Thread by @morganlinton on Thread Reader App
Short summary
This article stitches together several Twitter threads: it starts with a plain-English analogy for what model weights are, then shows how to iteratively refine code using Perplexity AI (with a stock portfolio analyzer example), and ends with practical steps for running LLMs locally on Apple Silicon via tools like Ollama.
- Model weights are literally just a file of decimal numbers (floats) that get loaded into memory and multiplied against input tokens to produce output.
- Iterating daily on prompts (breaking big tasks into focused, checkable segments) rather than stopping at the first working result yields dramatically better outputs, as shown with a stock-portfolio analyzer.
- A single detailed "autonomous team" prompt generated a 4,500+ line full-stack fund operating system (Thesium.finance) theoretically enabling 1-2 people to replace a 10-person analyst team.
- Apple Silicon's unified memory architecture makes it the easiest, most performant path for running LLMs locally, with tools like Codex Desktop launching on Mac first.
Full summary
Imagine you’re gauging whether you’ll make your flight on time: you look at traffic, departure time and distance, then give each factor a “strength” before multiplying them together for a final score. In LLMs, those strength values are the model’s weights. They’re just decimal numbers—floats—saved in a file. Load that file into memory, feed in your text tokens, and the model spits out its output. That file of floats is literally the model you’re running.
Since Perplexity.ai launched its Computer product, its author has been building and refining projects every day rather than stopping at the first working prompt. He uses a stock‐portfolio analyzer as a running example, showing how he tweaks prompts and workflows to squeeze more out of the system. His process involves iterating on code, breaking down big prompts into focused segments and checking each piece for reliability before moving on.
One standout build is a full-stack, AI-powered fund operating system—over 4,500 lines of code that, in theory, lets 1–2 humans replace a team of ten analysts. He prompted the system to act as an “autonomous engineering, product and research team” to create an MVP called Thesium.finance. That prompt spells out end-to-end goals—research through execution—so the AI agents can maintain live investment theses while humans supervise the interface.
On the hardware side, he swears by Apple Silicon for local LLM work. The unified memory architecture boosts GPU performance and simplifies setup. He points to tools like the Codex Desktop app, which arrive first on Mac, and argues that for anyone wanting to run models on their own machine—no deep AI background required—Macs offer the smoothest path.