Did anyone have any success with WebLLM? Which model worked for tool calling for you?

I am developing a pocket AI team. Its a PWA so you can access the power of your AI team anywhere. I am not doing fully active work on it, because its a fun project, but I have a business to start. So dont expect a fancy tool too fast. great good morning

Its crazy that... Nah, I just wish you all a great day, its coffee time here! ☕🏞️🤙 View quoted note →

Use AI as a tool to reach the goal, not to build a tool for a goal.

I am as close to have the best meal planning and nutrient app as never before. The goal has never changed, only the tools available, and my mindset how to use them. The goal is to have a meal planning app, that always adjust itself to me, but improves week by week, month by month. No sweating calorie control. And in the current state, it uses the data I am comfortable measuring. No necessary calorie counting. Build with what we have, and improve. great good morning

Thats interesting, and kind of efficient to build system with AI where the same AI is part of the system. So basically AI does the job, and then you do retros with the AI to do the job better. It is a lesser complex idea, than building a tool with an AI, that has to produce an output. Because there you have 2 outputs: the tool and the tools output. So you have to improve your AI, your tool, and your results. 3 layers, instead of 2. But if you cut out the tool layer, you kind of simulating the manually doing something instead of automation. And actually this is how you automate properly. You define first the process, fine tune the process, then automate. This way, you will be faster reaching your result, and fighting less fights.

I am aiming to improve my health with data. So I have an AI team to research, build, measure, reason about data. My current stance is, backed by data, that you shall pay attention to food variety but also ingredient amount. So I have now a table that summarizes my nutrients goals and variety goals for my generated meal plans. Both paint a different picture. You can eat in great variety, but still overeat, undereat, etc. Variety + the right amounts matter.

Who would be interested in a project, where you run AI teams in your browser? No server, just purely in the browser, your data stays local. AI team for - personal development - nutrition/meal planning - sport - business administration Thinking about a webapp, what runs locally, can download teams locally, and then the data stored by the team stays local, stays in your browser, can be downloaded.

AI agents == flexible workflows They dont have 100% deterministic output, but also do not need 100% deterministic input.

I advanced yesterday my coach team. Its quite awesome. The "meal-planning squad" can now: - plan meals from my recipes - add random filling ideas from non existing recipes - automatically registers new ingredients - compares the weekly plan against variety goals and nutrient goals It is quite simple though, bunch of md files, I sync across my devices, and a python script to run some deterministic work. No app, no database, no hard dependencies yet. Its so flexible (because of AI), that it can adapt easily to new ideas. I am planning to go from AI to semi-AI, and make most of the things automated, so flexibility stays, but uses less tokens, and be more deterministic.

I made yesterday a small visualization of my work llm credit spending. I wish I have done it earlier. great good morning

Imagine you download a social media app and upon signing up it opened a bank account for you. 😂😂

When netflix started, they shipped dvds to test their idea. What is the equivalent of this for AI development? Let the AI do manually what you want, and build the infra if you really need it.

I am wanting to build a meal planner app, that plans your meal around the goal of healthy nutrient ranges. My last attempts all were facing a lot of issues in a sense, that it would have to bridge many hard gaps, so I always started easier, and not meeting the original goal. I learned and improved with every trial. The last trial, when I made an AI coach team to be the app was the most successful. The whole app-experience is mainly markdown files plus 1-2 small python scripts, but achieved 100x more in a few hours as ever before. The learning: You can use AI to build an App. You 10x your output. If you make the AI do what you need, you shrink the improvement cycle, and 100x the output. The cool thing is, that you can prototype an idea by the AI doing it, then automate the steps. Then you dont need to implement a single line of code until you know this step makes sense. As it should be. great good morning