introducing model inversion greptile detects if your PR was generated using GPT or Claude models, and uses the other model for review. rohan and rodrigo from our research team share why we do this:

Very cool

a more technical deep dive:

I’ve been doing that with local reviews just by a weird intuition. Nice to have some data to back it.

legendary

Stop leaking our alpha

why not add a LLM label to the commit/pr?

Why is this a product?

@ycombinator Why not just do two reviews? Token spend?

That’s dope

this is badass

Cool to see actual research on this - confirms a paradigm that feels like a prudent thing to do anyways

that's sick

No hate, but this video literally made me thought it was a YC video

I like this automated model reviewer routing!

I’ve a question. If you generated thousands of working PRs using Claude/codex and then the inverted models found more bugs, in that case were those PRs 1. Buggy when they were made 2. In the uninverted(vanilla) case, that is right now, are codex/claude letting bugs pass in? I want to understand what does catching “more bugs” mean and how do you quantify whether a comment added by Claude/codex is actually a bug find that would have otherwise broke the PR

Which model is used in such cases:

super cool video

Something @vaibcode said early that nails this pattern and has proven quite useful: "fight slop with slop." Excited to try Greptile out again! It's a very interesting finding that these models consistently produce classes of bugs that the others miss.

Interesting approach to reducing blind spots in agent code review

I literally do this, Have an adversarial pr skill, that as claude code to use codex cli to make the review. This is awesome greptile!

@ycombinator Smart

related pro tip: cross check a model's work with another model before submitting a PR. super fast (sol/composer/grok) and super lower effort "hey cross check please"

Great product, now I patiently wait for an excuse to use “greptile dysfunction”

bruh...so the same shit we've all been doing? wtf.

if anyone wants to do something similar while coding, try this: i always get my fable / sol / grok to talk to each and audit each other's work until both are satisfied and then hand over execution to a cheaper model

does inversion still help when the authoring model changes between commits in the same PR?

This is really thoughtful.

That's very suggestive way of showing data on charts. Perhaps a bit too suggestive, lol.

Nice video

Very interesting and makes a lot of sense. I am curious about a) whether it detects the model used to make the PR and b) what it does if a model that isn't Claude/GPT makes the PR, for example, Grok or Kimi.

sounds clever. but is this really necessary? feels like chasing ghosts instead of solving real pain.

That's genius

One non-deterministic model checking the code of another non-deterministic model to find out issues. That is a far fetch dream. You gonna bring in algorithmic verification, be with an agent or without an agent.

smart solution

woah!