Live data from Hacker News

Launch HN: Bloop (YC S21) – Code Search with GPT-4

news.ycombinator.com

111–120 of 131 posts

Re: Launch HN: Bloop (YC S21) – Code Search with GPT-4

#111
post #90

This is awesome and has huge potential to improve developer productivity! The only thing that it gives me pause is that it requires a very broad Github authorization. Do you really need to "be able to read and write all public and private repository data"?

For the desktop version we implemented GitHub OAuth using the device flow, so you can hold credentials on your local device. The tradeoff is there's no granular control of permissions, it has whatever access your account has.

On bloop cloud we use the GitHub App permission system which is more granular and only request read access.

Re: Launch HN: Bloop (YC S21) – Code Search with GPT-4

#112
Ok, I'll be the bad guy and say that the linked demo on loom is just not impressive at all. It's just 10x slower than what a simple regex query would've answered. Again, I'll be the bad guy in this comment section, but ML-based search methods for code are just not that useful. They do sound nice for non-coders, that's about it. I also feel like I'm qualified to say this because I have first-hand experience building code search tooling at scale.

Re: Launch HN: Bloop (YC S21) – Code Search with GPT-4

#113

Earlier quoted context omitted.

The point is that the interface to Bloop (or whatever 3rd party SaaS product) is just plain text. In programming parlance I would call this a facade [1] on top of GPT-4. The benefit is that you, the end user, can just continue typing things like "Where do we create our database pool?" in a search box, but under the covers Bloop can use GPT-4, but then eventually upgrade to GPT-5, GPT-6, or some other much better curr…

Except in the case of GPT-x, the facade builder must take extra pains to sanitize the inputs, so it doesn't become just a free GPT account for the user to interact with on unrelated topics. That's much more challenging due to the open nature of the prompts (the selling feature). One really lame solution is to ask the user to use their own openai API key.

I don't think that is a lame solution, it would work pretty well if OpenAI had better key management, especially spending limits per key.

Re: Launch HN: Bloop (YC S21) – Code Search with GPT-4

#114
post #46

I don’t understand these new saas that rely on GPT. It surely is my lack of understanding, but since the interface to GPT is mainly text: why any company would pay for N different GPT saas (e.g., one for searching for code, another for searching for documentation on confluence, another for allowing your business analysts with their queries, etc.)? Wouldn’t it make more sense for any company to have only ONE interface…

You are not wrong on this one. The moat they are going to build is going to incredibly ... low. The main intelligence is going to be provided by GPT models.

Agree. The most useful moat you build as an AI-as-a-service company is detailed data on how good each of your past outputs were. E.g. Github presumably keeps track of which copilot completions end up being used and which don't. If you are just using the OpenAI API, you have basically no way of making use of that data for finetuning. You could use the data in order to finetune your own model in the future though, maybe that's the plan.

Re: Launch HN: Bloop (YC S21) – Code Search with GPT-4

#115

Would be interesting to see precision and recall numbers - do you have any?

To evaluate semantic search we generated synthetic questions on our own repos with davinci-003. We can probably generate tougher and more realistic queries with GPT-4, would like to re-create this and open source.

We don't have precision/recall numbers for CodeSearchNet which is probably the biggest eval in this area.

Re: Launch HN: Bloop (YC S21) – Code Search with GPT-4

#116

Earlier quoted context omitted.

Except in the case of GPT-x, the facade builder must take extra pains to sanitize the inputs, so it doesn't become just a free GPT account for the user to interact with on unrelated topics. That's much more challenging due to the open nature of the prompts (the selling feature). One really lame solution is to ask the user to use their own openai API key.

I don't think that is a lame solution, it would work pretty well if OpenAI had better key management, especially spending limits per key.

If you're forcing your user to get their own OpenAI account, you're pushing them away from your product, at least somewhat. IMO the integration needs to be seamless for the illusion of "we're providing value here, not just wrapping GPT" to hold.

Re: Launch HN: Bloop (YC S21) – Code Search with GPT-4

#117

Tangential note to the person doing the video voiceover: pls process the audio a bit (high-pass filter at least?), or use higher a quality headset.

what is a high-pass filter?

Sorry I meant a low-pass filter - a filter that cuts out high-frequency sounds, like hisses and hard reverbs/echos

Re: Launch HN: Bloop (YC S21) – Code Search with GPT-4

#118
post #112

Ok, I'll be the bad guy and say that the linked demo on loom is just not impressive at all. It's just 10x slower than what a simple regex query would've answered. Again, I'll be the bad guy in this comment section, but ML-based search methods for code are just not that useful. They do sound nice for non-coders, that's about it. I also feel like I'm qualified to say this because I have first-hand experience building c…

> nice for non-coders, that's about it.

I'm not a coder but I code; not sure if that makes sense. I sort of know a bit of regex, but find it utterly painful and not worth the time.

I'm an amateur, effectively, with little time. I like AI tools for coding, because I can input a request, and get sample code. I know enough to be able to read most of the code that I get back, but not enough to be able to easily write such code on my own.

These types of tools, in my opinion, have the potential to make people like me, and even people who are much less knowledgable than me, productive programmers. This could be transformative.

Sure, we won't be as good or productive as real programmers, but that's beside the point.

In other words, what you're referring to with "that's about it" could still transform quite a few lives.

Re: Launch HN: Bloop (YC S21) – Code Search with GPT-4

#119
post #46

I don’t understand these new saas that rely on GPT. It surely is my lack of understanding, but since the interface to GPT is mainly text: why any company would pay for N different GPT saas (e.g., one for searching for code, another for searching for documentation on confluence, another for allowing your business analysts with their queries, etc.)? Wouldn’t it make more sense for any company to have only ONE interface…

There's a lot of work you could theoretically put into automating many of the queries you would want to do with a GPT chat interface.

For example, I'm no a medical diet and need a meal plan. This diet has quite a few restrictions.

I could set up automated queries that produce a weekly meal plan, and grocery list, and also checks all of the ingredients against the allowable / not allowable food lists. This could include multiple queries, one to get an initial response, another to validate it, etc. Everything could be organized nicely, so I wouldn't have to input any text to get what I need -- I'd just click buttons. There's lots of ways this could be further customized to be more useful, such as saving recipes, generating new recipes based on old favorites, generating recipes based on foods already at home, etc. (Heck, you could wire it up to a camera in the fridge.)

Anyways -- text input to get text output is the most basic form of this technology. There's a lot you can do with such text to make it easier to interact with for specific use cases.

Re: Launch HN: Bloop (YC S21) – Code Search with GPT-4

#120
post #19

From the Codeium ( https://www.codeium.com ) team here! Exciting to see more teams take on the code search problem - modern pure regex definitely is a suboptimal solution. Just to objectively highlight some of the similarities and differences between Bloop and Codeium Search (Codeium also provides Autocomplete) to a user: - Both are free for individuals - Both use a mix of AI and regex - Bloop has both free local and…

Posting about your company when relevant is one thing, advertising it another’s launch thread is another, and it’s pretty gauche… especially when, in this very thread, one of your testimonials is saying he doesn’t actually use your product: https://news.ycombinator.com/item?id=35236557

I don't see the issue with this. It happens on most of these Launch threads and is a common occurrence. I don't think there are rules that prohibit it either.
Post reply on HN