Latest / Agentic Conversations (formally mlops.community) / LLMs in Biomaterials Production // Pierre Salvy // #193

Transcript

LLMs in Biomaterials Production // Pierre Salvy // #193 · Agentic Conversations (formally mlops.community)

0:00
My name is Piano. I'm the head of engineering at
0:02
Cumbria and the amount of times in my life where I've ordered
0:07
something other than the double espresso is less than five.
0:10
What's up everybody? This is Dimitrios and you are
0:14
listening to another edition of the ML OPS Community Podcast.
0:17
Today I'm joined by another none other than my man Stephen.
0:21
Where you at bro? Hey, how you doing?
0:24
Yeah. Dude, so today was some sci-fi
0:28
type conversation because we got into biotech and I did not
0:31
realise the advances and the type of cool things that are
0:35
going on. Pierre was kind enough to break
0:38
down all of the different. Initiatives that his company is
0:43
doing and how it is helping move the ball forward.
0:48
What were some key takeaways from your side?
0:51
It was like, well, first, everything is related to beers.
0:54
That's like, that's my thing. We did keep coming back from
0:57
that, yeah. He just talked about ohh, this
0:59
is how you make that. It's just like beer, except you
1:01
change this molecule. And then suddenly you have jet
1:04
fuel, obviously, yeah, no, but the main one is like they have,
1:08
they do like some crazy things. And I actually I generalised
1:10
that, you know, you could actually create.
1:12
Teens based on like, you know, he talked about collaged and
1:15
like how they created out of the lab and I thought there was none
1:18
possible, you know and and it was like that's like that was
1:22
real sci-fi thing to me being like, Oh yeah, actually we can
1:25
create it, you know, and then use LMS for that.
1:27
Not only that, but how he has like an API that he hits to
1:30
order DNA. Yeah.
1:32
Ohh. Yeah, true.
1:33
And it's like then the week after they have DNA in the
1:35
office, you know, in the lab. I don't know.
1:37
OK, that's yeah. And he did mention, I think my
1:40
favourite thing was a lot of times.
1:42
Talk to industry professionals that are doing things and
1:46
they're constraints are hey, how do I serve models at low latency
1:50
or how do I serve thousands of models or how do I make sure
1:54
that these models are constantly retraining.
1:56
And I think in your scenario, what you're doing at your job is
2:01
an ML platform that services data scientists.
2:04
But the needs of these data scientists and the needs of your
2:08
platform that you've created are much different than the needs
2:12
that Pierre. Just talking about and the
2:13
platform that he created, because he is not necessarily
2:17
servicing data scientists, he's servicing researchers.
2:20
And there's different constraints that he was playing
2:23
with, which is fascinating to me because he doesn't care about,
2:26
hey, can I serve 1000 models a second or whatever it is QPS,
2:31
what's my QPS like? Billions.
2:32
Of just my type of token? What's my?
2:36
None of that matters. It's all out the window because
2:38
what she is really constrained on and what he mentioned was one
2:43
of the. These things, as these
2:44
experiments are very expensive, so he has to make damn well sure
2:50
that when they are going to do an experiment, there's no data
2:54
leakage. That was one huge thing that he
2:57
said. Like if I have a data pipeline,
2:58
that pipeline needs to be working because if all of a
3:01
sudden we lose some data to the cloud and we don't find it ever
3:08
again or it doesn't get through to the experiment, that
3:11
experiment has not been leveraged to its full potential.
3:15
And therefore, if we have to do it again, it is very expensive.
3:18
So that was interesting to think about from a business aspect of
3:21
business and ML type aspect. And then there was also this
3:26
idea at the end that he talked about how large language models
3:30
and protein AI models are so similar.
3:33
And that was fascinating too because he's like, yeah, you
3:36
know, like proteins are they consist of lots of letters
3:40
because that's usually how we identify letters.
3:43
And you know what else consists of lots of letters?
3:45
Nerds. And so look at that.
3:49
There are some definite overlaps there.
3:52
And he talked about how they're leveraging that and what they're
3:54
doing with their products. This was an awesome
3:58
conversation. Yeah, I'm so, I'm so hyped by
4:00
the conversation. Every time, you know, I I hear
4:02
him talk about the products they do and everything.
4:04
And I'm always like. While they actually do things,
4:07
you know, they like, they use AI and it's not like they actually
4:10
create products. They actually make change, you
4:12
know, And that's always fascinating.
4:14
So, yeah, and he's so passionate about it.
4:18
It's like they can talk about it for hours and hours.
4:21
I just like also passionate about beers.
4:24
That, I think, was my that's true.
4:26
Like my favourite part of this is just how stoked he is and how
4:30
he really understands the impact he's making by doing his job.
4:35
Because he gets tangible things that you can touch and you can
4:40
experiment and. Arguably they might make the
4:43
quality of your life better. He's dealing with cosmetics
4:46
right now, but as he mentioned, they have plans to go into other
4:51
areas and maybe maybe in the future as he said.
4:56
Make. Less harmful Plastic or plastic
5:00
that isn't plastic and that can just biodegrade, which would be
5:04
awesome for anyone who has. Been fed up with plastic, They
5:09
know. So anyway, dude, let's get into
5:11
it. If you are enjoying this episode
5:14
and you listen to it and you think, wow, that was awesome, we
5:16
have one ask for you. And do you know what that one
5:19
ask is? Steven.
5:21
I think it's shared to one friend.
5:23
Just one friend. That's all we ask.
5:25
Before we jump into this, let's hear a quick blurb from our
5:29
sponsors. Alright, let's take a minute to
5:31
talk about our sponsors weights and biases.
5:34
It seems like everyone's working on large language models right
5:37
now, but understanding what to build with LMS and how they work
5:42
under the hood can be tricky. Thankfully, weights and biases
5:46
can help. You might know about weights and
5:48
biases. Flagship experiment tracking
5:51
product in the LM space. Teams like Open AI and Cohere
5:55
rely on weights and biases to train and debug their models.
5:59
But weights and biases. There's plenty more to offer
6:03
you. If you're thinking about
6:05
training, fine tuning or prompt engineering in LLM on your own
6:09
Weights and Biases Traces lets you uncover granular insights
6:14
about your LLM's, providing a visual representation of their
6:17
execution flows, colour coded by each component in the chain.
6:22
This helps you understand each step your LM takes on its way to
6:27
generating its output, giving you a deep, nuanced view of its
6:30
behaviours and letting you iterate and debug much more.
6:34
Quickly Weights and Biases LLM production monitoring enables
6:39
the creation of custom dashboards to monitor your LLM's
6:43
activities in real time, including throughput, token
6:48
usage, cost, latency, and errors.
6:51
That sounds so good, and if you need help getting started with
6:55
weights and biases, they can help too.
6:57
On that front, they've released multiple free classes for
7:01
machine learning communities like ours, the ML OPS.
7:05
Community, teaching them how to build LM apps intelligently,
7:10
approaching prompt engineering, and a whole lot more.
7:15
Whether you're a seasoned pro or just getting started on your LM
7:19
journey, you can find what you need at Weights and Biases.
7:23
You can get started today going to one MDB dot ME/ML OPS,
7:30
underscore traces, or free today.
7:34
Get after it. 1B dot ME/ML OPS underscore traces.
7:42
Let's get back into the podcast. see you've got a lot of ski
7:53
poles and skis in the background.
7:56
What's up with that? Where's the story there?
7:59
Actually, before living in Berlin, I used to live in
8:01
Switzerland. I was finishing my PhD, and so
8:03
that's where I actually learned skiing, I think.
8:07
Like, despite the fact that I left the country, I didn't leave
8:09
the passion, so I just brought this case with me.
8:12
Yeah, and honestly, I've taken them out once only.
8:17
They're basically desert decoration.
8:19
No, exactly. I'm guessing.
8:21
Yeah. Which ski mountains did you go
8:23
to? Not supposed to do any
8:27
advertisement for any resort, right?
8:28
But like the General, Swiss Alps are generally a nice place to
8:31
be. I've heard the French Alps are
8:34
nice too. Actually, the one place that I
8:36
will never forget about in my life when it comes to skiing is
8:42
Shamani. Yeah and the valley.
8:44
Ohh man. Very offensive defensive
8:47
managers going to. Shadow hanging around my thing?
8:53
Yeah, no, the the French shops. Also nice, but this was helpful
8:55
like literally 19 minutes away from my door, right?
8:58
So. Wolf.
8:59
Wow, that is spoiled dude. That is really spoiled.
9:03
So we started off this call with me hearing you to speaking in
9:08
French and I said I know a little French, I know how to say
9:10
it's just we false date. Will one of you French speakers
9:14
explain what the hell that means?
9:15
Cause I just keep saying it and people look at me weird.
9:19
Yeah, I can imagine thing that your French with like such a
9:23
strong accent, I think is probably like me claiming that
9:26
I'm British with my French accent.
9:28
So. But I'd like what you said just
9:30
means that yeah, you probably like high.
9:33
Basically, ohh, this was supposed to be too much or
9:35
something and then it's like, yeah.
9:37
It's smoked too much, Main. In some of the, you know, the
9:41
top of the mountain. That's it.
9:43
Yeah, there we go. I'm high skiing, I'm skiing
9:46
high, but alright, so we didn't come here to better our French
9:50
skills, or at least I didn't. You guys don't need that.
9:54
We came here because I want to know all about your story and
9:57
what you're doing as the head of engineering.
9:58
It came in and I also would like to just start with your
10:02
trajectory. You went to school.
10:06
At some point in your life, you got out of school, you started
10:09
working. Tell us about that.
10:11
Yeah. And that was already a pretty
10:13
long part of my life, right? I think let's start with the
10:16
basics. I was.
10:17
I think it was born in big fat nerd.
10:21
Nothing about science since like forever since the first memories
10:24
I have. So yeah, I studied classical
10:29
scientific trajectory influence. We have this system called
10:32
consent code, right? It's like basically fancy
10:35
engineering schools. So I went through which is like
10:39
basically preparation for competitive exam throughout
10:42
France. And then depending on your
10:44
ranking you can get into different schools.
10:46
I landed into the school of mines in Paris, which is
10:49
actually not just about digging holes in the ground, right, but
10:53
actually pretty good school for like applied physics and applied
10:55
mathematics. So I'd say like maybe the the
10:59
fourth first years after my high school, I just spent a lot of
11:03
time studying pen mathematics and applied physics and anything
11:06
from like material science to nuclear physics to
11:10
biotechnology, economics and stuff like this.
11:14
The whole day was just like playing math too, but of
11:16
different things, which made me very happy.
11:19
And then basically between my two years and masters I went to
11:23
the US and walked is total new energies at a company called
11:28
Amyris was basically making jet fuel out of genetically modified
11:32
bias thread. So process was pretty fucking
11:35
called. You take the same process as
11:37
making VR, except that the yeast that you have it removed the PR
11:41
making part and you are the making part.
11:46
We actually managed to make it work to like, you know, like we
11:50
literally producing tonnes of quantities of this, right?
11:52
And my role in this gigantic project was only to make
11:56
mathematical models of how this worked, right.
11:59
So it's like, yeah, I think I like biotechnology.
12:02
I think I like it because it's really complicated science.
12:05
Math never works. We want to play through it.
12:08
So I want to do a bit more of that, right.
12:10
And then I went to Switzerland, to EPFL to basically get a PhD.
12:15
Think that Fields and I studied there for like, beaches and four
12:20
years, got my PhD. And then basically think two
12:24
days before defending, There's like these random dudes eating
12:28
me up and LinkedIn, like, hey, we like what you're doing.
12:30
Do you wanna tell us about it? And, you know, an excitable
12:33
scientist and just like, yeah, sure, I'm going to tell you
12:35
about Cheers. Yeah, turns out it was the first
12:38
phases of an interview process. And so I defend my PhD.
12:43
And then they called me like, hey, do you want to come to
12:45
Berlin to see if we could work together?
12:48
And at that point I'm like, well, I just defending my PhD,
12:50
so I might as well just, you know, travel a bit and made you.
12:54
This was like in between COVID, August 2028.
12:58
So I go to Berlin, meet these people and you know, just like
13:02
nice five fast thinking people that obviously don't know as
13:06
much as I do on my particular topic, Right.
13:08
But always ask the right questions.
13:11
And I was like, you know what, maybe it's just sleep suits and
13:14
then come to build and try to make something with them, right?
13:16
So that's how I how I was hired as entry #2 at Candriam with at
13:22
the time, the title of the head of computational biology.
13:25
And it was just like 5-6 people sleeping bat cave trying to
13:29
figure out what we could do with with our brain and our
13:32
technology, right. So that's how it all started.
13:35
Right. So you were hired as like
13:36
employee number two then, right? And so then how did you go from,
13:41
like, being in prayer #2? Hiring A-Team as well and being
13:45
able to to produce actually a real product.
13:48
I think the, the way the founders wanted to hire was
13:53
initially to get basically the leads of the different
13:55
departments, right. And then depending on like at
13:57
the time we didn't run product right, we did not even really
14:00
know what we wanted to do. We just knew that we wanted to
14:03
do something sustainable related to biotech and AI, right.
14:06
And so they just hired a bunch of people that basically could
14:09
become heads of departments in the course trajectories.
14:12
And then naturally like as our use case clarified as what we
14:17
wanted to do became more clear. And then actually we all hired
14:21
our our different teams, right, and different departments.
14:24
So I'm the head of engineering. So that means we do in
14:28
everything in the related computers and robots.
14:30
And then there's you seeing the head of RE, there's like actual,
14:35
that's actual biology in the lab, right?
14:37
And of course you have like the finance and the step and the
14:40
sales and so on. So talk to me about what?
14:45
Your evolution had to look like going from employee number two
14:49
to now head of engineering. And how that?
14:54
Like what traits did you have to learn?
14:57
What is different about your day now than it was then?
15:03
Yeah, that's a that's a lot of questions to impact.
15:05
I will start with a quote that Rubin, my my colleague likes to
15:11
remind me is like, I think one day I was getting a bit angry
15:14
and I was like Ruben, I study computers that people I don't
15:17
know how to handle that, right. And so I think that's a good
15:20
summary of like 12 things I had to learn, which was like when
15:23
you have, when you're a manager, you need to understand how
15:25
people work and you need to understand what are good ways to
15:28
react to situations that you might not be prepared, right.
15:31
And as much as I had some some aspects, elements of the
15:35
training in my engineering school, it's always different
15:38
when you actually have, you know, people working in your
15:41
team, right. So I think that's definitely the
15:44
unexpected aspects of things that I learned, but I'm probably
15:48
also the one I'm most thankful for it.
15:50
Yeah. Are there things that you
15:52
specifically like? I I notice with myself and
15:56
having to. Lead teams.
15:59
One thing I had to get good at was having hard conversations
16:02
and that I don't necessarily like to do because I'm all
16:06
peace, love and tie dye and. Then when you do have to have
16:12
these hard conversations, it's like you cannot really, at least
16:15
in my experience, I've found that I can't grow without saying
16:19
things that are uncomfortable. And if I don't say the things
16:24
that I want to say, it eats at me later on and then it just
16:28
grows. And so bringing these things up,
16:31
I've learned more and more. When I'm with teams, I need to
16:35
bring them up right when they happen.
16:36
Otherwise, later on it's going to just fester.
16:40
And get bigger and become much more overblown than it
16:44
necessarily needs to. Are there other tricks that
16:47
you've learned as it comes to leading teams and leading
16:51
especially engineers in your in your tenure?
16:55
Yeah, I mean the point you made about like addressing problems
16:57
earlier, I think it's definitely one of the most important.
17:00
I also really dislike an easy conversation thread.
17:05
But as you said, sometimes you need to have them.
17:07
And I think the best way to have them is to make sure that the
17:12
person you're gonna have them with is following the same
17:14
journey as you have been following, right?
17:16
Like, at the end of the day, you're working with these
17:18
people, you've hired these people.
17:20
It's for a reason, right? So to some extent, you can trust
17:23
their brain power, their understanding ability.
17:26
And so I think what I always try to do is if you're not agreeing
17:30
on the situation and we need to have this conversation, what is
17:33
it that I do not understand from your processes and what is it
17:37
that you do not understand from my thought processes?
17:40
And once, and it's really nice because it's a very analytical
17:42
problem, so there's like no emotions involved there.
17:45
Everybody just lays, they're called on the table.
17:48
Everybody looks at what's there and like, OK, this is what is
17:51
disconnected and this is when we need to agree on, right.
17:54
And I think so far this, this allowed me to basically handle
17:59
most of the difficult problems we've had come here.
18:02
And because you know, it's a startup, so of course you're
18:04
going to have like stressful moments.
18:05
You're going to have high stakes decisions, you're going to make
18:09
mistakes. You always have these sort of
18:11
situations that pop up and these other situations around which
18:14
nuclear, you should need difficult conversations, right?
18:17
But again, I think the first thing to do is you have to trust
18:20
your hire. That person is for a reason that
18:22
clever. You have to understand what it
18:24
is that you don't understand from their brain and vice versa.
18:27
There is something that for me it feels like.
18:32
When I'm talking to you and just hearing your experience, you
18:37
seem to have. A lot of interest and a lot of
18:40
curiosity on many different topics and so I would love to
18:45
know how it went from. Like being in.
18:51
An academic situation where you're really studying in depth
18:56
biotech type things or just biological things and then being
19:01
more on the engineering side and needing to learn about.
19:05
The computers and I know you've studied that also, but it feels
19:08
like. What are some of the learnings
19:11
that you've had since you've been in industry that you?
19:16
Can can relay. Also, I think there's the
19:20
general question of and, and I think it's also like a stressful
19:24
question when you come out of academia of how do you stay up
19:28
to date with what's happening, right?
19:31
And especially not necessarily like I didn't.
19:33
But basically when when you come from a place where you've been
19:37
always at the bleeding edge of what technology can do or what
19:41
science can do or whatever, right?
19:42
Like moving back to a position where you have to manage
19:45
projects and talk to people and and do podcasts, right?
19:49
All of a sudden you don't have the time to stay technically up
19:52
to date with technology. How to handle this?
19:57
I think for me it was a combo of.
20:01
Still spending some time, you know, getting my hands dirty,
20:03
right? Like, you know, when people talk
20:04
about LMS and so on. Like I am happy that I took the
20:09
time to just like lay with long chain plays with different
20:12
backends and see what works. See what doesn't find hilarious
20:15
failure modes where like charged with this convincing me that it
20:17
was alive or something. You know you have to interact
20:20
with the technology. To to kill touch, but you don't
20:25
have to be playing with it every day to stay at the bleeding
20:28
edge. You understand what I mean?
20:29
And so that's element #1 #2 is like, you don't have to yourself
20:33
be at the bleeding edge if you trust the people you hire to be
20:36
themselves at the bleeding edge, right?
20:39
I think that's something I'm super thankful for is like, my
20:42
team is absolutely fucking insane.
20:45
They're really, really good at everything that to him.
20:47
And so for me, I'm like, even if I spent like 10 additional hours
20:52
a week trying to learn what they're doing, I could never get
20:54
to like 50% of what they're doing.
20:57
So, So you know, it's not, it's not even an opportunity.
20:59
I'm just like and let the expert do the exact things and I'll do
21:03
what I know what to do, which is let's talk at podcast and and
21:07
escalate conflict. I mean, yeah, for sure you can
21:11
talk at podcast and everything, but I kind of feel like as well,
21:13
you're still very technical as well at the same time, right.
21:16
You like you, you're like head of engineering, so handsome
21:19
manager. But also I feel like you still
21:21
want to get your hands dirty as you said and you played around
21:25
with their limbs a lot. So, so like are you the one that
21:29
would you like for example, pushed for the use of Element
21:31
Cambrian or how did that came to happen as well?
21:34
Can I actually talk about that other story before answering to
21:36
your question? Yeah, sure, It happened that.
21:40
I mean, it's true that I've actively get my hands dirty and
21:43
there's like one super cool thing that we developed which is
21:47
hard to pronounce, but some super core which is a protein
21:49
programming language. And I can talk about this more
21:52
details later on. But basically I think this is
21:54
one of my earliest contributions.
21:55
It was like we're going to design proteins.
21:58
There is no constitutive way to design proteins that you know is
22:04
permanent. Makes sense for everyone.
22:06
It's always like a little bit like wishy washy magic as I
22:10
could. You know what's a good way to
22:12
blend natural language and computer language in a non issue
22:16
magic programming languages, right?
22:18
And so we developed like the first version of this protein
22:21
programming, which is like a proper compiled programming
22:25
languages showing complete and so on.
22:27
And with it you can specify your proteins and so on.
22:30
Nailed me. Was really happy about this.
22:32
Then it evolved in many different ways when I gave it
22:34
into the hands of my engineers. I think this is, yeah, an
22:37
example of how in small scale startups like like ours, right?
22:43
Like an engineering manager can still have an impact in shaping
22:46
the direction of where the technology should go.
22:49
It's basically a often built MVP.
22:51
Some of them die, some of them you know, machine like and
22:55
basically develop them right? What was the question I was
22:59
supposed to answer with this non answer?
23:01
It was. It was around the use of LMS and
23:04
was that one of your MP's that you're playing around with?
23:07
Yes. So I think it's yet another
23:09
example of a beautiful discussion between two months
23:13
that did not agree. So Charlie's my Chief Scientific
23:18
Officer. So he's the guy who is setting
23:21
the vision. And I'm the guy who's trying to
23:23
execute. And there's like this natural
23:25
dielectric that happens in between the two.
23:27
Right. And I love it.
23:28
So the first time he told me, yeah, we should look into Geneva
23:31
and so on and so on. And I was like, oh, yeah, yeah,
23:35
height. I don't have time for height.
23:39
But then again, what is it that he sees that I do not?
23:43
And we took the conversation a bit further and then we saw
23:46
pretty obvious use cases like 3 or 4 projects that came from,
23:50
from, from this initial discussion, right.
23:52
And then yes, I was the first one to to play with it and then
23:56
I just showed my team. So like hey look into this, look
23:59
into this particular stack, look into this technologies, maybe
24:03
you can have some elements of your work streams, right.
24:05
And and so yeah, we have a bunch of exciting projects that range
24:09
from you know basically LLM based research scientist
24:13
assistants that's on Slack and they can help us design our
24:16
experiment and summarise our results to an actual proper LLM
24:21
size type of machine learning training for generating proteins
24:25
out of nowhere. Wait, so break these down.
24:27
This first one, let's start with the easy one and then get
24:29
complicated as we go along and see if I can follow you along
24:33
the way. You've got a Slack bot that
24:36
someone can interact with and says, hey, I want to create
24:41
something, whatever it may be some kind of experiment and can
24:45
you help me? Is that how it looks or or give
24:47
me the breakdown that's the vision that's not, that's not
24:50
yet but basically it's it's like a long chain stack right with
24:54
like document trouble at the end of the day.
24:56
So it's like a bot that you ask it questions and it's going to
24:59
fetch in our experimental data, our select messages or notion
25:04
pages and so on basically all the relevant elements to answer
25:07
your questions governance and give it to you and you can ask
25:09
it to, I mean I've asked it to write LinkedIn posts for me, you
25:14
know about our latest product and it's able to say like oh
25:18
your latest product is actually Nova call and you read it in
25:21
April and so on. So that's you know outcome, but
25:24
then you could also ask it hey, what is the latest high
25:27
performing protein that we've produced in the lab and it you
25:30
know gives you like the entry and you have like the protein
25:32
sequence and so on edges basically a very quick glorified
25:38
search engine with like really, really useful for us, right.
25:41
And that was, I think, like the first MVP we made.
25:43
But I love that it's like what you said is like very quick and
25:46
easy thing. But also I feel like that's what
25:49
you describe is like what a lot of companies are trying to sell
25:51
at the moment. You know, it's like their
25:52
product there was like, you know, it's rag, it's like a
25:56
search engine and everything and you're just like, yeah, it's
25:58
very quick and easy. That makes me mad that these
26:01
people are raising so much money for this stack.
26:05
And I mean to be honest, it's also like because the stack is
26:07
really well made and I'm really thankful for the engineering
26:10
that went behind it. But like when I talk to
26:12
investors, OK and this is the rage moment, right, when I talk
26:15
to investors and they're like asking me why can't we make like
26:18
a B2B SAS with like an alarm front end to just like give
26:22
requests then answer the technical requests and biology
26:25
to people. And like, so first show me how
26:27
we make value. Because with like a rule of
26:30
three, I do not see how you make any fucking value out of.
26:32
The second is like, this is not how you're going to change the
26:35
world. This is not how you source
26:36
sustainability. This is not how you know you
26:38
make things happen. And ultimately I think that's
26:40
what's important to me is like, we want to change the status
26:42
quo, we want to solve sustainability at the material
26:46
level, at least for us. And you won't do that with an.
26:50
And you feel like you are doing that because of some of the
26:53
other stuff that you're working on.
26:55
Tell us like, about the thing that you feel like are changing
26:59
in the world. Tell us like about the product
27:01
and like the things you actually created as well.
27:03
Yeah, it's too bad. I don't have samples here.
27:05
But basically, yeah, we use AI to design proteins that end up
27:09
being materials, right? So we literally using generative
27:11
AI to make a material difference in the world.
27:14
Something you can touch, something that is actually
27:17
outputted and these materials are just hallucinated by AI and
27:22
then you go, hey, that might actually be useful.
27:24
Or how to That's the intermediate steps.
27:27
But that's the general, that's the general process and also the
27:31
first part. So, OK, let's take some cool
27:34
facts. We have one product already, so
27:37
come with me, three years old, we got incorporated in September
27:41
2020. In less than two years we
27:45
managed to commercialise a biotech product and material
27:48
biotech product and that's called tobacco and it's
27:51
basically vegan human collagen that used for cosmetics.
27:55
Why is it important? Collagen is basically what
27:59
constitute most of saturated the, the most important protein
28:02
in your body in terms of weight. That's when you have the most
28:04
of, it's what your skin is made of right.
28:08
And when you age, basically collagen concentration use skin
28:11
decreases and that's why you get recalls that ring basically
28:15
Satsky. Dude, it's so crazy.
28:18
Somebody was telling me yesterday how ohh if your joints
28:21
are all fucked up you should have collagen and I am yeah it's
28:25
am a vegetarian leaning towards began.
28:29
And I'm like, ohh, can you get because they were like yeah you
28:31
can get it in bone marrow and all that.
28:34
I was like oh I'm not having that and we're like, ohh make
28:37
sure you you know like go and find that.
28:41
So all right you you sold me on it.
28:43
I'm going to find where you guys are selling that shit later.
28:46
Exactly how you get in your knees.
28:47
Directly, no. But then you summarised it
28:50
exactly, right. Like there were people that were
28:52
supplying collagen. But yeah, it all comes from dead
28:54
animals, right? And admittedly, people don't
28:56
want to rub dead animals on their faces.
28:59
So if you can come up with like a reading solution, easy.
29:02
Problem is, collagen is literally an animal protein,
29:06
right? So how do you do a vegan animal
29:08
protein? And the answer is with genetic
29:10
engineering. So we take the collagen sequence
29:12
or a bit of it, right, and we put it into our nice little
29:15
yeast and then we give them here.
29:18
And then instead of again making, they're going to produce
29:20
collagen or extract it, package it from terrify it and then send
29:23
it to us, OK. And basically the people that we
29:26
sell it to package that into a cosmetic and then sell actual
29:29
cosmetic product. I love how everything with you
29:31
is coming back to well, it's just like making beer, except.
29:34
That you? One step and then it's jet fuel
29:37
or it's, yeah, but that's that's the magic of industrial
29:41
biotechnology. And I think this is also like
29:42
something I think people trained like me have a duty to do, is
29:46
like demystify GMO's and genetic engineering.
29:49
And GMO's are not just Monsanto. GMOs are wet, you know, allowed
29:54
insulin to be produced without killing pigs.
29:56
GMOs are what he's basically making the enzymes you have in
30:00
your washing powder for your washing machine a basic building
30:06
block of really important scientific progress in the
30:09
modern world. And today still for most of the
30:12
population it is like OHH Monsanto bad right?
30:15
And I think you know it's time we move past past this because
30:18
it is putting a brake on innovation in Europe in
30:21
particular. Example is you've vegetarian
30:25
leaning vegan Dimitrius, right. So you probably have this food
30:30
based meat replacements, right? Which taste like not me.
30:37
The reason why is you know meats has some particular molecules,
30:41
particular lipids, particular proteins and other given its
30:44
particular phase, right. So we have all the tools to
30:48
actually make actual lab grown meat.
30:50
Right now it's a bit expensive, but you know following the
30:54
technological developments, the price will drop.
30:56
The problem is the regulation in York does not allow you to
30:59
actually commercialise, commercialise this in
31:01
supermarkets, right. There's a super nice company in
31:04
Berlin called Formo. They make molecular cheese, they
31:07
literally reproduce the cheese molecule by molecule and they're
31:10
just really good. No animals in the process.
31:13
They can't fucking commercialise it in Europe.
31:16
And it's coming from a French person as well.
31:18
Like you said, the cheese is good, so you trust you on that
31:20
one and lived in Switzerland so you know it's like you've you've
31:25
tried founder Swiss. That's why I trust them.
31:27
OK. Human.
31:29
That's the other. Friends.
31:31
Yeah, there you go. So you actually leading French
31:33
Down in this podcast? Yeah, exactly.
31:35
That's just to finish this story.
31:37
Where do they commercialise? First Singapore and then US, and
31:40
then only Europe, hoping that the lobby will manage to
31:43
actually make GMO more accessible supermarkets in the
31:46
next 10 years, right. But that's basically putting
31:48
Europe behind in terms of like progress with respect to food
31:51
tech, right, Which is super important in Britain in the
31:54
sustainability context for the next two years.
31:57
But also, I mean you touched on that point especially in regards
32:00
to regulation in EU and you know, we also hear it a lot at
32:04
the moment, like with the EU is trying to put like regulation in
32:06
place for AI because. Like people are a bit scared
32:10
that we're going to have robots killing everyone to know or
32:13
whatever they think. And they really think that LM0
32:16
right to the end of the world. But is it like has it had an
32:19
impact for example for Cumbria Man like different things, you
32:22
know, different products you wanted to try, maybe different
32:24
products you wanted to build as like regulation you feel like
32:27
slow you down a bit. Maybe in the process it didn't
32:30
slow us down, but it's definitely a bunch of
32:32
constraints we took into account when we decided our trajectory
32:35
product wise, right, because we're not going to make
32:37
cosmetics for the for the rest of our lives.
32:39
Languages like you start with markets that have the high
32:42
innovation potential. They also like a high price
32:45
elasticity and then you have a bunch of markets including
32:48
cosmetics there. And then the idea is like as
32:50
your technology becomes more mature, your cost production
32:52
decreased, then you watch your way down to actually replacing
32:54
plastics, right? One question we got a lot at the
32:58
beginning was like why don't you do Biomedical OHH or like, you
33:02
know, medical devices and so on. And the answer here is like
33:06
regulation makes it so that the life cycle of your first product
33:09
is at least going to be 5-8 years before you actually get to
33:12
market, right, with the funding you have in the Europe, In
33:14
Europe, you can't survive a startup for 5/8 years like this.
33:17
It's unless you have like special special incubators,
33:21
special investors, it's really complicated, right?
33:23
So we went for cosmetics. And again, I think it's really
33:26
important to stress how cosmetics is such an interesting
33:30
industry, right, because it's it's not quite medical, but it
33:34
still has like medical implications.
33:37
It's not as heavily regulated as again like medical and so on
33:41
South that allows you to have faster products, faster product
33:44
cycles. But also it's just like it's
33:46
always so hungry for innovation because it's the main motor of
33:49
the main engine of their marketing campaign.
33:52
There's this started basically every two years you have like a
33:54
new hero ingredient, you know, like a year, or in the case,
33:56
said JoJo Bauer. It's such a beautiful place to
34:00
innovate. And it's also the industry that
34:03
did save biotechnology after the oil crash in 2015, right when
34:06
everybody was in Again, biojet fuel, bio oil and so on matters
34:11
in 2015. And then all of a sudden the
34:14
price of the old crash below $60.00 a barrel, right?
34:17
Then what happened? BP just sold their biotech
34:20
division, so the I don't wanna Miss Mouse and they're gonna,
34:26
they're gonna give out the names.
34:27
Act check you some other big, I'm pretty sure that was the
34:31
last one to to get rid of of there by a few project right.
34:35
And then basically what did the biofuel companies turn towards?
34:39
Cosmetics was the only viable option.
34:41
And you know why in particular for the case of the job that we
34:45
are making, take two molecules of jet fuel, put it like this
34:49
and join them. Boom, you get a cosmetic product
34:52
called Squealing. Thought it was beer.
34:55
I was like, Dad would they be like this and it just shows that
35:00
you put them together and you have like he's cleaning, which
35:03
is a really good cosmetic product again for your skin and
35:06
what is the main tool of squealing in the world.
35:08
Sharks reports those fishing sharks getting them from their
35:11
liver, pressing their liver and putting them back in the sea,
35:14
really unsustainable and all of a sudden you can take jet fuel
35:16
make cosmetics out of it. And that's how Amir is managed
35:20
to survive the crash right. Fortunately did not survive the
35:24
second by the winter which made them actually get bankrupt I
35:28
think two months ago. Ohh, we just got a second.
35:32
Like honestly one of the flagship biotechs that really
35:35
were showing the example of how to do things in the last years.
35:39
So you say there is a second biotech winter or does that mean
35:43
that we're in one right now? If they just crashed, what is
35:47
it? What is the 2nd biotech than
35:49
Economics? OK.
35:50
So, like, I was like, but like basically, yeah, the the general
35:55
crash of the price of oil in in 2015 killed a lot of biotechs.
35:59
And then basically what happened between COVID and the Ukraine
36:02
war and so on, like, also a lot of biotech, it's like bitter
36:05
dust, right? This is a very, it's a very
36:07
difficult climate for us because we're doing material stuff.
36:13
You have really high CapEx, you have like you literally have to
36:15
buy robots, machines land, right.
36:17
So these are like difficult investments when you are receive
36:21
that's used to you know again like cut prices and and so yeah,
36:26
a lot of you know biotech startups and scaleups haven't
36:29
really easily survived the last 2-3 years we managed.
36:35
Yeah, because also you have a product, I mean you have like a
36:37
product that is working but also like you just mentioned, you
36:40
know like you need to buy robots, you need to have like a
36:43
lot of investment first. So it's also like when you use
36:46
an levels, you're going to try to put it into production.
36:48
I mean we're talking at the equity podcast.
36:50
So usually you try to put it into production and you might
36:54
have, you might make errors and then if you have a product then
36:57
I mean it's, it's fine, you know like your client won't be happy,
37:00
but I guess for you because of error is very different, right.
37:03
Yeah. And so I think our two
37:05
dimensions, I was like what is the cause of error of just
37:07
having a service not working right and what is the cost of
37:10
error of having wrong predictions from your machine
37:12
learning algorithms and elements For us as you are mentioning the
37:16
material earlier like we don't need the low latency service
37:19
serving like millions of people who are customers as they
37:22
generating. The customers are scientists as
37:24
you know, literally people we drink beer with and our friends
37:27
with. So you know if the service is
37:29
down, they're not going to insult us and they're definitely
37:30
not going to go to a competitor. They're just like, hey, can you
37:34
just make sure that whatever, this gene sequencing service is
37:38
up again? And that's nice because
37:41
basically that makes the constraints on our operations
37:47
much lower, right? Like you don't need an increase
37:49
like you crazy infrastructure team to make sure that
37:51
everything is working like 24/7 and so on.
37:56
Your service is down, just like we started.
37:59
That's nice, yeah. But then also like, I mean it's
38:02
like you create proteins. So I mean I have, I have no
38:04
background in biology, so I'm going to ask stupid questions.
38:06
But so you create, you create proteins and like maybe you
38:09
create a wrong protein or like something like that just like
38:14
that happen for example. And then in the end, your
38:15
product might not be as good or might be like maybe.
38:18
It did feel. Again.
38:21
The wrinkles that adds to him. Yeah, no, that's a that's a good
38:25
funny. That's basically the second part
38:27
of the question. I'll read them all.
38:28
Interesting ones like what is the cost of failure, right.
38:30
And again counter example, imagine we are in the medical
38:34
fields, we are making proteins that are supposed to cure a
38:38
disease, right. And your protein fucks up
38:40
something is not good and you know you're having adverse
38:43
reactions. The cost of failure here is
38:45
really high because you're going to have patients that are not
38:47
going to be right for us. Opportunity doesn't performance
38:50
like, you know whatever material we're making love is not from
38:53
like maybe just like the thing we're testing just like breaks
38:56
and cost of failure very low. I mean short it's like there's
38:59
someone like no lighter stakes or anything.
39:02
Again, like in our configuration what these machine learning
39:06
models and general computation is good at?
39:08
Is there risking experiments, right?
39:10
We come from a place where like without computers, we're making
39:12
a lot of experiments and maybe you know 1% is successful.
39:16
And if you we use computers including machine learning
39:19
elements and so on, like all of a sudden we managed to crank up
39:21
to like 10% success, 50% success, 75% success.
39:25
And that's just like save money, right.
39:26
So far as the cost of failure is very minimal and the gain that
39:30
all these computation and operations can bring is
39:33
definitely massive. And that's also how we managed
39:36
to basically take your product to market to in in less than two
39:40
years, right. O also known as I mentioned this
39:42
is that's the absolute record at least in the UK biotechs and I
39:46
think also in US biotechs. Why is it that we managed to go
39:50
that fast? Because like every single
39:52
checkpoint, every single, Yeah, verification step, we one shot.
39:58
Because everything was dressed in a computer, right?
40:00
And that's also the power of AI for us.
40:03
So you mentioned that there's not the constraints that you
40:08
have when it comes to like serving super low latency or
40:12
millions of requests per second or trillions of requests per
40:14
second, that type of thing. But there are different
40:18
constraints that you have. So what are those constraints
40:21
and what is what are you putting your head up against and trying
40:24
to figure out like the engineering problems that you're
40:26
encountering? I think that the biggest salient
40:30
point is the price of data, right?
40:32
Like a single biological experiment costs so much money
40:35
that every single data point you can have you want to save,
40:38
right? And so if you have a pipeline
40:40
that fails and like, you know, data disappears in the cloud
40:43
limbo, that's actually pretty dramatic.
40:46
So I think this is the one thing on which we are.
40:48
Really pent up is like making sure that we don't lose any form
40:51
of data. And then of course, because your
40:54
data is important, you also want to make sure that you have the
40:56
right data models, and that tails back to the problem of.
41:00
You know the data sources, right?
41:01
And then you have basically two types of data sources in in our
41:05
situation. One is the real machines, right?
41:08
And that's you know at least you can pause and then the goods.
41:11
And the other is basically the experimental notes of our
41:15
scientists. And I think this is something,
41:17
especially in biology, we've seen the curriculum in
41:21
universities has been a bit. Delayed in a taking, you know,
41:26
the digitalization of processes and so on and so.
41:30
Training scientists who think in a data-driven way takes time.
41:34
And it's really key to actually instrumentalizing a whole
41:38
machine learning stack, right? Like if your data source is not
41:40
is not your contrasted. Then your machine needs to use,
41:45
that is not going to produce anything interesting, right?
41:48
And so I think, of course when we hire our scientists, we make
41:51
sure that they have the right mindset.
41:53
But also just adapting your lab practises to this mindset and
41:58
it's as simple as. I don't want to say dumb, but
42:01
like as simple in concept as you know, you're eating things, but
42:05
you have to type stuff on your computer to say how much you
42:07
type it in. How do you handle this, you
42:09
know? And the answer is like you have
42:11
to bring robots so that you can free the hands of the scientists
42:13
so that the scientists would say this is the.
42:16
There's a setting that said then that Robert the robot do the
42:19
thing and so on. And so all of a sudden you have
42:20
like all this conflation of a data-driven paradigm that
42:24
changes your way of hiring, changes your way of conceiving
42:28
your lab and changes the way you do experiments.
42:30
And then you have to teach all your scientists to think in this
42:34
way that has like a non zero and say learning debt right.
42:38
And this is something you have to take into account and it's
42:40
also bunch of companies trying to solve this problem, but
42:43
there's definitely definitely a problem is not solved and.
42:47
I'm gonna add, we were lucky because we are building
42:50
something from nothing so we could hire the right people.
42:52
But when you are a gigantic biotech company, I'm not going
42:55
to be, I'm not going to give names, but there's a bunch in
42:59
Berlin, in Europe. In general, you have your team
43:02
of several hundreds of scientists, and you have 20
43:05
years of paper notebooks with the exponential data in there.
43:08
What is the price of like digitalizing this?
43:11
What is the price of like training these people with their
43:14
habits into a data-driven paradigm?
43:16
The answer is like genomes. And so that's where they're
43:18
actually spending a tonne of money trying to get this done.
43:21
And that's also where us, the middle child startups, we can
43:24
actually make our way and basically beat them sometimes.
43:28
So one thing that's fascinating that you said at the beginning
43:31
there is that the price of an experiment is so expensive.
43:36
What makes it so expensive? Is it just the computational?
43:39
No, it's really like ordering the DNA is expensive.
43:43
Handling DNA is expensive, robots are expensive,
43:45
consumables are expensive and you would not expect.
43:48
But also building our lab during the COVID times, there was a
43:52
general competition between the COVID testing centres that
43:55
needed basically the same supplies as we did.
43:59
And then you have also like an increase in the price of the
44:02
surprises. This is like.
44:05
It's all it's all sci-fi tag, so it's all expensive.
44:09
Yeah, it's like you just, you just said like you ordered DNA,
44:12
for example. It's something where I was like,
44:13
yeah, you'll get it. Like, you know, you get it, you
44:16
get it, unfold. And yeah, you have it on your
44:18
door. It's actually.
44:20
It's actually both easy and difficult at the same time.
44:23
But yeah, I mean we we have an API to other DNA.
44:26
So like we we have we have a machine learning algorithm
44:30
that's like concepts using like send them through the API and
44:32
then we get the DNA in the in the post office like a week
44:35
after that is sci-fi. Yeah, that's like, dude, let's
44:40
talk a little bit about large language models versus protein
44:43
models and how you see them being similar and different at
44:46
the same time. I'm gonna start with my
44:48
favourite analogy. Proteins are basically an
44:51
assembly of a bunch of building blocks.
44:53
These building blocks are called amino acids, and you might you
44:56
might remember this from your GP in high school or somewhere in
44:59
the book. These building blocks designed
45:01
to assist, they all they have like 20 of them so we usually
45:05
match them to a letter. If you look at this under this,
45:09
under this angle, basically a protein is a string of different
45:13
letters. What else is a string of
45:15
different letters? Natural language, right?
45:17
And so the same way, if you just take letters of the English
45:21
alphabet, randomly put them together, you have very few
45:25
little chances of actually getting an English word.
45:27
Even fewer chances to get like a proper poem or like a proper
45:31
essay, right? It's exactly the same with.
45:34
If you take these amino acids, put them together randomly, you
45:37
just get. Like spaghetti bowl, maybe yummy
45:39
one, but you know nothing. Nothing would come out of it.
45:43
But if you're very smart, like if you know the inner grammar
45:46
syntax of this fourteen language, you can put them
45:48
together. So you can, you can get
45:49
function, you can get the function that gives the
45:51
character in your hair, the flexibility, the function that
45:54
gives the elastic in your skin, the elasticity, the function
45:57
that gives haemoglobin in your blood, the oxygen carrying
45:59
capacity, right? And that's the whole challenge.
46:03
And he's been the challenge of by of protein engineering for
46:07
maybe 60 years, right. It's like to understand a little
46:09
bit this syntax and this grammar.
46:11
And so there was a bunch of breakthroughs when these efforts
46:15
then you might have in 2018 and then 2020.
46:18
But basically from the string of letter, we were able to predict
46:22
with a reasonable chance of success the shape of the
46:25
protein. And shape is very close to the
46:27
function of the protein, right? So.
46:30
That was good. We could proofread little bit
46:31
our manuscripts if you want with the technology and what generate
46:35
API in proteins is bringing us now and I'm talking about
46:38
technology we're using now, right, We can ask?
46:41
Machine learning models to write an essay for us, to write the
46:44
protein for us, to generate the protein for us with particular
46:47
functionalities and that's. That's dishing channelling
46:51
technology that actually took us to the next level in terms of
46:54
like capacity to create the capacity to engineer new things
46:56
and materials for the future. And so, yeah, that's basically
46:59
the technology we're developing and working with right now.
47:03
OK. So so like you mentioned like
47:05
you know you have your technology and your product that
47:07
that is now outside like it's out in the world.
47:10
Have you used like realms to be like, OK, I want to create a
47:14
protein that are going to remove wrinkles from people and then
47:16
they'll have is like, OK, here it is.
47:18
There's the shape. So Novak always developed before
47:22
the big LLM boom, So it was basically like a younger version
47:26
of this. It was using the protein
47:27
programming language. But basically what we said is
47:29
like, OK, here are a bunch of options you have.
47:33
We want to block. Has this property.
47:34
We want a flexible block here and then we want this optimise
47:38
for a bunch of. Of constraints, including we
47:41
want a peptide that is not allergenic, we want something
47:45
that's compatible with the formulation pH of the creams, we
47:48
want something that has maximised potentially in effect
47:52
and so on and so forth, right. So it was not quite an elegant
47:55
yet, but we already had the language transform.
47:58
The design of this portion into a mathematical problem which.
48:02
Pretty sure nobody else had then.
48:05
Yeah, because that's what I was gonna ask is like, it seems
48:08
like. That's the key basically to this
48:10
success because like was like not like big companies, you know
48:13
have that like Something Seminole to be like, yeah, we
48:16
want. Well, in something that is like
48:18
no allergies, you know, everything.
48:20
Like, is it because the like the project and language that you
48:23
have? Like that's basically the secret
48:25
to the success? I think it allowed us to sift
48:27
through orders of magnitude is 10 to the 18 different possible
48:31
points, right? And we got a top 100.
48:35
And other top 100 we tried maybe a bunch in the lab and then
48:38
that's where we got a product from, right without.
48:42
This technology that we use in order to choose the right
48:45
sequence that has like the right properties ratio will not have
48:48
one shotted the feasibility guys like you know from like
48:51
scalability of operating to like proper behaviour on the skin and
48:54
so on. So in that sense, not answering
48:58
the question yet, but in that sense like the computation
49:00
really was what's enabled us to basically be ahead of the game
49:06
and now we are using an element for next product which I can't
49:09
talk too much about, but basically it.
49:11
Is. Getting to what you're saying.
49:13
Using like delaying type of decision models, just like make
49:16
proteins that particular properties that we can code in
49:19
mathematical loss functions and it's really cool, like you can
49:23
really see the protein coming out of a blur of atoms and
49:27
falling together. So exciting.
49:29
Wow. Yeah, way different than what I
49:36
would have ever expected you to say on this call that seems
49:39
like, again, total sci-fi stuff. And you did mention how you're
49:44
like, right, We start with some cosmetic products.
49:48
Eventually we want to get to regenerating plastic, but not
49:52
plastic. Like a I guess non harmful
49:57
plastic. Non harmful for the environment
49:59
plastic. Is that the goal, the ultimate
50:02
goal? Or is that just like you see a
50:05
vision of going towards things that are trying to help the
50:08
planet more. Not saying that the the vegan
50:13
collagen isn't helping the planet because as if you are
50:18
talking to somebody that is in a vegetarian.
50:21
I understand how like not doing anything to animals is very
50:26
helpful for the planet, but it feels like there you you kind of
50:30
have in your head a bit of a a road map of where you want to.
50:35
Go and this again like a lot of this one might be sort of
50:38
country infrared, but like the general idea is very accessible.
50:41
It's like what are the intermediate markets that get
50:43
you from like specialty high-priced chemicals to
50:46
commodity low price chemicals? And there's a bunch of like
50:50
stuff that's in between. Ohh yeah, she mentally I'm.
50:53
I'm always using the example of plastics because they are
50:56
everywhere, right? Also in this fake leather
50:58
jacket, but you have a bunch of like, materials that people
51:02
rarely think about that are actually pretty unsustainable
51:06
when it's concrete. Concrete creates a shit tonne of
51:09
CO2 every time you you prepare it, right?
51:11
There's also like really cool biotech companies that make
51:14
chemicals out of the CO2 emanation of concrete
51:17
production. That's not what we do, but it's
51:20
really cool people doing this. So yeah, you want to display.
51:23
Yes, but you know you can't like sample size.
51:25
You talk to a concrete company, they're like, oh, we need
51:27
samples to make tests, Can you make US100 kilogrammes?
51:30
Footings and questions like. Expensive, right?
51:34
So you need me something to bridge in between an interesting
51:37
case that the lens about, well, the company was like Blues?
51:41
Used to stick things together. That's so complicated, also so
51:46
unsustainable. And there's like this use case
51:48
that we love talking about because we can never.
51:52
Happened but like space loop so sticking things in space.
51:57
Why is it a problem and why is it not sustainable when you do
52:01
things together? When you glue, your glue dries,
52:03
U shows it off gases a bunch of things right?
52:06
And this strings of gases very often not really good for your
52:09
health. So now imagine you're in a tin
52:12
can outside in space and you have off gas or something.
52:14
It's not really good for your help.
52:16
Pretty bad, right? Proteins are actually really
52:18
good glues. The muscle like the the thing
52:21
you can eat, At least we eat it in France has some of the best
52:25
animal glue you can find in the world to actually stick itself
52:28
to rocks underwater. Can you imagine like a water
52:31
resistant glue? That has like Jag Antic force.
52:36
You could use that in space. The problem is.
52:39
That never thought about that. It's the kind of random shit you
52:42
learn at my job, and I fucking love it.
52:47
Yeah, so close in space. Second problem with glues when
52:50
they burn also really harmful toxic chemicals.
52:52
When you burn proteins, it just smells like.
52:55
Bones, here you see this admittedly not a nice smell, but
52:58
you know it's not actually right.
53:00
So there's actually this intermediate step for like we
53:03
could make space glue not too expensive, but it's also like
53:05
not dirt cheap as like a pretty interesting use case, really
53:09
good marketing. The problem is the time to
53:13
market for this like 15 years because it's space industry.
53:18
So, yeah, there's like a lot of materials we don't really think
53:20
of on a day-to-day basis that actually can help us get down to
53:24
replacing commodities with sustainable.
53:26
And actually, that's the short story.
53:28
Now that's like. Sci-fi things we talked about it
53:31
like crazy. It even rockets, you know you
53:33
end rockets at the end we like. No, I was just expecting you to
53:36
say like, Oh well, turns out you can make glue by the same
53:41
process as making beer. You just replaced one molecule.
53:47
And in the end of the day, that's what you would be, right?
53:49
Like we take all, we take our bug, we remove the DNA that
53:52
produces the collagen, put the blue DNA and just ask it can you
53:55
make blue cheese? Downstream processing would be
53:57
one key but at the end of the day that's how it would that's
54:00
how we would turn out right. Ohh I love it dude.
54:02
Well this has been awesome. I think we've taken up a tonne
54:05
of your time we gotta ended there.
54:06
But for anyone that wants to learn more about what you are
54:10
doing I encourage them to check out the links in the description
54:13
about Cambrian you all hiring right now.
54:17
There we go. If you ever, if you want to go
54:19
save the world and work with Pierre soon, that's when it's
54:22
happening and I think we'll call it that Philip Fogging.
54:30
Hey, I'm Vishnu. I'm a data scientist at first
54:33
hand and I definitely think that you should subscribe to the ML
54:37
OPS Coffee Sessions podcast. It's the best podcast out there
54:40
to stay on top of what MLP's actually is, to talk to the true
54:44
thought leaders in the space. And oh by the way, Demetrius is
54:47
absolutely hilarious. What a weird guy.
54:50
You should definitely subscribe to the podcast.