1:59
Mhm. >> We can do this.
Search this video
3:19
We can do this all night. We can do this. We can build a future. We can do this.
3:40
We can do this all night. We can build a future.
3:50
We can We feel the future. >> You'll be all right.
4:54
I see multiple journalists on the horizon. Stand by. You're watching TBPN.
5:00
>> Today is Tuesday, November 25th, 2025.
5:04
We are live from the TBPN Ultra Dome, the Temple of Technology, the Fortress of Finance, the Capital of Capital.
5:11
Uh we had a we had a thing yesterday was Anthropic Claude 4. 5 day.
5:14
We had a lot of fun talking to Sholto about that.
5:17
You should go check it out.
5:20
We wrote a little write-up.
5:21
Uh I collaborated with Brandon and Tyler to kind of give our thoughts on the state of the AI race with regard to uh OpenAI and Anthropic and what makes Anthropic special.
5:33
The things that stuck out to me, I mean, the thing that went viral was just the fact that apparently Dario goes around Slack and writes essays every single day.
5:38
Uh and everyone was like, "Give me the essays. Turn it into a book. Paging Stripe Press.
5:42
We got to get Stripe Press to turn into a book."
5:46
Yeah, I not well, that, but I was also thinking of of potentially a risk factor for them being like the the Dario files, you know, a disgruntled employee that saved them all.
6:00
>> And then leaks them all.
6:01
>> I imagine >> Because he even when he's on mic, he's known to say some things that he uh uh at times uh sort of People people I feel like people take him out of out of uh out of context a lot. >> a lot.
6:14
Like I like he will say he will say >> of the final boss he's being like >> doesn't Yeah, if this doesn't go well, we could lose 50% of white-collar work or entry-level white-collar work and and people will be like, "Anthropic's stated mission is to destroy jobs." Take your father's job.
6:31
>> Yeah, yeah, it's it's rough, but uh Ramp. Time is money. Save both.
6:37
Easy-to-use corporate cards, bill payments, accounting, and a whole lot more all in one place.
6:43
Um Timeline was in turmoil over the weekend and yesterday.
6:46
We covered a little bit about the Nucleus uh dustup on the timeline.
6:50
Uh Kremyau will be coming on the show at 11:45. And then Fast follow.
6:57
>> Kion, uh the CEO of Nucleus, will be coming on the show at noon.
7:00
So, we will kind of have both sides.
7:02
Then we have uh Joe Weisenthal joining from Bloomberg.
7:07
Um And uh and then we have Who else do we have today? Kratios.
7:12
Kratios is coming on to break down uh Project Genesis, which we're very excited about.
7:16
Uh anyway, let's uh run through uh what other uh news stories were at the top of the timeline.
7:22
While we pulled those up, let me tell you about Restream.
7:25
One live stream, 30-plus destinations.
7:27
If you want to multi-stream, go to restream. com.
7:28
Oh, yes, the biggest news in tech in AI is that uh the uh Ilya Sutskever Dwarakesh Patel podcast has dropped.
7:39
Do we have Hit the timeline.
7:43
>> the Do we have the opening clip?
7:43
Because the opening clip is is iconic.
7:45
Uh it's it's it's uh it's very funny.
7:48
It's a It's a bit of a hot mic moment for Ilya, and I think we should we should pull it up and uh and play it because it has a fascinating uh just insight into It feels very like, "Oh, this where this is the real Ilya.
8:01
He's not He's not even thinking that he's on camera."
8:04
And he gives his real um his real feeling.
8:08
So, let's let's play this from the very start.
8:11
>> Listen, but all of this is real. Yeah? Meaning what? Don't you think so? Meaning what?
8:16
Like all this AI stuff and all this Bay Area Yeah, that it's happen like Isn't it straight out of science fiction? Yeah.
8:25
Another thing that's crazy is like how normal this slow take off feels.
8:29
The idea that we'd be investing 1% of GDP It's like hasn't even been discussed. They're entirely real. You know?
8:35
Where right now it just feels like We get used to things pretty fast, turns out. Yeah.
8:40
But also it's kind of like it's abstract. Like, what does it mean?
8:44
What it means that you see it in the news Yeah.
8:45
that such and such company announced such and such dollar amount. >> Right.
8:50
That's That's all you see. Right.
8:53
It's not really felt in any other way so far. Yeah.
8:55
Should we actually begin here?
8:56
I think this is an interesting discussion. >> Sure.
8:58
It's one of the greatest podcast intros of all So good. So good.
9:02
Uh anyway, we're we're not going to watch the whole show.
9:06
>> that's going to be a new meta. Yes. Yes.
9:09
I mean You you can't you can't fake that. It's amazing.
9:11
Also, it's just funny because uh uh you know, he it's effectively getting caught on a hot mic. But we were just joking.
9:17
I was like, "Of all the things that you could say on the hot mic before you sit down, oh okay, we're actually recording."
9:23
Uh his is just completely reaffirming everything we know about Ilya Sutskever.
9:29
Like, it's just completely the same.
9:31
Like, okay, he's a He is a true believer.
9:32
It's not like he was sitting down and being like like New York cash, like we got to go uh on my private plane.
9:38
I just sold so much secondary.
9:40
It's crazy what's going on with this stuff.
9:42
Like, if people really think this AI thing's going to pan out, I'm making billions of dollars. I'm I'm cashing out.
9:49
I'm I don't believe any of this stuff is real.
9:51
No, he wasn't caught on a hot mic like that.
9:53
He His hot mic moment is like, "Wow, it's exactly like science fiction.
9:57
Everything is It's all real." Yeah, which is iconic.
10:01
Well, you can go and listen to that in the Dudeswarcache tell our RSS feed and on the YouTube channel and on X.
10:06
He put the full thing up. It's 95 minutes long.
10:11
Uh Tyler, did you have any other takeaways from your speed run?
10:13
You're listening to it at 5x, right?
10:17
>> Well, on X you can only do up to 2x. I was on that.
10:19
So, I still have like 10 minutes left. Good.
10:22
>> Um but yeah, a bunch of good stuff in here.
10:23
>> Does he pop the scaling bubble?
10:23
Does he get a bearish take about at any point?
10:29
>> Um While you're thinking about that, let me tell you about Gemini 3 Pro.
10:33
Google's most intelligent model yet.
10:34
State-of-the-art reasoning, next-level vibe coding, and deep multimodal understanding. Continue.
10:37
Um so, I wouldn't say he's like anti-scaling, but he does kind of give this interesting take.
10:43
Um which he basically says that like now um AI companies like there's too few ideas for the amount of companies and and for for the the scale that we're at.
10:53
Um where he basically like you can think of AI progress as being in these kind of distinct uh ages, right?
11:01
So, he says 2012 to 2020 was like the age of research, where um you're trying all these like different ideas and and the scale of things is very small, right?
11:10
Like, to train the original um AlexNet was like two GPUs to do the original uh transformer was like eight, maybe 64, but like, you know, very small amount of GPUs.
11:20
Um and then once we kind of figured out that that uh transformers work, we entered this age of scaling.
11:24
And that's basically from from 2020 to 2025.
11:27
And now we're basically at this point where like yes, you can keep scaling and models will get better, but even if you scale 100x Mhm.
11:36
um like are we really going to get super intelligence?
11:38
It like it'll get better on the benchmarks. >> Yep.
11:42
Um and they'll become more useful, but it's not like this he doesn't think that just raw scaling alone is basically what's going to bring us there.
11:48
I mean, this has been echoed by a lot of people.
11:51
Um like this was I think uh Karpathy Karpathy said this where we still need a couple different kind of paradigms for this to work. >> Yeah.
11:58
Um and he said he like this is even kind of what um Schulman said yesterday, which is like pre-training is it's not dead, but it's like the reason that that Opus 4.
12:10
5 was better is not just cuz they scaled pre-training. It's scaling generally. >> Yeah.
12:15
Um but even then like the the scaling has gone from pre-training and now it's RL.
12:21
And so we basically we need to find another paradigm. Mhm.
12:22
And the way you do that is just doing like research.
12:24
And so he talks about SSI is basically being this it like return >> Return to research. Return to research.
12:31
Yeah, it it's small kind of training runs.
12:34
Uh even though you know, they only raised 3 billion, which is like small compared to other Sure.
12:37
to to other uh research institutions.
12:41
Um the fact that they're basically putting it all on these kind of I mean, I don't know if they're moonshots, but they're these small training runs where they're doing experiments. Yep.
12:49
And then they're going to scale it up eventually. >> Yep.
12:52
Um but they're not just basically trying to win the AI race by just scaling up and doing the same thing as everyone else. >> Yeah.
12:58
Yeah, they're trying to find a new a a way to actually bend the scaling curve, find a new scaling law, or find a new technology that like that they can scale against.
13:08
I was thinking about uh Ilya's talk at NeurIPS last year.
13:10
He pulls up this chart of uh the relationship between mammal the mammal's mass and the brain volume, and it's a pretty linear graph.
13:22
And so like the elephant is a lot bigger than the mouse and so it has a proportionally larger brain to its body volume.
13:30
And and it's this perfect it's this perfect linear curve.
13:35
I should I should just try and figure it out if I I can maybe text it in. I took a picture of it.
13:43
Cuz it's a very it's a very cool chart. Here it is. Where do I send this? The timeline? Let me see. Ship it. >> Uh share. Let me see. Timeline. Sorry. Um Boom.
13:57
So basically the the mammals have this like very clear linear trend.
14:02
But then the non-human primates are a little bit higher up on the chart and they're just doing a little bit better.
14:11
But then hominids the actual humans have a different there's a very distinctly different curve.
14:18
And so there's this interesting like it was making me think like like maybe that's what we're supposed to see when we think about yeah this this.
14:27
It's like it's like when we say like straight lines on log graphs when we say we are seeing scaling happen with the current architectures which line are we scaling against?
14:42
Are we are we actually scaling on the on the human curve or are we waiting for divergence from that current scaling law?
14:51
Um Yeah, he he has this good quote where scaling has taken all the air out of the room. Right?
14:55
Where like basically like we have more than enough compute to try these like different ideas.
15:01
But they're just all going straight into to training the next big model using the next paradigm and maybe it's slightly different right?
15:06
You have a different way of doing RL or whatever.
15:07
Um but it is still fundamentally the same thing, right?
15:12
And he talks about maybe continual learning is really the the better approach, right?
15:16
We we've we've been in this era of like having a pre-training thing for so long that we think of like AI is like you train this thing and then you release it and it's like done.
15:25
And RL's like a little bit different now because there's this idea of post-training and you can kind of integrate different things to it.
15:33
>> the interesting thing was with with pre-training you use the whole internet so you don't have to decide anything.
15:38
You're just applying this algorithm to just all the data, all the compute and there's no decisions but then with RL you have to decide, okay, we're putting in these math equations and we're maybe not putting in something else cuz we're actually creating the data and we're and we're it's not just this Exactly.
15:51
Like this is maybe why we see these kind of like models that are super well they do super well in evals but not Yeah, somebody overfitting and >> reason is because the data that we choose is not the correct data. >> Yeah.
16:04
Because researchers are basically being reward hacked maybe into like just solving for benchmarks. >> Yeah. It's interesting.
16:13
It it it it's interesting to hear this uh like like the conclusion is we need another breakthrough and and then simultaneously consensus be like but like we're definitely going to get that breakthrough in like the next decade.
16:28
It's like it's hard like like these breakthroughs it's hard to predict.
16:30
I I feel like it echoes a lot of even what Mike Newp has been saying right? We need new ideas.
16:36
Saying this for for months I and >> way it's way harder to predict the rate at which breakthroughs will arrive as opposed to like you can actually chart out, okay, the formation of capital, the time it takes to build a data center, how long it takes to do you know, manufacture a bunch of GPUs, rack them, run the training run.
16:53
Like that's much more predictable than like human came up with new algorithm.
16:57
Like that's sort of random.
16:59
>> Yeah, and he brings this up as as the reason why you see companies doing this because it's just if you're raising money, it's so much easier to to justify the raise by saying we're going to buy this data center and do this training run.
17:10
It's going to cost exactly this much.
17:11
It's going to monetize this way.
17:13
Yeah, and then the model will be this good and then we can use use to monetize this way.
17:15
Where if you're just saying like, oh yeah, we're just going to pay a bunch of like really smart researchers to do a bunch of research and then they'll they'll figure something out."
17:22
That it like you can't really do that.
17:23
>> Yeah, in some ways it it feels like SSI is set up for like somewhat of a mini AI winter or like at least riding the hype cycle down cuz it doesn't sound like he's sitting there being like, "We raised 3 billion and we're spending it in the next 12 months." It's like >> 2. 9 was that.
17:39
No, not not it not No, no, no, no. That's the point. It's not. It's like equity. It's just sitting there.
17:42
It's like he can either eat Yeah, it's like he's going to give each researcher researcher or different teams like a shots on goal.
17:49
We're going to keep taking those shots until obviously he'd be able to raise like another 10 billion dollars whenever he wants, especially if he has like a key breakthrough insight and they can be first to scale that. Yeah.
18:03
Well, let me tell you about Cognition.
18:05
They they make Devin, the AI software engineer.
18:07
Crush your backlog with your personal AI engineering team.
18:12
Um Nvidia has posted >> the timeline. >> the timeline.
18:16
Break this down for me, Jordy.
18:17
They said, "We're delighted by Google's success.
18:19
They've made great advances in AI and we continue to supply to Google.
18:24
Nvidia is a generation ahead of the industry.
18:27
It's the only platform that runs every AI model and does it everywhere computing is done.
18:31
Nvidia offers greater performance uh ver uh versatility and fungibility >> which are designed for specific AI frameworks or functions.
18:43
Uh That is a crazy thing to post.
18:47
>> Crazy crazy crazy thing to post.
18:47
You can't say I don't know, boys, but having the largest company in the world sending tweets to defend their main product is not very reassuring.
18:58
It yeah, it's just an odd I feel like this would be so much better delivered.
19:03
I actually don't have that much of a problem with the actual text here.
19:06
I just think this should be delivered by Jensen with some nuance in a conversational setting, it just hits a lot different when this is in at exactly 9:00 a. m.
19:18
Like clearly scheduled, clearly typed out in a document, you know, it's like it it it feels like a press release, uh which is just an odd odd thing when in when it should be, you know, this should be an answer to a question, which uh someone Bobby Cosmic in the chat was saying like, "Oh, the mainstream media is just now picking up on the Gemini 3 story."
19:40
And there's articles in the Wall Street Journal and other places saying like, "Oh, maybe Google's back.
19:43
Like, you know, buy Google.
19:45
Like, it's very exciting."
19:46
Um and um And so, Nvidia feels feels the need to respond to that, but uh it's a lot different when it's actually a response instead of just like a we're putting out a press release. Like, who knows why?
20:00
Like, as opposed to like Jensen saying like, "Well, since you asked, you know, talk show host or news anchor or whoever he's top podcast host, whoever he's talking to, Dwarkesh, you know, uh whoever he's talking to, maybe us. We'd love to have him.
20:14
I can ask him that question. He can defend this here. Yeah.
20:18
Uh well, the timing is is uh seems important cuz they are coming under a huge amount of pressure right now.
20:23
There was an article in Barron's this morning by Tae Kim. Yep.
20:26
The headline is not what Nvidia's comms teams would have liked it to be.
20:31
The headline is Nvidia says it's not Enron in private memo refuting accounting questions.
20:39
>> That's a crazy thing to say.
20:39
And of course it's not Enron, but Let let me get let me get into the to the coverage.
20:45
So, Tae says a series of prominent stock sales and allegations of accounting irregularities have put Nvidia in the middle of a debate about the value of artificial intelligence and its related stocks.
20:54
Now, Nvidia is pushing back in a private seven-page memo sent by Nvidia's investor relations team to Wall Street analysts over the weekend, the chipmaker directly addressed a dozen claims made by skeptical investors.
21:04
Nvidia's memo, which includes fonts in the in the company's trademark green color, begins by addressing a social media post from Michael Burry last week, which criticized the company for stock-based comp, dilution, and stock buybacks.
21:18
Uh Burry's bet against subprime mortgages before the 2008 uh financial crisis was depicted in the movie The Big Short, of course.
21:24
Nvidia repurchased 91 billion shares since 2018, not 112 billion. Mr.
21:30
Burry appears to have incorrectly included RSUs.
21:32
Uh RSUs, taxes, employee equity grants should not be conflated with the performance of the repurchase program, Nvidia said in the memo.
21:41
Uh employees benefiting from a rising share price does not indicate the original equity grants were excessive at the time of issuance. That makes sense.
21:49
Uh Barron's reviewed the memo, which initially appeared in social media posts over the weekend, uh and confirmed its authenticity.
21:54
Burry told Barron's he disagrees with Nvidia's response and stands by his analysis.
21:58
He said he would discuss the topic of the company's stock-based comp in more details.
22:02
Burry is, of course, now over on Substack.
22:06
He's charging $380 a year.
22:09
Uh and uh if you are a perma bear, I can't This is like Christmas coming early.
22:16
Uh Nvidia uh didn't respond to Barron's uh for a request for comment, uh but they also responded to claims that the current situation is analogous to historical accounting frauds, Enron, WorldCom, and Lucent, that featured vendor financing and SPVs.
22:30
Nvidia does not resemble a historical accounting frauds because Nvidia's underlying business is economically sound.
22:37
Our reporting is complete and transparent, and we care about our reputation for integrity.
22:42
Unlike Enron, Nvidia does not use special purpose entities to hide debt or inflate revenue.
22:46
There's no mark-to-market.
22:48
It It's like There's 25 examples of how this is not the same.
22:53
I don't even know why they're Nvidia also addressed allegations that its customers, large technology companies, aren't properly accounting for the economic value of Nvidia hardware.
23:02
Uh some of the companies use, we've talked about this, use a 6-year depreciation schedule for GPUs.
23:04
Burry said he believes the useful lives of the chips are shorter than 6 years, meaning Nvidia's customers are inflating profits by spreading out deep depreciation costs over a long period.
23:15
Nvidia's customers depreciate GPUs over 4 to 6 years based on real-world longevity and utilization patterns.
23:23
Older GPUs, such as A100s, continue to run at high utilization and generate strong contribution margins, retaining meaningful economic value well beyond the 2 to 3 years claimed by some commentators.
23:36
So, again, under under fire on the TPU front and, uh, from the Michael Burry camp, but, uh, again, I think I think they're their answers are totally valid.
23:50
Uh, Matt, uh, over on X said, uh, had a post here, he said, "The TPUs equal bad for Nvidia take is up there with the dumbest, maybe worse than Deep Seek, as it completely misses what actually happened in the last 6 weeks.
24:02
And I will remember who is who in the zoo, my view.
24:06
One, demand for AI is bananas, no one can meet demand, everyone is spending more.
24:13
Google said just yesterday they have to double capacity every 6 months to keep up.
24:16
Two, scaling laws are intact."
24:16
He's, uh, referencing Gemini 3.
24:19
"The flywheel is about to speed up.
24:21
Somehow the mid-curve crew thinks this is zero-sum competition.
24:26
None of this suggests that.
24:28
If you think the race is hot now, wait until you see what comes out of, uh, large coherent Blackwell clusters.
24:34
All the magic from the quote god machines is pretty much still Hopper-based.
24:37
Lastly, a quick GPU-TPU lesson, the cost and performance specs on the box aren't what you get in real life.
24:43
Uh, and Google is going to get a fat margin, too, doubled up.
24:47
What matters is system-level effective tokens to watt to dollars and TCO.
24:52
Nvidia GPUs have higher FMU because they are they're already embedded in workflows / the ecosystem is massive.
24:59
Uh by the way, this is a good test.
25:02
If you have an opinion on this topic, but you have to look up FMU, then perhaps create better shorts. >> MFU What?
25:08
MFU MFU Is what I said FMU? >> Sorry. Sorry. Sorry.
25:13
Uh the above effective token what gap also likely widens with Reuben.
25:17
Add in that Jensen can actually deliver volume in a tight market.
25:21
Uh plus future flexibility, multi-cloud capable, programmable for paradigm shifts, and he'll sell every GPU he makes for years.
25:27
And Google will too since everyone wants a second supplier and TPU is a fantastic chip.
25:34
But this is as far from either/or as it gets.
25:38
The one benefit of this confusion is that it is likely to give Google a brief stint as the world heavyweight champion, the most valuable company.
25:44
I would guess the midwits put the strap on them in less than 2 weeks.
25:48
So, he Put the strap on them? What does that mean?
25:50
Just like Like like pile in?
25:56
Is he saying just like So, he So, he Is he It seems like he's predicting that that uh people will overplay the Nvidia bear take.
26:09
And o- overplay the Google opportunity and that will result in Google becoming the most valuable company in the world.
26:16
Uh and uh he uses the phrase put the strap on them in multiple in less than 2 weeks. Um interesting post.
26:21
Um Uh in other news, David Sacks has hit the timeline.
26:28
He says according to today's Wall Street Journal, AI-related investment accounts for half of GDP growth, a reversal would risk recession.
26:37
We can't afford to go backwards.
26:37
Um we will Uh the article is how the US economy became hooked on AI spending.
26:43
And uh we will be chatting with uh Cratios uh in about an hour on this uh on this very topic.
26:55
Um so we can get into a little bit.
26:57
Well, before we move on, let me tell you about Adio, the AI native CRM.
27:02
Adio builds scales and grows your company to the next level.
27:05
Uh fact sheet from the White House President Donald J.
27:08
Trump unveils the Genesis mission to accelerate AI for scientific discovery uh today and this is yesterday.
27:15
Today Trump signed an executive order launching the Genesis mission a new national effort to use artificial intelligence to transform how scientific research is conducted and accelerate the speed of scientific discovery.
27:27
The Genesis mission charges the Secretary of Energy with leveraging our national laboratories to unite America's brightest minds, most powerful computers, and vast scientific data into one cooperative system for research.
27:41
The order directs the Department of Energy to create a closed-loop AI experimentation platform that integrates our nation's world-class supercomputers and unique data sets to generate scientific foundation models and power robotic laboratories.
27:52
The order instructs the Assistant to the President for Science and Technology to coordinate the national initiative and integrate an integration of data and infrastructure from across the federal government.
28:01
The Secretary of Energy, APST, and the Special Advisor for AI and Crypto will collaborate with academi academia and private sector innovators to support and enhance the Genesis mission.
28:12
Priority areas of focus include the greatest scientific challenges of our time that can dramatically improve nation's national, economic, and health security including biotechnology, critical minerals, nuclear fission and fusion energy, space exploration, quantum information science, and semiconductors and microelectronics.
28:33
Uh next harnessing AI for our national security and economic development.
28:36
With the Genesis mission, the Trump administration intends to dramatically expand the productivity and impact of federal research and development within a decade.
28:47
Um so uh there's a one more note here on strengthening America's AI dominance.
28:53
Trump continues to prioritize America's global dominance in AI to usher in a new golden age of human flourishing, economic competitiveness, and national security.
29:02
And so we will get into more of uh of this with Kratsios.
29:06
Yeah, I'm I'm very interested to hear how uh like how the public-private partnership actually works here.
29:14
There was a time when um every basically every cool technology was coming out of DARPA, coming out of the US government.
29:21
Uh the US government landed on the moon and since then, you know, I I think a lot of people in technology have lost faith in the US government uh overseeing the development of technology. Even academia.
29:34
I mean, the people people think like, you know, AGI will emerge from a private C-corp.
29:42
Like that's where people believe that the best work will be done with, you know, give Ilya Sutskever, give the best scientist $3 billion, let him go cook.
29:51
Like that's the thesis currently in tech.
29:53
Uh this feels like somewhat of a rejection of that in some ways.
29:57
There's obviously lots of different places where uh having AI resources, having science and technology resources within the government make a ton of sense.
30:05
Um but it'll be interesting to see like where are the uh interfacing points between the two uh between the two categories, the the public and private sector, because uh by default, I think most people in our audience in technology uh would would say, "Hey, like let's leave the uh let's leave the space travel and the and and the AI research to the to the private sector."
30:31
And uh and and this is, you know, uh potentially a different direction, potentially just a very synergistic.
30:35
So, be interesting to see where it breaks.
30:40
Uh well, should we uh run through the Astral Codex 10 piece on trait-based embryo selection to tee up our discussion with Kermit and Keon from Nucleus and go through that.
30:51
So, um uh this is from Scott Alexander in Astral Codex 10.
30:56
He says, "Suddenly trait-based embryo selection."
30:58
When a couple uses I uh So, in 2021, Genomic Prediction announced the first polygenically selected baby.
31:07
Uh When a couple uses IVF, they may get as many as 10 embryos.
31:10
If they want one child, which one do they implant?
31:12
In the early days, doctors would just eyeball them and choose whichever looked the healthiest.
31:17
Later, they started testing for some of the most severe and easiest to detect genetic orders like disorders like Down syndrome and cystic fibrosis.
31:27
The final step was polygenic selection, genotyping each embryo, and implanting the one with the best genes overall. Best in what sense?
31:33
Genomic Prediction claimed the ability to forecast health outcomes from diabetes to schizophrenia.
31:40
For example, although the average person has a 30% chance of getting type 2 diabetes, if you genetically test five embryos and select the one with the lowest predicted risk, they'll only have a 20% chance.
31:50
So, you get a 10% bump there. That's nice.
31:52
Since you're taking the healthiest of many embryos, you should expect a child conceived via this method to be significantly healthier than one born naturally.
32:00
Polygenic selection straddles the line between disease prevention and human enhancement.
32:06
In 2023, Orchid Health, founded by Noor, who we've had on the show, uh entered the field.
32:10
Unlike Genomic Prediction, which uh tested only the most important genetic variants, Orchid offers whole genome whole genome sequencing, which can detect the de novo mutations involved in autism, developmental disorders, and certain other genetic diseases.
32:26
Critics accused GP and Orchid of offering designer babies, but this is only true in the weakest sense.
32:33
Customers couldn't design a baby for anything other than slightly lower risk of genetic disease, you're basically just selecting out of what you already got. Yep.
32:40
They're not editing the genes.
32:42
They're They're merely sequencing them and then allowing you to select.
32:45
Um these companies refused to offer selection on traits, the industry term for the really controversial stuff like uh height, IQ, or eye color.
32:53
Still, these were trivial extensions of their technology and everyone knew it was just a matter of time before someone took the plunge.
33:01
Last month, a startup called Nucleus took the plunge.
33:03
They had previously offered 23andMe-style genetic tests for adults.
33:07
Now, they announced a partnership with Genomic Prediction focusing on embryos.
33:11
Although GP would continue to only test for health outcomes, you could forward the raw data from GP to Nucleus and Nucleus would protect predict extra traits including height, BMI, eye color, hair color, ADHD, IQ, and even handedness.
33:23
And it's worth noting that Nucleus is now being sued by Genomic Prediction.
33:30
Even though they have this partnership.
33:34
I I'm assuming Assuming the partnership is no longer. Yeah. Well, we can ask. Yeah.
33:38
Uh but I'm assuming it's no longer because one of uh GP's co-founders left uh left the company >> prediction to join Nucleus? Interesting.
33:48
and allegedly uh turned off all the security cameras the the >> that's metaphor?
33:56
Or is that actually Uh the lawsuit alleges that the that he turned off all the security cameras on his last >> That's not a metaphor for like, you know, sharing a Google Drive of PDFs.
34:06
You literally mean >> his last day at work Okay.
34:07
and he was allegedly like rounding up >> so he turns off the cameras allegedly and and the implication is that maybe he was rummaging around like literally taking documents or something like that.
34:21
That's at least what the timeline is.
34:22
what the the lawsuit alleges. >> Okay, wow. That that That's wild.
34:24
I I did not know that that was a literal uh accusation.
34:28
And then another part of it uh apparently nucleus uh it's new people at nucleus were emailing uh uh the former uh co-founder uh at his old email address uh evidence of them violating the the the agreement that they had.
34:51
So, anyways, it's very very very very messy.
34:55
We can ask >> Yeah, there's like four or five companies involved in this uh >> And all of them are controversial because this is the most I think the most controversial probably like category that you can be in.
35:05
Yeah, it's it's certainly up there.
35:09
Health is already like one of the most controversial topics.
35:12
And as a health influencer they've gotten into various >> Yeah.
35:18
Um and also there's just like the there's just it's so easy to throw I mean in the same way that people are throwing Enron at at Nvidia like it's so easy to throw Theranos at any biotech company that's not you know that's accused of anything.
35:33
Uh and and also biotech it's like it's it's pretty hard to understand the underlying science.
35:38
It's not it's not a it's not as popular as okay like does the website work?
35:42
Does the business make money?
35:44
You know, what's the cash flow like?
35:45
It's way more complicated and so uh it it it does attract even more attention.
35:50
So, one of the other companies in the space is Hera site and uh Astral Codex 10 continues here.
35:57
Uh they entered the space with the most impressive of disease risk scores yet, an IQ predictor worth six to nine extra points, and a series of challenges to competitors whom they call out for insufficient scientific rigor.
36:10
Their most scathing attack is on Nucleus itself accusing its predictions of being mis misleading and unreliable.
36:15
Let's start with the science and then move on to the companies to see if we can litigate their dispute.
36:20
In all theory in uh in theory all of this should work.
36:25
Polygenic embryos polygenic embryo screening is a natural extension of two well-validated technologies, genetic testing of embryos and polygenic prediction of traits in adults.
36:34
So, genetic screening of embryos has been done for decades, usually to detect chromosomal abnormalities like Down syndrome or simple gene editing disorders like cystic fibrosis. It's challenging.
36:46
You need to We've talked about this before.
36:49
You need to take a very small number of cells, often only five to 10, from a tiny protoplacenta that may not have many cells to spare and extract a readable amount of genetic material from this limited sample.
37:02
But, there are known solutions that mostly work.
37:04
And so, the companies that we're talking about today aren't necessarily doing like the fundamental lab equipment development, building the machine, figuring out how to sequence data from the first It's more about the analysis that happens on top of the results.
37:20
>> And the recommendations. And the recommendations.
37:22
>> Which is which is probably which I would say is the most controversial part of this.
37:26
Which is Uh I don't I don't know that any of them are recommending, "Hey, we think you should take you you we think you should pick this baby."
37:33
They're more just saying like, "We think that according to the data, this baby might be taller than this one."
37:40
If you're giving if you're Yeah, but that's not a recommendation.
37:43
Like, if I tell you this car is 700 horsepower and does 0 to 60 in 2 seconds and this one does 800 horsepower and does 0 to 60 in 2. 4 seconds.
37:48
This one's faster in a straight line, this one's faster on the curves, and then like you pick.
37:53
Like, I didn't make a recommendation, I just told you the stats. Right?
37:57
Yeah, but but from when you look at the these companies from a from what they're marketing to consumers of of what of why you should care about the service. >> Sure.
38:08
And then the way that they deliver the information, if they're advertising we can effectively advertising we can help you have a smarter, healthier baby.
38:17
And then they're saying like, "Hey, we think this direction is going to get you a higher IQ.
38:22
How is that a I don't think it's a recommendation.
38:26
>> It's not an explicit recommendation, but I think people are trusting the service to try to get them what was marketed to them.
38:33
Uh yeah, people want the data and they want the data to be accurate because they're going to make a decision based on that.
38:39
But I mean, here Scott Alexander actually gets into some of the uh some of the complexity of uh of the actual trade-offs because there are So, uh most traits are polygenic requiring information about thousands or tens of thousands of genes to predict.
38:57
These are too complicated to understand fully at current levels of technology, but some studies have chipped away at the problem and gotten to a partial understanding.
39:03
Often, this looks like being able to predict a few percent of the variance in the trait to determine whether someone's genetic risk is slightly higher or lower than uh average.
39:13
And uh so, some people might genuinely want to select on a single condition.
39:18
For example, people with a strong family history of schizophrenia might want to minimize their chance of their children getting the disease.
39:24
For these people, reducing schizophrenia risk by 58% while keeping everything else constant sounds pretty good.
39:32
Everyone else probably wants a genetically healthy um generically healthy embryo with low risk of all conditions.
39:38
Exactly how this works depends on the customer's own value.
39:42
Would they prefer an embryo with lower cancer risk to one that will have fewer heart attacks?
39:47
Like that's a trade-off that you have to pick.
39:49
Um and the exact benefits will depend on how parents make that decision.
39:53
Genomic Prediction and Hera site try to help by providing semi-objective measures of which embryo is overall healthiest according to different conditions, effects on longevity, and patient-rated quality of life.
40:06
For genomic uh prediction, that's the embryo health score.
40:09
This is uh you know, that's close to a recommendation.
40:13
I think you're you're you're getting close.
40:14
Yeah, and and Nucleus's uh Subway campaign is have a healthier baby. Yeah.
40:22
Yeah, it's it's the the the marketing claims are are a big big piece of this.
40:28
Uh I think I think the scientific claims are are potentially uh just as important, but the the it's both the the they're both uh understanding where the science actually is both broadly and then also within the companies and then how it's marketed.
40:42
Like all of that is important uh to to get like a complete picture of what's going on here.
40:46
So, uh for Hera site, it's a polygenic longevity index.
40:48
They don't give exact risk reduction numbers for each disease saving that saying that it depends too much on a couple specific family history, but say that most people gain 1 to 4 years of healthy life.
41:00
When I test uh it on a set of 20 embryos, the healthiest gets an extra 1. 66 years.
41:09
Um And so, how much would you pay to give your children an extra 1 to 4 years of healthy life?
41:13
This is no longer healthy a a hypothetical question. Here are the costs.
41:18
Genomic Prediction uh is around uh $3,250.
41:23
Orchid is around $12,500.
41:26
Nucleus is around $9,249. And Hera site, $53,250.
41:34
That is expensive compared to the rest, five times the price. Uh is it worth it?
41:37
Well, if you're already doing IVF, the claimed risk reductions are accurate, you value your kids' health as much as your own, you have low time discount rate, you're well off enough that these aren't extraordinary sums of money to you, you're okay using expected utility calculations where 50% chance of preventing X is half as good as fully preventing X, then I'll go out on a limb and say, "Yeah, it's obviously it's worth it."
41:58
Uh consider Genomic Prediction which costs $3,500 for five embryos and claims to lower absolute risk of type 2 diabetes by 12%.
42:05
That implies that not getting type 2 diabetes is worth $27,000.
42:12
Uh ask anyone dealing with regular insulin insulin injections, let alone limb amputations, whether it would be worth $27,000 to wave a magic wand and not have type 2 diabetes.
42:21
It's not a hard question.
42:23
And that's just one of a dozen conditions you can lower the risk for.
42:27
Other ones, like not getting breast cancer, might be so valuable that it's hard to even attach numbers. Um So, what about IQ?
42:34
Six extra IQ points, which is Harrisite's estimate with five embryos, is about a quarter of the gap between the average person and the average Ivy League student.
42:42
The benefits of of intelligence are hard to quantify, but it's been shown to have probably causal positive effects on income, mortality, and achievement.
42:51
Probably the income effects alone make up for the cost of the intervention.
42:57
Again, assuming total parent-child altruism and a low discount rate.
42:59
So, if we accept all of these claims and assumptions, the choice seems obvious.
43:05
It probably even accounts It's probably even obvious for governments to pay for all citizens to get these, given how much they'd save on health care costs, says Scott Alexander.
43:14
Uh but in practice, it's complicated.
43:16
Critics have raised both scientific and ethical objections to polygenic embryo screening.
43:21
Most significantly, it's been condemned by various bodies, including the Society for Psychiatric Genetics, the European Society of Human Genetics, and the Behavioral Genetics Society.
43:32
These Their statements are not good.
43:33
They tend towards vague language about how people are more than just their genes, or how no genetic test can be perfect, or how embryo screening is not exactly the same thing as some other form of screening, which uh has a longer history and more proponents.
43:48
Although quote, "Although in general, higher scores mean you are more likely to have a condition, many healthy people will have higher scores.
43:54
Uh other might Other people's Others might develop the condition with even with a low score," says the Society for Psychiatric Genetics, as if they have just blown the lid off of some dastardly conspiracy.
44:07
Uh screening embryos for psychiatric conditions may increase stigma surrounding those diseases, they continue.
44:10
An objection which, taken seriously, could be used to ban every form of medical treatment.
44:14
Because if you take care of something, you remove them from the population, that might increase the stigma, but we should still treat these.
44:20
So, he says, "We will mostly ignore these people and try to imagine the implications of uh, the objections that mildly competent critics might raise, some of which will coincidentally overlap with the content of the non-hypothetical statements.
44:33
The big question he wants to answer is scientific objection, the scientific objection around efficacy.
44:41
Does this Are we sure with this works at all? Are we sure this works?
44:45
So, a typical polygenic score is created by collecting thousands or millions of adult genomes, then matching genetic information with surveys about who has the trait/condition of interest.
44:57
Reputable studies then test these scores on holdout samples, adults who were not used to make the score, to see if they still accurately predict who has the trait/condition.
45:06
Polygenic embryo selection depends on an assumption that the scores, which work in these kinds of retrospective tests, will also work on prospectively on embryos.
45:17
This assumption hasn't been formally proven in studies, which would require years or decades to conduct, um, but seems common sensical.
45:26
The strongest challenge to the application of polygenic scores for embryo selection comes from a recent body of research showing that most scores combine causal genetic effects with population stratification, and therefore can be expected to lose much of their predictive power when comparing two members of the same family.
45:44
Uh, there is an increasing agreement in the field that unless scores are validated within families, headline results like decreases risk of X by Y percent will be large overestimates.
45:56
When I talked to company representatives, they all said that they took accuracy extremely seriously and had various white papers and journal articles where anyone could double click could double check on their methodology.
46:06
But I attended an industry conference a few months ago and the gossip level was comparable to a high school cafeteria.
46:14
Minus the sex rumors, most of the attendees were having their kids via IVF.
46:19
Everyone had some story about someone being careless or fudging their numbers.
46:27
Some of the conflicts broke out into the open on Wednesday when Hera site left stealth and published a white paper and associated blog post.
46:33
They criticized Genomic Prediction for reporting between family rather than within family results.
46:40
And Orchid for smuggling a term for age into their Alzheimer's predictor.
46:46
Unsurprisingly, this makes it work better.
46:51
We'll get to their accusations against Nucleus below.
46:54
Note that this was recent enough that competitors haven't had time to air their own criticisms of Hera site.
47:00
If this happens, I'll try and keep you updated.
47:04
>> and to be clear, this article is from around five months ago. >> Yes.
47:08
And since that time Nucleus has been accused of plagiarizing the the paper. From Hera site. From Hera site. >> From Hera site. >> Yeah.
47:19
And then also accused of stealing IP from Genomic Prediction.
47:23
>> So there's again bunch of different accusations. We'll let Yeah.
47:26
Keon So yeah yeah I mean the goal here is is to just give an opportunity for you know, Kermit and Keon to to answer some questions, try and contextualize it, try and make their case to a broader audience.
47:44
Um I there you know, I've I've read through as much as I could I I can but without actually getting in the lab and rolling up my sleeves, I don't think I could come to a firm conclusion here but I can certainly talk to them on this show and hopefully get some more information that the community can do with what they will.
48:03
So, Sky Alexander concludes this section talking about his strongest opinion of the scientific criticism.
48:12
He says, "Authorities on all sides have cited Alex Young as an authority on how polygenic scores can be confounding or misleading.
48:19
Last week, Alex Young revealed that he had been working with Haris Sites while it was in stealth mode and endorses their research. Three, lol.
48:28
Probably that means Haris Sites products are okay.
48:31
That serves as proof of concept that this technology can work and means other companies claims are at least plausible.
48:39
So, lots of back back and forth and we will be joined by with by Kermit in just a few minutes.
48:45
I actually need to message him and make sure that he has the information.
48:52
Is there anything else that you think would be worthwhile to discuss before we hop on? Let's see.
48:59
>> Yeah, I can I can just go through.
48:59
I mean, the original accusations came from an account called Sichuan Mala. >> Sichuan Mala, yes.
49:07
Uh who wrote an extremely lengthy blog post on on a bunch of the issues that they felt they had found uh with with Nucleus.
49:18
Nucleus ended up firing back and saying that Sichuan Mala was uh or or sort of implying that Sichuan Mala was funded by a competitor or competitive service as well as making these allegations with Kermit.
49:35
They go into uh issues around potentially fictitious customer reviews, which we'll ask Keon about.
49:41
Uh AI generated blog posts, accusations of intellectual property theft, uh saying that the the Nucleus Origin white paper is plagiarized. Yes.
49:54
Saying it that has a bunch of errors.
49:54
Um Nucleus has responded already to Yes. a lot of this stuff.
50:02
Well, our first guest of the show is here.
50:04
Let me tell you about Linear, meet the system for modern software development, purpose-built tool for planning and building products.
50:11
Uh we will bring in Creme from the Restream waiting room into the TVP and Ultra Dome and have him set the table for us.
50:17
Creme, how are you doing? Welcome to the stream. How we doing? Glad to be here, guys. Uh thanks so much. >> me? Good as always. >> Looking good.
50:25
Uh do you do you want to >> I can go face docs if you want. >> Let's do it. Let's do it.
50:29
We can we we can show his actual video this time, which is great.
50:33
We've had him welcome to the show. >> Hey, hey. >> Good to see you. Hey, Don.
50:36
Uh Uh what what what actually kicked this off for you?
50:40
Do you know Sichuan Mala uh separately, independently?
50:44
Did you know that this was coming? Um set the table for us.
50:48
Like, why did Nucleus come to the top of your mind?
50:52
So, can I actually go back to uh Hereticon Please. with this? Yeah. All right.
50:58
So, for about a year, we've told Nucleus about issues with their products.
51:04
Uh it's Can I just actually give a big like I I can I can monologue on this for a minute.
51:08
I can tell you a lot of the details if you want to go into it. >> Go right ahead. Okay.
51:13
So, one of the early things that really peeved a lot of us who are aware of how this tech works is that Nucleus claimed to provide parents with information about rare variants based on microarray files.
51:26
Their website's wording is incredibly ambiguous.
51:28
So, the excuse when I raised this to Kean in person was that they were referring to imputing a child's embryo or a child or an embryo's microarray based data with parental whole genome data.
51:40
But, this is not sufficient for rare variants like they claim it is, and it only works out in very well uh in narrow, well-behaved cases.
51:45
It's not a clean substitute for sequencing an embryo or child. Mhm.
51:48
The rare variant information they can offer is limited and their claim is highly misleading because it sounds like they can achieve coverage of de novo rare variants and ultra rare variants reliably for children and embryos, but they cannot do this because of like crossover that happens during recombination inside the haplotype blocks you're using to do the imputations uh and because of mutations.
52:08
They can't They can't even get high confidence coverage of rare variants more generally, which is what their big claim is about uh in specific wording um with the imputation-based methods they claim to be using.
52:17
So everything they say has a huge error bar on it and it's They shouldn't be advertising it basically.
52:23
It's something incredibly misleading.
52:24
Uh and to give you an idea of So so So just to like actually zoom out, like what do you think they are capable of doing?
52:33
Uh I think they are capable of microarray sequencing.
52:37
I think they're capable >> I mean I I mean I mean in terms that you could advertise to a parent.
52:43
Like you could like like like if Keon went to a prospective parent, the parent says, "I'm doing IVF.
52:49
Uh what can you do for me?"
52:52
>> even ready for large-scale out-of-home advertising?
52:55
Do you think Do you think we're >> It It absolutely is.
52:57
Um the issue is that Nucleus should not be doing it because Nucleus has produced scores that are invalid.
53:03
Uh they're incredibly invalid.
53:05
Like for example, they used an ADHD score um that can included 12 single nucleotide polymorphisms.
53:13
Uh it They claimed it explained 4% of the variance in ADHD, which means it's a pretty good predictor and you can use it to get some improvements on the margins.
53:20
But 12 SNPs means that there's just no way.
53:24
Uh they could explain less than 1% of the variance and the best current ADHD polygenic score uses more than a million SNPs and explains about 1% of the variance.
53:34
So they just basically lied.
53:36
They made up these numbers um and Harris had to go through and throw that, "Oh, actually a bunch of their numbers are made up.
53:43
Yeah, but I mean, how do you know for sure that they're I mean, like that claim, you know, the state of the art was was a million for 1% now it's 12 Nucleus claiming 12 for 4% that seems like a huge exponential you know, growth in efficacy of that particular test, but uh just an exponential progress it is not necessarily evidence of malpractice, right? This isn't progress.
54:09
They've They've never been able to show that they can use 12 snips and no one can. It's impossible.
54:16
They don't explain this much variance.
54:19
It's It's literally physically impossible.
54:20
There's no It's not possible in any way.
54:22
They should never have claimed it.
54:23
They should never provided score reports to people based on it and they did.
54:26
They provided customers with score reports that have to be fake.
54:30
Not fake per se, but they have to be incorrect. They have to be wrong.
54:34
There's no way they could stand up to scrutiny.
54:35
What I'm saying is that after they made this 12 snips to 4% variance claim people have looked and they've shown the latest ADHD polygenic score which is more recent than what they've claimed to have on offer uses more than a million snips and explains about 1% of the variance.
54:49
It's just that's the state of the art.
54:52
They are claiming to have better than the state of the art with completely implausible parameters. It is impossible.
54:59
Okay, so so so back to the original question like what do you think they actually could offer even if they say, hey, you know what?
55:09
We're not We're not necessarily state of the art.
55:11
We're partnering with a lot of different labs.
55:13
We're standing on the shoulders of giants.
55:15
We're using the tools that are available.
55:19
What is a reasonable claim that if they made it, you would be like, yeah, that sounds that sounds reasonable.
55:25
A reasonable claim from Nucleus is that they are pulling polygenic scores from the polygenic score catalog, a publicly available resource that lists a bunch of different polygenic scores from different genome-wide association studies.
55:35
That would be reasonable, but that is not it Why would you go with them then?
55:39
They have nothing unique to offer.
55:41
What they have offered seems to be unique are claims that don't hold up to scrutiny or which are clearly plagiarized from one of their competitors.
55:48
competitors. Yeah, well there I mean there's there are plenty of there are plenty of services that are like, you know, in you know, effectively wrappers around Quest Labs and you know, I get a better UI and it tells me that my cholesterol is in the range and yeah, I know it's just looking up the range from the the data but it has a nice UI and a
56:07
reasonable billing system or whatever and people pay for that and they get you know, they make decisions about their diet based on that and I don't think that there's anything necessarily wrong about providing something that's a commodity or not a scientific breakthrough as long as you're up to as long as you're honest about what you're doing and you're not inflating the results. Yeah,
56:27
Yeah, the problem is that they're inflating the results.
56:29
The problem is that they are effectively making up the results.
56:33
They have actually they've in their latest report claims to basically match Harris I not claimed it directly.
56:39
They've copied their very unique citations which no one else has made.
56:42
They've copied their method and copying their method is incredibly weird.
56:48
You mentioned there that Scott Alexander noted that Alex Young is like seen as a big authority in this space and it's true.
56:55
Alex is seen as like the go-to guy.
56:57
If you want to learn about family-based sampling, trio imputation, if you want to learn about within family validations or imputations or quality control, you ask Alex.
57:05
That is the thing to do and they apparently copied Alex Young's quality control pipeline and it's very unique.
57:12
So doing this is unusual, it's unheard of, it's not likely to have actually been done.
57:20
It is like saying that you woke up one day and here's my morning routine.
57:22
I woke up and today I was John Coogan and I brushed my teeth exactly three times each time around and I went and prayed to my household God.
57:31
I did like all sorts of things.
57:32
It's just it's incredibly specific in a way that is very unlikely to be real.
57:35
They've very likely just kind of mimic exactly what they said, and they said they did it on additional data, but the results came up very close, which is not likely.
57:45
We have strong theoretical genetic reasons to expect that if they had this additional data that Harris site did not, the results would have actually looked different.
57:54
So, they have all these signals that they didn't do the analysis they said they did, and lots of indications that they plagiarized.
58:00
And when Sichuan Mala called them out, they responded.
58:05
The response they generated was amateurish.
58:08
It was kind of embarrassing because they admitted simultaneously to denying they did the plagiarism.
58:15
They admitted to doing the plagiarism.
58:17
They copy They admitted they copied things directly from Harris site.
58:20
They admitted they used resources that are unusual to use the and they effectively got the data from them, and they did nothing that was really unique. Okay. But they did Oh, yeah.
58:31
So, so, so copying might be, you know, looked down upon.
58:33
In tech, people copy stuff all the time.
58:37
There's, you know, stories was copied into Instagram from from Snapchat.
58:44
Different different machine learning architectures are being copied constantly.
58:46
Some stuff can be patented.
58:48
Some stuff is can be trade secrets.
58:51
Are we looking at anything that's that's that goes beyond just like, yeah, it's kind of bad form to copy?
58:55
Or or is this something that's actually like a a problem beyond that?
59:01
Yeah, would you would you be less angry if they copied it perfectly instead of like sort of copied it directionally?
59:10
They copied a lot of it perfectly, and this leads to weird results because they should not have done that given the cohorts they used.
59:15
They used separate data for their validations, so the results should have been different.
59:19
They copied details that made it apparent that they were We're using the data they claimed to, or they were just fudging the results.
59:27
Uh it has to be one of the other.
59:29
There's no way they could have gotten the results they did with the data they did and the copying they did.
59:33
The copying tells us that they lied about something somewhere along the way.
59:37
There is something fishy here, and we don't know what the exact error is.
59:40
We just know they have to have made an error because there's no way that the additional data they had access to was identical to all the data their competitors used. >> Mhm.
59:48
It would have delivered different results. Yeah.
59:50
Have you So So I wanted to ask, they released uh open weight models, the origin models.
59:54
Part of Part of their Part of their They didn't Part of uh Well, part of their pushback against uh Sichuan Mala's critique is that anybody can just download the models and play around with them.
1:00:08
Have you Have you downloaded them?
1:00:10
Have you >> you have to ask for access, and they do not give out access.
1:00:14
We've had multiple people go in and try and ask for access, and they've not received it at all.
1:00:17
And one person who asked for more information was told, "Stop contacting us." Mhm.
1:00:24
So No, they are not open at all.
1:00:24
They're open in the sense that OpenAI is open. They're not very open. Okay. Got it.
1:00:27
Uh shifting gears to the marketing claims, um what what uh stuck out to you there is uh is particularly in need of addressing?
1:00:40
They very much need to uh address the fact that they seem to have fake reviews.
1:00:46
So when they started Nucleus Embryo, they launched it in June.
1:00:49
Um they weren't offering any sort of embryo screening services beforehand.
1:00:54
And if they were, then it would have been How?
1:00:55
They would have had They have to specify what lab they used for all this stuff.
1:00:59
It's There's a lot of details that should go into this that they can't actually specify because they didn't do that.
1:01:04
So they claim to have had customers that have already been served by this.
1:01:07
Well, as everybody knows, it takes about 9 months to serve a customer in this at minimum. Yeah.
1:01:13
And uh it has not been 9 months since >> in 1 month, you know, it takes 9 months to make a baby.
1:01:19
>> So I'd love to see these 3-month-old babies that came out perfectly and were uh you know, made their customers so happy, but uh I don't think they exist.
1:01:27
I think they're not real.
1:01:29
Um so, why do they have these reviews? I don't know.
1:01:32
And the reviews are also they have a lot of fake elements.
1:01:34
There are some that are clearly fake.
1:01:35
So, they use stock images in these reviews to to show the customers who are happy.
1:01:40
>> Which should be clear.
1:01:42
You can imagine a scenario where they use stock imagery and fake names and they put an asterisk and say, "Due to HIPAA compliance reasons, we're not just publicly displaying, you know, the names of any of our clients."
1:01:55
But There's There's no issue revealing this stuff and other companies do actually reveal their real customers.
1:01:59
So, Orchid has revealed real customers.
1:02:01
I've been introduced I've met uh Harris site customers like that.
1:02:04
>> Didn't Jason Carman do a whole video with Norn the first uh in the first uh Orchid baby and this has been like basically making a documentary about that person's journey.
1:02:15
Like, there's no like Yeah.
1:02:15
And And I feel like I see in drug commercials all the time and they'll be like, "This is a real customer who loves this hair loss meds."
1:02:22
And they're like, "Yeah, it looks great."
1:02:23
Yeah, but >> And like there's regulation It seems much harder to get It's probably much harder to get somebody to opt into that than just opt into generally providing Sure.
1:02:32
Uh but clearly clearly clearly it's possible.
1:02:34
I guess >> Is Is Is that Is that Is that legit, you think?
1:02:38
Uh no, they still could not have had the babies in time.
1:02:40
Uh it doesn't fit with the time.
1:02:41
The chronology doesn't work here.
1:02:43
These customer reviews are not really physically possible and I'd like to see an explanation from them because it doesn't make any real sense to me. Mhm.
1:02:49
>> Um they could be I don't know, making some sort of representative review that they maybe hedged in some fine print on some page somewhere, but I haven't seen it and their entire site has been archived now.
1:02:59
So, if that page exists, they'll have to show it to you on the archive.
1:03:03
One thing Jordy and I were debating was uh this this big question of like, is Nucleus making a recommendation or not?
1:03:11
I don't know if this is relevant, but I would love your take on this.
1:03:14
This idea of like uh you know, I see a I see an ad that just says like height is 80% DNA-based or genetically inheritable, and then you go into the dashboard and it just says here's the predicted height and here's the predicted, you know, IQ, and then you make the decision, and it's not necessarily that they're recommending one or the other.
1:03:35
It's more of a diagnostic and you can do with that information what you want.
1:03:38
Does that Does that absolve them of some sort of responsibility there? No, it doesn't.
1:03:43
So, if you still provide wildly inaccurate scores, then it doesn't matter what you're recommending.
1:03:49
You are effectively just recommending something that doesn't matter.
1:03:50
I mean, you are giving them incorrect recommendations.
1:03:53
You have to give some range of uncertainty within the best of your ability.
1:03:58
You have to give them something that is to the greatest extent knowable reliable, and they can't have done that.
1:04:05
They've provided scores that they know They must know were incorrect.
1:04:09
You can't explain 4% of the variance with 12 snips.
1:04:11
It's not really feasible.
1:04:13
It's There's no possibility of it, really.
1:04:14
Um There's also no possibility of getting high-confidence coverage on rare variants to make from the imputation methods they described.
1:04:22
So, they can't make parents certain about this stuff.
1:04:24
Like, they're saying lots of things that are that would require them to basically advance the science 20 years.
1:04:28
They would have to be leaders in innovations here.
1:04:34
And I just kind of doubt it.
1:04:35
The people who are actually leading on the innovations here are Helix.
1:04:39
Orchid is doing so as well.
1:04:39
They're doing a lot of Orchid's whole genome sequencing of embryos is the only one of the industry available to do this.
1:04:47
And Helix's innovation is that they made this stuff low-cost for what they have admitted is a reduction in quality if you get PGTA-based imputation.
1:04:55
Now, but another thing I would do want to mention though is that Nucleus might and we're still look I'm still looking into this.
1:05:01
I'm collecting some patient reports.
1:05:03
I found one for my friend Dylan.
1:05:04
She got a report from Nucleus and she got one from Invitae.
1:05:08
It was a whole genome sequencing thing done, and Nucleus, and I did see the report, nucleus said she had a a a Mendelian disorder, you know, a monogenic disorder caused by one gene, and Invitae said she didn't have it. Okay, this is weird.
1:05:21
So, what's the inconsistency? We don't really know.
1:05:25
Uh Nucleus recently changed this result.
1:05:28
Now, per the law, you have to notify your customers if you change their sequencing results.
1:05:32
This is a clear regulation, the CMS regulates this, and there are that is a potentially major issue that they might be out of compliance with.
1:05:39
And I understand that startups are often out of compliance, and a lot of them fake it till they make it in terms of following the letter of the law, but I'm of the opinion they really shouldn't, especially because this is serious tech with major implications for people, their families, and I mean all future generations of their families.
1:05:57
This is as I think Jordi called it bloodline optimization, and we have these rules in place for a reason.
1:06:03
You need to show you're abiding by them, but they don't have to.
1:06:06
>> do the other players in in the category or how much are they worried about Nucleus just kind of setting back the entire category?
1:06:16
So, I have gotten it's on video, and because Kean tweeted about being in it, I'm allowed to say that we were at a private conference, and I did lead a panel on this topic where Kean was one of the panelists, and the heads of some other embryo selection companies were also on the panel as well.
1:06:31
And I will say that everyone was worried.
1:06:35
Everyone was worried about one of the major uh missteps that actually has already been addressed by another one of the companies, and the series of major missteps made by Nucleus that they have failed to address.
1:06:45
I warned Kean about this stuff many months ago.
1:06:48
I warned him about problems for more than a year now, and he has simply not addressed them.
1:06:54
They have They're still there on the website. You can go find them.
1:06:55
I've archived these pages a few times to see it Are they getting to them?
1:07:00
Are they getting to them? The answer is no.
1:07:01
They are still making claims that are either not possible, not possible with the current tech, uh might be possible in 20 years, or just don't seem consistent with the evidence that they provided to their customers.
1:07:13
Okay, last question, uh and then we're going to hop on with Keion.
1:07:16
Uh I'm a big believer in redemption arcs.
1:07:18
You're obviously unhappy with uh Nucleus's behavior in the uh in the industry.
1:07:24
Uh what does redemption look like?
1:07:27
What would Keion have to do or show you to uh get back in the good graces of the industry in in your good graces?
1:07:37
So, the whole industry knows that Keion has been a problem for a while.
1:07:40
And we actually a lot of people have tried to give him a redemption arc already.
1:07:43
They have tried to come to him.
1:07:45
They have given him clear advice on what he needs to do.
1:07:50
They have told him, "You need to stop making XYZ different claims."
1:07:52
They have told him, "You need to offer scores that are vetted.
1:07:56
You need to be open about your vetting.
1:07:58
You need to be open about your sources.
1:07:59
You need to qualify everything like the other companies do, but you haven't."
1:08:04
He has been told this for a long time.
1:08:06
Um I sent you some pictures earlier today.
1:08:08
You can see the panel we're on.
1:08:09
Uh where clearly like this has been a thing that's come up a lot.
1:08:14
And we wanted to give him the arc already, but it came to this.
1:08:16
It came to a person going online, a blogger deciding, "Hey, I'm going to look into this."
1:08:20
after reading the various blog posts and saying, "Well, shoot."
1:08:24
So, to make up for it, I think they would need to be incredibly open.
1:08:28
And they need to apologize.
1:08:28
And they need to admit to what they did wrong.
1:08:31
And they need to say that they were out of compliance with Clea rules and regulations.
1:08:34
And they need to say, "Hey, we uh might have misstepped here or there.
1:08:38
We didn't know we were doing this." or whatever.
1:08:40
Like whatever it takes to be accurate. Document everything.
1:08:43
Tell us everything you've done.
1:08:45
Tell us all of the missteps. Don't exaggerate. Do not lie.
1:08:49
Just be upfront with everyone and submit yourself to regulation.
1:08:52
Not in the sense that you have to go and tell the regulator you want to like implement whatever new rules and regs, but submit yourself to openness. Be really open.
1:09:01
Stop this whole thing about not telling us your methods, which they've done and Such Malah has documented that in the release blog post. And give us your data. Give your data out.
1:09:10
Stop provide like saying it's going to be available upon request.
1:09:12
It doesn't matter if your competitors have it.
1:09:15
Offer them offer better pricing or something.
1:09:17
You beat them on the margin because we shouldn't have to compete on trusting you.
1:09:20
I'm saying we, I'm not talking I don't have a company in this space, but um everyone should be trusted.
1:09:25
All the companies in the space need to shape up a little bit and they need to be a little more open.
1:09:29
And Keon needs to do that the most.
1:09:31
Thank you for coming on the show and breaking it down for us.
1:09:33
We appreciate you taking the time and walking through all of that.
1:09:38
Have a great rest of your day.
1:09:38
Who knows you might be on the show very soon as this debate continues.
1:09:43
So we really appreciate you taking the time. Thank you. Talk soon. You too guys. Bye.
1:09:48
Before we bring in Keon, let me tell you about Fall.
1:09:50
Build and deploy AI video and image models trusted by millions to power generative media at scale.
1:09:56
And we have Keon from Nucleus in the pre-stream waiting room.
1:10:00
Let's bring him in to the TVP Ultra. Keon, good to see you.
1:10:07
Wish it was less dramatic circumstances, but you know, it makes for good TV and we're happy that you're here and we can chat about this and I mean I'd love to just give you the floor.
1:10:16
I'm sure you saw you know, some of the early segments.
1:10:19
Where do you think it's important to start?
1:10:21
Where do you think it's most important to set the record straight as a first point and then I'm sure we'll have a bunch of questions.
1:10:28
Well, I didn't see what Kremer said.
1:10:28
I was busy helping a patient, but I think the key thing to remember is that Kremer and I we are definitely aligned on doing great science.
1:10:38
At the end of the day that's what we want to do.
1:10:38
We want to serve the patient.
1:10:39
We want to do amazing science.
1:10:41
I think what we're not aligned on is Kremer basically for several months has not disclosed that he's been affiliated with a competitor.
1:10:51
And you know, that wouldn't be so much of a problem unless they're basically concerting together.
1:10:57
And so that's on the Cremo side of things, but honestly that's like the less important thing to me. Yes, I agree.
1:11:03
I think that's less important.
1:11:03
I I So, I I've seen him post positively about your competitors.
1:11:08
I don't I I've not seen any proof that he's actually being paid or has equity in that competitor, but to me it almost doesn't matter.
1:11:16
It could every single post from him and Sichuan Mala could literally be from Nori, Orchid, or someone of one of your competitors.
1:11:23
You still need to address it, right? 100%. Okay, cool.
1:11:26
And so, first and foremost I'm going to say that our science is completely public and it's been completely public.
1:11:31
So, one thing that is like really important to say is that anyone And by the way, we back this point of shared our models with over 15 different entities, which includes by the way people affiliated with our competitors, several of them.
1:11:41
Um that means >> Just to be Just to be Just to be clear, I mean, Cremo said that a number of people that he's aware of have requested access to the models and not been given access and been told to uh stop reaching out.
1:11:55
And so, I do think uh I don't know who's who's utterly inaccurate and false.
1:12:01
There's not one person who has filled in the Nucleus Origin Typeform, which is a Typeform, filled in a Typeform, that has not gotten access to our model weights. Okay.
1:12:11
And by the way, that includes people affiliated with the competitor. Okay.
1:12:14
And so, I think what's really important here is the science is public. Mhm.
1:12:18
The message to the community is go and test it.
1:12:22
In fact, our science is public. The competitor's is not.
1:12:25
So, what I would propose is they should make their science public, and let's have a third-party independently evaluate the rigor, the quality of the science, and let's do it for everyone to see.
1:12:35
Instead of he said, she said, they tit for the tat, you know, put the science out there, have a third-party independently evaluate them.
1:12:45
That is my message to our competitor.
1:12:45
We are happy to stand behind our science, and we know that it's the highest quality science that can exist today.
1:12:53
But by the way, John, that's not even the point either. Please.
1:12:56
>> know what the point actually is? What is?
1:12:57
The point is about the patient.
1:12:57
It's about having the empathy with the patient so that they can know when they do embryonic selection, they can feel comfortable and confident in the results. >> Yeah.
1:13:07
And this Twitter back and forth, this tit-for-tat, this oh this person's race changed on the Nucleus landing page, it's ridiculous. Yeah, yeah.
1:13:16
>> It's really ridiculous.
1:13:16
Well, well >> And so that's my message.
1:13:18
>> Speaking of the the the the patients on the landing pages, uh what about the what about the the chronology here?
1:13:24
This idea that that that there's a review of a baby with 3 months old takes 9 months to work through.
1:13:31
Baby should have happened a year ago.
1:13:32
Was the service available a year ago?
1:13:35
How do you square that particular allegation that the the review the timing of the review just doesn't line up with what Yeah.
1:13:41
must have happened in the real world had they used your service?
1:13:45
Well, there were several claims about the reviews. Let me address each. Please.
1:13:51
First and foremost, obviously as a HIPAA covered entity, we cannot disclose patient name, much less their picture. Okay?
1:13:58
If a patient chooses to, they can publicly endorse the company and they can put their name and their picture.
1:14:04
Otherwise, a patient can submit an anonymous uh anonymous review and then we'll put that according to the landing page.
1:14:10
And so maybe perhaps the people on the Twitter timeline, maybe they never run a company that involves any protections to the patients.
1:14:16
Maybe they don't know about this.
1:14:18
Maybe in like the broader tech community, it's like unfamiliar.
1:14:19
If you've run a software company, wouldn't make sense necessarily to not disclose the patient name.
1:14:24
And I'm going to say to John, this is really important.
1:14:26
Yeah, do you think you have to disclose the fact that you're using an AI generated image?
1:14:30
Is that best practice or is that legally required?
1:14:32
You know, what I think we should do though is now that the community gave us this feedback is we should update and make it more clear.
1:14:40
Hey, this is clearly not a real picture and this is also not a real name. That's reasonable. Okay? Now, fraud, this, that. Guys, really? Come on. Okay, we'll update it.
1:14:51
We'll make sure that the picture and also the names are more clear.
1:14:55
But again, we're HIPAA covered entity.
1:14:56
You can imagine when you launch an embryo product specifically, people do not want their name to be affiliated with it.
1:15:01
I mean, you have anon accounts that don't want their names affiliated with these things.
1:15:04
Imagine a patient actually underg- underwent Nucleus' services.
1:15:08
So, now regarding the timeline thing, that's the second thing you mentioned.
1:15:11
I want to directly address that as well.
1:15:12
Obviously, a company like Nucleus can start providing services to patients earlier than we publicly launched a product.
1:15:18
Moreover, you would imagine the services that you provide to patients would be the ones that actually the the beta services you provide to patients would be the ones you have reviews for.
1:15:27
Obviously, cuz they do the services prior to the company actually launching uh the services publicly. So, that's it. Um that's the answer.
1:15:36
So, you So, you were you were using the service before maybe a year ago or something, then you announced the service, uh and that's why the the the We We had a request, John, to do embryo analysis probably 3 years ago. Yeah.
1:15:50
I would wager that that, you know, that was actually was probably one of the first times before any of these companies to actually provide a sort of this sort of services.
1:15:58
So, we didn't think about this for a long, long time. Yeah, of course. Yeah, it makes sense.
1:16:02
It's a very logical place to go.
1:16:04
It's also a very competitive industry.
1:16:06
There's a bunch of reasons why you'd want to play in that space.
1:16:09
What about the the del- the delta or the perceived gap between the marketing claims, what's on the billboard, what's in the New York subway?
1:16:18
I'm seeing 50% IQ, 80% height.
1:16:23
They feel like bold claims.
1:16:23
What's actually possible?
1:16:25
What What can customers actually get from Nucleus today?
1:16:30
What could they get a year ago when you were beta testing the product?
1:16:34
What What I I want to I want to interrogate the the the gap there.
1:16:39
So, to be clear, there is no gap. Right. Have a healthier baby. Have a taller baby. Have a smarter baby. IQ is 50% genetic. Height is 80% genetic.
1:16:51
These are just facts of the matter.
1:16:54
The latter two are heritability estimates. They are what they are.
1:16:58
The former two are basically describing what you can do with nucleus. Yeah.
1:17:02
What what I think is interesting here is actually broader commentary.
1:17:06
Nucleus is bringing the science mainstream.
1:17:09
We are taking it out of the little echelons of the rationalist community, the little echelons of the the the the the scholars going back and forth at it, out of the Twitter alleyway, and we're bringing it to the actual people who will benefit and use these services.
1:17:20
I cannot tell you when you actually talk to a patient, not somebody on tech Twitter.
1:17:26
When you talk to a patient, they have no idea this technology exists. Yeah.
1:17:29
And the first time they discover it is when they go and they actually see the campaign in the subway, and now I think it's personally good for the entire industry, right?
1:17:36
Where you actually bring broader awareness.
1:17:38
It lets everybody, us, our competitors, and makes this actually more and more into a space. It's very early.
1:17:44
My my advice to our competitors is focus on serving your patients.
1:17:49
Because at the end of the day, the market's huge, and this market is extremely in its infancy. Mhm.
1:17:54
Um And I and I and I think they would push back and say the industry can't afford to be sloppy, and I think that the >> the I think it's a fair allegation that some of the ways that Nucleus materials have been presented have been sloppy.
1:18:17
Uh Do you I guess one question I have is >> Jordan, what are you specifically talking about? What has been sloppy?
1:18:24
Uh specifically the like the reviews the reviews are sloppy.
1:18:27
I I I totally understand using using an AI image or using a stock image without was it clear that that this is anonymized because of HIPAA.
1:18:37
If it just said if it if it said at the bottom and there's a little asterisk that said anonymized because of HIPAA, I think everyone would be like, "Oh yeah, that makes sense."
1:18:44
Like they made a choice and I think that's like the first thing that I would count as like sloppy. That's that's fair.
1:18:52
We're going to update that.
1:18:52
Do you do you think that, you know, that's proportional to the temperature on Twitter?
1:18:57
I think I think this is the most politically charged category in technology.
1:19:03
And you can't afford to you can't like basically the industry as a whole I don't think can afford to make a lot of mistakes, right?
1:19:10
These are people These this is the These is the going to involve the health of of the children of all the industries, you know, clients.
1:19:22
There were allegations too that you guys were sort of had updated a test result on the fly.
1:19:26
I have no idea if this is true, but I saw the claim going around somebody had gotten a certain test result and then it had been updated two weeks later.
1:19:36
>> Yeah, what's going on there? >> What's happening?
1:19:37
That's super interesting.
1:19:38
Nucleus, remember guys, unlike these other players, we've served thousands of patients.
1:19:43
Anyone can go on our website right now and use our product.
1:19:45
They can see our services, right?
1:19:46
I mean, I think it's really funny what's flying around when someone can just go and buy a DNA kit and see the product for themselves. Yeah. Okay.
1:19:52
So, the idea that we'll be updating a model, we've updated models for the last several years.
1:19:57
I mean, results will change. We make that very clear.
1:20:01
And by the way, any embryo selection company, Nucleus is full stack.
1:20:04
We do adult DNA testing analysis.
1:20:06
We also do the embryo and also do a full end-to-end IVF experience.
1:20:08
We're kind of multi-product.
1:20:10
But these models will evolve.
1:20:12
One one point here is an important John Venture.
1:20:15
People have a very good intuition when it comes to AI that ChatGPT is going to be better next week.
1:20:18
Grok's going to be better two months from now.
1:20:20
The same expectation has to be communicated to the genetic optimization industry.
1:20:24
The main limitation for building polygenic predictors are is data. It's a data problem.
1:20:31
And so, what's going to happen is all the different polygenic predict predictors are going to be approximately equivalent, okay?
1:20:39
Until we get more and more and more and more and more data, or people get more and more access to data.
1:20:43
That is the fundamental bottleneck of the industry.
1:20:47
In other words, similar to actually the AI situation, all the value is going to shift downstream to the application layer.
1:20:53
The reason why the reason why people are so upset is because Nucleus has excellent science, rigorous science, and we have >> Mhm. >> Mhm. >> Mhm.
1:29:06
>> here actually being get let the models in everyone's hands because that seems to be a fundamental disagreement here.
1:29:13
Uh you're saying that you'll give them the the model.
1:30:20
>> wasn't filled in my nucleus.
1:30:20
I'm going to do and the last thing I think which is really important as sometimes you know I on Twitter this this one personally hurt me. for six years.
1:31:02
I don't know how many years trying to get that right.
1:31:04
When you see the embryo on your smartphone, it's easy to look at that and just write it off, right?
1:31:10
I encourage everyone And also honestly, I think I need also more mindful.
1:31:42
I mean clearly, I've inadvertently I've pissed some people off.
1:31:48
And that's understandable.
1:31:48
That that feedback's >> the time to hop on and and and set the record straight, uh, give your side of the story and explain, uh, what you think folks are getting wrong on the >> They just spammed real life.
1:35:49
They totally They spammed the physical world.
1:35:53
>> And it was getting worse and worse.
1:35:53
Uh They were They were spamming Los Angeles.
1:35:55
And then I live in a suburb of Los Angeles, Pasadena.
1:35:57
And then one day, I'm driving around my hometown, which is very quaint and sort of out of the loop.
1:36:03
It's not It's not the San Francisco hubbub. It's not Teapot. It's It's Pasadena. It's very chill suburb. And I see a friend. com ad billboard.
1:36:10
I'm like, I can't believe you followed me here.
1:36:14
It's following me everywhere.
1:36:14
It follows me on the internet. Follows me to LA. Follows me to Pas