LLM

I’ve built a chrome extension only with AI tools. I used Cursor as my primary code editing IDE and sometimes switched to VS Code with Github Copilot. With cursor, I mostly relied on Antrophics Claude 3.7 . Overall, my general assumption was confirmed that most models only help for smaller task and aren’t (yet) capable of understanding overall concepts of an application. For me, these tools are helpful to write some boilerplate code, rename or refactor some function or write simple scripts without any side effects.

Meta has announced the release of its Meta AI app. This is a trend all of these LLM players are doing. I’d expect it generates more traffic since it’s easier for a user to access ChatGPT, Gemini, etc. via its companion app instead of using the web interface. Also, from a user perspective it’s easier to upload a photo from your camera roll and using your smartphone’s microphone as well feels more natural.

I think about how much more a LLM knows about me compared to a traditional¹ search engine. I provide the llm way more information compared to a search engine since I use it way more interactively. A search engine does not provide me with results based on a previous query. The queries are always isolated and not based on each other². That’s different with llms. I could ask for a list of restaurants in my area. And then follow up with filters for this list, e.g. asking for restaurants (from the initial list) that are open this evening and provide vegan dishes. Using a search engine, I’d start with the same first query, asking for a list of restaurant recommendations, but then I would look at the recommendations one by one by myself. I definitely could also ask the search engine in the first place for “a list of vegan restaurants” but as soon as I’d ask for “a list of vegan restaurants opened tonight” some - if not all - search engine won’t be able to provide me with valid results. Thus, the llm learns more from me than the search engine (in this example, it would know that I might eat a restaurant tonight). This is just a simple example but I think interacting with the llm let them learn so much more from us than we think. Probably, you can compare it to humans interacting with each other. If you have a conversation, reacting on the things the other person is saying, you learn more than just telling / listening without a reaction.

WhatsApp has a new feature called “Meta AI” which is a Llama LLM model. The AI can be used in a chat manner (you can archive or delete the chat) to ask questions (just like ChatGPT and others). It can also be triggered via WhatsApp search. This feature is so underwhelming. Not the way I had expected it to be integrated. I had expected it to be way more similar to my “What if?” series: