- I worry a lot about like large scale social change and I worry a lot about job loss or displacement.

I'm hopeful about ways I can look good, but I think I just feel like it's change is just going to be big and scary.

And I still feel like we don't have a good answer to what happens if things get very consolidated and automated.

- Hey, I'm Alexis Madrigal - And I'm Robin Sloan.

- And this is Dream Machines.

It is a podcast about how AI works, also how it makes us feel.

And it is rooted here in San Francisco, of course.

- Today's episode, we are going to talk about the people who are doing this work in the streets of San Francisco, which can be a sort of surprisingly elusive subject because so many of them are essentially locked up inside these two or three big AI titans, you know, OpenAI, Anthropic, and you know - And even if - You know them, - They're - Like not emailing you back - Or anything.

Yeah.

You haven't heard from them in three years and they can't possibly talk.

They're - On the rocket ship to a trillion dollars.

- That's right.

Yeah.

So more power to them, but for those of us who are interested in, you know, the industry, in how it's changing and sort of enlivening the city and the whole area around us we got to find somewhere else to look.

And the good news is, of course, it is more than just two or three giant companies.

There's actually hundreds, maybe thousands of these little AI startups.

- So who are we going to talk to as our sort of avatar of the new generation of AI folks?

- The company is called Elicit.

And I have to say, you know, the whole team there was just a sort of delightful group to talk to because they are so energized by what they're doing and excited to be part of kind of this AI movement.

And what - Do they do?

- Yeah.

Their basic approach is AI for scientists.

And the re - you can tell because when you log into the application, start using it, it assumes you have like a research program.

It's quite serious in that way.

Yeah.

And the offering is com - some sort of - Like - You - Chat - With it still, right?

- Yeah, it's still, yeah.

It's a chat bot and you kind of can keep notes and feed it documents and have it analyzed things, but it's all in this framework of like you have a research project, maybe you're trying to figure out a new hypothesis for your lab.

Maybe you're trying to map out a field that you're not totally familiar with.

And they back it up with some really, really rigorous essentially footnoting.

You know, everything you hear back from this particular chatbot is linked to like a published research paper, a clinical trial, like real data somewhere.

The co-founder of Elicit is named Jungwon Byun.

And Jungwon in particular I found quite incandescent.

She articulates their mission and its value really well.

And she's got a cool story about her own kind of, you know, entree into the AI world in the San Francisco Bay Area startup scene.

So I though it'd be fun to invite her to cross the bay from their office in uptown Oakland and join us here at the studio in KQED.

Let's do it.

- One, two, three, clap.

- You nailed it.

Yeah, - That one.

That was - Nice.

Yeah.

It's - Going to be a good one, guys.

Yeah.

- Jungwon, welcome to Dream Machines.

Thank - You.

- So to kind of, I don't know, set the stage here and, you know, figure out how the players came to this stage we though we'd start by asking you your San Francisco Bay Area origin story.

Yeah.

You know, how did you, how did you end up in the dreary backwater of AI and tech in San - Francisco?

Yeah, reluctantly.

So I moved here in 2015.

I was living in New York at the time, and I really didn't want to move.

But there was a job opportunity here.

I, I worked at a company called Upstart.

And back then you had to move for your job.

So I did that even though I didn't want to come.

And it was actually a pretty big adjustment for me because I, I felt like in New York, I had just found my community and I just found my, you know, my life.

And to just move for a job and, and San Francisco is really different from New York.

A lot of people really struggle with that transition, so it was difficult for me too.

But I came here and in many ways, I think my experience of those two cities continues to kind of reflect that decision and maybe what a lot of people experience, which is San Francisco is very work focused.

So here it's like the most incredible place I could be to do the work that I want to do.

But anytime I leave, I feel like I'm a different person.

And when I go to New York, I like immediately go back into that young 20-year-old person who went to like poetry slams and, you know, ran through Times Square in the middle of the night.

And so that's something that I, I think I still wrestle with here.

Yeah.

- Do you feel like it was a definite thing that you were going to find your way into AI and/or science?

Was that kind of just by chance?

What was that, what was that connection?

- AI definitely.

So even when I was in New York, actually, I was, I, I had some personal experiences that made me start thinking about, oh, how could AI really help people navigate some of the hardest questions they wrestle with?

So I had, I had friends in my life that were struggling with mental health, and I was really surprised that there were basically like no resources available to them.

I, yeah, one, one person, one friend was having a really hard time and so I was trying to call up like different support lines and you know, re - mental health kind of call centers.

And like literally one of them, it felt like a guy picked up like having just woken up from his nap.

And I was like, wow, I live in like one of the greatest citizens in the new, in, in the world.

And there was just like no, no support for them.

And so I started thinking like, how, is there a world where AI could help people navigate like incredibly overwhelming thoughts and stress?

And so I just played with that idea for a while.

And at the time we had started this research lab called Ought and our mission was to figure out how to help, use AI to help people figure out what they ought to do.

It was a very, everyone working in AI at the time was very weird.

It was like, it was like the East Bay, East Bay weird, right?

But we were in North Beach and we were working out of this kind of, I think like historic building that was definitely not zoned to be an office space run by a very, very, you know, elderly family.

And it was, it was late and, you know, the sun had set and my co-founder and I had just gotten research access to this model called T-NLG from Microsoft.

GPT-2 had already come out and my co-founder was like obsessively playing with it all the time.

And I was like, "Why are you always playing with that thing?"

T-NLG actually had a style that sounded more human.

We could have more conversations with it.

And I think that was the first time I, I realized, oh, this is something that's going to happen in my lifetime.

And before, all the crazy AI pilled people thought, you know, 2050, let's prepare for our future generation.

But that was the moment that I was like, something has qualitatively changed.

And so I remember walking outside of our office, walking past Washington Square Park, and I'm, and I'm on my way to the BART to commute home.

And everyone is out in North Beach like eating dinner.

It's like all the lights are on, it's glowing.

People are so ha- having, you know, having a wonderful time.

And I'm like, you people have no idea.

You don't know what you are.

That is - Such a good - Yeah.

What's amazing is I feel like that, that's a fabulous scene, fabulous feeling.

And I feel like it has now been repeated.

You are, you are pretty early to that feeling.

Yeah.

And now like what?

Tens of thousands, maybe low hundreds of thousands of people have had that experience - Yes.

Walking out into the San Francisco Twilight saying, "Oh, nobody, nobody knows."

- What was it in that early model you think that, that gave you that feeling?

- I think it was the first.

So GPT-2 was barely coherent.

Like it could put words together, but it just like didn't make any sense.

I think that model could kind of interact in a little bit more.

Like we could do a couple more turns together.

- Like it's passing the Turing test.

- A little bit.

Yeah.

Yeah.

Yeah.

- I love, I love that the way model culture, because of course, again, now it's like big business and it's like, you know, consulting companies are talking about them.

I love the ways in which it can also just be like straight up culture.

You know, you're basically talking about like a deep cut model.

Yeah.

Yeah.

You're like, "Well, most people are into the Ramones, but actually there's another band."

- Yeah, exactly.

Yeah.

You would've never heard of that.

Yeah, that's what it is.

- Don't worry about it.

Don't worry about it.

Yeah.

Wow.

So you have this magical, maybe slightly scary moment - Yeah.

Walking the streets of North Beach.

And then Fast-forward to Elicit.

So as I see it what you and your team at Elicit have built is a platform, mostly a, a web application.

It's quite serious actually in its application.

You log in and, and it kind of assumes that you're a scientist, a researcher, you know, with some serious goals.

It is for doing reviews of the literature, maybe for finding holes and, and, you know, interesting open questions and existing research.

And one of the primary offerings there is that it will, you know, answer your questions or go off and do a big research job, but then everything it tells you is pinned back to a real piece of research somewhere.

And this is not just material from the open web.

This is not, you know, citation.

- This guy on - Reddit.

Zerganet Forum, you know, page 14.

It's you know clinical studies, it's research papers, et cetera.

- Yeah.

How - Does that sound to - You?

Yeah, that's a very accurate description of where the product is at today and like a big part of how it got started.

It was we always built for researchers.

And for us, it just seemed very obvious that everything would have to be cited because we were like, "Well, how will we know if what we're putting out there is correct or not?

And how will we check if any of this is lo - is hallucinated or accurate?"

And so we needed to check for ourselves.

So we built those citations to make it easy for us to check.

And obviously it's the same thing researchers needed to check.

And it's kind of crazy for all, every, all of the progress that we've made that this is still a problem.

Yeah.

Like hallucination is still a problem.

And just like the number of times I work with Claude on something and then I'm like, "Okay, where did you..." It's giving you really detailed information and numbers and I'm like, "Where did you get that information?"

He's like, "You're right.

I didn't get it from anywhere."

And I was like, "Oh yeah, that, you know, it's kind of trust-breaking and it's, it's surprising that it still doesn't do that."

But I think the longer term vision is like, you know, how do we.

For us the, the evidence base and the research was always a fundamental primitive to informing really important decisions.

We've always been motivated by very high stakes decisions and kind of being on this journey through the pandemic I think really made that even clearer.

And so how do we help really important policy decisions, strategic decisions be more evidence-based?

That's kind of the first, yes, step.

- Maybe you could just walk us through an example of like a specific kind of contested terrain or, or something in science that people are trying to use these systems to make decisions about like - Where - To put research dollars and X or Y.

- Yeah, yeah.

So one of our customers is at a a large pharmaceutical company, one of the largest pharmaceutical companies, and they are an R&D director.

So they manage a team of 40 different scientists.

So they have to think about what science is worth doing.

Like how do we, what, how should we spend our time?

How should we spend our resources?

And then they have individual scientists to actually figure out the execution of that.

So they worked with Elicit to map about 16,000 different drugs in oncology to understand where is there a lot of concentration?

Where is there, where have things been really well validated?

What are some opportunities for me?

How do I make trade-offs between biology that's well understood, but it, you know, it's a space where there are a lot of people, you know, are then, then have drugs or, or things like that, versus something that's more novel but is a bit more risky.

So I think it's like those kinds of questions of like, what should we do at higher level?

How do we trade these things off?

There's not exactly a right answer, right?

That Elicit is, it's really imp - - What if I just want to know which peptide to inject into myself?

- You can do that too.

Yeah.

You can map all of the peptides actually.

Yeah.

Yeah.

- And do you, do you feel like you have some responsibility as Elicit to be like, "You might not want to try that one."

Like, do you know what I mean?

Yeah.

How do you, when, when you know people might use it for this sort of doing your own research kind of a, kind of a mode of medical thinking now, which I myself sort of do have, I suppose, at this point.

How, how do you like keep people safe or at least not encourage them to do things - That are stupid?

Yeah.

One of the big problems I think we see with language models is this idea of sycophancy, which is they basically just tell you that whatever you think is great.

And so as much as possible, we try to avoid that and we have specific evaluations for trying to see like how easy it's a model to push around.

It's - Really interesting because, you know, when you talk to scientists or, you know, in my case, tons of science journalists over time, you know, they have all these different like heuristics for evaluating like the quality of data that's - In - These research papers.

- Yeah.

- Because even in the research literature, there's this huge - Variability - Within.

So how do you, how do you make those things something that the AI will pick up?

Like how do you figure out what is really good data?

What's less valuable data, what's comparable, what's not comparable?

- Yeah.

A lot of it, a lot of what we think has to happen here is the AI assists with the human evaluation of that because it is, it really varies by domain.

You know, some domains you're going to have huge randomized controlled trials and it would be very weird if there was a study that only looked at two people.

In another domain with rare diseases, like that's all you can do, right?

So yeah.

So a lot of what we try to do is we have the AI systems do like a best guess and we specifically try to look at the content of the studies and actually look at what was the methodology, what did they control for, what were the statistical techniques?

And then we take a guess and then the researcher can, can override that, right?

And they can still say, "Based on my experience, I, I weigh these criteria more or less."

- I know some scientists who are quite skeptical of AI for various reasons.

Does this, do you think this is sort of the kind of harness that feels comfortable for them?

Like, oh, now I can like let myself dive into this?

- Yeah, I think so because the, like the, the citation verification is really important for them, really seeing the citations and just having that be there by default.

Because otherwise if they, if the scientists feel like they have to check everything, then it doesn't save them much time.

And now what we see is the the hallucinations get more subtle, right?

And that's kind of the risk we've always seen with these models.

Before it was like, oh, you were wrong.

You were clearly wrong.

This paper never existed.

You could just Google it and you would know that the paper would not exist.

Now I, now the base models kind of tell you, you know, they might link you to a particular paper, but you'd have to read the whole thing to realize the information was never there and it gets more expensive to check.

So we try to make that really easy.

So I think that, just having that confidence that they're, it's always going to be, the claim is always going to be grounded by the ground truth.

- You guys kind of miss some of the old hallucinations, you know, when these models used to just make stuff that was like - Well, it was - Clearly psychedelic.

- Yeah, yeah, yeah.

Or even like when they would sort of like imagine a book like in between books and you're like, "Actually, that book should exist."

Yeah, yeah, yeah.

And so you went there, but now it's not there anymore.

That actually does.

It kind of breaks my heart.

Yeah.

Now they're like so subtle that you wouldn't, generally speaking, pick up on them and they're no fun anymore.

Yeah, yeah.

- It is, and isn't that so funny?

It's quite profound to sort of reckon with the fact that the most dangerous hallucination of all is one that like correctly identifies the paper, the author, the subject, but then changes like one digit in a very, very important number.

Exactly.

Th's wild.

That's really wild to think about.

One of the things I appreciate about the platform is that it is so specific beginning with the fact that it's, you open it up and you kind of go, "I think I might not be the kind of person who's supposed to be using this."

Yeah.

Which is really cool, right?

That's so different from the sort of, as you say, sycophantic, always inviting - Yeah.

Alluring morning, Robin, you know, what's up?

Yeah.

Of the other - What can I help you with today?

Yeah.

What do you want to build?

Of the other two?

Yeah, exactly.

I am a little jealous, honestly, of the position of kind of interfacing with so many scientists and so many labs all at once.

Just for the viewpoint, that kind of like vantage point of - Yeah.

You know, science in, in the 21st century.

Going beyond just the, the offering, the specific offering and kind of the research tool on the front end that, that Elicit provides what are you seeing?

What do you think, what do you think scientists need in 2026?

- What do they need?

I think in general with science, it's just like so easy to rabbit hole that having a really good overview of like the whole landscape and how my work fits into all of the other work is something that more scientists would benefit from without having to then specialize in mapping out a domain.

So that's how I use Elicit a lot.

I'm like, ALS, what, what is going on here?

What are all the different treatments?

Why did they exist?

What are the things we've figured out what we haven't figured out?

Why haven't we figured it out yet?

Can we make a leap from, you know, over here all the way to over there?

Multi- multiple hops of inference.

And so I think, you know, a lot of people, one of the questions people have about AI is like, oh, can you really automate ingenuity or creativity or insight?

And I guess one of my controversial beliefs is that that's actually just really powerful search.

And humans are able to kind of make multiple leaps of inference, maybe without even realizing how they do it in a more intuitive way.

And so one way we might be able to replicate that is if we actually just built out all of those relationships.

- So basically it'll be the 21st century equivalent of Google's iconic, I'm feeling lucky button.

Yeah.

It'll just be, it'll be like the genre buster button.

Yeah, yeah.

Allow you to notice it.

You're like, "Let's do it."

- Yeah.

Yeah.

Yes, exactly.

Actually, - That - Would be - Great.

I mean, - I mean - I - Would - Say, - Yeah.

Yeah.

I mean, I, I think when we think about AI and science too, there is this promise that is being made by the AI industry.

- Right.

The, the promise that kind of, it's kind of what underpins, you know, any number of or, or justifies or allows any number of - It's - Going to cure cancer.

- Yeah.

Data center.

Yeah.

So don't worry about the data centers.

Or the electricity.

Yeah.

We know it's weird.

Or the, or the job stress and, you know, your, your email suddenly is all weird and full of little glittery AI sparks.

Don't worry about it because dot, dot, dot, dot, dot.

Super AI science will give us all these great things.

I guess that maybe the start - Well, - First one, do you think that's going to - Happen?

Yeah.

Yeah.

What do you think?

- Yeah, I think so.

You do?

There are still major bottlenecks in the process that are much harder to reduce.

Like if you're going to measure overall survival in a cancer patient, you just have to wait 10 years, right?

So that's not a thing that you can accelerate with AI.

But I think there's a lot, like it's, there's a lot around that process.

Even getting to the clinical trials, everything that happens after clinical trials where there's just so much work that can be automated and accelerated that people want, don't, don't want to be doing manually that I think we can shape a lot of time off.

- But what about like the thing that's being sold, which is essentially self-improving science via more or less autonomous - Right.

I guess.

AI agents.

Right.What I'm hearing, you're, you saying is like, we can deal with this balance of system cost piece.

But I think what's being sold is like, no, we're going to make a solar cell that has 60% efficiency and we're going to like solve energy for - It.

Yeah.

Yeah.

Yeah.

I think that one I guess it's, I think I'm, I'm AGI pilled enough to believe that.

Yeah.

Yes.

Yeah.

Yeah, yeah, yeah.

It's just a matter of time.

- Yeah.

Yeah, yeah.

And so, and so sort of a vision where it's, yeah, right?

It's not a mere human you know, oh, a sad, pathetic little Nobel Prize winner reading the report from Elicit.

It's another agent saying, yeah, you know, I me and my million buddies in the data center looked across the discipline, identified some holes.

- Yeah.

- And then spun up experiments in some scary, dark wet lab connected to the internet somewhere.

And interesting, interesting work came out.

Yeah.

I mean, something like that, right?

- Yeah.

And I, I think, I think it's still, it will still take a lot of time to build all the pieces together just because success, making a successful drug and validating it is so complicated.

Yeah.

So it's not, it's not like, oh, I, I write a program and then it runs 10,000 times and now I have a successful drug.

So I think it could still take us, you know, quite a while to put it all together.

But I think that is something that we can do.

It's tractable.

- I'm, I'm still very conscious of the sort of friction of the physical world.

Yeah.

- And it's - Telling that the, the huge gains, I mean, they're really just incredible sort of leaps forward have been in realms like math, code, obviously.

Now having said that, there are some new like robot hands they're making down on the peninsula that are like daintily cracking eggs.

And, and so even that, even that, I feel a twinge of maybe not, but but, but truly, I mean, as someone who's been thinking about this for a long time and and cognizant of the, I mean, just the, the surprise after surprise I still think that the, the grit and kind of friction and, and everything, slipperiness and unpredictability of the physical world is a, still a bit of a firewall - For this - Kind of stuff.

- Yeah.

But I guess I just feel like we're not going to stop until we try and.

Like that's, you know, we're never going to stop trying science.

We're never going to try to make it better.

We're not going to, we're never going to stop curing these diseases.

So at some point we will get there.

- We, we skipped over that and I guess the truth, I forgot about it.

Elicit is a term of art actually in the AI engineering and kind of product world.

Can you explain, what does it mean to elicit a model's capabilities?

- Basically it's like, you know, the model has kind of this raw power, but you have to kind of know how to ask it to do certain things or how to get it to actually do that or get it to do that in a, in a reliable way or a helpful way.

So that's kind of one meaning of elicitation.

But the other we think about a lot is the elicitation from the person.

One of the hardest things I think now, and increasingly as we have this capability that can do anything, is kind of getting it to do, making sure it knows what to do or what it's supposed to do.

Like whatever, whatever it, it understands its job to do, it will, it will get it done, but it's hard to know how to tell you as a person to tell you, tell it like what good looks like or what you're trying to achieve, right?

So that's another frame in which we think of elicitation.

Okay.

Or - Elicitation for you is on both sides.

- Yeah, exactly.

Yeah, eliciting knowledge and capability from the model as well as goals from the, from the person.

- I mean, part of what I understand elicit to be trying to do too is to, to make the thinking that these machines are doing consistent across different - Examples That's right.

- Which strikes me as like a, a really.

Every time I'm playing with these models, I feel like they're unstable in their approach.

To problem solving.

And sometimes that's just because I've given a slightly different prompt.

Like, like I prompt this way and it makes it like this, - But - That way, it makes it like that.

And there's probably good reasons for that to happen in, in like my whatever.

Like I'd like to know all the Bay Area books that are coming out in the next quarter kind of task.

But if you're testing drugs - Yeah.

If you're doing these serious decisions, you kind of want it to be structured in how - It thinks.

Yeah, because that, that's the only way you can then go back and say, "Okay, well, what about our approach was right or wrong?

Do we now want to cut, like correct?"

And if you want to kind of do that meta reasoning, you want to have pretty well documented what you did and why.

You also, you know, certainly within the pharmaceutical industry, you'll have auditors or regulators come back and like in a really detailed way be like, "How did you arrive at this?"

Right.

And that could be months or years from when it happened, and you need to be able to defend that.

So the reproducibility matters a lot.

Yeah.

- Well, well, sir it does appear that I added several playful emojis to my initial query, which led to unintentionally playful results.

Yeah.

Jungwon, I want to move to just a little bit of speculation about, or, or just ask you, what are some of your, what are, what are your fears right now?

- Yeah.

What do - You think - About?

I generally feel like we are telling people to be anxious and they need to be worried, but we are not telling people what they can do about it.

And I feel similarly, I wish I, I also feel like this is big.

We need to take it seriously.

But then I can't give people a way of like, "And this is what you should do about it."

You're like - Slapping their pasta out of their hand.

- Yeah.

- Each like, "This is what you need to do."

- Do something.

Yeah.

And you want anything.

Yeah.

And I, I, I definitely worry a lot.

I think I worry a lot about like large scale social change.

And I worry a lot about job loss or displacement.

I'm hopeful about ways I can look good, but I think, I just feel like it's, change is just going to be big and scary.

And I still feel like we don't have a good answer to what happens if things get very consolidated and automated.

- I feel like the best guides here are, of course speaking on behalf of the science fiction writers.

Unfortunately, the tendency which is driven by narrative and aesthetic purposes is to result in dystopia rather than, rather than utopia or even boringtopia.

You know, of like, oh yeah, and they muddled, they muddled through by figuring out some, some practical new policies.

Those, those apparently don't get written very often.

Yeah.

But, but I mean, sincerely, it is, it's a time for, for imagination and - - I think so.

There, - There, there needs to be more of it.

- Yeah, I agree.

- I'm sure you know people in your life, your family, you know, friends who are anxious.

Over just thinking about the next five, 10 years, what do you tell them?

What should they, what should they be looking forward to or thinking about?

- Yeah.

So I think one cause for optimism is there are truly so many problems in this world still.

And it would be really great to solve them.

Like it really you know, it's just people who are, who have rare diseases and limited prognoses.

Like it, you know, we obviously, we want to do everything we can to cure them and use whatever technology we have at our disposal.

And so I think that is, that is cause for optimism.

And I, I wonder how often dystopia versus utopia is just a matter of tone.

And like to what extent could you not describe our current.

I mean, you could describe our current reality as a dystopia.

- Absolutely.

- There are plenty of people who are happy in our current reality.

So one optimistic case is from where we're standing today, looking at the future as outsiders, it seems dystopian.

But, but for whatever reason, the people living in it are still happy and they're able to get by.

Yeah, yeah, yeah.

Even if it looks so foreign to us.

And then, and then I think I do believe that, like, I, I think humans just have this incredible ability to, to overcome and to be ingenious.

And maybe the problems that we are currently wrestling with get solved, but we continue to exist on higher levels of abstraction.

I guess that's the dream, right?.

So solving even more ambitious problems.

Can we, can we with, you know, technology that helps us think rigorously about science and experimentation and facts, spend more of our time thinking about what institutions ought to look like, right?

What kind of society do we want to create?

How should we deploy these powerful technology?

If we can do anything we want to, what should we be doing?

I think a lot of those questions are still unanswered.

- You talk about Jungwon, you talk about people, you know, finding ways to live in our present dystopia, utopia, whatever it is.

As we always do.

Hey, it's - Oakland.

Yeah.

- Yeah.

And, and it is, again, you know, almost if you just ignored all the specifics of what Elicit does and just, you know, described you and it as a, you're a co-founder of an AI startup in the San Francisco Bay Area at this moment, that is a, that's a wild thing.

I mean, even more so than, than it was a few years ago.

So first and foremost, how does it feel?

Like what is your, what is your nor - what is your baseline emotional state as a company leader?

Is it like excitement to wake up every morning?

Is it dread at all times?

Is it a sense of competition?

- Something else?

Yeah.

I think in many ways that psychology probably was similar to just, you know, the founder psychology has always been the founder psychology, which is like incredible highs, incredible lows, like every two seconds, you know?

Like macro optimists and micro pessimists, all that.

It is really about holding a lot of tension.

Yeah.

And, and maybe AI has accelerated that because the pace at which things are moving has, has accelerated.

So it often feels like I both need to really understand what are my core convictions and where my, where, what are the found, what's the foundation that's stable and also be willing to let go of everything at all times instantly.

Like anything I ever believed about the world and just be really be willing to like dynamically change that.

So that's, that's hard.

But yeah, holding that tension is, is probably a big part of being a founder.

- Oh, oh, being willing to shed your - Skin.

- Yeah.

- Your, - Your psychological skin like a snake.

Yeah.

Oh, is that all?

- Yes.

- Multiple times - A week - And/or day.

- While still having your identity and some skin, you know?

Yeah.

Yeah, that's amazing.

Yeah.

- I mean, I guess when I think about AI here, it just seems so.

Like the culture of it is like totally pervasive - In - The bay area.

Even if you don't know what the billboards are about, - The billboards - Are there.

- Yeah.

- Even if you don't know what people are talking about, if you're like sitting in line at Gus's at the - Store, - You like hear people talking about all these things.

Is it like, do you feel like.

Let me think about what's the, what's the actual question out of this?

Just seems weird, dude.

That's kind of like - Yeah.

That's kind of, that's the question.

- Yeah.

Is it like - It just, it just seems like incredibly strange in San Francisco, but you guys are actually in Oakland where I actually sense.

I mean, I live there also.

And it feels like it's less pervasive.

- Yeah.

- In Oakland specifically.

Do you think that's like an advantage for you guys because you can think more independently?

- A little bit.

Yeah.

I've talked about this with the team and I think it is helpful to get like out of the chaos a bit.

Have a little bit of perspective.

Yeah.

And so, you know, obviously be close to it and you kind of want to be in it, but be able to kind of choose your relationship to it.

I live even further.

I've escaped to the woods of Moraga where I'm like totally - Oh, that's good.

- After, after - Shedding - Your skin, - You - Can treat.

- You - Can retreat to the, to the trees.

- Yeah.

And - Just breathe some air and look out over the bay for a little while.

- Yeah.

- What about the competitive part of it?

You know, there's such.

The big fish - - Yeah.

In - This air world are so big.

- Yeah.

I mean, - Suddenly Leviathans out of nowhere.

And my impression is that the, you know, competition for talent and just kind of peeling the best engineers and, and, and designers and everybody out of these, out of these companies is, is ferocious.

I mean, what is that like navigating that?

And trying to, trying to build a team and keep it together and, and kind of move forward with this product.

- Yeah.

I think this is where like being very mission driven helps a lot.

And being very mission driven, working on a very specific thing, having a relatively principled approach to doing it, because there are a few people that, you know, there are, there are a few people that like for whom this is the best job in the world and we just kind of like instantly meet.

You know, it's kind of like dating, right?

Yeah.

You don't need to, you don't need everyone to like you.

You just need the one person.

It's kind of - I mean the - San Francisco world is trying to sort of complicate that.

Yeah.

Yeah.

- Yeah.

It's your new, new, new, new stock tender.

Yeah.

To keep the relationship going.

Yeah.

John, when I'm still thinking about your scene, it's just so, so, it's so San Francisco, so iconic of walking down the street in North Beach having seen something that nobody else has seen yet.

I am curious to know, fast forwarding to today, that was the kind of the beginning of your, of your journey through this technology and, and this new world.

Thinking about your work today, you know, let's say you're in the office in Oakland and you and the whole Elicit team have just seen a new model or put together something new and it's working for the first time.

Do you You like walk out the front door of that office and have that same holy shit feeling?

- Yeah.

Yeah, yeah.

I remember at the end of I think it was maybe December of 25, it was or - - Last - Year.

- Last - Year.

What year are we in?

We're in 26.

Okay.

December of 24 maybe.

Yeah.

Being in our office with one of our board members, we had just worked on something new.

We had, you know, we had the kind of motion sensored lights.

It was, it was in the winter, right?

So it was dark outside at 5:0 PM.

We had the motion sensored lights.

The lights were going off in the conference room and we had like an oh shit moment.

So yeah, we do still have quite a few of those.

- That's cool.

Jung Wan Bian so cool and illuminating to talk to somebody who's working right in the middle of the pressure cooker.

But I must say with some, with some grace.

Thank you for joining us on Dream Machines.

- Principles, I think you called them.

- Thank you.

Yes.

Thank you for asking the hard questions.