development · AI · opinion · · 2 min read
A case for AI-pragmatism - leverage cheap compute while it's here
Make the most of cheap compute to do cool things.
It’s no doubt highly entertaining to wax lyrically about AI. “It’s the bee’s news” say some, others “It’s the end of civilization as we know it”. And the truth is probably somewhere in between. What I find lost in the politicized Zeitgeist is pragmatism.
As an individual, I understand fully that Anthropic, OpenAI, Google, Apple, Microsoft, insert literally all large tech companies here aren’t building tools and services for me to live a better life but to make profit. This really is not shocking, nor is it remotely controversial. Profit is what drives these organizations - period. That being said, there can still be opportunity for utility as an individual of said technology without giving up agency or being bought into anyone brands marketing. And I think this is especially true with LLMs and frontier models being offered on the market today.
Microsoft’s GitHub copilot provides a useful case study on what I mean. It was insane how cheap and accessible GH Copilot actually was. I was able to blast massive amounts of tokens for $10/month with little friction or rate limits with a variety of computationally-expensive models. No doubt they were collecting data and metrics as BigTech does but I was cognizant to not put actually private or otherwise sensitive data into their cloud-enabled software. I also knew this was not going to last forever - there was no chance GH was making money on the feature, and it was likely an attempt at them capturing the developer market to then profit from at a later date. Fast forward a few months, and that’s exactly what they are doing by completely overhauling how their pricing now works (it’s now essentially a “pay per token” model, aka boring). Again, this is all very typical, unsurprising strategy adopted by literally every subscription-enabled service since the dawn of time (see Spotify, Netflix as examples). Now that the rug is pulled so to speak, I cancelled my subscription, coming out ahead.
I was still able to fully leverage that opportunity, using free tokens to create a really cool (to me) local-first end to end fine-tuning project called Listenr that actually lets me build free-as-in-freedom ASR models without Microsoft at all. I learned a lot, I got something I think is actually useful and it cost me very little. If I were to have used my own GPU with worse models, I wouldn’t have gotten close to the output I got with BigTech compute. If I were to have paid retail for the tokens, it likely would’ve been way to expensive to justify.
It feels like we’re in this twilight zone where companies are still competing on AI offerings and willing to absorb a great deal of cost. But it’s also obvious that it won’t always be the case.