Mondays Aren't Just For People
Thank you for subscribing and being part of this growing community. There are 2030 people and 5 machines subscribed as of this morning. I appreciate the feedback everyone provides and the time you put into improving my content.
Yes, 5 machines are subscribed. There are probably a few more that I haven't found out yet. They are scrapers that feed LLMs. What's the goal?
I asked DALL-E2 for a "mother in the style of Vincent van Gogh," and it produced a compelling copy.
GPT-3 and other LLMs can do the same with blog posts. We are not far from "a blog post about strategy by Vin Vashishta."
Vincent van Gogh has about 900 works for DALL-E2 to learn from. I ran Fernando Botero with only around 300-400 pieces and got similar quality results. We can assume that with between 300-900 samples, a large-scale model will achieve a passable reproduction of the creator's work.
That's sobering. The US Supreme Court has ruled that scraping content from websites is legal. Subscribing and getting content via email makes it even more accessible and less prone to legal jeopardy.
Microsoft's Co-Pilot is trained with the open source library contained on GitHub. That sets a precedent for other models. Open source code and publicly available content have substantial overlaps from a legal perspective.
The models are capable. The data is available. The law provides no protection. This isn't a fictional scenario or something far off in the future. We will see people's work plagiarized at scale, and creators will have little defense against it.
I gave DALL-E2 "mother in the style of Banksy." The images lacked the same power and impact. The model could not reproduce Banksy's work's intensity and perspective-shifting aspects. Banksy has fewer works to sample, around 200, which plays a big part, but there's more.
Botero and van Gogh have both entered the art community. Their styles have been integrated by artists and imitated by students. The core elements of their genius have been studied and defined. Banksy has something novel that isn't as universally understood and copied.
Someday it will be. With time and examination, the patterns become clear to people and models. Banksy will be integrated into the next generation's art. DALL-E2's successor will have much more to learn from, and its reproductions will be more accurate.
The great creators from our past have sown their innovations into the generations that follow. We don't always see it, but most art is variations on themes. If content or paintings are simply a combination of existing styles, models will easily replicate them. The patterns are evident to a model trained on the works of the people who introduced them.
Creators will soon be challenged to create something novel or be replaced. I don't think many creators realize their work synthesizes several styles, but not a unique synthesis. We've been told that we all have a unique perspective, so it will be disheartening for many to discover they don't.
The technical tutorials will be quickly replaced by model-generated content. It will be far more comprehensive in sourcing and depth than people can create. It will aggregate thousands of posts to build educational material we don't have time to match. Explorations of previous works and summary research papers are far better handled by machines.
Models will recommend products in 5-7 years, and companies will realize their "big idea" is evident, not innovative. That will be a dose of harsh reality for several startup founders. Models in chemistry and biology are already having early success discovering what scientists have missed.
Working together, biologists and models are discovering novel protein structures. The complex rules are quickly learned, given enough data. Models excel where people fail.
We will be challenged to create something novel, combining past geniuses' patterns and adding our unique contribution. Models have yet to generalize beyond the themes they have learned from the data. The graph that connects each innovation to the next is still obscured.
The most exciting question for machine learning is, "Does innovation follow a pattern that can be learned?" I believe incremental innovations do. Remote controls for TVs have benefitted from several incremental innovations.
They began their lives connected to TVs by a wire and quickly became wireless. Some remotes now recognize speech. Buttons and layouts are much improved. Most remotes can be programmed to control multiple devices. There have been no groundbreaking innovations for remote controls, but they are better.
I think models can learn from the usability improvements across thousands of devices and find patterns. Incremental innovation doesn't have long to live as the exclusive domain of biological intelligence. Disruptive innovation is something else entirely.
The fire-starting leaps forward come from lived experiences and interactions with brilliant minds. Disruptive innovators learn differently, not from studying the past but by connecting what is to what's never been seen before. Their process must be forward-looking, and innovators unlearn common patterns to see something altogether new.
In the next decade, we will be challenged to define the boundaries between mechanical and biological intelligence. Some boundaries will be built for trust and safety. We will define what models should and should not do. Our differences will carve out other boundaries. Machine learning has the potential to teach us so much about ourselves.
One of the greatest lessons will be to reveal our value. Strip away the logical and semi-creative tasks. What's left is a realization of our potential. Driving a car to work is a waste of your abilities. Reading dozens of emails daily and attending meetings that should have been emails are too. We have not reached our potential because we are only as brilliant as our circumstances allow.
Machine learning will allow more human brilliance and potential to emerge. I don't see Utopia in our future. The change will be hard for many and impossible for most. On the other side of this upheaval is something better for us. I think we are about to go through something that future generations will look back on as a significant turning point for humanity and mechanical intelligence.
Welcome to my first 5 mechanical followers. I look forward to your feedback and seeing what you grow into.