Comments URL: https://news.ycombinator.com/item?id=44494709

Points: 3

# Comments: 1

Comments URL: https://news.ycombinator.com/item?id=44383356

Points: 2

# Comments: 2

I've been an active Leetcode user for a long time and while I think it's a great platform to test your skills, I've never really considered it a good resource for learning. Whenever I'd get stuck on a question I'd just look up the answer then try to understand it via reverse-engineering which never seemed very efficient to me. That's why I built HinterviewGPT.

HinterviewGPT started as a side project where my initial idea was to build a UI that looked a lot like leetcode with the question on the left and the editor on the top right, but have a chat UI on the bottom right where you could chat with an LLM-based "tutor". The "tutor" would be aware of the question and your current solution in the editor and it would be instructed to answer your questions when you are stuck by giving hints and specifically NOT instructed to give you the answer (otherwise it wouldn't be a very good tutor).

I built this side project and it worked surprisingly well. The tutor's hints guided me towards the correct answer conversationally rather than me having to look up the correct answer when I got stuck. This really made my learning more efficient and cut down on my study time per question. It also promotes "active learning" (learning through hints) rather than passive learning (trying to reverse-engineer the solution) which I think made the learning stick more. As a result I decided to build this into an actual product.

For the product I added a "question generation" feature. Via another chat UI you can describe the kind of question you want to practice (based on industry, role, topic, job req, etc.) and HinterviewGPT will generate a question for you which you could then practice. So you don't only have to study leetcode questions with HinterviewGPT, you can really study any kind of interview question. You can also study both code-based and text-based (e.g. behavioral, system design, etc.) questions. Or if you have a particular question in mind that you want to study, you can manually enter the question as well. Once you're done practicing the question you can submit it for a final evaluation and your solutions and evaluations are saved for future reference (again, like leetcode).

Here's a demo of the full workflow i.e. question generation, practicing with tutor, submitting your solution: https://www.youtube.com/watch?v=Yj6qvEQYWi0

All the LLM-related features use OpenAI models under the hood specifically gpt-4o-mini, gpt-4o, o1-mini, o3-mini. You can choose when to use which model. gpt-4o-mini does suprisingly well in a lot of cases but obviously for more complex questions the more advanced models are preferable.

HinterviewGPT offers a free trial (with only gpt-4o-mini is available unfortunately).

Note: don't take this as me endorsing leetcode-style interviews, I actually don't like them at all or think they are effective.

Comments URL: https://news.ycombinator.com/item?id=43730150

Points: 1

# Comments: 0

I've created Diagify, a CLI for converting natural language to technical diagrams. There are a bunch of diagram-as-code tools out there and I had the idea that it should be possible to use LLMs to generate the diagram code based on a natural language description, execute it, and return the generated image. This is essentially what Diagify does.

More specifically, Diagify generates code for the Mingrammer library which is specifically for technical diagrams. It also uses the OpenAI API to generate the Mingrammer code from the natural language description.

The workflow is first OpenAI generates the Mingrammer python code based on the description, then it does some basic error checking. It was found that often the generated Mingrammer code had incorrect imports so this is checked. If incorrect imports are found, OpenAI is called again for correction with the incorrect imports identified and also some suggested imports to replace them with. Then the Mingrammer code is executed. If it is executed successfully the corresponding image is generated. If there is a runtime error then OpenAI is called again with the identified error in an effort to correct it. It's become somewhat reliable at this point.

My reasons for creating Diagify are two-fold 1) to see if it even would work and 2) creating technical diagrams by hand or even using diagram-as-code tools can be tedious so using a simple natural language interface could be helpful.

Comments URL: https://news.ycombinator.com/item?id=42420049

Points: 5

# Comments: 0

What were the layers of meaning? To me, the shining is kubricks most straight-forward movie.