__0x01

Agents are very good at making us think the tower is rising, when in fact it is falling beneath our feet.

Thanks for this. I especially enjoyed the section on word counting software and word linking games.

From the companion paper:

> The argument relies crucially on ideas that may, at least in retrospect, be attributed to Ellenberg-Venkatesh, Golod-Shafarevich, and Hajir-Maire-Ramakrishna.

Can someone please elaborate on this?

> A problem with this is that in order to confirm the findings, you’ll need an expert human. But generally expert humans are busy doing other things.

The article suggests using LLMs to identify and fix UB. However as per the above, I think the issue is that we need more expert humans.

LLM generated code will eventually contain UB.

EDIT: added "eventually"

> Correctness: 1,778 test cases from the official jsonata-js test suite + 2,107 integration tests in the production wrapper.

The AI generated code can still introduce subtle bugs that lead to incorrect behaviour.

One example of this is the introduction of functions into the codebase (by AI) that have bugs but no corresponding tests.

EDIT: correct quotation characters

Can a datacenter practicably upgrade to the next generation of GPU every year?

What was the interesting problem that Swiss watchmakers could have focused on, instead of pursuing brand?

Please can you elaborate? We are more intersted in "the end" in what sense?

Is "AI code review" a correct term?

A code review requires reasoning and understanding, things that to my knowledge a generative model cannot do.

Surely the most an AI code review ever could be is something that looks like a code review.

> This is nothing but speculation

Did you read the paper?

I also worry about a centralised service having access to confidential and private plaintext files of millions of users.

Often when I look closely at the output of LLM generated code, I see repetition, redundant logic and deeply hidden bugs.

Notwithstanding the above, to my understanding LLM services are currently being sold below cost.

If all of the above is true, at some point the degredation of quality in codebases that use these tools will be too expensive to ignore.

These LLM tools appear to have an unprecedented amount of access to the file systems of their users.

Is this correct, and if so do we need to be concerned about user privacy and security?