# How I Use AI: Mid-2024


I've been in a mode of trying lots of new AI tools for the past year or two, and
feel like it's useful to take an occasional snapshot of the "state of things I
use", as I expect this to continue to change pretty rapidly.

-   **Claude 3.5 Sonnet (via API Console or
    [LLM](https://github.com/simonw/llm))**: I currently find Claude 3.5 Sonnet
    to be the most delightful / insightful / poignant model to "talk" with. It
    excels at complex reasoning tasks, especially those that `GPT-4` fails at.
    For example, I tasked `Sonnet` with writing an AST parser for
    [Jsonnet](https://github.com/google/jsonnet), and it was able to do so with
    minimal additional help. I don't subscribe to Claude's pro tier, so I mostly
    use it within the API console or via Simon Willison's excellent
    [llm](https://github.com/simonw/llm) CLI tool. The
    [Artifacts](https://www.anthropic.com/news/claude-3-5-sonnet) feature of
    Claude web is great as well, and is useful for generating throw-away little
    React interfaces.

-   **GPT-4o**: This is my current most-used general purpose model. The most
    powerful use case I have for it is to code moderately complex scripts with
    one-shot prompts and some nudges. `GPT-4o` seems better than `GPT-4` in
    receiving feedback and iterating on code. I also use it for general purpose
    tasks, such as text extraction, basic knowledge questions, etc. The main
    reason I use it so heavily is that the usage limits for `GPT-4o` still seem
    significantly higher than `sonnet-3.5`. And the pro tier of ChatGPT still
    feels like essentially "unlimited" usage.

-   **GPT macOS App**: A surprisingly nice quality-of-life improvement over
    using the web interface. Having the ability to `⌥-Space` into a ChatGPT
    session is super handy. I don't use any of the screenshotting features of
    the macOS app yet. They're not automated enough for me to find them useful.
    If there was a background context-refreshing feature to capture your screen
    every time you `⌥-Space` into a session, this would be super nice.

-   **Github Copilot**: I use Copilot at work, and it's become nearly
    indispensable. I recently did some offline programming work, and felt myself
    at least a 20% disadvantage compared to using Copilot. Copilot has two
    components today: code completion and "chat". I find the chat to be nearly
    useless. It has "commands" like `/fix` and `/test` that are cool in theory,
    but I've _never_ had work satisfactorily. The chat model Github uses is also
    very slow, so I often switch to ChatGPT instead of waiting for the chat
    model to respond.

## Use cases

-   **Docs/Reference replacement**: I never look at CLI tool docs anymore. LLMs
    have memorized them all. Whenever I need to do something nontrivial with git
    or unix utils, I just ask the LLM how to do it. I very much _could_ figure
    it out myself if needed, but it's a clear time saver to immediately get a
    correctly formatted CLI invocation.

-   **Limited Scope Refactorings**: Copy/pasting a small chunk (<100 lines) of
    code or SQL, and asking it to perform some transformation (i.e. "Make the
    query return weekly data instead of daily data", "Change this function to
    work with Fizz protos instead of Buzz protos") tends to have a high enough
    success rate that it is a time saver.

-   **General Knowledge Conversations**: I've enjoyed using the original ChatGPT
    voice chat feature during my commute. It feels like talking with someone who
    has read every Wikipedia article ever. As of 2024, the
    ["new" voice chat](https://openai.com/index/hello-gpt-4o/) feature powered
    by GPT-4o hasn't landed yet, so I don't have any experience with that.

## Things I Haven't Had Time to Try

-   [Gemini Pro/Advanced](https://deepmind.google/technologies/gemini/pro/), or
    its related tooling like [NotebookLM](https://notebooklm.google/). The
    coolest part of the recent Gemini models is their extremely large context
    window (2M input tokens). In my limited testing, Gemini seems "good", I just
    haven't had enough time tinkering with it to see where it exceeds the
    capacities of the OpenAI/Anthropic models.
-   [Deepseek Coder V2](https://github.com/deepseek-ai/DeepSeek-Coder-V2): An
    extremely powerful open-source model for coding. This one looks pretty great
    by the benchmark results Deepseek have posted. However, I tried playing with
    the quantized model locally and was disappointed. The full model is rather
    expensive to host locally, which has been a barrier. Deepseek also offer the
    model via an API (at quite low cost too), which I hope to try eventually.

## Resources

-   [Zvi Mowshowitz](https://thezvi.wordpress.com/)'s weekly AI posts are
    excellent, and give an extremely verbose AI "state of the world".
-   [Simon Willison](https://simonwillison.net/)'s blog is also an excellent
    source for AI news.
-   [The Cognitive Revolution](https://www.cognitiverevolution.ai/) podcast
    hosts some pretty good interviews that I find to be high-signal-to-noise,
    and is much less hype-driven than many other AI-centric podcasts I've
    attempted to listen to.