Anthropic上場、ChatGPTに財布機能、AIによる創薬
Anthropic goes public, ChatGPT gets a wallet & AI designs new drugs
Anthropic上場、ChatGPTに財布機能、AIによる創薬
Mixture of Experts
要約
IBMの専門家3名とホストのDavid Zaxが、Anthropicの上場とその資金の使い道(多様なAIチップへの投資)、OpenAIのChatGPT Walletに代表されるエージェントによる購買、Isomorphic Labsの50億ドル調達とAI創薬の現在地を議論した。ハードウェアの多様化、ワンタイムカードによるリスク限定、シミュレーションによる創薬の成功率向上が主な論点である。
この回のポイント
- ●Anthropicの上場資金はハードウェア供給量を上回る規模で、各社のチップを幅広く使う動きが業界の多様化につながるとSridharは述べた。
- ●AIによるコード最適化で、これまで使い切れなかった多様なハードを活用できるようになったという見方が示された。
- ●ChatGPT Walletはワンタイムカード番号とパスキーによる承認で、エージェントや店舗に実カード情報を渡さない設計だと紹介された。
- ●最大の懸念は不正利用で、Kushは「レモン市場」の問題を挙げて慎重な姿勢を示し、Sridharは短時間で失効する認証情報で被害を限定できると述べた。
- ●Isomorphic Labsは50億ドルのシリーズCを発表した。シミュレーションで候補を絞ることで、費用の大半を占める治験の成功率が上がるとの指摘があった。
- ●創薬では候補分子の生成より、毒性や溶解性などの予測と制御された生成が難しいとKushは説明した。
章立て
Grandthropicと AIとゲームの歴史
GTA VIとAnthropicの上場が重なった話題から、チェッカーやDeep Blue、NVIDIAの出自まで、ゲームとAIの関係を雑談する。
上場資金とハードウェア多様化
資金がハードウェア供給を上回る中、各種チップを使い分ける動きと、AIによる最適化の役割を議論する。アイスクリーム屋台の比喩も登場する。
垂直統合と循環的な資金
AmazonとGoogleの出資比率、小規模プレイヤーへの影響、OpenAIの垂直統合戦略に触れる。
ChatGPT Walletとエージェント購買
ワンタイムカードと人間の承認を備えたChatGPT Walletを紹介し、パーソナルエージェントの普及と不正リスクを語る。
各自のエージェント利用と安全性
Claudeに購入を任せた体験談、レモン市場への懸念、短時間で失効する認証情報の考え方を共有する。
Isomorphic Labsと創薬の加速
50億ドルの調達を受け、シミュレーションによる絞り込みが治験成功率とROIを高める仕組みを解説する。
生体シミュレーションと制御生成
AlphaFoldや分子生成を例に、毒性や溶解性の予測と制御された生成が鍵であることを説明する。
解説記事
IBMのポッドキャスト「Mixture of Experts」の今回の回では、ホストのDavid Zaxと、IBM FellowのKush Varshney、ソフトウェアエンジニアリング部門ディレクターのSridhar Rajagopalan、IBM ConsultingのAlexis Zamkowが、Anthropicの上場、エージェントによる購買、AI創薬の3つを取り上げた。
Anthropic上場とチップの多様化
番組は、Anthropicの上場日(11月19日)がGTA VIの発売日と重なり、ネット上で「Grandthropic」と呼ばれていることから始まった。ゲームとAIの深い関係として、機械学習という言葉がチェッカーのプログラムから生まれたこと、Deep Blue、そしてNVIDIAがゲーム向けGPUから成長したことが語られた。
本題でSridharは、資金量がハードウェアの供給量を上回っているため、できるだけ幅広いハードを使うことが重要になると述べた。以前は多様なハードを使い切るのは難しかったが、AIによる最適化が得意とする領域であり、規模が大きくなれば小さな最適化の積み重ねが効いてくるという見方だ。Kushはこれを「ハードウェアのアイスクリーム屋台の全フレーバーを買う」と表現し、NVIDIA、AMD、Amazon、Googleのチップを使い分ける姿を描いた。
Alexisは、AmazonとGoogleの出資比率(それぞれ14%、13%と紹介)に触れ、循環的な資金の流れや、小規模プレイヤーへの影響を懸念した。またOpenAIのSam Altmanが、自社チップ・自社LLM・自社プロダクトという垂直統合を差別化点に挙げていることにも言及があった。
ChatGPT Walletとエージェント購買
OpenAIはChatGPT Walletを発表した。エージェントが商品を探してカートに入れ、購入まで行う機能で、購入ごとにワンタイムのカード番号が使われ、エージェントも店舗も実際のカード番号を見ない。さらにパスキー(顔認証や指紋)で人間が承認する仕組みだという。
Alexisは、この分野は新しくないとして、Facebookのパーソナルエージェントが先行していると紹介した。ダウンロード数は1090万件、ユーザーの25%が購買に使っているという数字を挙げている。一方、最近の調査では最大の懸念は不正利用だったとも述べた。チケット販売のようにエージェント同士が競う世界では、エージェントなしでは買えなくなるという見方も出た。
DavidはClaudeに依頼してケーブルや850ドルのプロジェクターを購入した体験を披露した。これに対しKushは、売り手も買い手も相手を知らない状況は経済学の「レモン市場」に近づくとして慎重な立場を取った。
認証情報を限定するという考え方
Sridharは、5分で失効する使い捨てのクレジットカードなら被害を限定できるため、積極的に使えると述べた。これはSRE(サイト信頼性エンジニアリング)チームが、長期有効な認証情報を避けるために使ってきた考え方であり、クレジットカード業界にも最近導入されたという説明である。すでに実証された技術が新しい領域に移ることで、利用範囲が広がるという見方を示した。
Isomorphic Labsの調達とAI創薬
Google DeepMindからスピンアウトしたIsomorphic Labsは、11月17日に50億ドルのシリーズC調達を発表した。資金は創薬エンジン、生物学シミュレーションモデル、初期候補の臨床試験に向けた基盤に充てられるとされる。
Sridharは、AIの効果を2点で説明した。第1に、膨大な選択肢を絞り込む作業を、ウェットラボではなくシミュレーションで行えること。第2に、シミュレーションの精度向上により、選んだ候補の成功確率が高まること。費用の大半は人を対象とする治験にあるため、成功率が上がれば投資対効果が大きく改善すると述べた。Alexisは、バイオマーカーによる被験者選定などで治験の第2段階も速められる可能性があるが、まだ実証されていないと補足した。
Kushは、候補分子を大量に作るのは簡単で、難しいのは選別だと指摘した。がんを殺す物質は多いが、人も殺してしまう。そのため毒性や溶解性を事前に予測する必要がある。IBM Researchのモデル「Molformer」は、溶解性が高く毒性のない分子だけを生成するよう制御する手法だという。原子や結合の物理シミュレーションだけでなく、他の分子の大量のデータから性質を予測する方法もあると説明した。
まとめ
今回の議論は、AIの資金とインフラの競争、AIの自律的な行動への信頼、そして実世界の科学への応用という3つの流れをつなげて見せた。日本の読者にとっては、特定ベンダーに依存しないハードウェア活用の重要性、エージェントに権限を与える際の「短時間・限定的な認証情報」という設計原則、創薬での「選別精度の向上が治験コストに効く」という視点が注目点になる。いずれも出演者の見解であり、今後の実証が待たれる。
文字起こし(英語・自動生成)
Previously, you had a lot of different kinds of hardware, but it's very hard to actually use it all. But that's the sort of thing that AI is very good at optimizing. All that and more on this week's Mixture of Experts. Hello and welcome to this week's Mixture of Experts, a podcast for IBM in which we gather a few IBM experts to talk about the week in AI news. I'm your host, David Zacks, a senior writer at IBM Think, and I'm joined today by three special guests. We're joined by Kush Barshley, IBM Fellow at IBM Research, Sridhar Rajagopalan, Director of Software Engineering, and Alexis Samco, Intelligent Marketing Global Offering Lead, IBM Consulting. Thank you for joining this week. We're going to look at three stories today. We're going to talk about drug discovery and AI. We're going to talk about agents and money, but we're going to begin with the big story of the week, Granthropic. This is a massive technological story, a massive financial story, a massive business story, a massive cultural story.
I mean, a few times in the history of technology have we reached a milestone like this. I'm referring, of course, to the release of Grand Theft Auto 6, which I'll be playing on my PlayStation 5 very soon. I don't know if this is a trillion dollar product, but I'm sure it should be. Incidentally, a minor AI company is having its IPO this week. It's called Anthropic. I don't know much about it. But, of course, the Internet is making much hay out of this guest. You know, it's only been a few years since we had great fun juxtaposing the frivolity of Barbie with the seriousness of Oppenheimer in a cultural event that the Internet dubbed Barb-enheimer. Now the Internet is dubbing Thursday, November 19th, Grandthropic. Raise your hand if you are a gamer on this panel. Am I the only one? No. Okay. Sorry. No one cares but me. I'm the only one torn about the significance of this. So I will park most of my jokes about the play styles of the various anthropic founders
and the jokes I've seen online about Dario Amode very patiently following all the rules, refusing to blow through any of the traffic lights in Vice City. Yeah, but the topic of guardrails, David, I mean, that is the absolute alignment in guardrails is the huge topic that everyone is talking about. Grand Theft Auto, there's no guardrails. And people are really concerned about where all of these frontier models are going with regards to alignment guardrails. So I think, you know, it is kind of timely, right? That's true. I think we need to agree there should be pacing agreements about how fast we're all willing to play Grand Theft Auto VI, because I can't turn to it for a few weeks, so I'm worried that if we don't pace the frontier of this game, it's going to get spoiled for me, and I think it's really unfair and unsafe. A few other parallels we could draw between gaming and AI before we talk about the anthropocopio proper. But I do think, firstly, it is interesting to note that in the history of AI, I read Empire of AI, the book that came out last year or a couple years ago,
and so much about the early quest to make artificial intelligence was about simulating game playing. And, Kush, maybe you can speak about this, but increasingly also there's an interest in, from Jan LeCun and others about models that don't just infer things about what may or may not be the world from language, but that are increasingly run through training meant to simulate something like an understanding of the world. Yeah, I mean, the history of AI is very much a history of using games, creating worlds, and kind of doing things in those. So, I mean, here at IBM, of course, the word machine learning was developed as part of a system to play checkers. And then Deep Blue, of course, was a great success in AI for playing chess. And then things have gone on from there. But, yeah, I think whenever you're kind of in this other world,
it's the place where you can experiment, where intelligences of different kinds can kind of take over or be used in different ways. And I think that's exactly the appeal of a lot of the games as well. And if you think of the start of NVIDIA, I mean, it was the graphical chips, his ability to help gamers, you know, play interactive games with great fidelity. And that was the start of Jensen's success. So the technology foundations have been built a lot around what was needed for gamers to have a really immersive experience. I really think you, when you think about it, who is the Oppenheimer in this world? Is it Dario Amode for building Claude? Or is it me in 1996 playing 007 James Bond and creating a market for the GPUs that would yield the industry? I think it's a tough case. Who really gave us AI? I think David Zux wins.
I certainly think I would like to persuade Dario that I or, you know, whoever's running NVIDIA now that I deserve a chunk of that cash. Sweetheart, moving more squarely into speaking of cash, Anthropik's about to get a lot of it. And I gather that you have been keeping an eye on you have an interest in how they might spend that and specifically some of the hardware investments that they've declared that they're planning to make. Can you tell me a little bit about what you're tracking there? Yes, I mean, I think partly, you know, there's a lot of money, not just in this funding event, but other ones, right? And the amount of money that there is exceeds the supply of hardware there is, right? And therefore, being as wide as possible in the use of hardware is very important, which is why there's diversity there. people are acquiring and using different kinds of hardware for purposes which are tuned for those kinds of hardware.
And the amazing thing is that at the same time, previously, if you had a lot of different kinds of hardware, it's very hard to actually use it all. But that's the sort of thing that AI is very good at optimizing. And when you're doing things at this scale, hundreds of small optimizations can add up to be a very, very important thing. And in fact, we see that a lot of the code that's being written is very tuned to the hardware-software interface. And is exploiting all the corner cases of that interface and doing it really really well And so I think we are in a world where things that were not possible a while ago are more than possible now, in fact, are absolutely essential to do because of the combination of economic and supply chain as supply chain issues
and the possibility of optimizing things at the end of this, which, I don't know, 20 years ago, we couldn't do at all, I think, the last time we faced something like this. So, David, this was an interesting bit of information. I was just looking at ownership of Anthropoc and where it's going forward. Amazon, 14% ownership now in this new model. Google, 13% ownership. And it's all about the compute that's going to be needed to support their massive growth. And it's interesting because then you start talking about circular finance and you talk about competitiveness. And I guess some of the concerns being, you know, are we setting ourselves up at a very small ecosystem which might just sort of dominate the industry? How did that affect the smaller players? I mean, it's fabulous that they're coming out and being the beacon of innovation and success. At the same time, this compute issue is going to continue to be something that we all have to think about.
Regular listeners of the show know that I love to ground these things in what Reddit calls Eli5 terms, explain like I'm 5 terms. So just, we've spoken to the experts and semi-technical people in terms of compute and chips and so on. Sweetheart, you had a wonderful way of saying this yesterday on Slack. and Thropic is planning to buy all the flavors on the entire hardware ice cream truck with their IPO money. It's going to be interesting to see how they use the capacity. Can you help me unpack what are these flavors? We have chocolate, vanilla, strawberry. What do these map to for the Explain Like I'm 5 audience? So I think of all these chips, right? I mean, the chips that come from NVIDIA. Chocolate, there's mint chocolate. AMD is selling mint chocolate and Amazon is selling its own flavor of strawberry, as is Google. and so on, right? So, yeah. And they're all slightly different, and some of them are very different. The amount of money you have is such that,
you know, you can't spend it on only one thing, right? And so you're going to be spending it on all of these things. And then that's basically, I think, what's going to happen. And it's a good thing, right? At least it helps the industry diversify more. And I'm sure there are going to be things like other anthropic ice cream trucks. So I think overall, I think it's good for the consumer to start seeing this variety and having to deal with it. So when you talk about folks being in a chip space with vertical integration, Sam Altman was on DealBook recently. yesterday talking about how his differentiation is that they're going to build their own chips, they're going to have their own LLMs, they're going to have their own products, and consumer and business both in line.
So this whole vertical integration is a very interesting competitive positioning to have. They're all going to eventually have their own views of it, but to the point which you made about everybody having their chips, there will be a lot of diversification outside of what NVIDIA is offering. Great. Wonderful. With that, I will move us to our next topic. So OpenAI launched something called ChatGPT Wallet this week on the 13th. So specifically, it's letting its so-called dots agents, these cute little agentic instances of ChatGPT, find products, fill a basket on Amazon.com or whatever it may be, and check out for you. um and now as i understand that each purchase uses a one-time card number so it's a bit of a kind of constrained and like cryptographic component here uh so the agent and the store never see your real card um and also a human is in the loop you need to approve every purchase with a pass key like a face id moment or a fingerprint something like that um i haven't tried it out yet
um but alexis you were pointing out that this is um not entirely new it's new to open ai and to ChatGPT, but agentic commerce and this trend of people and firms allowing agents to make purchases for them that a few firms have gotten down, you know, a little bit down the line here. So maybe you could kind of summarize for us what's going on in this space and what you're thinking about it. So this is a recent development of personal agents. I mean, we've been talking about AI search agents and we can speak a little bit more about that in a bit. But moving to the personal agents, So the ones that are literally on your desktop managing your environment for you. The first out of the gate was news. And no surprise, as you know, Facebook is very consumer-focused. And they came out with their personal agent news back in October. And it was an incredible launch. There's right now 10.9 million downloads globally. 10 million of those are in the U.S., very dominant in that market.
and folks are using this agent to organize their lives, to make reminders of what they needed to do to purchase products. It was recently mentioned that 25% of users are actually using it for purchases, which I have to admit is aligned to what we're seeing sort of in general in the market and the way of acceptance of agents purchasing. There's just a recent study that came out just this week that talked about, you know, what was the biggest issue? Why are people, you know, concerned about these agents buying products for them? And it is fraud. No surprise. And the issues are not just on the consumer side. It's also on the business side. And think of what happens to kind of experiences like buying tailors with tickets. I mean, these things were selling out in 10 minutes. Now with agents shopping on your behalf, I mean, they'll sell out in, you know, three seconds or something. And in fact, you won't be able to get tickets unless you have an agent for them. My agent will compete with your agent, Streeter. The scarcity, the constraint that is Taylor Swift's supply is well known
and is something that I hope that the AI labs are working on by creating digital twins, triplets, and so on of Taylor Swift. Kush I curious to know sometimes I like to be nosy and proud of people agentic behaviors I share mine before I ask you to you know consider sharing yours Like here where I am on agentic commerce I run my life increasingly through Claude to a degree that I hope my wife isn't fully aware of. Thankfully, she does not watch this podcast. But like I, in the last month, I've begun letting Claude buy stuff for me. I had an interesting experience the other day where I was in my car, I needed a certain cable, I needed Claude to help me decide what cable to get. And then I realized that I had a connection. I was on my iPhone, but I realized that a connection live to my laptop. And I realized that my laptop had Claude in Chrome, and that my Chrome was signed into Amazon. And I realized this because Claude realized this. And Claude was like, hey, if you want, I can just I can put this in your cart and buy it. Just say yes. And I'll click yes. So I didn't have to do a separate face ID moment.
And it bought me the cord. Lo and behold, the cord came a day later. It was exactly what I I needed. I sort of feel like the fraud detection and alert, I feel like Chase would alert me if that had happened, right? So I've made my peace with this. I was so empowered by this that I had to make an $850 purchase the other day, which was to choose a new projector to run in the home. And I basically allowed Claude to make that decision for me. And that projector is right here as we speak. I've yet to test and make sure that it suits my needs. But, you know, among the things I had to do was to run the risk analysis of like worst case scenario, you can turn it back, run it back in 30 days. And I did fact check that. I went on eBay to visually confirm that that was true. So that's where I am here. Kush, are you experimenting with this at all? And yeah, everyone on the panel, I'm just curious where you are and how you're experimenting with this, if at all. Yeah, no, I'm very old fashioned. So yeah, like I drove to Home Depot a couple of days ago to like actually buy wood or whatever, like I mean, that sort of stuff, right?
So, yeah, no, I think it's like there's this whole concept in economics called the market for lemons. And I think we're kind of getting to that point a little bit in the sense that, like, I don't know what it is that I'm going to get. And the seller doesn't know who I am. And I think all of this can kind of just devolve into like really, really bad stuff happening at some point. And it's just like the world is moving too fast for me personally. So, yeah, I'm kind of like too cautious, I would say. Our other esteemed panelists, are you willing to divulge? Does anyone want to come out, you know, and fight Kush on this? Or should I have to do it? No, I'm actually, I'm very tech forward in these terms. But I'm also very security conscious. So in some sense, I'm an early user of both the tech and of the things that stop the tech doing bad things.
So in particular, in this particular story, one of the features that I like is I can issue a very short-term credential to do a very specific thing, which is this instant credit card which has only one purpose and is limited to a five-minute window. After that, the credit card doesn't exist and can't be abused anymore. And that gives me the confidence to go ahead and be tech forward in a way because I know that my damage is limited. Yes, of course, there are lots of issues, but I tend to be lazy enough not to chase after refunds when I get a bad product. I just bin it and I get on with my life. But if I know that there's a limit to the amount of damage that can be caused, then I'm more confident with my lifestyle of being a bit brave. So you're referring to the one-time card number that is part of the architecture of this DOTS infrastructure, right?
And I think, correct me if I'm wrong, Sarah, with your background in data and security, do they call this like just-in-time credentialing? Conceptually, there's a term of art for this notion of really limiting the scope to a point-in-time, one-off disposable credit card, and that is a security best practice that's at work here. Yes. Those who operate large systems, SRE teams, have been using technology like this for a long time so that they can make changes to services which impact all of us. Because you do not want to have long-term exposed credentials that allow people to make changes to services like Google, for example. It will cause a big impact. The credit card industry has not had the ability to do these kinds of things until fairly recently. It is there now, and people are beginning to learn the lessons that were originally learned in kind of IT operations, if you will, and transport them into this new world.
And that allows all of us to start behaving in a slightly different way because not only is there monitoring and ability to kind of alert you when something bad is happening, but there's also the ability to limit the damage of a bad thing. And I think this is technology that has already been invented and proven, being transported into a new world, and therefore kind of enabling a broader set of use cases. Well, thank you. I will report back on whether this $850 projector I bought was a lemon or not, Kush, and then you can laugh at me if I guess. Yes. This week, some news. It's some funding news, but it's an interesting milestone. Isomorphic Labs, which is the drug discovery company spun out of Google DeepMind, announced a U.S. $5 billion series C on November 17th,
co-led by Greenoaks and Sequoia, major VC firms. Now, the money will go to its AI drug design engine and more advanced biological simulation models. as well as the infrastructure to move its first candidates into clinical trials. It has reported some positive early results in cancer and immune system programs. Now, of course, even if you could just kind of rocket blast the drug discovery and drug candidate pipeline and condense it to five seconds, you would still have a slow testing regime, FDA, regulatory controlled, there are certain things that AI cannot speed up. But my question for you, the more technical and frankly smarter guests on this program than your host whose one skill is handing it up for the camera Help me to understand what component of drug discovery exactly and again if you can
and kind of explain like on five terms, like what exactly is AI able to do faster that you couldn't do 10 years prior, five years prior in kind of wet labs? What is the bottleneck that is most getting un-bottlenecked here? And maybe we'll start with you, 3R, because in our previous conversations, you were helping me to understand this well. Yeah, I think there are two things, right? One is the design problem itself, which is accelerated a lot by being able to search a large set of options and come up with confidence of this particular option is worth chasing and that one is not worth chasing. and so being able to pool down a very large search space into a smaller set of candidates and then only work on the smaller set of them. Previously, all of that had to be done in the wet lab. Now it can be done in a world of simulation. The second thing is that
as simulation technologies get better, the confidence itself that you have in your winning solutions out of the gate is much higher. So in some sense, the chance that something that you come up with, which is a selection from a thousand things of the most possible and most relevant candidate, making it to the end gate is much higher. So you're preconditioning to a much higher likelihood of success by doing this. And really, that is triggering a much larger return on investment because the bulk of the cost is in the human trial phase. And if you know that your chance of success in the human trial phase is much higher, then it makes the investment more worth it. So I think that we're going to see the same amount of money in drug design go a much longer way because of all these...
acceleration in terms of the number of things you can do, but also acceleration in the sense that we can take more short of bets going forward. That's fascinating. I've never heard that said before. So not only does the first stage go faster, but because it's fundamentally going better, we have a hunch that when we do get into the slow, wet, meat-based trial, that we'll have a higher kind of candidate success rate. And so effectively, that part that I posited was resistant to acceleration also may get accelerated. I think in addition, so absolutely agree with Shredar that it's going to reduce the failures up front in that first phase of discovery. So we know they can discover the drugs to a certain extent, but will they work? And that's that second phase. If you start thinking about AI beyond the actual drug itself, but looking for the right biomarkers for the right individuals to participate in these trials, Finding the right trial audiences is important, the biomarkers, but also keeping the folks that are the legislators and the folks that need to be calm and confident about the outcomes, recognizing that you're putting as much synthetic insight into every step of the process with the due diligence of analytics, I think could also increase the speed of that second phase.
but it hasn't yet been proven out, but I do think there's the potential. Kushan, I'm wondering if you can help me understand the, like, biological simulation component of this. I mean, for my naive mind, the temptation is to just say, oh, Demis Hassabis of DeepMind, he got the 2024 Nobel Prize in Chemistry for AlphaFold, which predicts the shapes of proteins, so maybe the AI is just better at, like, permutating the shapes of proteins that could work on drugs. But there's this whole other component, which is not just about drug discovery, but it's about, with increasing fidelity, simulating the biological environment in which drugs work, namely the human body. Could you help get me a little bit smarter in the Eli 5 way that I love that you're so helpful with? What is the frontier in simulating, creating digital simulations of the body, essentially? Yeah, so like you're saying, I mean, it's actually, and Sridhar said this, right? I mean, it's really easy to create quadrillions, quintillions of possible candidates, right?
And even like 13, 14 years ago, we did this project called Chef Watson and it was for recipes for food. But same idea, it's computational creativity. And the selection part is the hard part, as Reader was mentioning, right? And with drugs, efficacy is usually not the hard thing. Like there's tons of stuff that will kill cancer, but they'll also kill you. Right. So the problem is how do you know in advance or have some prediction of the toxicity of the drug as well? And also other things that make it more drug like solubility so that it can actually be administered and this sort of stuff. And so, again, we had a model ourselves at IBM Research called Molformer, and it's called control generation in the sense that you're not just creating these quadrillions or quintillions of molecules. You're controlling them so that you're only getting likely things that are soluble, that are non-toxic and so forth.
And it's all about, I mean, kind of how can you have data or other sort of ways of predicting these other properties of the molecules that will get generated. So some of it is, I mean, you can think of it as like a physical simulation, but other things, it's just a lot of data of other molecules and kind of physical chemical properties that they might have and how that relates to these other properties. And then just go from there. So it's not necessarily that you're like actually like having these models, like in a chemistry model sense, like these atoms and bonds and stuff, and you're doing something to those to know what's going to happen. But just with enough data, you can kind of make the same predictions. This is a wonderful, enriching discussion. I learned a lot. If you are watching this on YouTube, folks, it really helps if you can hit like, comment, subscribe. Thank you so much to my guests, Alexis Rehart-Klisch, and we will see you next week on IBM's Mixture of Experts.
番組の概要欄(原文)
Visit Mixture of Experts podcast page to get more AI content → https://www.ibm.com/think/podcasts/mixture-of-experts Grand Theft Auto VI and Anthropic's trillion-dollar IPO have one thing in common: November 19. The internet has dubbed it “Grandthropic.” But beyond the memes, what does Anthropic's public debut mean for the AI chip market? This week on Mixture of Experts, host David Zax is joined by Alexis Zamkow, Sridhar Rajagopalan and Kush Varshney to discuss this and more. Next, OpenAI...
出典・この記事について
- 番組
- Mixture of Experts (IBM)
- 元の回
- Anthropic goes public, ChatGPT gets a wallet & AI designs new drugs
- 配信日
- 要約作成日
- 公式サイト
- mediacenter.ibm.com/playlist/dedicated/403892542/1_x5yvmanw
この要約・解説記事は、元の音声を文字起こしした内容をもとに AI(Claude)が日本語で作成したものです。 固有名詞や数値に誤りを含む場合があるため、正確な内容は元の音声でご確認ください。制作方法について
関連エピソード

Gemini 4 Neon を試すべき理由——Fable 派にも「セカンドオピニオン」としての価値
The AI Daily Brief: Artificial Intelligence News and Analysis

GTA VIは人間の手による金字塔だ
Hard Fork

OpenAIの新モデルAstraは評判が散々、でもGemini 4 Neonは意外と良い?
AI For Humans: Weekly AI News, Tools & Trends

Anthropic歴史的IPO、ジェンセン復帰ならず、Gemini 4 NeonがGPT 6.2 Sol/Astraに迫る|ThursdAI 11月19日
ThursdAI - Highest signal weekly AI news show

Huang騒動、Anthropic上場、そしてMetaのMuse――EP99.40
This Day in AI Podcast

Anthropic の1.4兆ドルIPO:人類史上最大の富の創出イベントか
Moonshots with Peter Diamandis