Big Tech has an AI crisis that needs journalist translators

#281. AIs don't like the way we humans write our stories, and that realisation is about to spawn a multi-billion-dollar new industry...

Image: the.factory.nft via Midjourney

A mate mailed me this morning with the weirdest coincidence. A scientist with my name Richard Sutton had just won the Turing Award.

It’s awarded for people who can make AI tech exhibit intelligent behaviour that’s indistinguishable from a human.

What’s even more weird is that Sutton and teammate Andrew Barto bagged the gong for reinforcement learning - something I used at my AI start-up Oovvuu a decade ago.

I called it Loop.

Think of it like evolution. Every living thing learns from its mistakes. Ouch, that’s hot and burns, let’s not touch that again, for example.

Sutton and Barto did this for computing - using positive or negative rewards to train AI machines to be smarter.

Oovvuu figured this out too, and we used humans to teach the machines how to optimise for the best outcomes, and learn.

Reading up on the award today - thanks for the heads up

Mark Pesce

- it triggered a memory of a post I wrote back in August 2023 predicting this future.

This is that post, updated for the developments since.

Welcome to Future Media and new subs overnight from the C-suite at Britain’s largest commercial news publisher Reach, The Times in London, Stanford University, Unicef, the Daily Mail, a sudden influx from Experian in New York, anti-piracy outfit ContentHound in Seattle, Deloitte Digital in Australia, the Menzies Research Centre, Issue Media in Detroit, Campaign in Hong Kong, Australia’s peak science body CSIRO, the New York Post, New York’s digital advertising club 212NYC, and media buyers Horizon Media in Los Angeles, among others.

And if you’re looking for an AI career pivot, I’ve got one. Let’s go…

---

Big Tech is discovering a new problem with AI, and it has the potential to unlock a multi-billion-dollar opportunity for journalists and editors.

Silicon Valley is learning that the way we humans create stories, videos and music is not in a format that AIs want to consume it.

And that’s proving costly and painful and expensive, especially at Google where they burned $70 million doing a deal with Reddit only to find it taught its AI Gemini nonsense.

The problem is that humans and technologies commonly view the same things very differently - like two people speaking different languages.

At Oovvuu, we unearthed the problem in 2017. We dubbed it the French / Swahili Conundrum.

A human naturally reads a story as saying:

Australia’s Matildas are through to the next round of the Women’s World Cup.

But an AI seeing the same words and pictures will interpret it as a string of numbers from which it has to infer a sense of its meaning.

It takes tons of computer processing to do what the human brain does naturally.

That’s why Big Tech is spending hundreds of billions and sucking up all our energy and water.

We needed to address this problem at Oovvuu because our mission was to automatically match articles with videos.

Articles are made of words and pictures, while videos are audio and visual. It means they were not the same. French and Swahili. Geddit?

We solved it with an AI we wrote called Compass. We translated words and pictures into strings of numbers using an early LLM. And we did the same with videos.

The outcome was that an article about Livrpool beating Newcastle in the English Premier League written on the day of the game would be converted into a number.

Let’s say it was 123 456 789 10. (The number was actually far more complex, 63 lines of 20-plus numbers.)

While a video about the game and produced within a few hours of the final whistle might be translated into, say 123 456 789.

The result was that article and the video were 99.9999999919 per cent similar to each other, so they were almost certain to be relevant to each other.

Therefore, the video should be embedded in the article. Voila. Job done!

But, if the video was a year-old clip of the same game last season, the identifier might be 123 567 9012.

That's 64 per cent different, so the conclusion was that the article and video were only peripherally related, and not contextually matched.

And the video should not be embedded.

It led us on a seven-year journey of mathematics, trial and error, journalistic judgement, and a scientific interpretation of the difference between explicit, and implicit context.

What we learned was what created Oovvuu’s AI superpower.

Big Tech is now discovering for itself the extent to which people and technology speak discernibly distinct languages.

And that’s terribly inconvenient for their explosive AI expansion plans.

Their challenge is immense.

They must reassemble a planet’s worth of information and repurpose it for a chat bot, and then stir in a cocktail of ever-changing demands from four billion people with individual likes and biases, all within a geopolitically charged environment.

But one thing I am certain of is that Big Tech will never get there if it cannot solve the fundamental issue of how differently people and machines behave.

This post proposes a model where publishing and tech can align and win together.