Testing ability of GPT4 and open-source LLMs to detect C++ bugs | Hacker News Reader

Testing ability of GPT4 and open-source LLMs to detect C++ bugs

catid.io

56 pointspanabee21 comments

I don't understand how the author can write "No, they cannot.", while the page says "GPT 4: 0 false alarms in 15 good examples. Detects 13 of 13 bugs."

Edit: the post says open-source models.

Was confused, too. The non-local models should be denoted.

Was confused, too. The non-local models should be marked.

It says

> Can open-source LLMs detect bugs in C++ code?

> No, they cannot.

GPT-4 is not open source.

Imagine if these were fully open source LLMs. None of them are

Then you could fix them

I don't get those 'Open source not good enough' posts. However good openAI can get, their secret sauce is known and open source will get there even with a delay. The don't have an exclusivity to the data on the internet

How would you fix them in an LLM?

You can provide additional examples of these specific issues, sure... but that doesn't mean it'll catch ones you haven't provided. It's not reasoning about what makes the patterns dangerous.

We’d need to know know Open AI does that, maybe they do indeed reach out to different models for coding ?

Few short learning + chain of thought would let it identify patterns in its gaps

There are fully open source LLMs. For example GPT4All GPT-J edition, Bloom, and there are bunch of code specific ones as well.

"open-source" is the key phrase here -- the author finds that GPT 4 can detect C++ bugs:

"GPT 4: 0 false alarms in 15 good examples. Detects 13 of 13 bugs."

It is amazing to see how GPT is so much better than other LLMs. Should this be attributed to OpenAI's better training data?

Presumably. Vicuna was fine-tuned on ChatGPT responses, but from a quick glance through the examples in the blog post, there are none asking it to find a bug: https://vicuna.lmsys.org/

I've found Vicuna 13B to be quite capable at _generating_ code. But not as good as GPT-4. Model size probably also plays a large role: 13B parameters runs nicely on normal-ish home user hardware, GPT-4 not quite.

Its possible that LLaMA 65B would be stronger with some instruction tuning, but it looks like its only been done on smaller LLaMa models.

I really want to see an instruction tuned LLaMA 65B. Hope it will happen.

Tested on toy examples: fibonacci numbers, quick sort and so on. Give it a pile of some C++ mess from any well known project.

Regardless of the current SOTA, this is a task where I'm fairly confident future LLMs can get very good at, unlike the "code in natural language" nonsense.

I'm thinking something like a linter with access to the AST, that can produce warnings like "you forgot this corner case", "you are not freeing resources. Is this intended?" and so on.

Probabilistic linters does seem like a fruitful realm for them, yeah. A lot of code shares strong similarities with others in small scales, thanks to stack overflow and various other "help me solve X" -> "try Y" answer pairs.

I do wonder how to reliably tell it to ignore noisy, incorrect warnings though. They're potentially sensitive to any new input / weight / random-seed changes, so it seems like literally every LLM upgrade will run the risk of ignoring existing suppressions (or you say "ignore this whole line" and miss useful warnings) due to small perturbations...

I just ran these past OpenAssistant and I'm seeing a much higher error detection rate. I've tried 6 so far, and it's caught bugs in about 4. OpenAssistant actually had a very good technical explanation of the bug it found in the BubbleSort example.

Didn't test SantaCoder or the salesforce models.

Would be interesting to see how they do.

> The answer may surprise you.

May. This is by far two most hilarious bytes I ever heard! :D :D :D

return (c_plus_plus->is(FREAKIN_HARD) || GPT->is(FREAKIN_LAME));// non-mutually-exclusive

Reply on news.ycombinator.com