“Welcome To Costco, I Love You”. Idiocracy (2006)
It has been almost four years since ChatGPT was released to the public in November 2022.
For human development, 4 years is a substantial time period.
- High school is 4 years long, the formative years of teenage development.
- College is 4 years long, the socially accepted period of evolution from a dependent on society (student and learner) to a contributor to society (entering the workforce).
- At around 3 to 4 years old, babies begin to speak in complex sentences and tell stories.
Universities and employers are about to receive a shock to their system.
Not only will this new cohort of users expect AI to be easily accessible, but they will likely be in the top ~10% percentile of AI users (compared to their surrounding environment) because they have learned how to use AI as their daily tool for the last four years.
This fall, students starting university will have used AI throughout high school. This summer, graduates are entering the workforce with the habit of reaching for AI as their first-choice tool for most tasks and settings.
There are some risks to RSI (recursive self improvement), most specifically risks to our next generation’s ability to learn.
Risks that lie ahead: flashing warning lights
First, the following risk narratives are not inevitable, but they are increasingly probable scenarios. Obviously, there will always be a distribution of outcomes; it will not impact everyone uniformly. Some individuals and teams will find it very easy to adapt and overcome, while others will suffer the worst of all possible outcomes.
Unfettered dependence on AI
The timing is not great. With students still reeling from COVID-19 lockdowns and the pandemic’s effects, AI’s arrival compounds the symptoms of both phenomena. A year of remote learning and digital-only interaction trained students to rely more heavily on digital tools than previous cohorts did.
Now, it is easier than ever for students and workers to build stronger and stronger dependence on AI. Not only is AI literally trained to convince the user the outputs are useful, the users themselves are becoming more convinced the act of querying and retrieving information from AI systems is effectively learning and productivity.
Early warning signs are already appearing at top universities, and that is likely the tip of the iceberg that is to come around the world.
Avoidance of frictions in learning and risk-taking
Trial and error is the old cliche of learning. Each person has their own way of learning, through different mediums and cadences. Rote learning also has its place. Learning to learn is to find the least painful method for an individual to make mistakes through trial and error.
Universities and classrooms are mostly risk-free settings for most students to make these mistakes and develop their personal learning processes and habits.
Once in the workforce, workers have to balance the social requirements of their job with their functional requirements. In other words, making mistakes not only hurts the worker, it will likely hurt the hierarchy above and around you. It leads to less risk-taking behaviors.
Now, none of the above has changed with the introduction of AI. Learning how to learn and making mistakes along the way is a critical friction in obtaining mastery.
But AI has become a very easy escape-hatch to lean on to not make mistakes. Over time, this escape-hatch will bias users more and more toward avoiding the friction of learning. Eventually, it can lead to a loss in mastery learning overall.
Deepening generational angst
Incentives drive behaviors. In most of the working world, society is hierarchical and seniority based.
First, imagine a manager that has spent the last 15 years building their career and mastery of their job.
Now, a generation of entry-level employees claim to be able to use “AI” to do a vast portion of the job function.
At the same time, the manager is at a stage of their life where they prioritize security and not development. It is very costly to learn another new set of skills and tools, as they will make mistakes and look probably quite foolish (at the beginning).
To add insult to injury, this new generation of workers communicate vastly differently. The mediums and cadence of communication, more direct and immediate, come off as offensive.
Soon, the shared understanding starts to break down. Less communication leads to more information asymmetry and speculation.
Distrust breeds and deepens generational angst.
Humans is the original recursive self-learning species
We live in a world with a “country of geniuses in a data center”. This world is here. Today.
Yes, today’s models have intelligence that is jagged, and rough around the edges. Honestly, I think you can probably admit that about most actual human geniuses.
Abstractly, we need to think about how to train the next generation of models to also prioritize “learning outcomes” and “user skill mastery”, exactly alongside other outcomes that our models optimize towards today (i.e. usefulness, helpfulness, harmfulness, etc…).
Tactically, the only way models “learn” to achieve anything is via datasets. We have to invest in environments that demonstrate human skill mastery in action.
Some principles for the path forward:
- AI proficiency should be an expectation, treat it as a first-class skill and learning outcome
- There is a wrong way to use AI, be opinionated about this invariant in different settings and skill levels
- Bring the agent to class, that’s how we check the homework. The process is just as important as the final output.
What’s even better than a country of geniuses in a data center?
An actual country of geniuses running those data centers! Computers will get smarter; it’s on us to get smarter, too.
That’s a future worth building.