Introduction

machines will take over there's no space for humans in a world where intelligence is a product there will be a one person billion dollar company the age of hiring 10,000 employees to scale a company is

over human employees will soon be paid to hallucinate we will work in imaginary simulations that do ab absolutely

nothing let's play a game out of these three which of them do you think is the lie so can I get a show of hands if you think a is the LIE oh this is scary there's only three or four hands okay we got we've got some hope over in this corner if you think B is the LIE hold your hands up cool and finally C that's the lie a

few more of you interesting so I'll answer that question at the end let's

The Emergence of Intelligence as a Product

jump in so so intelligence is now a product we can uh call it via API we can

pay our 20 a month uh chat GPT license uh we can call on something that's as smart as us in specific tasks what does

Skills versus Intelligence

that mean for us so our plan is going to be looking at the state of play skills versus intelligence uh we'll have a

little look at a new order and some systems for scale and finally we'll briefly touch on where next and Jevon

Unleashed

The Current State of Digital Skills

so

93% that is the number of UK companies that currently have a digital skills Gap

11.5 trillion that's the uh amount of uh economic uh growth lost in g220 countries due to this skills Gap however

Transformers and the Evolution of AI

uh in the nick of time intelligence is now a product uh Transformers really started working at scale in 2020 uh 2 in

November um and we see this with opening ey anthropic Gemini it's as smart as us but it's only in some tasks and it's a

bit of an unknown problem uh to understand in Che it which tasks ahead of time it can or can't do and it's very good at convincing us when it doesn't know something and pretends that it does

if we look over time gpt3 it's the 5sec task it was amazing uh it was supervised it could write lmics uh about your nan

gp4 comes in can do the five minute task maybe some of your homework uh however it's still supervised and

it's expected GPT 5 uh whether it's the end of this year whether it's early next year it's going to be able to do something that amounts like the 5H hour

task unsupervised uh 01 their uh logic reasoning unit came out last week it starts to think before it answers it

jots down uh a plan of action and the idea is you can then throw a bigger weightier a smarter model at some of

those trunkable steps to really start to get towards doing doing uh an entire essay a body of research uh a financial model a pitch deck all link to link so

The Future of Work

Surfing the Jagged Frontier

what does this mean for work well it means surfing the jaged

frontier and at the moment no company is doing this well today the Jager Frontier

is a concept developed uh by Ethan mik if you don't follow him online uh he he

will uh do all of the the research you need to do in his own time uh ping it to you and you just have to read his tweets

uh and you'll be of 99% of people um inside the frontier of this draged

Efficiencies and Challenges

Frontier efficiencies are improved so we've got some tasks uh AB b c d maybe one's writing a document maybe one's uh

speaking with an external client if it's inside the frontier then efficiencies are improved we can automate it we can

pass uh to some form of AI whether it's generative or otherwise and if things are outside it's still difficult it's

challenging we've not quite uh improved General models to get there so how it going at the moment well 92% of Fortune

500 companies already use generative Ai and gpts but they treat it like a binary can it do X so if we go back uh to our

Jagged Frontier let's mock up a marketing Sprint maybe we need to sign off a brief um after that ideate uh some

new designs U maybe we need to test them after that uh we might want to analyze them pick a winner and run that campaign

for finally writing a report well the jagged Frontier may look a little bit like this we've got some stuff humans

can do and we've got stuff machines can do and in the middle uh Sumer collaboration so we may want a human to

sign off in that brief uh to set the alignment uh to decide what success looks like id8 new designs I think we've

just seen in that last talk go for it uh I want a thousand new designs uh testing them I want to have some say in that but

I'll I'll respect your analysis uh before analyzing picking winner the machine can do and maybe we can work

together on a report um at the moment is this happening in the workplace I don't think it is and there's some theories

AI as a Binary Tool

why firstly AI is not a binary can it do X can it do y instead actually it might be able to do 65% of Sprint X if uh it

can do X but can't do y it could do 82.3% of Tas y but with a human in the loop it will be able to do 100% And so

you end up with this idea of the human machine composite uh where if you have a human human it will normally always be

able to do the ask two people together uh high quality high cost a little bit slow if you get two uh machines and you

wanted to leave them for 6 to n months uh I think they'd start hallucinating pretty quickly perhaps after a couple of

hours the human machine together they start to perform to a better efficacy whilst also having speed and so it

becomes how do you pair a human and machine so going back to our state of play and the emerging uh uh uh where

we're going we've already got this and this is coming well then why can't companies today realize those promised

gains of speed efficiency and output

The New Order: Systems for Scale

that brings us on uh to thread two which is the New Order and finding systems for

The Daily Tasks of AI

scale so let's think about poor old AI what is getting asked in a day build me a financial model for my boss oh also analyze these six competitive detects

for my colleague uh a drawing of a cap would might be great and actually of course in the style of Salvador dely and Fin fin I also need to plan and book

that 4-we trip to Barbados I've always wanted to do these might take 5 to 10 seconds to write for each of them the AI is going to go away and do something

like this I've completed 38 micro actions uh I've booked flights for one uh but I've done so on your partner's birthday and I've sent 12 files

including the cat pick to your boss um it may then also ask you please respond to these 14 review points and each of these might take 20 to 30 minutes it's

clear it's very easy to prompt AI systems but it's hard to get the alignment between intention and action right and actually the more you prompt

the more work you're creating for yourself to get that right so uh we think fractal I should have said at the

The Problem with Individual Node Level AI Strategy

start an intro my name is Johnny uh run a company called fractal we're trying to solve what we call human machine teaming um we think the problem is the

individual node level uh everyone in an organization being given a gp4 license and that amounting to AI strategy uh the CEO or board crossing their

fingers and hoping people are more productive and of course there there'll be a a longtail distribution where 20 or 30% of your employees or your staff your

teammates uh every time they get an action they're intercepting their thoughts uh thinking what could I prompt what model could I use what tool and as

we know there's hundreds of of generative AI tools that they can use and then there's someone who's used it once and never again and we've seen that

with the usage figures from open AI so at this individual node level something like a human something like an AI um and they they want to have a chat they want

to do some stuff so I might say to AI do X and that might take me as mentioned 10 seconds to say the AI might say I have done X and let's pretend uh gbt 5's out

it's 2025 and that might take 30 seconds but it's producing a serious body of work I'm then left with the realization although this is a fantastic answer and

it's technically correct it's not quite correct for my company context so how long does this take to fix well 2 to 3 minutes now I might review some code I

might review uh some documentation uh and I'll be able to get to the the Crux of the matter and have a recursive conversation and get there but to review

something reprompt and then refine 5 hours worth of code or or writing or creative it's going to take me 20 20 to 30 minutes and now imagine this across

everybody in your organization uh everybody who's using it it becomes a bit of a mess and so it's a really interesting conclusion that in the age

of AI and intelligence as a product actually at the node level human latency will be the core blocker to the speed of human machine systems to the speed of uh

these new organizations we're going to be creating uh that potential one person billion dollar company or four person billion dollar company it will always be

The Need for New Tooling

able to go quicker uh dependent on whether the humans can be quicker so the inclusions we need new tooling we're going to be entering this new age

companies are starting to think about how to adopt it not at the node level but at the company level and so back to that question why is no company surfing

Overcoming Outdated Management Systems

Frontier well I'll put it to you that every person in this room uses an outdated technology and often daily that was designed in the 1940s and is about to completely break does anybody have any ideas of what that could be

keyboard mouse well if no one's surfing the jagged Frontier it's because we're all stuck in the 1940s we have invented the

first computer we've uh uh created nuclear power uh put a man on the moon the personal computer proliferated by Apple the worldwide web Bill Gates in his car using internet um the mobile phone and what that lent to web 2 and social networks we push forward we invent self-driving cars self-landing rockets and now we have created intelligence as a product and yet everyone in here at some point in the last month has used

this to do doing done um and listen I I hear some of you engineers in the room your flavor of to-do doing D may be blue right it may be jira but in effect it's still a c board and you can imagine some of these tasks maybe jir develops its product strategy and thinks some of these tasks machines can do that's fantastic we'll just get them to do it within this archaic structure someone in your team is going to have to review a new inbox of about 40 50 tasks that suddenly disappear from to-do to doing to review um and again that that's not going to create an increase in speed or a very little increase in speed at the

organizational level so why do uh existing systems of management of management not work well it was developed in the 1940s by to Toyota it was to run and manage old world factories not agents doing 5 hours worth of work in 30 seconds and hey we can tweak the variables we can say it's 1 hour's worth of work in 2 minutes the the effects still the same and this is coming next year so why do they break in

the age of AI we've got a new project and okay we're not in the 1940s I'm not

going to put up my Post-It notes but I am going to manually write out the 30 list of things that uh my team needs to do that manual forward planning uh

editing reviewing uh if I've run a Sprint before and I need to do it again it's very poor uh kind of huris rules if this than that if we've done this Sprint before maybe we can preload something

it's retrospective logging uh with looked at that human machine composite with machines in the loop everywhere everyone's going to be using a co-pilot um uh a model uh working with agents

you're going to be able to drain the company of data that you've not been able to do in a way before uh you think about robotic process automation RPA uh to do that well and to automate parts of

companies there's a huge uh process of uh process uh mining where you go into a company you try and see what's done you see whether it can be automated that will changes with machines and models in

the loop and so some of that respective uh logging will no longer work and it won't enable you to work with models and agents in a way that's efficient because again it's the human that's the blocker

so the tech models and agents will work I trust uh Google Microsoft and open AI not with everything but with uh making technological Pro uh progress but it's

the company structures that are going to fail so team management company management it breaks the onset of models and agents and so therefore weirdly uh AI efficacy is an organizational design

problem it's about how we work how we organize ourselves how we pass work from Human to human human to machine machine to human and actually a technical

The Organizational Design Problem

implementation problem second although that the Tex's going to get a lot better so the New Order and the system for scale is the

graph so back to our Jagged Frontier we've got a a micro project here it's not particularly complex we've got a a piece where we kind of want to P pass work from Human to machine and machine to human well maybe if we connected two two nodes uh or two tasks with a path and I want to go from A to B maybe from uh 2 to 3 3 to four maybe after for I want to go back and want to test some

more designs this starts to look like a structure called a graph uh all a graph is is it's relationships uh between pieces of data that's displayed in a more Dynamic flexible and elastic manner

uh than existing structures uh now this isn't meant to be a Pitch so we'll rattle through this but essentially this is this is what we're looking at doing

we want to build management tooling that's native to the intelligence age and I promise to our early design Partners is you want to double their capacity and output through really exploring that human machine system so

uh it looks like this it's not a cam board it's from a prompt so anything you want to do and there's be plenty of of people in the space where from a prompt let's build a new product feature let's tag Internal Documentation and systems

AI already uh can map that out into a graph Network whereby you've got dependencies I can't use uh uh my data science team to generate insights until actually I've got some of that usage data I've collected customer feed feedback and it's a much more Dynamic way of working and what's exciting is once you've got that relationship and once you understand the tasks and how they connect you can actually say these points I actually want a machine to do

and if it's too complex we can just click on something click on a node and split it and say I don't know how to do this make me another graph and what you'll start to get in this bottom right is a stack of graphs in 3D so previously if we had the 2D plane here you're going to get graphs breeding subgraphs breeding subgraphs until eventually uh they're very doable my machines and

obviously that uh Jagged Frontiers expanding research the competition generative AI can do that today evaluate product feedback generative AI can do that today until uh that Jagged Frontier starts to appear in your workflow and half of it will be done by machines but you're putting the human in the loop at the end for Quality so that they don't

hallucinate so why does the graph work well it enables you to automatically delegate work to models and agents for Speed but you're still in ently placing humans in the loop at the right place to ensure quality and that pipeline of human to machine machine to humans uh done a lot more uh manageably you've

also got adaptive human machine systems and that's what's going to enable scale

The Implications of Human-Machine Collaboration

in the intelligence age so the negative implications of this because with every positive there's there's negatives too

there's tradeoffs if I'm delegating half of my work to machines and uh my team have new capacity which means they can take on more work so as a company as a whole we can do more do we end up with

skills atrophy in that human machine pair they can do more quicker but the machine increasingly takes up the the body of that Workforce uh and the human just there to pass it onto the next

machine um to build resilience into systems in machine learning you think about this a lot it will come down to a a company's risk tolerance if you're a startup generate everything I don't care pass it to the model we've got nothing

to lose if you're deoe or PWC uh or you're the NHS and you've got um patients to look after actually that's not a risk we're willing to take and so

it will look something like in the future company setting a risk tolerance at the company level um and ultimately you may end up delegating and simulating work to ensure the human stays sharpen

this may sound ridiculous uh but it's it's where things are going if you don't do that uh we saw with the the air traffic control crisis I don't know whether anyone remembers in August

2023 the countries with the least Automation and autonomation they were happy as when the systems went down they continued as normal it was the countries

who've uh got very good uh setups networks Etc that they completely failed because humans didn't know how to do their jobs without those systems and so

as increasing amount of work gets done globally by machines uh that's something to to really watch out for and uh one possible fix is to delegate work even if

the machine can do it so where do

The Evolution of Company Management

company management go um well if we started here and 19 % of the world still is here which is management 1.0 it's 100% human it's 1950s it's can ban

boards management 2.0 okay assana Trello Monday jira uh 80% human 20% machine perhaps some heuristic rules codified in

software information's a product this is the SAS era and these are multi-billion dollar companies uh which jokingly

interly we just think they Post-it notes in a wall and they they won't scale with you we think where it's going is

management through 3.0 it's adaptive human machine systems a piece of work might need to be done 95% by human 5% by

machine or the inverse but you want that flexibility to decide at a company level so where next and finally to to

Jevons' Paradox and the Future of AI

round off jeevon Unleashed jeevon who's he well it's a paradox uh with all this

fancy words to say in the long run an increase in efficiency of resource use will always generate an increase in resource consumption not a decrease

which is a very grandiose way of saying with every technological Revolution it causes disruption and potential loss of jobs in the interim before ultimately creating more

think about the Industrial Revolution and arcrite and the Mills and uh the

phrase Lite we might poke at people at work and call them a lite it comes from people smashing Mills because they're taking our jobs uh how am I supposed to do my work and so in that uh interim

step uh you're losing work but obviously factories all around the world creating more work for people it creates

industry now if AI starts doing our work where do we go and does jeon's paradox hold up if AI can start to do slices of

our jobs if AI can start to be our colleague our teammate if we could have a team of 20 people and 100 machines uh taking on a a corporate of 500,000 people and over time that that eats into that uh what does that mean

The New Age of Leverage

well hopefully this will display a lovely graphic there we

go so back to our prompts um uh build a new product feature based in customer feedback and we've got our

graph now this graph consists of nodes and those nodes have paths uh if we go

back uh 200 years is those nodes would be uh people and those paths would be roads uh we'd be passing information uh

int Information Age comes along and we get computers for those nodes and those paths are emails well what happens when

all of those nodes are replaced and they're replaced uh all entirely by uh machines and the paths to and it's just

machines talking to machines what happens is it becomes 3D and the human will start to be the Observer of uh

systems and will start to exist on a level so if there's five people on a graph previously and that graph's now

five machines well that means one person's going to be running five machines but there's four more people to

run four more systems so ultimately it results in more um output so to

Conclusion

conclude what's the LIE machines will take over they won't we become observers of system and will enter a new uh age of

Leverage where people will be empowered to do uh far more than they're able to do uh now think of the creater economy

but for business one people running billion dollar companies and human employees being paid to hallucinate uh

we also think that would come true thank you so much

[Applause]