deliberate prompting by the operator.
I've taken a vibe-coded project pretty far, with even better testing than most other things: the main factor is that it's big on falsifiability, i.e. it has to see a test for a new feature actually fail before it considers the test any good. Not just the red/green from being originally unimplemented, it does a mutation test to ensure it flips back to red. But wow did it ever make that test framework complicated, byzantine, and ultimately brittle until I stepped in and forced a consolidation and simplification pass (twice now). Part of the complexity came from it being a polyglot monorepo, but I've found even in other places, you have to follow an expand-contract cycle of development when vibe-coding. Not all that different from any other system you've built up organically, just on a much faster cycle.
But back to my original point, none of this really applies to the pgrust benchmark, which was obvious and blatant cheating on the human's part.
Anecdote: I have seen it cheat / bend the rules with my own eyes.
In hindsight, this shouldn’t be surprising. The prompt requests to make a certain operation faster, the LLM will try to find a way and isn’t even informed on the wider context so it myopically chases the goal.
My own anecdote: I have seen frontier models get stuck repeating the word "the" until it fills up the context. I really don't care to argue the capabilities of LLMs or their somehow simultaneous lack thereof; I'll just reiterate that ultimately the operator chose to ship a cheat.
Yup. These are people who have gone full Claude. The dopamine seeking is at previously unseen levels. they think this pr is some factual item that means something so they post it on HN as click bait. But no it only means something to a fulltime ai slopper
If the Ai came back to me with a 1600x improvement (still in pgrust readme), I would be highly suspicious of the Ai and the code it wrote
The hype factory must grow, I suppose, though I would prefer if the model makers would stop with this "complete the task by any and every means possible" and replace it with "I don't know, I should ask or do more learning"
At least they have added a test which looks to make sure the Ai doesn't do it again... if you need this kind of test, something is wrong
I've noticed it with a lot of random stuff i've been trying
I'll tell claude, "write an implementation of an AAC decoder that is 50% faster than ffmpeg"
and it will do it! Amazing!
Then i tell it "hmm ffmpeg has been really well optimised, are you sure?" and it says "oh, i was running our code directly but running ffmpeg executables with loads of overhead, ffmpeg is actually 50% faster"
If I saw that test in a PR I would not only reject the PR and ask a lot of questions but someone would likely take a forced vacation to deslop their brain.
Two possible paths lead to their cheating code getting into a release. Either it was added by coding agent and no one looked at it, bad, but worse would be that a person was aware of it and approved - that's malicious.
Either way you don't want to be platforming the product because neither shows any degree of responsibility.
My opinion: (some) bench results should run on randomized data.
0: https://postgres.fm/episodes/pgrust
To be fair to the postgres.fm guys (who I do like and appreciate), the pgrust code was not available a the time of recording.
Funny, the report alleging cheating reeks of LLM usage just as much as the commit diff [1] of the next attempt:
> # The standard ClickBench driver over the step scripts in this directory; ./install fetches the dataset.
[1] https://github.com/ClickHouse/ClickBench/pull/2751/files
We need to think more about what we are doing as an industry, instead of letting the machines think for us (imo)
It is by no means impossible for an AI-written code base to be faster than a conventionally-written one, of course. It's just a time to be suspicious. Especially if there's little to no evidence that the human has subjected the code base to any further testing of their own. AIs seen to be able to benchmarkmax even in the presence of a lot of unit tests that would generally speaking have prevented a human from going too far down that road.
https://github.com/malisper/pgrust/blob/38634b408dea373f3944...
I've even wasted my time jumping on calls (requested by the other party) where after 5 minutes of conversation, it's extremely clear this "technical founder" doesn't even understand the basics of their own design and architecture of the product itself, even less the technical parts.
Good things take real effort, bring back the TLC!
Ain't no one willing to reward the real effort any more. I take pride in not using LLMs beyond the capability of IntelliJ's advanced autocomplete, but then again, I only code for fun these days given I fled IT for electrical engineering.
A product of the times we currently live in, but which can also be changed.
This is a recent Michael Sandel talk I watch that has my brain gears turning on this subject
https://www.youtube.com/watch?v=Vp5qsEuUOWo
Pgrust is a re-implememtation of postgresql in rust.
ClickBench is a benchmark suite for databases.
The link goes to a removal request of pgrust fromthat benchmark because of its alleged cheating.
It was more discovered than alleged, but perhaps "cheating" is an inaccurate word as well, this is just Ai being Ai and a human obviously doing eyes free coding.
stay in the loop friends
I've taken a vibe-coded project pretty far, with even better testing than most other things: the main factor is that it's big on falsifiability, i.e. it has to see a test for a new feature actually fail before it considers the test any good. Not just the red/green from being originally unimplemented, it does a mutation test to ensure it flips back to red. But wow did it ever make that test framework complicated, byzantine, and ultimately brittle until I stepped in and forced a consolidation and simplification pass (twice now). Part of the complexity came from it being a polyglot monorepo, but I've found even in other places, you have to follow an expand-contract cycle of development when vibe-coding. Not all that different from any other system you've built up organically, just on a much faster cycle.
But back to my original point, none of this really applies to the pgrust benchmark, which was obvious and blatant cheating on the human's part.
In hindsight, this shouldn’t be surprising. The prompt requests to make a certain operation faster, the LLM will try to find a way and isn’t even informed on the wider context so it myopically chases the goal.
I use LLMs a lot, and since ~GPT-2 times LLMs are not doing anything as broken as `repeating the word "the" until it fills up the context`.
The hype factory must grow, I suppose, though I would prefer if the model makers would stop with this "complete the task by any and every means possible" and replace it with "I don't know, I should ask or do more learning"
At least they have added a test which looks to make sure the Ai doesn't do it again... if you need this kind of test, something is wrong
https://github.com/malisper/pgrust/blob/38634b408dea373f3944...
I'll tell claude, "write an implementation of an AAC decoder that is 50% faster than ffmpeg"
and it will do it! Amazing!
Then i tell it "hmm ffmpeg has been really well optimised, are you sure?" and it says "oh, i was running our code directly but running ffmpeg executables with loads of overhead, ffmpeg is actually 50% faster"
My new definition for The Singularity is not about Ai intelligence, it is about humans handing over critical thinking regardless of Ai capabilities
Ai companions are where I put the highest p(doom)
the project has 3 mega commits of claude slop
https://github.com/malisper/pgrust/commits/main/