Here’s some solid advice in case you are thinking about building (or even purchasing) an AI product:

AI products are typically an AI model wrapped in some packaging software. You can improve their performance in two ways:

1. Through engineering effort: using domain knowledge to build constraints into the packaging software

2. Through better models: waiting for AI labs to release more capable models

You can pursue both paths, but here’s the crucial insight: as models improve, the value of engineering effort diminishes. Right now, there are huge gains to be made in building better packaging software, but only because current models make many mistakes. As models become more reliable, this will change. Eventually, you’ll just need to connect a model to a computer to solve most problems - no complex engineering required.

That last sentence says it all.

Link to article.

Sam Altman, OpenAI’s CEO, just posted on X that the company is losing money on their pricey $200/month/seat ChatGPT Pro subscription – apparently due to subscribers using the product more than anticipated.

This might come as a surprise to some or many – especially given that ChatGPT Pro is ten times as expensive as the standard subscription. We can safely assume that OpenAI is not making money on the standard plan – their losing money on the Pro plan is indicative of how expensive it is to run a frontier AI model and a good indicator of how unsustainable the current wave of frontier AI companies really is.

I guess “to be continued”…

Link to story.

What happens when you give people near-perfect information about the future (a proverbial “crystal ball”)? It turns out we still make bad decisions.

In November 2023, we ran an in-person, proctored experiment involving 118 young adults trained in finance. We called the experiment “The Crystal Ball Challenge.” We gave each participant $50 and the opportunity to grow that stake by trading in the S&P 500 index and 30-year US Treasury bonds with the information on the front page of the Wall Street Journal (WSJ) one day in advance, but with stock and bond price data blacked out. The game covered 15 days, one day for each year from 2008 to 2022. The players in the proctored experiment did not do very well, despite having the front page of the newspaper 36 hours ahead of time. About half of them lost money, and one in six actually went bust. The average payout was just $51.62 (a gain of just 3.2%), which is statistically indistinguishable from breaking even. The poor results were a product of: 1) not guessing the direction of stocks and bonds very well, and 2) poor trade-sizing.

Here is the full article – a good reminder that it is not enough to know the future… You still have to (consistently) make good decisions about it.