Can AI models create highly technical, personalized solutions for institutional clients, such as integrating into a firm’s tech stack?
I’m not an expert in crypto, but models will fail a lot—about 70% of the time. The goal isn’t to prevent failures but to make them happen quickly so we can learn. Models are highly parallelizable; you can spin up dozens, test many approaches, discard what doesn’t work, and feed the results into the next model. By iterating fast, you eventually get a solution that works for complex, technical use cases.
It's easy to get started with AI communication because the companies building these models intentionally make them accessible so anyone can get decent results without being a prompt engineering expert. However, the real advantage comes from skill: due to a power‑law relationship, improving your prompting ability from the 90th to the 99th percentile can roughly double the quality of the AI’s output. In other words, small gains in skill produce large gains in results. While the basics are simple, mastering AI prompting is hard, and because AI is a meta‑tool that enhances all other tools, becoming proficient at it makes you effective across many domains.
You want to wrestle a little control back from the big monolithic AI companies, and I can help with that. There are two kinds of improvement happening right now: hardware (massive data centers letting labs train bigger, and per the bitter lesson, generally smarter, models) and software/algorithmic improvements (inference optimizations, quantization, and other hacks that let you run faster or smaller models at similar intelligence). The hardware path isn't very accessible if you're not capital-rich — big models just cost more to run and labs will charge accordingly. But the algorithmic path doesn't require massive hardware. China, rate-limited on compute compared to the US's roughly 50-100x compute advantage, has poured its effort into this path — models like GLM 5.2 and Kimi K2 have gotten almost an order of magnitude more efficient purely from software optimization. As a result, you can run a model like GLM 5.2 locally right now, 1-bit quantized, on a MacBook with 256GB of memory. It's not unreasonable to expect that within about 6 months we'll be running something similar (or a bit dumber) on our phones. So while the fear of big model companies taking over is understandable, it seems very likely everyone will soon be able to run models much smarter than what we have today on compute as modest as a phone — you won't get the frontier models doing galaxy-scale planning, but you'll have genuinely capable local assistants you don't depend on any one company for.
local modelsopen source aivendor lock-inquantization
I understand the concern about big model companies dominating, but there are two main types of improvements happening right now: hardware and software algorithmic improvements. Hardware advances mean larger server clusters and compute resources, which let companies train bigger models faster. Software improvements—quantization, inference hacks, and other algorithmic tricks—make models run faster and smaller. Because of these algorithmic gains, especially from teams that lack massive capital, we can run efficient models locally. For example, the Chinese models like GLM 5.2 are highly optimized and can be run on a consumer laptop: a 1‑bit quantized version runs on a MacBook with 256 GB of RAM. In the next six months we’ll likely be able to run comparable models on phones. So the fear of being locked out is overstated. The simplest way to get started is to download Ollama, a platform that lets you run open‑source models locally. You can pull models such as Ornith, GLM 5.2, or Kimmy K 2.7‑code, and many community members provide quantized versions that run on modest hardware. Even if a model is too large, you can pay a few dollars for a hosted service to help set it up, and you’ll be fine.
In general, the best niches for an AI service are high ticket, so these are services that themselves get a lot of money when they close a deal. They also tend to be digital, meaning services that you can scrape online fairly easily. And finally, they tend to be really low compliance need — services that don't require you to check a bunch of HIPAA boxes or do a bunch of SOC 2 compliance in order to enter the space. The reason why: high ticket means if you get them a single client, you pay for yourself for a long time. Digital means it's really easy to scrape, so low friction. And low compliance needs means you're not going to screw anything up — there are a lot of hidden risks that people constantly consider before accepting a project with an outside vendor if they are in a compliance-heavy industry.