[w4]

I agree that Google search is very vulnerable. The results are spammy and the current ad volume badly degrades usability for me. In my case, I don’t use Google search much anymore. Previously it was my one stop for finding information online.

Instead of Google search:

I’ve switched almost exclusively to Kagi for vanilla search. It feels like Google in 2015, and the forums/small web filters are great. I do rarely use !g, and usually I'm disappointed.

If I just need a quick question answered I’ll usually use Perplexity.

For coding questions, I mostly use copilot and documentation these days. I barely use search for coding questions, unless I run into a weird edge case, or need to find Github issues for a project. There’s too much link spam to wade through, and copilot is built right into VSCode, so I don’t have to context switch.

I was a big user and early adopter of ChatGPT, but I don’t use it much anymore. Special use models are better for specific use cases (e.g. copilot for coding), and if I want to learn general information about a topic or gather some opinions, I much prefer to search forums and blog posts or watch some YouTube videos (this is the one major bright spot for Alphabet imo). If I do need AI to do something for me, I prefer to use an API or local model. ChatGPT is inferior to the API/local products, especially if you know some Python. This is even more true when you take into consideration copilot (and starts to hint at the compounding opportunity of these tools).

If I want to do a deep dive on something I don’t even bother with most web content these days unless I'm researching a cutting edge topic. The general web is too noisy and inundated with inaccurate and low quality content. Most of the time I either buy a book or directly visit an authoritative source (e.g. the SEC, the courts, Wikipedia, etc.). Less frequently I'll find a niche forum or blog via Kagi's filters, or hit up Google Scholar/Arxiv (another bright spot for Alphabet, but very niche).

The most interesting thing to note here is that I’m paying for almost all of these things, rather than using free Google search, because the Google search user experience has degraded so badly. Kagi, copilot, LLM APIs, books, they all cost money. That doesn’t bode well for the Google search product long term.

[hackerlight]

> ChatGPT is inferior to the API/local products, especially if you know some Python.

Why is this? GPT-4 outperforms smaller models I can run locally. Is GPT-4 via API better than GPT-4 via webapp?

[w4]

Yes, for a couple of reasons:

- I control the system prompt to guide the model to do exactly what I want, rather than begging and cajoling ChatGPT. I will often use the API playground for this reason where someone else might use ChatGPT.

- For more complex problems, having the input and output in Python gives me a ton of options for data manipulation, storage, etc.

- Using an API call lets me daisy chain different models and other APIs/tools together. ChatGPT can sort of do this via plugins, but I haven't found it to be a great experience.

GPT-4 is definitely better than the local models by itself. It's not a clear winner when you start to get into fine-tuning/ensemble workflows. It's also less attractive in situations where I want to chew through a ton of tokens for a personal project and GPT-4 would get expensive.

[hackerlight]

> daisy chain different models and other APIs/tools together

> ensemble workflows

This is intriguing, can you explain in more detail what this looks like?

[CuriouslyC]

Look up langchain. You can have most models format output in such a way as to facilitate programmatic execution based on their contents, and you can let the model just run that way with some checks to keep it from losing the plot.