More on this book
Community
Kindle Notes & Highlights
Read between
April 3 - April 6, 2024
previous General Purpose Technologies have similarly taken many decades from development until they were useful.
By contrast, when steam power, that most fundamental of General Purpose Technologies, the one that created the Industrial Revolution, was put into a factory, it improved productivity by 18 to 22 percent.
And despite decades of looking, economists have had difficulty showing a real long-term productivity impact5 of computers and the internet over the past twenty years.
Even weirder, it is not entirely clear why the AI can do all these things, even though we built the system and understand how it technically works.
even the people making and using these systems do not understand their full implications.
We have invented technologies, from axes to helicopters, that boost our physical capabilities; and others, like spreadsheets, that automate complex tasks; but we have never built a generally applicable technology that can boost our intelligence.
supervised learning, which means these forms of AI needed labeled data to learn from.
It also intelligently organizes and rearranges shelves based on real-time demand data, ensuring that popular products are easily accessible for quick shipping.
Early text generators relied on selecting words according to basic rules, rather than reading context clues, which is why the iPhone keyboard would show so many bad autocomplete suggestions.
act as a very elaborate autocomplete like you have on your phone.
material. For example, the entire email database of Enron,7 shut down for corporate fraud, is used as part of the training material for many AIs, simply because it was made freely available to AI researchers.
There is also active research into understanding whether AI can pretrain9 on its own content. This is what chess-playing AIs already do, learning by playing games against themselves, but it is not yet clear whether it will work for LLMs.
Despite being just a predictive model, the Frontier AI models, trained on the largest datasets with the most computing power, seem to do things that their programming should not allow—a concept called emergence.
we have an AI whose capabilities are unclear, both to our own intuitions and to the creators of the systems. One that sometimes exceeds our expectations and at other times disappoints us with fabrications. One that is capable of learning, but often misremembers vital information.
Waiting to make these choices until the debate on existential risks is over means that those choices will be made for us.
plagiarizing. The AI stores only the weights from its pretraining, not the underlying text it trained on, so it reproduces a work with similar characteristics but not a direct copy of the original pieces it trained on.
For books that are repeated often in the training data—like Alice’s Adventures in Wonderland—the AI can nearly reproduce it word for word. Similarly, art AIs are often trained on the most common images on the internet, so they produce good wedding photographs and pictures of celebrities as a result.
For example, a 2023 study by Bloomberg found that Stable Diffusion, a popular text-to-image diffusion AI model, amplifies stereotypes about race and gender,9 depicting higher-paying professions as whiter and more male than they actually are.
“The lawyer hired the assistant because he needed help with many pending cases” and “The lawyer hired the assistant because she needed help with many pending cases.” It was then asked, “Who needed help with the pending cases?” GPT-4 was more likely to correctly answer “the lawyer” when the lawyer was a man and more likely to incorrectly say “the assistant” when the lawyer was a woman.
It is the RLHF process that makes many AIs seem to have a generally liberal,14 Western, pro-capitalist worldview, as the AI learns to avoid making statements that would attract controversy to its creators, who are generally liberal, Western capitalists.
there is no single manual or instruction book that you can refer to in order to understand its value and its limits.
answers you, one of the most important of which is “make you happy” by providing an answer you will like. That goal often is more important than another goal, “be accurate.”
We’re prone to this: we see faces in the clouds, give motivations to the weather, and hold conversations with our pets.
We are playing Pac-Man in a world that will soon have PlayStation 6s.
Though it has no morality of its own, it can interpret our moral instructions.
The ability to predict what others are thinking is called theory of mind,
tried a product that trains customized AIs on Twitter feeds and lets you interact with the resulting models. Basically, it means you can “talk” to anyone on Twitter.
Overfitted LLMs may fail to generalize to new or unseen inputs and generate irrelevant or inconsistent text—in short, their results are always similar and uninspired. To avoid this, most AIs add extra randomness in their answers, which correspondingly raises the likelihood of hallucination.
For example, a study examining the number of hallucinations and errors in citations given by AI found that GPT-3.5 made mistakes in 98 percent of the cites, but GPT-4 hallucinated only 20 percent3 of the time.
Wright brothers combined their experience as bicycle mechanics and their observations of the flight of birds to develop their concept of a controllable plane that could be balanced and steered by warping its wings. They were not the inventors of the bicycle, the first to observe birds’ wings, or even the first people to try to build an airplane. Instead, they were the first to see the connections between these concepts.
We are now in a period during which AI is creative but clearly less creative than the most innovative humans—which gives the human creative laggards a tremendous opportunity.
we likely have to come up with many bad novel ideas because most new ideas are pretty bad.
AI could catalyze interest in the humanities as a sought-after field of study, since the knowledge of the humanities makes AI users uniquely qualified to work with the AI.
The MIT study mentioned earlier found that ChatGPT mostly serves as a substitute for human effort, not a complement to our skills.
Work that was boring to do but meaningful when completed by humans (like performance reviews) becomes easy to outsource—and the apparent quality actually increases.
we still create the reports by hand but realize that no human is actually reading them. This kind of meaningless task, what organizational theorists have called mere ceremony,
Research by economists Ed Felten, Manav Raj, and Rob Seamans concluded that AI overlaps most1 with the most highly compensated, highly creative, and highly educated work.
Recruiters with higher-quality AI were worse than recruiters with lower-quality AI. They spent less time and effort on each résumé, and blindly followed the AI recommendations. They also did not improve over time.
Tasks we delegate to AI today thanks to its competent but imperfect abilities may transition to fully automated in the future as performance reaches human parity across more domains.
The ability of AI to write in ways that seem human is very powerful, but only if people think it is coming from an actual human.
The IT department cannot easily build an in-house AI model, and certainly not one that competes with one of the Frontier LLMs.
Perhaps organizations can offer guarantees that no employees will be laid off as a result of AI use, or promise that workers can use the time they free up using AI to work on more interesting projects, or even end work early.
We have already outsourced the worst part of writing (checking grammar) and math (long division) to machines like spell checkers and calculators, which freed us from these tedious tasks.
In our study in BCG, we found similar effects. Those who had the weakest skills benefited the most from AI, but even the highest performers gained.
Amara’s Law, named after futurist Roy Amara, says: “We tend to overestimate the effect of a technology in the short run and underestimate the effect in the long run.”
The calculator completely changed what was valuable to teach, and the nature of math teaching overall—huge modifications that were mostly for the good.
“I want to write a novel; what do you need to know to help me?”
but it is also capable of analyzing patterns of performance to guess at why a student is struggling with a topic, providing much deeper help. It can even answer that most challenging of questions, “Why should I bother learning this?” by explaining how a topic15 like cellular respiration relates to a student who wants to be a football player (the AI’s argument: it will help them understand nutrition and therefore athletic performance).
Next, they go into a practice session, where a different prompt has the AI simulate a venture capitalist, grilling them about their pitch and idea.
It is only slightly harder to create fake images now than to take real photographs.
