Thread by @morganlinton on Thread Reader App

Open original

This unrolled thread covers four topics: a plain-language explanation of model weights, strategies for refining code with Perplexity AI’s Computer, an AI-native fund system built end-to-end, and tips for running LLMs locally on Apple Silicon using Ollama. It walks through each use case with examples and practical advice.

  • Perplexity's Computer platform generated 4,500+ lines of full-stack code for Thesium.finance, an AI-native fund aiming to replace a Bloomberg terminal and ten analysts with two humans overseeing automated trading agents.
  • Getting usable output from AI coding tools requires iterative refining through multiple prompts, not just a single initial request.
  • Apple Silicon's shared CPU/GPU memory makes it well-suited for running LLMs locally, letting users host models and keep data/inference private without needing to understand the underlying math.

Think of model weights as sliders that tune how much each factor matters when a language model makes a prediction. If you’re guessing whether you’ll reach the airport on time, you’d weigh traffic, departure time, distance and so on. In an LLM, each connection between neurons carries a decimal value—a float—that multiplies an input signal. All those floats live in one big file. Load it into RAM, feed in tokens, and the model spits out text. That file of numbers literally is the model. Around late March, the author jumped into Perplexity.ai’s “Computer” platform and built several projects, starting with a simple stock portfolio analyzer. They emphasize that the first prompt isn’t the end—every build needs iterative refining. Using a step-by-step workflow, they tease out better code and sharper prompts. For their AI-native fund system, Thesium.finance, they had Computer generate more than 4,500 lines of full-stack code, aiming to replace a Bloomberg terminal and ten analysts with two humans supervising automated agents that research, assess risk and execute trades. Finally, they walk friends through running LLMs on local hardware, declaring Apple Silicon—and its shared memory between CPU and GPU—a game changer. You don’t have to know the math under the hood. Install a Mac-first tool like Codex Desktop, download your model weights, and point the app at them. With a few commands, you’ll host a transformer on your machine and keep both your data and inference private. It’s about moving from “it’s magic in the cloud” to “I control every float in that weights file.”

Ask about this article