The emerging world of dual-sided AI
Regression to the algorithmic mean, or the world where algorithmic content generation meets algorithmic amplification and content curation
The Wapo published an incredible interactive depicting the “gargantuan amount of text, mostly scraped from the internet” that powers generative AI chatbots such as ChatGPT and large language models (LLMs), based on this fantastic paper.
What will be the use of this enormous resource? One immediate consequence will be the ubiquitous content industry. Kate Eichhorn says the quintessential example of content is the Instagram egg—an image that imparted no information or knowledge and circulated simply for the sake of circulation. Kyle Chayka, in a series of essays for The New Yorker has explored what it means that the Internet has turned us all into content machines.
Garbage in, garbage out may well become “average garbage forever”, transforming our aesthetics and consumptions habits forever.
The result may well be that algorithmic content creation and algorithmic recommendation will push us all into some sort of regression to the algorithmic mean.
How does this affect the future of work? We already have some idea about discriminatory pricing enabled by algorithms. However, as workers adapt to the new algorithmic realities of algorithmic pricing, we would see more and more strategic behavior by individuals (and perhaps algorithms competing with algorithms in the world of high frequency trading)
Let me illustrate this with two examples
The first example is from Gad Allon’s substack where he discusses his recent paper examining how gig economy workers may be strategic in their response to algorithmic pricing by the platform.
“Multi-homing refers to workers accepting jobs from more than one platform, e.g., driving for both Uber and Lyft on the same day. Repositioning involves workers choosing a new physical location to become eligible for more lucrative jobs, e.g., an Uber driver intentionally driving to the airport for their next pickup.
However, these decisions can be complex due to their anticipatory nature. Both multi-homing and repositioning involve trading off a worker's immediate earnings for future earnings that rely on the expectations of future demand.”
What, I cannot help but wonder, will the impact when gig economy workers can use sophisticated algorithms in making decisions about multi-homing? While most of them may not possess the means or anticipatory behavior, can we be moving to a world where a new generation of algorithms will assist gig workers?
This brings me to my next example: the world of high frequency trading
In the world of finance, high frequency trading has almost completely removed the human from the loop. However, trading algorithms that are intended to benefit from micro-seconds of arbitrage opportunities are especially vulnerable to adversarial attacks through strategically placed buy and sell requests.
What comes next? We already know that ChatBots are vulnerable to prompt injection attacks. A Stanford student demonstrated how we can use prompt injections to discover Bing Chat’s initial prompt. Prompt injection is a method that can circumvent previous instructions in a language model prompt and provide new ones in their place, which works as a sort of social engineering hack.
Some more thoughts on a recent episode where I was on the Hamilton Mann Conversation, discussing some of the implications of our new algorithmic era.
So what’s next? We are in a world of dual-sided AI characterized by regression to the algorithmic mean, algorithmic feedback loops and prompt injection attacks! Watch this space for more..