I first wrote about AI and its potential to augment human creativity a little over 18 months ago. It’s quite staggering how quickly things have evolved since then. The release this summer of Stable Diffusion felt like a watershed moment, a material advancement in both the power and the accessibility of AI tools.

But it’s worth thinking how we got here, and where we might go. Is the power of AI destined to be controlled by a handful of companies, or might it be more broadly available and shaped by more organisations?

This week’s article

Commodity AI

Is AI likely to be controlled by a tiny handful of companies, or will it be something accessible to the many?

Within the last couple of years, it’s become increasingly clear that useful AI is basically upon us. Not some general-purpose, smarter-than-human intelligence that would take over the world, but AI models that nevertheless can do useful work and make things easier, augmenting work that is still fundamentally directed by humans.

What hasn’t been clear, though, is exactly how the market for AI would shake out. Would we end up with AI controlled by a tiny number of companies who would rule the world? Or would AI be somewhat more democratised, accessible to the many rather than the few?

The three factors that mattered were:

  1. How much computing resource would be needed in order to train useful models?
  2. How much expertise and knowledge would be needed to design and fine-tune them?
  3. How much would having high-quality, proprietary data be an advantage, compared to data that was available to anyone?

The potential outcomes for AI, then, were likely to be one of two scenarios.

The first scenario was that successful and useful AI models would require enormous quantities of computing power, cutting-edge technical know-how, and access to proprietary data that was expensive and time-consuming to collect. In this scenario, AI would be something controlled by a handful of companies. That would likely mean Google and Facebook and a perhaps a few others – the companies that had access to the greatest computing power, developer resource, and proprietary data.

The second scenario was that AI models would have a much lower barrier to creation, requiring resources that were accessible to a far greater number of people and organisations, and usefully trainable on freely accessible data (like that found by crawling the public internet). In this scenario, it’s likely that AI would become something much more recentralised, not owned by any single company, like many past technologies. It might be like the idea of an SQL database, something that’s provided by many different competing providers and that many different businesses make direct use of.

What’s become clear in the last couple of months is that we’re headed to the second scenario. The release of Stable Diffusion and Midjourney have demonstrated that, on all three counts. The amount of compute power to generate the models costs in the region of the single digit millions of dollars – expensive, but well within the reach of countless businesses and even, in Stable Diffusion’s case, wealthy individuals. The technical knowledge necessary is in the public domain; the ML community has shown itself to be remarkably open, and knowledge is shared freely, much more akin to then open publishing of academia than the walled-up trade secrets of Silicon Valley. And finally, the efficacy that can be achieved by training models on large quantities of “dirty” data, such as that found on the public internet, is more than good enough to achieve remarkable results. It’s not necessary to have access to, say, Google’s level of proprietary data.

So, what does that mean? Well, a number of exciting and disruptive things:

  • AI is likely to be everywhere – not the foundation of a small number of platforms, but a feature within lots of applications and websites
  • AI is likely to be a commodity – low value and accessible to millions, rather than the preserve of a precious few or the basis of significant competitive advantage
  • AI is going to be specialised – honed to the needs of specific industries and sectors
  • AI is going to develop faster than we can imagine, with contributions from all over the world

AI’s effect on industry, then, is like to be more like the advent of computerisation and automation than some kind of sinister, dystopian, Skynet-like consolidation of power in a tiny number of companies. It’s going to disrupt everything – but it’s more likely to disrupt individual jobs and roles than whole companies or sectors. Just like we have fewer secretaries, hot metal typesetters, and clerks now that we have widespread word processors, desktop publishing systems and spreadsheets, we’re likely to have fewer illustrators, writers, designers, and animators. These are scary times, but what’s clear from this summer’s developments is that they’re no longer just around the corner. They’re here.

Click here to read the article »

This week’s three interesting links

Buckle up, it’s time to learn incorrect things about someone you’ve never heard of

There’s a particular phenomenon on Twitter: the “buckle up” thread, in which someone grossly simplifies a historical issue in strident terms. Rosa Lyster suggests these threads are typically:

“bewilderingly irate, laden with a combination of baroque linguistic flourishes and performatively subversive swearing, assumption of complete ignorance on the part of the audience, fondness for the word “gaslighting,” a powerful youth pastor-like eagerness to “meet people where they are,” high likelihood that it will be retweeted by people who refer to themselves as “Scolds” in their twitter bios, strong urge to lay the blame for the ills of the 21st century firmly at the foot of a basically random actor or event, total erasure of most things that have ever happened.”

The main problem, Lyster argues, is that these threads are strangely popular:

“The thing about Buckle Up Twitter, hard as this may be for right-thinking people like me to accept, is that a lot of other people LOVE IT. They absolutely love to be told that they are morons and that all of this is actually Beau Brummell’s doing.”


What are Executive Off-Sites Good For?

A characteristically thoughtful post from Tom Critchlow on the phenomenon of the “executive offsite” and the annual planning day, with observations about why they so often fail and suggestions of how they might be made to work more effectively. #

Sorry, we wrecked the economy

A compelling website that captures the unbridled joy that the UK’s shadowy right-wing think tanks expressed at Liz Truss and Kwasi Kwarteng’s “mini budget” a few weeks ago, before the subsequent crash. #