Good at your job? AI is coming for you.

New studies say AI makes crappy employees better. Also: ChatGPT finds its voice, Danish researchers keep horsing around, AI Coke chokes

What happens when you hire cyborgs and centaurs as consultants. Source: DALL-E.

For the last year or two, people like me have been raising the alarm about AI replacing humans for tasks previously considered the sole province of semi-intelligent bipeds. [1] The idea was that machines would take on the boring/repetitive/simplistic stuff done by less skilled humans, without impacting people with greater experience and more advanced skills.

Turns out we may have had that completely backwards. Two recent studies have shown how AI can teach lower level workers to perform at a level comparable to their more experienced co-workers. They did it by using AI to catalog the techniques of more effective humans and teach them to less skilled colleagues.

One study, performed by researchers at Stanford and MIT, looked at call center personnel working for a Fortune 500 software firm. It found that using AI to make suggestions on how to handle callers made the support folks 14 percent more productive, but almost all of that productivity gain went to people who were not great at their jobs.

The second study focused on consultants. Researchers at Harvard, Wharton, MIT, and the Boston Consulting Group split them into three groups: Consultants with no access to AI, some that used AI some of the time, and the third group completely automated. Not surprisingly, AI improved the less skilled consultants’ performance by 43 percent, vs 17 percent improvements for more talented carbon-based life forms. Per the study:

Further, our analysis shows the emergence of two distinctive patterns of successful AI use by humans along a spectrum of human-AI integration. One set of consultants acted as “Centaurs,” like the mythical halfhorse/half-human creature, dividing and delegating their solution-creation activities to the AI or to themselves. Another set of consultants acted more like “Cyborgs,” completely integrating their task flow with the AI and continually interacting with the technology.

What the studies suggest (but do not directly say) is that organizations can use AI to transfer the institutional knowledge of their most expensive employees to their cheapest ones. The implications of that are pretty obvious: “Congratulations, you’ve been named Employee of the Year! Now clean out your desk.”

Talk nerdy to me

It was only a matter of time before generative AI found its voice. The next time you curse at ChatGPT for making shit up, it might hear you and respond. Our nonhuman overlords at OpenAI are rolling out what nerds call a ‘multimodal interface’. You’ll be able to talk to ChatGPT and have it respond in one of five human-sounding voices (their names are Juniper, Sky, Cove, Ember, and Breeze – this AI clearly had hippie parents). You’ll also be able to upload images, use a drawing tool to circle part of the picture, and ask ChatGPT about it.

These features will be rolled out to paid accounts first, but will eventually find their way to the rest of us mere mortals. The good news? After AI puts you out of a job, you’ll be able to chat with Alexa, Siri, or Google Voice while waiting on the unemployment line.

Source: The brilliant Elle Cordova

Just say Neigh

If you’re looking to train a Large Language AI model, you need to start out with a Large amount of Language. In English, that’s pretty easy; the researchers who built ChatGPT and its LLM cousins went out and scraped a few billion words off the Interwebs. [2] When AI researchers in Denmark wanted to teach their AI models how to speak Danish, they had a problem: Because Denmark has much stricter copyright laws than other Western countries, much of its Web was legally unscrape-able. [3] So they used the largest source of public domain conversations they could find: heste-nettet.dk.

Per Bloomberg:

Heste-Nettet, which translates to “the horse net,” is a Danish web forum created in 1997 for equestrians, breeders and other equine enthusiasts to talk about horses. It also happened to be one of the first Danish forums on the internet, and the focus of its discussions soon expanded to much more than horses: relationship dilemmas, pediatrician recommendations, high-school math problems, how many minutes one should soft-boil an egg.

Because if anyone should know how to soft boil an egg, it would be this guy. Source: Daily Motion.

Apparently, Heste-Nettet is like Yahoo Answers + Quora + Wikipedia, with a healthy amount of Bojack Horseman thrown in. More than 20 percent of the Danish equivalent of ChatGPT comes from the horsey set. So if you happen to be visiting Copenhagen and ask the AI how to prepare stegt flæsk med persillesovs [4], don’t be surprised if the recipe contains a few bushels of alfalfa.

Things go better with… AI?

Leaping with both carbonated feet upon the technology bandwagon, Coca Cola asked generative AI to help design its latest ‘New Coke’ flavor, designed to predict what the sugary beverage will taste like in the year 3000. The new Coke Y3000 is now on the market. The reviews are in, and they’re not pretty:

“The drink slithers to the back of the mouth and sits there, broiling with a numb triviality across your tastebuds.” – Gizmodo

“There’s a tasty berry vanilla flavor at first, but the aftertaste sometimes reminds me of buttered popcorn jelly beans.” – The Impulsive Buy

“Are you ready for Coca-Cola Zero Sugar Y3000? Probably not; it is, rather predictably, revolting.” – The Mary Sue

Nothing like some carbonated popcorn jelly beans to wash down your vegan cheese insect burgers. I’m glad I won’t be around to experience the year 3000.

Would you try AI coke? Share your thoughts in the comments below.

[1] In this context, ‘semi-intelligent bipeds’ includes most humans and large ape species, but excludes politicians, Instagram influencers, and daytime talk-show hosts.

[2] Teaching an AI model to understand language by scraping the world wide web is a little like teaching it to cook by rummaging through a restaurant dumpster.

[3] “Scraping” involves downloading the source data for each web page, removing all the HTML code and weblinks, and saving what’s left in a large database that can be used to teach an AI model about word adjacency and sentence structure. One of the reasons most of the LLMs don’t reveal their source materials is that they really don’t want us to know how that sausage is made. (Another reason is why they’re being sued for using material that was allegedly protected under someone else’s copyrights.) As I’ve written previously, an alarming percentage of the material used to train these things is trash.

[4] Crispy pork with parsley sauce and potatoes, officially regarded as ‘the Danish national dish’.