Home of Chris Adams
  • Home
  • About
  • Contact

Giving tool-calling LLMs scratch memory

  • cadams

Chris Adams

4 October 2024 • 1 min read

https://docs.anthropic.com/en/docs/build-with-claude/tool-use#single-tool-example Tool support · Ollama BlogOllama now supports tool calling with popular models such as Llama 3.1. This enables a model to answer a given prompt using tool(s) it knows about, making it possible for models to perform more complex tasks or interact with the outside world.

Launching AI Trix Consulting
Technology

Launching AI Trix Consulting

I always aspired to be an entrepreneur. Have a successful business and make money myself, no paychecks in sight.
7 March 2025 5 min read

Making apps that won't hook your kids

This post is about a superpower we gained in 2024. With a little work AI can build you whatever software you ask for. Because it’s made by you, it can work the way you want. The saying “there’s an app for that” is great marketing - now it can finally come true. Large language models like Claude and ChatGPT can write code! Just describe what you want and they’ll choose the software tooling, write the code, and explain it too. Then you can ask how to make the code work and use those instructions to try it out. When it’s not quite right (it never is the first time) you can also ask for changes! It’s not yet suitable for making anything complex unless you’re already a software engineer or you’re very persistent. That said, for little toys it works great. Why not make an app for your kid. Unlike most of the app stores yours can be free of addictive tricks; free of micro-transactions, have no ads, have no time-based unlocks, no in-app purchases, and no fake currency tokens. You could even bake in a screen-time limit - just ask for that feature when you specify what you want. The sky’s the limit. Maybe you want to create a story about your kids, or maybe a flappy bird clone. To prove the concept I went and created a pocket xylophone. I call it XyPhone. Farm animal noises apps on the app store overly monetized apps for kids/babies, let’s make our own! outline this post is about lLMs cam write code problem: spo many apps today are crap, overly monetized, and some are even worse - designed to hook kids show examples of how terrible the apps are sad state of the app store in 2024 is that nothign is free, everyone is trying to suck money out of you, and in a lot of cases the only move is not to play. making my own - xyphone - first cut - asked for an app - realized i could just do my own swiftui - process for self meta learning, trying ot make another app and found 12 hour course, came back and understood what to do to make my own thingy - generate audio - would’ve have to look up frequencies and how to make audio with AudioKit - saved a ton of time there harmonics & reverb would have had to play with values to get something good, instead got a great first cut right away. - tr-catch around call to sunthetizer.sendProbramChange - uint8 fix - same width bars (screenshot of original look) - ignore safe area - overlapping notes - popping when a button is pressed - tried attack/release, midi tempo - tweak instrument duration - why fuzzy at startup??? - tried to add a loading indicator - asked for it to generate the waveforms fo rthe nortes - didn’t work, it insisted I could just insert wav files of each note dragonbreath - wanted for a while to be abel to breathe on the microphone and have fire come out - claude version was crap (include a video) - grabbed a stock fire footage but it was too slow - asked for a ffmpeg command to speed it up 2x, that wasn’t fast enough,m went to 3x and then 5x then 10x. the faster the better up to a point, since the whole point is to have a snappy demo that gets the point across before people get bored, so went for 10x speedup - i believe this is fair use as it’s parody and being done for educational purposes to demonstrated the power of AI to write code - todo: add fire crackling sound effect from youtube. add it as a soundtrack to the video so it’s always in sync. - adding the sound - downloaded a fireplace crackling, firest5min of https://www.youtube.com/watch?v=xKFBYZNXqAk - and a fireball sfx https://www.youtube.com/watch?v=AHRf27GPhQc - sped up teh fireplace 16x to get more crackles - boosted the bass - trimmed the fireball to just a whoosh portion - sped it up slightly - overlaid then on top of one another - sped up the whoosh 1,5x so it was right length for the video and redid the overlay - added it as the sound for the video - rebuild the app and voila! - needed to play audio via speaker! - sfx: recap
18 November 2024 3 min read

AI cannot push the frontier forward

frontier people don’t notice the big changes happeniong, incremental progress adds111 up to a lot, over a long time Disagree on it being incremental. I need to write something about this. My argument is in short - that LLMs can synthesize anything “inside the frontier” eg. writing a poem (there’re enough examples to work with to make a novel poem) or implementing something that’s well documented / has loads of examples - but it can’t “push” the frontier forward (eg. can’t synthesize a large codebase as there are no examples to work from); that may change in future but for it to change it’d require hypothesis forming AND testing AND a self-iterating feedback mechanism to all be part of the agentic loop. Self researching agents - not a thing without ability to synthesize beyond what’s int he training data Our training data comes from our environment, and training is continuous There won’t be an “AI takeover”
18 November 2024 1 min read
Home of Chris Adams © 2025
Powered by Jekyll & Kaschber