Haunted by new AI predictions

Want to skip gloom? Jump down to the second story (about Microsoft)

In 2021 a researcher called Daniel Kokotajlo wrote some AI predictions of what would happen with the technology over the next five years. Within twelve months he’d been hired by Open AI.

If you’ve ever tried to make predictions you’ll know how humbling it can be, you either shoot for the stars and miss by whole galaxies or you play so safe that there’s no satisfaction in getting remotely close. Kokotaijlo’s predictions which pre-dated Chat GPT by 2 years, have proved to be astonishingly accurate.

Open AI has had such a soap opera style side drama that you possibly heard about Kokotajlo resigning halfway through last year. As he left he spoke out saying that the company wasn’t doing enough to keep its technology safe, you can see clear hints of what he thinks of Altman in some of the projections.

This month he published his next five years of predictions styled ‘AI 2027’. I’ve been haunted by them night and day since I read them.

The preceding week I’d been enchanted with a long post by economist Daniel Susskind which mused about what jobs humans could do once robots could do everything.

Susskind’s post concludes that there are some jobs where we will *prefer* a human (maybe like hairdressing or teaching), some jobs where we want a human making a *moral decision* (like a doctor deciding our cancer treatment) and a few jobs where a human is *just better* (like acting or performing). (Top tip, I printed it to PDF and loaded it into Eleven Reader text-to-speech app for free).

But Susskind’s work is rendered dusty in comparison to AI 2027. Daniel Kokotajlo (and his collaborators Scott Alexander, Thomas Larsen, Eli Lifland and Romeo Dean) take the jumping off point that the CEOs of the three biggest AI labs have all predicted that computers will surpass human intelligence within the next two or three years.

I’m pretty sure most of us hadn’t fully considered the implications of that milestone. We’re always reminded that the human mind struggles to understand exponential growth. I always think of the Wembley Stadium hosepipe example:

The watering system at Wembley develops a fault: in minute 1 one drop of water is released; minute 2, two drops, minute 3 four drops, and so on. Every minute the number of drops doubles. How long will it take to fill Wembley stadium?

The answer is that the stadium would overflow from the roof before half time, at 44 minutes.

But the power of AI won’t merely double, it will accelerate. In the forecasts Kokotajlo and co describe a projection for mid 2027:

[The unnamed leading AI firm] runs 200,000 Agent-3 copies in parallel, creating a workforce equivalent to 50,000 copies of the best human coder sped up by 30x.

Since the first arrival of Chat GPT we’ve been seeing a month-by-month improvement of the technology over the last 3 years. But that is already accelerating, by September 2027 the new predictions estimate that it will be delivering a year’s improvement every single week. It is way beyond the comprehension of those responsible, let alone the rest of us.

The forecasts end with two different scenarios. One where the technology continues unchecked. You’ll be astonished with the speed of what happens next, but let’s just say you shouldn’t plan a five year reunion with friends in 2030. The alternate scenario, that involves regulation, control and a very stressful stand off with China ends significantly more positively.

And by positively I mean the whole economy is dismantled by the arrival of robotic technology and the urgent implementation of a Universal Basic Income but we’re still alive.

Read more: NYT on AI 2027

It might seem a tough juxtaposition but the sense of scale of the story above might give you a better start point to reflect on what Microsoft is describing in its new Future of Work projections this week. Once you’ve contemplated that the scale of change that is coming is going to be seismic, now you can stop thinking about incremental shifts.

Alexia Cambon and team articulate a transformation from organisations being constructed around org charts, to being built around work flows. A movement away from long established functions, to something faster moving and more organic.

This isn’t just an updated version of PowerPoint. This is work reinvented. This chart starts to describe the stages. We’re already seeing stage 1 for firms who are open to it and it’s been proven by the recent P&G study.

Microsoft’s report recognises that not all organisations are ready for this and they describe the willing organisations as ‘Frontier Firms’. I’m really hoping I can go deep on this outstanding report with a conversation with someone from Microsoft this week.

The report has immense resonance with a brilliant conversation I had this week with Howard Lerman, the inventor of fascinating new platform Roam (coming as a podcast later this week).

Howard told me that for a long time we’ve fixated on ‘inbox zero’ but in fact we’re soon going to realise that the enlightened organisation of the future operates on ‘calendar zero’. Where there are no meetings on the calendar as we go into the week and that we work with agility to pull meetings together as we require.

Other headlines from the Microsoft report:

Slightly more old school but last week also saw the publication of the Gallup State of the Global Workplace report. UK employee engagement remains static at just 10% of the workforce. (Ireland is at 9% and France is at 8%, I always like to know we’re not the worst). Most of our engagement comes from our relationship with our manager (see last week’s newsletter and podcast about ‘mattering’ to explore this further).

Also notable:

UK workers rate their lives at 46 out of 100. (Finland is in first place with 81).
16% of Brits say they spent yesterday feeling anger for a lot of the day
26% were sad for a lot of the day yesterday (the UK was almost top in Europe here)
17% describe themselves as lonely
30% are actively looking for a new job