So apparently, employers are not working people. According to the wisdom of Treasury minister Emma Reynolds, employers are companies and working people are employees.
Can you really believe government leaders are saying things like this in today's age?
Minister for the Thick, more like it.
Speaking of intelligence, earlier today, a friend asked whether the frontier models are close to artificial general intelligence (AGI).
AI can make videos, slide decks, amazing marketing campaigns, production-ready apps and more.
But does it have AGI?
The answer is, it depends how you define it. For some, AGI means a machine that can do any thinking task a person can.
Others would argue it means it can handle the bulk of the work graduates and middle managers to do.
Demis Hassabis at Google DeepMind sets the bar higher still and wants it to produce new science and work in the physical world.
I don't think that answer is clear.
And it's also worth separating AGI and sentience.
Sentience is a machine with feelings. AGI is more about what a system can do.
A model could do most of the thinking work in your business without feeling a thing. There aren't reliable tests for machine consciousness yet, so the labs define AGI entirely by capability.
Testing
This alone is a cause for question - but in AGI testing there are questions over the reliability of answers too.
In September, OpenAI launched GPT-6 Astra and argued the AGI era has begun.
I thought that was punchy and premature - and it's important to keep your eyes on independent tests.
François Chollet (creator of ARC - an AI test outlined below) - defines AGI as a system that can learn new skills as efficiently as a person can.
ARC's test - ARC-AGI-3 - includes a set of small interactive puzzle games. When the ARC Prize Foundation launched it in March, human testers between them solved every game and frontier models scored under 1%.
Six months later Astra scored 99.9%.
But the people who built the test ran OpenAI's Astra on their own standard software, where it scored scored 62.7%.
It only reached 99.9% on OpenAI's own adapter, which lets the model keep its reasoning from one move to the next. In other words, it had a notepad to refer to for answers - and that's not how it works usually.
Google's Gemini 4 Argon did the same thing in October, scoring 49.8% on the standard version and 97.5% with the adapter.
ARC Prize called Astra a milestone but said it lacks the evidence to call it AGI, while OpenAI president Greg Brockman said that we are there.
Timing of comments
There is a sense of desperation in the AI industry to claim AGI is here to hold high valuations. It's those talking up AGI who are also the companies raising billions of dollars.
Hassabis has the best description I've heard for where we are now: he calls today's models jagged intelligence, brilliant at some things and oddly poor at others.
His biggest complaint is that models stop learning once they’ve been trained. A new employee gets better the longer they work with you. A model only gets better at the job if someone builds it a memory.
He puts AGI at around 2030 and thinks fixing that is one of the breakthroughs still to come.
So are we there?
A scientist would say we’re a few breakthroughs away. Yet - does that matter?
Your company needs to ask something simpler: can a machine do a big share of the work you want?
The models are already good enough for a lot of jobs. Whether they deliver depends on what you put around them - the data, memory, the checks on their work and the person who signs it off.
Personally, I’d swap the big AGI question for case-by-case queries.
- How long does it take a person?
- How would you know if the machine got it wrong?
- What does it cost you when it does?
What do you think?
Dan x
Cohorts
New: the AI leadership cohort
Your leadership team learns to run the business on an AI operating model over six weeks, with one live online session a week led by Ewan MacLeod and practitioners who do this work for a living.
Between sessions your team applies lessons each week to its own business, and in the final week you score your plan on the AI Strategy Triangle.
We are taking a small number of companies for the first cohort. Reply to this email or register your interest.
Courses
Our courses run in London and online.
- Working with AI agents, with Ewan MacLeod - 10 Nov, 18 Nov, 12 Jan - book
- Selling to the C-suite: Listen. Create. Engage. - 12 Nov, 14 Jan (London) - book
- A morning with Claude, with Ewan MacLeod - 17 Nov, 15 Dec (all online) - book
- Repositioning you - 17 Nov, 8 Dec (online) - book
- Build your message - 19 Nov, 11 Feb (London) - book
- Presence - 23 Nov, 10 Dec (London) - book
- AI governance - 26 Nov, 10 Dec (online) - book
- Automate your marketing - 26 Jan (London) - book
- Leading people and agents - 18 Mar (London) - book
See all the courses and dates here.
