If you think about the approximate lifecycle of AI that’s being built today, it goes something like this:
1. Write algorithms (eg, neural nets)
2. Scrape data (eg, text and images)
3. Train (1) algorithms on (2) scraped data to create models (eg, GPT-4, Stable Diffusion)
4. Use human feedback (eg, RLHF) to fine tune (3) models – including addition of explicit rules / handicaps to prevent abuse
5. Build products using those (4) fine tuned models – both end-user products (like MidJourney) and API endpoints (like OpenAI’s API)
6. Let users do things with the (5) products (eg, write essays, suggest code, translate languages). Inputs > Outputs
7. Users and AI owners then evaluate the (6) results against objectives like profitability, usefulness, controllability, etc. Based on these evaluations, steps (1) through (6) are further refined and rebuilt and improved
Each of those steps initially involved humans. Many humans doing many things. Humans wrote the math and code that went into the machine learning algorithms. Humans wrote the scripts and tools that scraped the data. Etc.
And very steadily, very incrementally, very interestingly, humans are automating and removing themselves from each of those steps.
AI agents are one example of this. Self-directed AI agents can take roughly defined goals and execute multi-step action plans, removing humans from steps (6) and (7).
Data scraping is mostly automated (2). And I think AI and automation can already do much of the cleaning and labeling (eg, ETL), in ways that are better cheaper faster than humans.
AI is being taught how to write and train its own algorithms (steps 1 and 3).
I’m not sure about state of AI art for steps (4) and (5). Step 4 (human feedback) seems hardest because, well, ipso facto. But there are early signs “human feedback” is not all that unique, whether using AI to generate synthetic data, or to perform tasks by “acting like humans” (eg, acting like a therapist), or labeling images, etc.
Step (5) is definitely within reach, given all the viral Twitter threads we’ve seen where AI can build websites and apps and debug code.
So eventually we’ll have AI that can do most if not all of steps 1-7. AI that can write itself, train itself, go out and do stuff in the world, evaluate how well it’s done, and then improve on all of the above. All at digital speed, scale, and with incremental costs falling to zero.
Truly something to behold. And in that world, where will humans be most useful, if anywhere?
Just a fascinating thought experiment, is all. 🧐🧐🧐
These times are only gettin’ weirder.