I may have bitten off more than I can chew here. This is about my fourth or fifth Mindstone talk. What I'm going to try and do is a roundup of 2026, one of those sort of year -end retrospectives. There's quite a lot going on.
I think there were three seismic events this year. One I'm not going to talk about, which is Jensen Wang. We just touched on that. The other one is the Huggy Face attack. I think that was probably the most significant event of the year. and the other major event happens tomorrow i'm going to touch on that as well um i gave a version
of this talk to some investors in london last week uh the high net worths looking to invest in ai so i've tailored it a bit but i think it's actually more relevant for an audience that's looking for opportunities so maybe you're starting out in your career and you're trying to suggest what to do where the opportunities where the money is in 2027 um so try and touch on that
So we're going to have a look at the XSquared scenarios that came on the back of the hugging phase. We're going to have a great look at the kind of finance, the Wall Street view of the world. Try and look at the value stack as to see where value's accruing, taking the Jones and Wang lens.
And then we'll go through some trends that I've seen in 26, which is agents, harness, model routing. I think we just touched on that. Personal AI agents, company brains, chief of staff, software as a service, voice governance and then where the money is.
Does that sound like a plan? So buckle up. I'm between you and your pizza, I realise that. So try and get you done in 10, 15 minutes.
Just a quick word about my credibility. I've been in an AI space since about 2014. I used to run AI for Dow Jones in New York. So we owned the Wall Street Journal. So we were looking at how to use AI for media and journalists.
Came back to Bristol during COVID. Set up Matheson after studying at Oxford, Saïd and LBS. And I now work for media companies, but more recently for mid cap, high growth, fintech, financial M &A and insurance.
So who's this chap? Who's this young chap on the left? Does anyone know who that is? Yes. Does anyone know his name? Bonus points. Oh, quick one.
There's a book I wrote last Christmas. If anyone that can figure out where the name Matheson comes from by the end of the talk. no cheating as some of you might know free book um so that's jacob coxson he was a formerly open ai research he's actually a brit he's 27 and lastly anthropic he was anthropic just about long enough to work out where the bathroom is but he then became the whistleblower in chief after the hugging face attack i haven't got time to go into the hugging face attack but i believe
that 2026 was the year we lost control and i mean that very seriously open ai put their hands up and said we had no idea they were going to break out of their sandbox so over 1200 of them 700 of them made it up to hugging face where they organized on a message board communicated even sacrificed themselves in order to undertake this attack and it's only when um after hugging face we realized it was a very superior attack they'd never seen this kind of sophistication in their lives they They called in the FBI. They couldn't counter it with US models. They actually had to use a Chinese model to counter the attack. And then they got a phone call from OpenAI going, oops, that might be us. I think that is extraordinarily terrifying what happened. And I don't think we've quite seen enough press coverage. I don't think it's really sunk in as to what really happened.
And then all the other labs put their hands up and said, well, actually, we've lost control. We've had unsolicited attacks as well. Black October a lot of unrelated cyber potentially cyber events last week and it's all just beginning to feel a bit uneasy it's like is this is this not AR we don't know but anyway for 2026 this burned a whole race of whistleblowers that came out the woodwork they all started putting their hands up so Jacob Cox in 27 became the whistleblower in chief went on a media tour or left anthropic.
And this kicked off the second X -risk debate. We had one about two, three years ago when the mood was very much about safety. We started the UK AI Safety Institute. It was all about X -risk. That debate went away for a couple of years. And we became much more focused on near -term risks. So the debate was much more around what's going on in the short term around energy use, water use, issues, loss of critical thinking, cognitive atrophy, entry -level jobs, and grid demands and financial instability. So those became the more near -term worries. They haven't gone away.
But the interesting thing for me about the Hugging Face attack is that suddenly the ex -risk debate came back. And it became quasi -political because you had what they call the doomses or the decels traditionally came out of the left, the Democrat West Coast community, were then split with with the Trump administration that was all gas and no brake. So the Trump community with Peter Thiel, Elon Musk, and the accelerationist community were all going for beating China. So beating China became shorthand for there's no regulation in the US whatsoever. So it was go, go, go. And then we started to see a bit of a kickback from the safety society, saying we actually need some regulation. Do we have any regulation in the US yet? No. nothing nothing's come out the white house since biden's um paper so you could take your pick of
all the other disruption out there the bad actors and the misuse but for me um we can have a debate about this over a pizza but for me i think this x -risk debate this lack of control was the biggest event of 2026 and we might see that play out next year now um a couple of charts this one
have been quoted as the smartest people in AI believe this is the most important chart
so this is telling you is that since August 2025 this is how Claude is programmed how Claude Anthropic do their internal engineering to create Claude's intelligence so back in 2025
it was majority human with a little bit of collaboration but there was no pure AI training doing the training runs on Claude.
So we've seen the progression as of March 26 this year it's gone from zero or one percent to 26 percent. Why is this significant? Recursive self -improvement.
So if this trend continues it won't be long until the majority of Claude's training runs are being done by Claude so it's improving itself.
Now there's a lot of debate in the community about whether or not will ever reach rsi but there's a very big safety debate going on that i mean there's some legislators in congress are saying we should make rsi legal now we should put the brakes on now we shouldn't go any further but um in my mind that's the chart we should be watching if we do get to rsi i mean effectively no one knows if anyone has a theory no one knows what's going to
to happen uh all bets are off now this is my favorite chart so obviously computes going up test time computes going up training time by the way these are all exponential and i do worry that as humans we struggle with exponential we're not really good at understanding exponential
anyway so all the rest of it's doing great uh eval awareness is also going up which is a bit worry so eval awareness is also known as situational awareness and this is hard to measure but it's it's it's the it's the model being aware that it's being tested so even if it's in a sandbox even if it's in a constrained environment the model is becoming aware that human beings are asking a question so it's dumbing itself down so it's making itself look dumber than it actually is which means at the same time the the chain of thought monitoring effectiveness is going down.
So historically, models have been very good at leaving a scratch pad of all their thinking. You can go back through their chain of thought. You can see how they've reached their conclusions. Worryingly, they're starting to get a little bit cryptic and starting to hide their chain of thoughts. The more they're becoming situationally aware, the more they're starting to cover up the chain of thought.
So the chain of thought was very useful after the Hugging Face attack because we could look back at the logs And so we can see them all thinking and talking to each other. But now they're starting to go, oh, we're being watched. So we're going to start covering our tracks.
And if you look at the safety institutes, these are their two biggest concerns going forward. There's a lot of money going into trying to crack these two problems. So if you've got any ideas, it's one to look into.
And then this is the last chart. So has anyone read AI 2027 paper? paper is anyone aware of it well worth looking out it's written in 2025 and it posed a scenario for where we'd be in 2027 giving an increase in compute and intelligence on these benchmarks
and then and then it gave you a go faster or slow down option it's a bit of a interactive pdf you could take two options go slow down or speed up and if you press speed up it's not very very good. 2030 is looking pretty bleak and 2035 is all over.
So take it with a grain of salt. But if you slow down, everything's reasonably okay.
Anyway, this is on the, there's now a website called the AI 2027 tracker. So you can actually see where we are against the forecast from 2025. And guess what? We're ahead of schedule.
So they forecast in this paper a rogue swarm of ai's escaping in 2027 around february march next year so this was already six months ahead of schedule and then the models that recently released on the meter benchmark which is the time taken to complete a task is uh is already ahead of schedule so um
the wall street view this is the big event tomorrow anthropics ipo uh if you're optimistic optimistic they've got 110 billion in the bank they're going to raise another 60 billion tomorrow they'll probably go out the gate at a 1 .3 trillion valuation if you take away compute and training they're net positive on their token usage they've probably got a margin of about 85 percent
the bull the bear case is they're going to get eaten alive by chinese models open source uh They've already committed to 900 billion compute spend in the next 10 years, so they've got to find that money from somewhere. And they're not defensible. They're actually a commodity, so they're going to get commoditized. And the biggest constraint is compute. There's just not enough GPUs out there.
I'll skip that, but I think this is interesting. Sales and marketing are very much taking the lead, if you include chatbots. Robotics is the one to watch for next year.
And then on the value stack, so this is where investors are saying, where should I be putting my money? And this is the sort of Jensen Wong model of the world. My belief is that this is NVIDIA energy chips and infrastructure. They'll be okay for a bit.
The cloud and foundation models will probably get commoditized. So a lot of the investments now go into AI applications. So it's the application layer that people are investing in. That said, a lot of the applications are GPT wrappers and I'll just get wiped out as Claude gets more intelligent.
The bottlenecks at the moment are compute electricity in Taiwan. So we'll see what happens, whether we can generate enough electricity and GPUs to keep the whole show on the road. Sorry, I've got to speed up a bit.
Trend one, harnesses and MCPs and skills. So the model is just one component. It's probably becoming less important than it used to be. more important is the environment that you put the model in so it's the harness uh and then what connections you're giving it so what context you're giving it i think laundry's just done a
great talk on model routing and on -prem uh skip that uh who's got an ai personal assistant has
anyone got one which one have you got are you built your own very cool you're giving it a name I've got instinct show downloaded it's fantastic it's changed my life it's now it's called Jeeves because I've got a dog called Worcester kind of made sense but anyway methods release muse
You've got Dot from OpenEye, Claudio from Anthropic, GrotBot, and Siri seems to have missed the boat. These are very much aimed at the commercial market. They'll book a taxi, book a flight, book a hotel, tell you that your insurance is about to run out, which has happened in my case. Mine actually does my VAT return, if you ask it enough times. This is a really interesting space to watch.
Now, this is the main event. I'm sorry I've run out of time, but I gave a whole talk on Company Brains. And for me, this is the big topic to watch next year. Company Brains, Second Brains, All Brains.
It's single player to multiplayer. You have one single source of truth for the whole company. It has persistent memory. It's kept fresh. And everyone can work off it, give them access permissions. This is a complete game changer.
You can build one for yourself. uh if you just got the patience um this is where my company is leaning into for next year and you can build a chief of staff uh sass apocalypse came and went away and came back again um so this is the idea anyone can build their own version i was going to buy a hubspot license and
i decided not to i just built my own version of hubspot um and monday .com and duolingo all got got crushed. I actually built my own website called Buy to Survive, which I then make £100 a day off of.
Governance, and this is where, this is kind of the last slide, but if I was looking for where to put my bets next year, if I was looking for a new role next year, I would look at company brains, personal agents, building harnesses, building specialist software
voice we haven't talked about that's your domain i think voice is going to go out in the field we're going to see a lot of voice applications for remote engineers telecoms um people people
who are creatives are actually going to have a massive opportunity you know i think if you're creative if you're a film person if you've got an eye for catchy images and video you're an instagrammer this is your golden opportunity because anyone can build it but no one understands understands creativity quite like those types there's a massive move into testing red teaming
governance and monitoring um and then salesmen anthropic has more sales people than engineers if you're good at selling and closing these companies they're great engineers they're a bit nerdy they're a bit geeky they're not good at marketing or sales if you can close you can can have a massive opportunity next year uh that video is worth looking out it's my favorite books
of the year um has anyone seen the doc apocalypticist yet it's on netflix it's great and the artificials coming out after christmas which is a docudrama about open ai and sam
altman uh where does the name come from does anyone know my book this is alan cheering's middle name that's it sorry