Introduction

Hi, everyone. My name is Mark Riley. I'm CEO of Matheson AI, based out of Bristol.

Thank you all so much for coming on this warm August evening. It's great to see so many people here.

I believe this is my fifth MindStone talk. It's probably the most ambitious one I've done, so please bear with me.

I want to introduce you to the concept of company brains. brains.

And this is exciting to me and it's important to me because this is very current. This is kind of the holy grail of where AI is right now. This is where a lot of interest is in the entrepreneur and the VC community.

And importantly, I think it's also where AGI is going to be born in the next two or three years. I think it's going to come out of a company brain.

But I'll try and explain that more fully in a second.

And if you've seen the film Odyssey I've been on a bit of an odyssey myself because it was back in 2014 I was running innovation for Dow Jones a company in New York which owns the Wall Street Journal with my colleague Chris over here and I was

introduced to a start -up an AI start -up in 2014 called you know a guy called Ruggiero grammatica who came over from Italy was actually living in San Francisco, but he came over to challenge us using a product called Factiva. I don't know if anyone uses Factiva.

Factiva is the world's largest news aggregator. It's got about 23 ,000 sources. And we took two terabytes of Factiva, which in those days was a hard drive, and we overlaid some very rudimentary knowledge graphs.

And the challenge was to see if we could then start predicting supply chain risk, geopolitical risk, M &A activity, and stock market moves using very rudimentary traditional ai back in 2014 and the answer was yes but it was slightly ahead of its time and it was too expensive we ran into licensing issues so this question over company brains has been vexing me for at least 12 years now and it's pretty exciting for me to see

it finally coming to life um before we move on a quick challenge if anyone can guess where the name Matheson comes from by the end of my talk, if you've been before, you're excluded. You win a Matheson baseball hat.

Talk roadmap: what we’ll cover

What I'll try and do in 15 minutes max is just go through a few personas, who's this for, how we got here, why would you want one, should you build one yourself or should you buy one, and then after that I'm going to try and get even more

ambitious and see if we can move from where we are today to where we should be tomorrow and make them behave better.

We'll have a quick look at who's investing and what, and then we'll look at six reasons not to do this, and then a final thought.

Who is a company brain for? Five personas

So in terms of five personas, these are five potential use cases for this application.

We've got the tinkerer, the vibe coder, someone who's at home building product for themselves.

We've got the AI wizard, who's probably the guy, the girl in the office who's the resident expert, or it could be the CTO.

This is super relevant for AI consultants, especially FTEs, Forward Deployed Engineers.

It could be an AI entrepreneur who's looking for a new startup idea. This is highly relevant, or it could be for an investor or VC.

It's a quick survey. Who considers themselves a tinkerer or a vibe coder? Great.

Who's an AI wizard? Who's the local expert in the office? A few of you.

Who's an AI consultant? Wow. Wow. Competition.

Who's an AI entrepreneur? Who are the crazies? Who are the crazy entrepreneurs? You're nuts.

And then any investors here? Wow. Okay. And anyone, none of the above?

Anyone hasn't put their hand up yet? Gotcha. Cool.

The vision: what if your company could remember and learn?

So what if slide? What if your company knew where the scattered critical information was buried?

What if it became better at learning from past mistakes and poor decisions what if it hidden memories instincts and relationships became a source of intrinsic value what if it became smarter and self -improved overnight every night what if it could see around corners and start planning for future scenarios and what if there's one unified source of truth for everyone including agents now i'm going to pause there for a minute

because i fear i may have lost half my audience already because this sounds a bit sci -fi but what we're starting to do now is talk about companies as if they're individuals it's almost becoming anthropomorphic and i don't know if there's any fans of yuval noah harari here but he's keen on this topic whereby we've actually given companies legal entity rights already

companies exist as legal rights incorporated they have lawyers they have spokespeople they have leaders they have committees they have strategies and this is why i think once they they've started to deploy AI really thoughtfully in these company brains, this is where AGI will emanate from in the next two to three years.

Why now: the rise of “company brain” as a category

So a new category is taking shape. This is Google Trends from yesterday. You can see this is a search for company brain. It's up 1 ,250 % in the last 12 months.

We're not talking about a rudimentary, glorified, sophisticated search application here. This is something that can think in rationale, sort, and give context to information in real time. Four things happened this year in quick succession that caught my attention and really got the ball rolling on company brains.

So back in April, Andrej Kapathy, who was one of the founders of OpenAI, is now defected to Anthropic. open sourced on github and LLM maintained wiki that compounds it's the first person to do this come up with a concept that you can have a live wiki using markdown files updates in real time

and can draw knowledge and be kept current Jack Dorsey of Twitter fame and block fame came up with a very controversial blog in March suggesting that we could do away with hierarchies because the the only role of mid -management is really to pass information up and down the chain, a bit like how they did in Roman times with centurions, was his argument.

So once you have this all -knowing company brain sitting in an office, you could actually do away with middle management altogether. And then intelligence replaces it, replaces the hierarchy.

And then Gary Tan, the CEO of Y Combinator, came out with his G -Brain in April, which was the first attempt using open claw and a live wiki to create a company brain and then white combinator Tom Bloomfield announced the company brains became a startup category

this is the first time I really saw it on Instagram I saw this post on Instagram and I was quite taken by it I was also taken by the thousands of comments underneath everyone claiming they'd already got one which I thought was impressive this was obviously very very current so he said the biggest

What a company brain is (and how it works)

blocker to automation is no longer the models they've got so good quickly the blocker is not now the blocker is the domain knowledge it's not a company wide search or chat bot it's a living map of how a company works the company became the missing layer between raw company data and reliable information I think every company in the world is going to need one so what is it what it is is an

ingests of all the documents the context the knowledge you have in your business the more information the more context you can feed it the more intelligent it will become I don't know who imagine most of you are now using transcription tools but granola is personally I couldn't live without it's not granola

has now become incredibly powerful transcriptions become incredibly powerful way of feeding those machines and giving them information from every meeting and hang out and Google meet that you're on you take all that information into a context engine and this is where the magic happens because

it forms a effectively a knowledge graph over the top and then it learns what's relevant it connects the pieces of information together it synthesizes them and it's able to update them in real time and verify what's true and what's

not true then produces a company memory layer and then your customers your use cases are both the employees the AI agents and the workflows and most importantly they talk to each other so you've got a feedback loop that goes

back into the shared memory and you have something called a dream cycle so every night this machine gets smarter because they'll go off at night and read everything's happened in the company during the day all the documents they've been updated so when you wake up in the morning it's refreshing it's even

Why you’d want one: practical benefits

smarter than when you went to bed why do you want one onboarding is an obvious use case so you get a new colleague come on board they no longer have to go around asking awkward questions they can get up to speed very quickly you'll see

CEO is a key person who may, if they leave, devalue the company. So you may want to download all the knowledge from the CEO or the leadership team so the company survives when people leave.

Speed. Speed is super important. You can no longer get stuck in bureaucracy trying to justify decisions. You can make decisions much, much faster.

The knowledge compounds, the value of your company goes up, and you enable your agents to operate off a truth layer, a layer that agents can actually trust.

Build vs. buy: approaches and tooling

so how do we go about this well if you want to build basic one yourself good luck knock yourselves out this is how you can basically do it for the very beginners out there essentially if you do a shared clawed project you've effectively got yourselves a shared company brain

just using clawed projects well good old custom GPT's is not a bad solution what a lot of people are doing is using obsidian memory vaults which is essentially a collection of markdown files that are accessible to the agents and then you plug in your model of choice where it's claw gemini chat gpt you can run graphify over the top which is an

elementary knowledge graph and then you can use an agent like hermes which is effectively an open claw version it's an open source from the company uh is anyone any hermes fans in tonight yes it's coming hermes is one of those words once you've heard it once you'll hear it about five times tomorrow it's a very very cool open source agentic layer but then don't forget once you set this up

you're still responsible permissions provenance and maintenance if you don't want to build it yourself you can always go and buy one there's plenty of startups now that are doing fairly good

attempts at off -the -shelf solutions to glean any glean customers in the house i'd love to get your or view afterwards, Slites, a new one, known as actually set up by an ex -colleague of mine from Dow Jones called Neil Mann, and known as being very, very good at veracity, at testing the truth, trusted layer of knowledge within the knowledge, within the company brain.

Or there's plenty of open source solutions out there, so G -Brain, we talked about, Super Memories, an open source one, Obsidian, Graphify, and Hermes are all open source.

The trust problem: relevance, time, and “garbage in, garbage entrenched”

but here comes a bit of a pivot in the talk so far so good but we have a trust but verify problem with these company brains so i know someone who is a glean customer and they're asking questions

about hr policy and it started spitting out covid policies that he should wear a mask to to work tomorrow so these models these company brains are good but they haven't really got much for a sense of time. They don't necessarily understand relevancy today.

So garbage in, garbage entrenched. If you train the model on your inefficiencies and your idle chit -chat and it picks up bad habits, those bad habits will get ingrained into the company brain.

So what I'm going to try and do in... How long have I got? Two minutes. Try and suggest a fix for this.

And this is speculative. This is a hard problem. No one's really solved this yet, but this is one way of thinking about solving this garbage in garbage and trench problem

A proposed fix: retrospectives, hindsight, and temporal truth

What you need to do in effect is go back in time and train the company brain on what works and what didn't work So I'm calling this building the knowledge through retrospective layer giving it context, which is essentially what happened Why did we make those decisions which ones worked and which ones failed?

I call that the debrief and then through that it gains experience which I call hindsight and once it's got the hindsight it can move into foresight this is hard because you're going to have to go back through company experiences now there's plenty of books have been written about why

companies failed why people make bad decisions it's often sociological issues and personality personality issues. It could be political issues. We're human.

We fail. We're full of hubris. We're too adventurous.

We're not adventurous enough. We overestimate. We underestimate. We have poor data.

What you really need to do is go back in time and be forensic about the rationale for a decision that was taken. Most importantly, the bad ones, which takes a degree of humility.

Look at the rationale. What was the expected outcome? Review what happened, and then look at what actually

happened and then how would we improve you need to feed that experience into the machine I put in my speaker notes don't share too much this point but if you want to have a crack at this go for it this is a hard problem I put prototype than asterisks this is more of a hypothesis than

a prototype at this stage essentially you need to capture the experience and then you need to build an ontology and a temporal truth. What that means is you need to figure out what questions to

ask, how to rank those questions and importance, and then you need to timestamp those experiences so the model can go back. It's not confusing decisions with different decisions.

You need to store that experience, connect all the experiences together, retrieve, and then have a human overview.

And guess what? There's an AI solution for this. So you can probably use Hermes to repeat the exercise and get this machine working for you so it can train

Quick case study: applying retrospectives to M&A decisions

the model quick case study got a client of ours is an M &A boutique M &A they probably do three or four deals a year they want to look back and see which ones works which companies they found that were completed a deal which didn't

help we then forecast which companies to look at in the future so we do the capture we go back we look at which companies they looked at which decisions they made they wish bad decisions they made do a retrospective debrief on the

on the buyer behavior the deal team and the process and then they can test the hindsight once they captured all that in a bottle they can move forward and look with foresight which companies to go out and buy tomorrow there is money flowing

Market momentum: who’s investing

into this this is the hottest ticket in Silicon Valley right now Ngram is the leader right now they were in stealth till about a month ago and they announced they launched with

17 98 million dollars viven's another one i found uh interloom is hot for tracks and center i won't

Six reasons not to build one (and what to do about them)

go into them now but please feel free to check them out um six good reasons not to build one

so security and permissions people are very worried about permissions about the wrong data the wrong information getting to the wrong hands in the organization time and

token costs yes right now this will be expensive in token costs and it is expensive in the time to commit to build one of these technical complexity there's

still worries about hallucinations is my data ready the smaller younger companies are a massive advantage here if you work for a legacy institution with 20 years of disparate data over 20 silos that's going to be tough but if you're young and you're nimble, and you've got small data sets that are tightly packed, it's going to be much

easier. People are worried about disruption, like the argument that middle management is no longer needed.

How is this going to impact my org and my people? And then there is an argument, an interesting philosophical one. Is this too powerful for me?

Do I want something in the room that's going to be smarter than anyone else, including the board and the CEO in the room? So just to

Mitigations: permissions, costs, scope, and change management

mitigate you can give access permissions that commute from your previous permissions so if you weren't allowed to see it in Google Docs you shouldn't be allowed to see it in the Google brain there's lots of ways to mitigate token costs you start narrow set budgets use markdown files use model routers to find

the best cheapest model open source solutions or you just wait a year these costs will come down by a factor of 90 % in the next 12 months then 90 % the next

year after that to avoid the technical complexity try and rationale one set of decisions before you move on to the whole problem the whole company again with poor data readiness start narrow and then move broad and then if you're worried about disruption there's an argument that people remain

responsible for intuition relationships ethics novel situations and high stakes decisions just reassure your colleagues with that one and then the powerful one i just think we need to get get over it.

I think we've just got to accept this stuff's coming. Every company's going to have one. And if you don't want someone smarter than you in the room, then tough.

Sorry about that.

Conclusion: adding outside-world context to company memory

My final thoughts are going back to 2012, 2014, when I was looking at Factiva.

The one other thing I think companies are also missing right now is understanding the outside world.

So a bit of a hypothetical question. What would happen if you trained these models to understand the news and the financial data was going on when these decisions were made so you don't only look back at internally in the company you look back at what was going on in the international weather at the time and then you can really start to look forward with confidence I'm over

time thank you very much that's Mark Riley at Matheson