なぜ Anthropic の IPO は失敗するのか――Gary Marcus が語る弱気論

Why The Anthropic IPO Will Fail — With Gary Marcus

Big Technology Podcast59分
なぜ Anthropic の IPO は失敗するのか――Gary Marcus が語る弱気論

なぜ Anthropic の IPO は失敗するのか――Gary Marcus が語る弱気論

Big Technology Podcast

0:0059:06

要約

NYU 名誉教授の Gary Marcus が、Anthropic の IPO に反対する理由を司会の Alex Kantrowitz と議論した回。約 9,000 億ドルの将来債務、Q3 の成長鈍化、最上位モデル以外に強みがない点、Chinese モデルとの価格競争などを挙げ、最終的には安値での買収が最も可能性が高いと述べた。一方で、Claude Code が神経記号 AI(ニューロシンボリック)を実装した点は評価し、AGI は次の 10 年と見ている。

この回のポイント

  • ●Marcus は Anthropic の製品は優れているが、約 9,000 億ドルの将来債務に対し利益が出ておらず、ビジネスとして成り立たないと主張した。
  • ●売上成長率は Q2 の前期比 147% から Q3 は 47% に鈍化した。Marcus はトークン消費を競う「token maxing」の沈静化と中国モデルの台頭が要因だと述べた。
  • ●性能向上のコストが指数的に増える一方、ベンチマークは飽和し、モデルのコモディティ化と価格競争が進むという見方を示した。
  • ●Kantrowitz は反論として IPO の大幅な超過応募、売上の約 10 倍成長、計算資源あたり収益の改善、蒸留対策、Jensen Huang の信頼低下による対中競争の緩和、製品企業としての展開を挙げた。
  • ●Marcus は Claude Code を「純粋なスケーリングの勝利ではなく神経記号 AI の勝利」と評価しつつ、検証できる領域(数学やコード)以外では依然に限界があると指摘した。
  • ●AGI は今後 10 年の間(2030 年代)と見ており、世界モデルや継続学習などの発見が必要で、AI 安全性は進んでいないと懸念を示した。

章立て

  1. 神経記号 AI としての Claude Code

    Marcus は Claude Code が条件分岐やループなどを組み込んだ神経記号 AI であり、純粋スケーリングの勝利ではないと説明する。

  2. 弱気論:債務と成長鈍化

    約 9,000 億ドルの債務、Q2 147% から Q3 47% への成長鈍化、token maxing の終わりを議論する。

  3. 収穫逓減とコモディティ化

    性能向上コストの急増、ベンチマークの飽和、中国のオープンウェイトモデルとの価格競争を論じる。

  4. 市場の空気と顧客集中

    OpenAI の資金調達の縮小、SoftBank の動向、Goldman などの大口顧客の利用制限を取り上げる。

  5. 反論:超過応募と収益効率

    IPO の超過応募報道、売上 10 倍成長、計算資源あたり収益の改善、蒸留対策や Jensen Huang の影響を議論する。

  6. 顧客拡大と製品企業化

    高額顧客の増加と Claude の事務ツール展開に対し、Marcus は ROI 研究や再現リスクを指摘する。

  7. 疑う側は間違っていたのか

    推論や数学での成果を神経記号の観点から整理し、検証可能な領域に限られることを説明する。

  8. AGI 時期と AI 安全性

    AGI は次の 10 年と見ており、教条主義の後退には楽観的だが、AI 安全性には悲観的だと述べる。

解説記事

NYU 名誉教授で AI 懐疑派として知られる Gary Marcus が Big Technology Podcast に出演し、Anthropic の IPO に反対する理由を司会の Alex Kantrowitz と議論した。今週は弱気派の Marcus と、後日登場する強気派の二本立てという構成で、Kantrowitz が反論役を務めている。

「純粋 LLM」ではなく神経記号 AI の勝利

Marcus はまず、自分が批判してきたのは「純粋な」大規模言語モデル(LLM)だと強調した。ニューラルネットワークと、条件分岐やループなどの古典的な記号処理を組み合わせる神経記号 AI(ニューロシンボリック AI)が必要だというのが、大学院時代からの主張だという。Claude Code はまさにその実装であり、純粋スケーリングの勝利ではないと述べた。Anthropic の技術を評価する一方、IPO には反対というのが彼の立場である。

弱気論の中身

Marcus が最も強調するのは財務面だ。ブログでは、Anthropic は最大約 1.4 兆ドルの評価額で上場するが、補助を受けた例外的な 1 四半期を除き黒字化したことがなく、約 9,000 億ドルの将来債務を抱えていると書いている。製品が良くても事業として成立しない例として、Juicero(高価な搾汁機)やレモネードの比喩を使った。

成長の鈍化も論点となった。引用された数字では、Q2 の売上は前期比 147% 増だったが、Q3 は 47% 増にとどまった。Marcus は、従業員にトークンを使い切らせる「token maxing」の流行が終わり、品質の高い中国モデルが台頭したことが響いたと見る。赤ちゃんの体重が倍増し続けて大学入学時には 1 兆ポンドになるという「1 兆ポンドの赤ちゃん」の誤謬になぞらえ、指数成長は続かないと述べた。

さらに、最上位モデルを作るコストは指数的に増え、改善幅は小さくなる。多くの課題は安価なモデルで解け、最上位モデルが必要な領域は狭く、しかもオープンウェイトモデルに追われているという。准教授と教授の比喩で、「教授級」に 1,000 億ドルを払う価値のある需要があるのかと問うた。市場の空気、OpenAI の調達が目標を下回ったこと、SoftBank の慎重姿勢、Goldman Sachs などの大口顧客による利用制限も根拠として挙げた。

Kantrowitz の反論

Kantrowitz は番組中盤で反論に回った。Bloomberg の報道では、IPO は複数倍の超過応募で、調達額は 644 億ドルを超え最大約 904 億ドルになり得る。Marcus はこの点は自分の主張の中で最も弱いと認め、市場心理次第なので予測が難しいと述べた。

売上については、Q3 2025 の約 14 億ドルから Q3 2026 の 170 億ドルへ約 10 倍に伸びた。計算資源 1 ドルあたりの収益も 0.4 ドルから 1.72 ドルに改善したと S-1 に示されている。Marcus は事実は争わず、効率化は競合も進めるので価格が下がる点と、四半期ごとに成長し続けなければならない点が問題だと答えた。

また Kantrowitz は、Anthropic が蒸留(他社モデルの出力で自社モデルを学習する手法)への対策を進めていること、Jensen Huang の信頼低下で対中強硬派が力を得ること、年 10 万ドル以上の顧客が約 950 から 1 万 1,900 超に増えたこと、Claude の事務ツール展開を挙げた。Marcus は、蒸留対策や対中規制は Anthropic に有利に働き得ると認めた。ただしスプレッドシートのような用途では幻覚(事実と異なる出力)が致命的になり、競合が数か月で追随すれば持続的な強みにならないとも述べた。ROI の欠如を示す研究が 12 件以上あるとも指摘している。

「疑う側は間違っていた」への応答

Kantrowitz は、Anthropic や OpenAI が純粋スケーリングにこだわらず手法を柔軟に取り入れ、推論でも成果を出したと水を向けた。Marcus は、推論モデルの成果は数学やコードなど、Lean のような記号的検証器で答えを確かめられる領域に偏っていると説明した。LLM が候補を生成し外部の記号システムが検証するのは神経記号の典型で、自分が言ってきたことが当たったと言える。ただ、台湾の防衛計画のような検証できない領域では通用せず、AI は依然として「狭い問題に対する狭い解」の段階にあると述べた。

AGI と最終的な見通し

Marcus の最も可能性の高いシナリオは、Anthropic が債務を払えなくなり、NVIDIA、Amazon、SpaceX、Apple などに 1 兆ドルを大きく下回る額で買収されることだ。破産もあり得るが、買収が先だと見る。AGI は世界モデルや継続学習の発見が必要で、2030 年代になると予想している。Rich Sutton らも「さらなる発見が必要」と言い始めていると述べた。教条主義が薄れたことには楽観的だが、AI 安全性は実質的に進んでいないと懸念を示した。

まとめ

今回の議論は、技術的に優れた製品と、持続可能な事業・妥当な評価額は別だという点を整理している。Marcus の弱気論は、約 9,000 億ドルの債務、成長鈍化、コモディティ化に集約される。反論側は、超過応募、急成長、効率改善を示した。IPO 初日の値動きよりも、成長率の維持と計算資源契約の負担に注目したい。なお、数字や予測は出演者の発言に基づく。

文字起こし(英語・自動生成)

Anthropics performance on the public markets will be a referendum on the state of AI and whether it's actually a business that's sustainable. And because it's going public this week, we're going to have two conversations on the show. One, a pessimist view of Anthropic and its business, and the other, an optimist view of Anthropic and its business. So we're doing double programming this week. For our pessimist view, we're going to have Gary Marcus. He is a professor emeritus at NYU and also the author of Marcus on AI on Substack. And then coming later this week, Aaron Levy, the CEO of Vox, will take the pro side. So let's start off with Gary. Gary, it's great to see you. Hey, great to be back. Fourth time on the show. Wow. Now, you've said a lot of bad things about large language models over the years, yet We're about to hit what's looking like a $1.4 trillion IPO for Anthropic coming up when this gets published. Before we get into the IPO, I want to clarify something. I said a lot of things about pure LLMs, pure large language models.

I said pure scaling would not solve things. We would have to go to neurosymbolic AI. We'd have to add in symbols in different ways. That's been the fundamental thrust of my research since I was a graduate student. And the interesting thing about Anthropic is they actually did that. Nobody talks about it, but Cloud Code is as neurosymbolic as you can get. It is a big, big, large language model. Scaling was part of it, but they added in conditionals, loops, tools, all kinds of things that are symbolic, that take abstract representations using classical AI techniques, and that's all integrated with the rest. And so when Cloud Code came out, I actually wrote a piece saying, this is great, this is actually neurosymbolic AI. So this is not a victory for pure scaling, the way that you kind of said and the way a lot of people said. Even if it flies, it won't be. It'll be a victory for Neurosymbolic AI. Nobody will admit that, but people who watch your show will know it. Yeah, of course. And I remember you and I spoke about this. I remember reading that piece. And yet I remember reading another piece that you wrote recently,

which is Just Say No to Anthropics IPO. I still don't like the IPO. I'm super thrilled that they're using Neurosymbolic AI, but I still don't think the economics makes sense. So today we're going to go through, I think I noted down maybe eight or nine reasons why you're bearish on anthropic and its ability to succeed on the public markets. Then I wrote down a few rebuttals, but I think it's important for us to at least go through the bear case. Because I read it and I was like, oh God, Gary makes some good points here. And then we'll talk through some of the arguments against your argument here. So, all right, let's just go through it. The first thing that I think you believe is that it's just a bad business overall. So this is the beginning of your post. You said it's going to IPO at up to $1.4 trillion, more likely a bit less, an astronomical sum for a company that has, with the possible exception of one anomalous heavily subsidized quarter, has never made a profit and has nearly $900 billion in future liabilities. So you'd think

unprofitable. It owes like trillion dollars to continue to build. To value a company like that at 1.4 trillion is lunacy because it's a bad business. Yeah. I mean, you can talk about what you mean by bad business, but the economics of putting in that much money and getting that little return makes it a bad business. So it's a good business in the sense that they make a good product. It's a good business in the sense that they have a lot of revenue. It's a bad business in the sense that it's really expensive to run this business. It's really competitive. So they have a lot of people doing similar things. They have a very narrow market niche, which is basically the highest end models and nothing else because they're getting beat on the lower end model. So, you know, it's not that they're not doing something useful. Bad business is, you know, a complicated term. They might mean a bunch of things. But what I specifically mean, that's your term, not mine, is I don't think that they're going to be profitable anytime soon. It's not clear that they will ever be profitable, and I think there's a real chance they will go bankrupt.

So in that sense, they're a bad business. More likely, I think, they will eventually be acquired, but they'll be acquired for a lot less than they're about to go out on the public market. Yeah, I mean, you could have a good product and be a bad business. Like, you can make the world's best lemonade with hand-squeezed lemons and sugar that's like, I don't know, basically you buy from the hands of the gods. But if you sell it for a quarter a cup instead of two bucks a cup, your lemonade might be good, but your business is bad. In that sense, it's a bad business. It's like Juicero. I don't think the problem with Juicero was that it wouldn't make the juice. The problem was people realized they didn't need it, and certainly not for the price that was being charged. All right, a callback to one of the all-time great tech scams. Okay, the second point that you make is that basically a lot of the discussion around Anthropik, including valuation, has been predicated on the idea that it would maintain a certain growth rate. And a lot of this growth happened in the token maxing era where companies would basically say that their employees use as many tokens,

which means spend as much money on AI as you can, and we won't even look at the numbers. Do you remember my tweet about that? No, say it again. My tweet about that was about six months ago. It was, token maxing is stupid. Prove me wrong. And no one ever did because it was stupid. Yeah, I don't think there's any argument against that. Although, actually, that's wrong. An argument against it would be you should encourage people to use, to pressure test these tools and use them as much as they can, because only then will you find the uses. That's the best, most optimistic case. That is a better counter argument than anybody made on my Twitter feed. Where were you when I needed you? I'm not a token maxing defender because the other side of it is. I mean, that would be the argument. When you gamify it, it's just like you're going to lead to people running loops, unending loops of token spend. Yeah, people were running 40 agents overnight because they got paid more money because they were running more agents. And so all that stopped, right? And so, I mean, going back to the more serious point, the reason people think that this company is valued at $2 trillion or $1.5 trillion

is partly because they saw graphs like the one Joseph Jax had, which got, I don't know, millions of views on Twitter. Joseph Jax had this plot. It has the weirdest y-axis I've ever seen. The numbers go up and then down and then up again, and it's, like, sort of trying to be logarithmic. Anyway, ignore the y-axis on this crazy graph. It plotted Anthropic making more revenue than Google or Alphabet by, like, 2028 or something like that. And just, you know, rapid, rapid rise in revenue. And what I said then was this depends on this fad of token maxine. This fad is not going to last. And so you have these projections. I always think of what I call the trillion pound baby fallacy, which comes from a tweet that I saw on Twitter that absolutely left. A guy had a picture of his baby and then his baby like a month later. And it was kind of a joke. He said, you know, my baby has doubled in weight in its first month of life or something like that. And I've taken the liberty of making some projections by the time he goes to college. He'll be a trillion pounds. And, of course, the point is that you can have exponential growth for a while, but usually it tails off.

And I said, look, Anthropics growth was exponential, like the trillion pound baby, but it ain't going to get to be a trillion pounds because there are going to be countervailing pressures. One of which I anticipated correctly. I said that token maxing would die. That's one of the reasons why revenue or growth has tailed off. Another that I didn't get precisely, but I had gotten the gist of a long time ago was in 2023. I said, everybody's building the same thing, and that means you're going to have a commodity where there are price wars and profits are going to be hard. What I didn't quite get was how fast suddenly China would be there, but it was a symptom of what I always said was going to happen. Large language models and even enhanced large language models, if everybody else knows the tricks, are going to be a commodity. There's going to be price wars, and it's going to be really hard to make profits. And sure enough, by the middle of this year, suddenly Anthropic was struggling to find a market niche at all because people were using the Chinese models. What it looks to me in the latest data that I've seen is that Anthropic took off like a bat out of hell for about a year.

And then around July, things slowed down. It's not that there's no growth anymore, but these projections of them making 147% growth every quarter are just clearly wrong. And if they're clearly wrong and they have $900 billion in liabilities, the math doesn't math. Indeed. Let's go to your post. You said quarter two growth was 147%. It was up 147% compared to Q1. But Q3 wound up being a mere 47% in growth compared to Q2. The decline of token maxing and the rise of good quality, if not quite top tier competitors from China, began to take their toll. So we're going to keep making the case here for why it might not work before we go to the rebuttals. But I do need to just say one thing when I was putting this together last night. It just strikes me that Anthropic grew. So in the middle of this token maxing era, they grew 147% in a quarter,

which would just reflect basically tokens gone wild within companies, uncapped token spend, token leaderboards, all these things. there has been a discussion about them reining that in about companies reining that in yeah but the quarter afterwards so it didn't maintain the growth rate okay but it didn't it didn't shrink like the reining in of token maxing led to uh the next quarter being a 47 percent increase in spend over the crazy quarter yep yeah no i mean it it didn't die completely it's just that he got $900 billion of obligation. I mean, another way to put it is it would be a good business if the liabilities weren't so high. I mean, maybe that's sort of the story of all businesses that fail. But, you know, they are making a bunch of money. They do have a good product. But they're not making profits off of this. And the revenue is not growing fast enough to catch up with what they've committed to.

Yep. All right. Diminishing returns, I think. We've talked about this on previous episodes. You bring it up again. Let me just set it up. You write, the cost of making better AI in Anthropoc's only real niche is models at the extreme top end, since they can readily be beat on price at the low end, is growing exponentially. So the cost of making these better models is growing extremely fast. Each new tiny improvement tends to cost far more than the last. and you basically showed that, shared this tweet, Anthropic is getting exponentially less capability per billion dollars spent over time. Yeah, they should be terrified about that. This is an issue because this is always like, for me it was like, okay, you know, once you make the model smart enough and then you run another multi-billion dollar pre-train, it's like the improvements might only be noticeable to, you know, the people who they really apply to. Like a PhD might notice, but, you know, me, an average user might not. That's right. I mean, some of it's about what the user notices. Some of it's about what appears on the benchmarks.

But the point is, if you have to spend a billion dollars to impress your users this year, $10 billion to impress your users next year, and $100 billion the year after, and that's kind of what it's looking like, you know, if Anthropic wants to say we did, you know, four points better on the benchmark, suddenly it's going to be $100 billion to them. And then, you know, ultimately a trillion. At some point, that is just not sustainable. I mean, also, this might be sort of a counter-narrative to say or whatever, but like the benchmarks are just like – so many of the benchmarks have been saturated that it doesn't really matter. And that's going to go to your commoditization point, which is just like the model – the base models have just become smart enough that like really getting to the frontier of the benchmarks at a certain point, it's like you're not going to have a meaningful difference if you like increase the benchmark by a little bit over time. Yeah, I mean, there's a separate issue of is the benchmark still valid, like if you get close enough to the top. But we'll put that aside for the moment. It goes back to this issue of market niche.

Like, who cares? So the way I think about it is there's three kinds of problems. There's the ones that you can solve with kind of any model nowadays. And so you don't need Anthropik stuff because you can go on Open Router and get it from someone else cheaper. there are the very few problems where really the best thing is what you need not even the second best and i think problems like that do exist and i think also people are still experimenting to see if any of those exist for them and that's where some of the rather and there are some that i think no current system can solve and that that doesn't help them so the only thing that that really sustains their business right now are these fairly hard fairly difficult problems that they're uniquely qualified to solve. And, you know, that's itself a moving target, right? Because the open weight models get better and better. And so, like, you can think it's like this little slice from this big graph and that little slice is like moving. Maybe it's moving up a little, but the bottom is catching up. And, you know, again, like you're making this bet if you buy this IPO that they're going

to make that little slice work, that the revenue on it is going to grow, that somehow they're going to be able to serve up that slice at a profitable level. Like, it's not impossible. I am not so pessimistic as to say there is zero chance that this will ever work, but I'm pessimistic enough to say that so many things have to come together, and some are already not coming together, right? The revenue growth that really was sustaining that notion of evaluation is already tailing off. And so, like, they're probably not going to get the perfect storm. And if they don't get that perfect storm, then I think that they either go bankrupt or get acquired at a discount. The way that I describe what you're hitting at is basically like if you're a hiring manager and you could hire a sixth grader or a freshman in college, you're going to hire the freshman in college every time because the sixth grader can't do the work. And that was sort of like the difference between standard models and frontier models maybe a year ago.

where, like, the frontier models were so much more capable than the standard models, you wouldn't even think to go to the standard models. But then when everything increases in intelligence, and now your difference in intelligence is, like, professor or associate professor, you're going to start going. Most people aren't going to care. Yeah, exactly. Especially if you associate this. Yeah. When you need the full professor to solve the problem like there are such cases And if it absolutely mission critical you know there might be good reasons to go for the professor that bills more than the associate but a lot of times you don need that And if Anthropic is spending $100 billion to make the smarter professor, then that's where you're going to get. Yeah, exactly. Is there $100 billion worth of revenue for having the slightly smarter professor? Yes. And this is important. There's still going to be some problems that even the $100 billion model is not going to be able to solve. There are still going to be hallucinations. We have still not really solved artificial general intelligence. And so you're going to be spending more and more money to find out if it can solve your super hard problem. And sometimes the answer is going to be no.

So you're going to spend premium prices hoping that you can solve this super hard problem, and then you don't. How many times are you going to iterate on that before you say, fuck it, I'm not paying the full professor prices. It works for me sometimes. I don't know, and it's expensive. I'll just stick with the associate professor and hope for the best. A bunch of people will conclude that. If not this iteration, then next iteration are the one after. All right. I have some thoughts to share on this one as well, but I'll hold them for the moment because we'll keep making the case here. We can tell the fun part is going to be the second half. I think so. But we've got to get everybody up. We've got to set it up. Then point five from your piece is the market is maybe souring on AI. There has been stories, more than stories, it's been true that it's been much more difficult to raise capital, to build AI data centers recently. and the market has been sort of unsure what to think about AI companies. I think even before the Jensen scandal, you had mentioned that over like a year or so time period, recently Krispy Kreme was doing better than NVIDIA.

Not certainly, although. Over the last year, Krispy Kreme is doing better than NVIDIA. And you also linked a piece from Adzitron in your newsletter, which I thought was very interesting, where he points out, and I think this is a story that no one really touched on, that OpenAI said they were going to try to raise $30 billion, which up until this point had never been a problem for them. They had just raised $122 billion. They ended up raising $11 billion in total. With some promise, there might be a little bit more. And the valuation, instead of being like $1.5 trillion or so, was $1.25 trillion. I think they went out asking $1.4 trillion and it wound up being $1.25 trillion. So they got less money at a lower price. Less money at a lower price. I don't think that had ever happened to them before, ever. That's the first, yeah. And so, like, that's another sign of ambivalence, right? The market is becoming ambivalent. It's not completely left the building. Like, you know, the market is still listening, hearing things out. You know, there's some indication that maybe the Anthropic IPO will be oversubscribed, but we'll see. Like, each day now brings, you know, new news, and there's the Jensen thing going on, and so forth.

So it's not that it's dead in the water, but there's definitely ambivalence, And you can see it in the credit rates. And I just saw SoftBank saying maybe, you know, we shouldn't be quite putting that much into AI. And SoftBank, you know, has really propped up OpenAI for quite some time. And so, like, when SoftBank, you know, Masa, who's, like, all in, is like, maybe we should think about this. Like, that's a real sign that the mood has changed. Like, it's a vibe shift, right? Yeah, I mean, you highlighted this in your post. they posted an unexpected $7 billion loss off Bank did. Yeah, they made some really bad deals, but we don't have to go into – I don't know all the details. They tried to basically hedge, and they got screwed on both sides of the hedge. Oh, concentration risk we should get into. The other part, when you think about Anthropics' business is – now, it was two – in last year, there were two of their customers that made up 25% of their revenue. Now they have five customers that make up 19% of their revenue, which is a little bit better than it was previously.

But then you look at the action that's happening, and some of those big customers are starting to try to move off of Anthropics tools. You cite Microsoft Meta, which have competing tools, but they are moving off Anthropics. I think Goldman Sachs was also... Yes, at Goldman, not everybody left, but they basically told their bankers and traders don't use this. They said the coder is fine. Keep using cloud code, but the rest of you cut it out. Yes, and so if they lose these customers, the business doesn't implode, but it could be very difficult for them to continue to grow the way they're growing. Yeah, I mean, the problem, again, is that they need to be a rocket. Like, they can't just steadily sell at the current levels because they're losing money at the current levels and they have these liabilities. It needs to stay a rocket. And so every time somebody gets off, my metaphors are getting a little complicated here, but every time somebody pulls back, then it becomes harder to maintain the rocket pace that they would need. It's like getting escape velocity.

You have to get the escape velocity. And each time a customer says no or limits who can use it internally, the chance that they're going to get that escape velocity diminishes. Yeah. And then you look at the market cap that Anthropoc was anticipated to go public with, and it's nowhere near what they were anticipating. So the talk was $2 trillion. Or even a little higher. $2 trillion, and now it's looking like it'll be $1.3, $1.4, although who knows what will happen on the first day of trading. Exactly. But when you speak about momentum, that is you don't really want the valuation to be going in the down direction as opposed to – As you IPO, right? You don't want to be IPO-ing as your valuation goes down, but that is what's happening. I mean, it's still a very impressive valuation if they achieve it, and they may sell out the IPO. The real question is, like, if people buy at the IPO, are they going to get screwed because it's not really sustainable?

The IPO day is a little bit of, like, almost like a meme stock thing. Yes. What's the vibe here? We saw that with SpaceX. We saw that with SpaceX. The long term is, like, is there value in the company? Is it returning on the investment? And so, you know, the real test here is time. Your bottom line, I would not be shocked to see Anthropic go out of business, nor would I be shocked to see if the IPO fails. Go out of business? I think the most likely outcome, actually, is that they will have to sell to somebody else. NVIDIA, although NVIDIA's got its own troubles right now, but I used to say NVIDIA, or Amazon, or Elon Musk, SpaceX, Apple, something like that, for a lot less than a trillion dollars. Why? At some point, they're not going to be able to pay their bills. They're going to have options at that point when they can't pay their bills. They have a lot of talent that has value. They have a lot of customers that have value. But at some point, well, I should say this more carefully, at some point it's possible that they won't be able to pay their bills.

Like that's what I'm projecting. My guess is at some point they won't be able to pay the bills. You know, $100 billion something a year in bills, and I think that it rises. I don't quite know the shape of the curve. You know, at some point, if they're not making profits, that's going to be a problem, right? And at that point, they could file Chapter 11. You know, I don't think they'll do that. I think instead they'll shop themselves around or other people will realize that they're in trouble and make offers. That's how these things go. And so somebody will approach them and say, you know, we'd like to buy you for $200 billion. They'll say, no way. There'll be some back and forth. And in the end, it'll be like $400 billion. It won't be great. But the alternative will be they can't pay the bills. You know, that's a problem. Right. The underappreciated part, everyone's been looking at the revenue escalate, like growing and all this stuff. But the underappreciated part of the AI moment is that the only reason why these companies, I mean, it happens with a lot of startups, but the scale is greater. The only reason why these companies have gotten to this point has been their ability, not the only,

but the companies would not be able to get to this point without their ability to raise gobs of money. Enormous gobs of money. The IPO is not the final. They're not standing on their own feet. They're standing on investors. ABC and big tech and NVIDIA money. And the IPO is not the final fundraising event. Like they could end up selling more shares or raising more debt. Yeah, they may do those things before they do what I think is going to be the outcome. Now, at a certain point, and you're in the life of any company, you have to be able to make a profit. It's one of those things about business. The lemonade needs to cost more than the ingredients and labor. You've got to sell it for more than it costs you to make it. And, you know, the costs are going up, too, for a whole bunch of different reasons. Like the resistance to data centers, we can argue about that, but it makes data centers more expensive, right? And that raises their costs. The fact that they need to make a bigger model to get the same improvement on a benchmark or whatever, that raises their costs. So they have a bunch of things that are actually raising the cost. There was also just new numbers from OpenAI, and they made more revenue than before, but their costs went up, and so they continue to lose money.

you say in your piece, it's not clear that the capital markets are really going to come up with the $64.4 billion or so they're asking. And that is where we will pick up when the rebuttal begins right after this. This episode is brought to you by LinkedIn Pulse. One thing I've learned covering technology is that the pace of change is relentless. Just when you think you've wrapped your head around one major shift, another comes along. And keeping up isn't just about following the headlines. It's about understanding what's actually changing and what those changes mean for businesses, the workforce, and the economy. Now, that's where LinkedIn Pulse comes in. LinkedIn Pulse provides unique insights on work and opportunity, informed by data from 1.4 billion members, 71 million companies, and 42,000 skills. These insights help us understand how the era of new work is taking shape as businesses evolve and the skills people need change.

And what's especially useful is that LinkedIn Pulse isn't just trying to predict what might happen next. It's helping us understand what's happening right now, which is essential when everything is moving this quickly. So explore LinkedIn Pulse at pulse.linkedin.com. Everyone's talking about AI, but there's a big difference between running a promising pilot and actually making AI work across an enterprise. That's why we're launching the Big Technology Guide To series presented by DXV Technology, available on YouTube. Throughout the series, we're tackling some of the biggest questions facing businesses today. How do you take enterprise AI from pilot to production? How do you navigate an increasingly complex cybersecurity landscape? What's the real opportunity with AI agents? And how do you calculate the return on your AI investments? On my Alex Cantor's YouTube channel, you can hear from experts like DXC Technologies Chief Digital Information Officer Russell Jukes. When you ask a question to somebody in your company, do you trust them? And what made you trust them? Do you trust but verify every single time?

Russell and I discussed the challenges and opportunities of building production-grade AI and what it really takes to move beyond the hype. Check out the Big Technology Guide To series presented by DXC Technology on my Alex Kantrowitz YouTube channel today. I know planning for the future isn't always the most exciting thing to think about, but it's one of those things that can really matter in the long run. I've always been a fan of anything that makes planning for the future a little easier. Ethos is a great resource for exploring life insurance options without making the process more complicated than it needs to be. Ethos makes the process fast and easy, and it's 100% online. you can get a quote in seconds apply in minutes and even get same day coverage there's no medical exam you just answer a few health questions online you can get up to three million dollars in coverage with some policies as low as thirty dollars a month take 10 minutes to get covered today with life insurance

through ethos get your free quote at ethos.com slash big tech. That's E-T-H-O-S dot com slash big tech. Application times may vary. Rates may vary. This episode is brought to you by Intelligence AI. Is your investment in AI coding tools helping your team ship better software? Intelligence connects your AI spending to code quality and engineering outcomes from individual sessions and pull requests to entire projects. You can see where AI saves time, where it creates rework, and how to improve. And its model router selects the right model for each task based on quality, speed, and cost. See how Intelligence AI helps enterprises optimize AI-powered delivery with connected visibility into cost, quality, and outcomes at intelligence.ai. And we're back here on Big Technology Podcast with Gary Marcus. Gary, fourth time on the show. It's always good to have you and get a chance to speak with you, especially in such a momentous news week. All right. Now that you've laid out your argument, let's debate it a little bit.

We left off before the break with your line saying it's not clear the capital markets are going to come up with the $64.4 billion or so. Anthropic is seeking. We're recording this, of course, the week before. But there's a Bloomberg headline that I'm looking at right now. I'm sorry, I'm sorry, I'm sorry, I'm sorry, I'm sorry, I'm sorry, I'm sorry, I'm sorry, I'm sorry. I'm sorry, I'm sorry, I'm sorry, I'm sorry, I'm sorry, I'm sorry. I'm sorry, I'm sorry, I'm sorry, I'm sorry. Anthropics IPO is said to be multiple times oversubscribed. oversubscribed. Anthropics initial public offering has drawn multiple times the orders for the shares available as demand accelerates for potentially the biggest first time share sale ever. The demand has been driven by so-called anchor size indications of interest. We also have NVIDIA that's going to put $10 billion in, assuming that holds. But instead of, your worry was $64.4 billion might not be possible. The Bloomberg story says It now looks like they might raise as much as $90.4 billion. And we'll have to see it. I think that's the weakest argument in what I put. It's the most dependent on public women's sentiment, right, for how much people are willing to put into this.

So the other arguments are about whether it's a profitable business. This is about how much people are willing to take a chance on it. Even I realized it was kind of a mistake, but there was a complication that I didn't get into in the piece, which is the 64 billion was one number, but there's a bunch of different ways to think about how much more they needed. So my point was basically I didn't know if they would get what they needed for the round. Obviously, if they don't, it looks bad, and then you wind up in a death spiral, right? And they certainly don't want that. Open AI has already taken a step towards that death spiral by not getting their terms for the first time. And it would look really bad for Anthropic if they can't fill out the round. Maybe they will be. Maybe I'm wrong. I mean, what I've generally been right about is technical limitations, and I feel like I've often underestimated the foolishness of people who want to keep doubling down. It's been a pretty rough year for the tech market as a whole. Like, as we are recording this, it looks like Core Wave is going to pop, but even after that pop it going to be down by 45 for the year or that You know the public mood and the sentiment is hard to predict People often tell me ask me like or I often think about you know if I lived in the era of the tulip craze in

you know, the 1630s in Holland, what would I tell people? Should you buy, you know, like people are buying tulips for the price of a house. And I would say it's crazy, but I don't know how long for sure it will take before people realize it's crazy. So there is some argument for buying it while it's going up, even if you know it's going to go down eventually. And so you have factors like that. Like some people might, you know, realize, for example, they know that this, you know, let's say more carefully, they share my opinion that this is not a sustainable business, but they think that there's going to be another six months of enthusiasm and they're going to buy it and dump it in six months. So I don't know all of the dynamics that lead people to actually, you know, buy an IPO. I think this is complicated. we will see what happens when it actually comes out, whether it's oversubscribed or not. I stand by all the rest that I really have concerns about whether they can be a sustainable, profitable business. If everybody buys the IPO, you know, valuation that ultimately gets back up to $1.5 trillion or whatever, I have serious doubts about whether, you know, that investment is going to return in the long run. Right.

Okay. So then when it comes to the business itself, you just, you know, the revenue numbers are absurd. They are crazy. I'll tell you a quick story. I mean, I was, and I told this on the show before, but I was interviewing Dario last year in like July, and he said, we're at like a 4 billion. We were at a 1 billion run rate last year. We have 4 billion now. We might be at 10 billion by the end of the year. And next year, who knows? And I've heard that speech from a lot of CEOs in the past. We have exponential revenue and it will continue growing. And as the numbers get bigger, The growth won't slow down. And I was like, okay, well, let's see what happens. The run rate as of October 2026 is $81 billion on the year. They went from – Can we pause there? I know you don't like run rates. I'm just going to say this one last stat. They went from $1.4 billion in revenue in Q3 2025.

So right around the time that I was speaking with Dario, their Q3 2026 revenue was $17 billion. So they 10X'd over the year. Yep. Go ahead. I mean, what's the argument against that? I mean, I'm not going to argue against the facts. Yeah. I mean, I don't know the exact numbers, but, you know, they did 10 or 12X or whatever it is, for sure. The problem is they need to keep doing that. I mean, that's my counterargument. And there's indications that in Q3 they stopped doing that, indications that it has slowed down. You know, there are many processes. They double, they double and double or whatever until they don't. It's incredibly impressive what they have done. But, again, when you owe or you have future liabilities of $900 billion and you're not making a profit, you need that to happen every quarter. You miss a couple of quarters and you're done. Darryl himself used to make arguments like I'm making. Do you remember when he used to do this? Listeners might not. He used to say, you know, we are threading a needle here.

I can't quote him verbatim, but he used to make these arguments. We're threading a needle here. If we make too much purchase orders on compute and we don't need it, we're going to be screwed. And at that point, he was basically making that argument against Sam. He was saying, Sam YOLO's everything. You only live once, and he's like buying wild amounts of compute, and we're not doing that. We're more sober. You remember those days when Dario. And I'm basically just taking the same argument. But he also wishes he bought that compute. He does. He does. But in the end, I think he's going to get skewered by the second side of it. I mean, it is a very delicate thing, And that's the problem, right? It's very delicate. And I think now he's gone too much in on the compute. There's too much commitments. I mean, another thing that will happen is they'll try to renegotiate out of those commitments, I guess. And, you know, we'll see how all of that goes. But I think in the end they're not going to need all this compute. And let me add one thing that I forgot to put in the article, which is that the other thing is the shift towards local from cloud, right? So the local models are going to get better and better.

And so people aren't going to need all these giant data centers. They're not going to need to go after the providers as much. They start running stuff on their own. So there are a whole bunch of different ways in which that growth can slow. Any of them basically kills them. That's what he was saying about Sam. It is true of his own company. Okay. On the profitability side, they have this chart in the S1 that I'm sure you've seen about compute efficiency. that in Q1 2025 for every dollar they spent building compute, they made 40 cents. In Q3 2026 for every dollar of compute they invested, they would make $1.72. So that would be the argument that if you're looking at Anthropik, their argument for why they're going to be profitable, they're making it work per dollar of compute invested, and that also suggests that they won't go bankrupt. I mean, I think the best hope that any of these companies have is that there will be efficiencies eked out. And that's a reasonable expectation. The way I think about it is there's a lot left to do, in my view, to get to AGI.

Some of it is science. We need whole new ideas around world models and things like that. And some of it's just efficiencies with things we already know how to build. That's more like engineering. And that is happening, which means that the cost to serve a given model is going down. It has gone down steadily. It will continue to go down. And that is their best hope. The trouble is all the competitors are working on that same efficiency. Like the DeepSeek moment was about DeepSeek finding a cool new way to serve models more efficiently. And so everybody's trying to do that. And that means that the prices that they can charge go down, right, because the competitors are doing it. If there's a way that they can make this work, I think it does rely on eking out those efficiencies. But because the competitors are doing the same thing, it's going to be hard to sustain the prices. and prices have gone down, what, by a factor of $100 or $1,000 over the last year. And so, you know, it's like if you were selling lemonade. You know, last year you could sell it for $3, but you ain't selling it for $3 anymore because everybody else figured out how to make it cheaply.

So it's great that you're now making it for, you know, $0.40, but now you can only charge $1 or worse. You're now charging, you know, you can charge $1, but it's costing $1.10. It's really more what it's like. And so, yes, you are selling for less. It costs less, but you need those things to intersect in the right way such that you actually make a profit. When everybody else has the same idea, it's hard. Let me give you my best argument in Anthropics' favor on this one. Let's see what you think. So you agree the rest of them weren't compelling? No, no. I just want to talk about this because it's very interesting and I think underappreciated to a degree. it barely made headlines that Anthropic was starting to identify the distillation impact on its distillation efforts. It was able to identify them, and it's been able to stop them to a degree. At least that's what it says. To the point that it's going to lose billions, I think, a year. By turning down the distillation. By turning off distillation queries.

So the biggest open source competition is in China. China Open weight, there's not much open source China has used distillation, maybe not as the key, but it has helped a lot in their ability to catch up with these models, right, because it's like six months so what happens? Distillation is part of it It's going to be more difficult for them to distill now. Now a new wrinkle even though NVIDIA is investing well actually it makes sense that NVIDIA is investing in Anthropoc because Jensen was the world's most effective advocate for, against export controls to China. You could argue that without Jensen's advocacy, the open-weight ecosystem in China would be much further behind because he's been able to deliver them the chips. There's going to be a lot of discussion about that point. That is going to be very, very difficult for him to do for obvious reasons. Jensen's lost his chance to talk about this. He's not credible anymore. He's not credible anymore.

And so that is, so before you make your argument, I'm accepting your premise. I think there's another premise we need to hear out. But I think it is right that China hawks are going to capitalize on Jensen's mistakes. Right. And there's going to be more discussion about, you know, whether open weight models can be served. I think I know where you're going with this. I think that that's, you know, whether Chinese open weight models can be served in the United States. I think I agree with your premise there. And I think that that is something that Anthropic can be happy about. Whether the world can be, it's a complicated question because prices might rise around it and so forth. But I accept your premise that the fallout of Jensen's escapades is going to be that the hawks are going to get more voice. I don't know what the ultimate fallout is, but I think what we're going to see when the Democratic Congress takes over in January There's a lot of investigations about what happened, and a lot of the hawks are going to say, why do we give all this stuff to China?

Yeah, and so if that is your number one competition in terms of making things, like you referenced DeepSeek, making things more efficient, et cetera, et cetera, maybe Anthropic does have a more sustainable advantage in that world. I think it is the best argument you've made. You do still have OpenAI. You still have Google. You still have Microsoft. You still have Amazon. You still have the problem that everybody's basically building the same thing, but it reduces the amount of your competitors, which reduces the pressure on the price wars and so forth. So in that way, I think Jensen's blow-up is helpful to Anthropic. Let's talk about growth in customers. I think that a year ago, maybe we spoke in March, I think March 2025 last, there was this open debate. Well, first of all, there was a debate about whether the technology is going to hit a wall, which we'll get to. But there's also a debate of like how many customers this is actually going to be applicable for and how many, how broad can you actually get? Or is it just going to be useful in like some very narrow areas like coding?

This is from Anthropics S1. From September 30th, 2025 to September 30th, 2026, we grew customers with $100,000 or more of revenue on an annual basis, 12-month trailing basis, from more than $950,000 to more than $11,900. That type of customer growth suggests that – I mean, some companies it's easy to spend $100,000. you need to go through approvals. You need to show something to get that type of money approved. That would suggest that some of the technical limitations are starting to be beat by model progress. And this technology is actually going to be, you know, broadly applicable to a lot of companies. And to, like, add further on, based off of what we were talking about, the average spend per seat right now in even startups, because this is from the Ramp Economics Lab, is like $21 or something like that.

So companies, by and large, are not even using the most advanced style of this, and yet the customer growth and the applicability is growing like that, which would suggest that this is the bull case for Anthropic. It would suggest that we're at the very beginning of this. There's already great momentum. There's already lots of different use cases that are starting to be valuable to companies that are working with this technology. and the opportunity for that to grow is enormous. So, I mean, I think that is where some of the play is. I think one thing you didn't mention is, like, more than a dozen studies have shown a lack of ROI. And so, and, you know, we talked about the big customers cutting down. And so, like, there's multiple dynamics here, right? People hear about this stuff a lot. They, you know, think that they need to do it. They often don't get the results they want. And it's a question of, you know, can you sustain that growth? I think that the models are better now because they are neurosymbolic. And things like loops and conditionals, which come from classical symbolic AI that people used to piss on for a number of years,

are now in widespread use, and that has made them better. It has not made them foolproof. We still have problems with hallucinations. LLMs are still at the core. So you now have these harnesses around the LLMs. of the LLMs still have some comprehension problems, still have some hallucination problems. And it becomes a kind of like case-by-case basis for does this really work in your business? And some of what we see is experimentation and some of it is, you know, long-term people are really committed to using it. And the question is, can they have this sustained, you know, massive acceleration of growth? And we saw some slowing down in Q3 and that, I mean, that is the dynamic that matters. Okay. Here's – I have two more points to make. One is in terms of what the arguments would be against your position about just say no to the Anthropic IPO. Yeah. One is that like up until this point of the discussion, we've largely been talking about Anthropic as a model company. Yeah. And I wrote down a note, they're not a model company. They are a product company that has models. Cloud Code, Cloud Cowork, Cloud Design.

just this week. Anthropic, this is from Bloomberg, Anthropic unveiled a suite of tools that weave the company's artificial intelligence capabilities into more everyday office work, a move that puts it into sharper competition with Microsoft and Alphabet. It has something called Cloud Sheets that resembles Microsoft Excel and Google Sheets. Now, I don't know if this is good for the world because you have this possibility that AI is just going to ingest everything, but they have shown an ability to build their own products and they become competitive with software today. And so someone within there might say, well, who cares if some of these model advantages are competed away? We're going to have the users, we're going to have the native integrations and we will be able to build AI products and basically upend a world of legacy software that's trying to bolt on AI and we're going to build it natively. If they can do that in a distinctive way, then that helps a lot, right?

I mean, a lot of the argument that I've made is that so far they don't have a lot of moat. The only moat they have is around these very high-end models, and even there, like the lead goes back and forth with OpenAI. And one week they're getting the revenue growth and another week OpenAI is. So I think it smart to try to find products that other people can build or you got enough got enough lead You know we saw with Cloud Code they actually did have an amazing product And then their cybersecurity sucked and the code leaked. And then, you know, Codex replicated it and so forth. I don't know if Codex replicated with the leaked code or not. But, you know, OpenAI wasn't very far behind. So you can ask the same kind of, well, around the thing that you just brought up, you can ask a couple questions. One is like a spreadsheet really needs to work. So, you know, problems of hallucinations and so forth are a problem. I can see some, or a potential issue, I should say. I can see some ways that they might do this to make it work, but it really has to work.

Like, you know, if you're relying on these monthly spreadsheets or whatever and it starts hallucinating, then people are going to get disgusted with it. And you can think about things like Copilot with Microsoft tried to do stuff a little bit like that and didn't work very well, right? And so people mostly got disgusted with it and didn't use it very much. That was either a technology problem or a Microsoft problem. I would argue it probably is the latter. I will refrain from comment on that. But, yeah, those are the possibilities. And I'm refraining because I don't really know. If Anthropa can do that better and uniquely better, then for sure that helps them. There's no question. If they can do it better and open AI and Google can replicate whatever it is that they do, you know, in three months, then it's, again, not a sustainable mode. So we have to see if it works. well, and then we have to see whether other people are able to replicate it quickly or not. Okay. Here's my last point, and don't hate me for this one, but I titled it, They've Proved the Doubters Wrong. Now, there's nuance here,

but... Yeah, which doubt? Well, okay. So, let me make the argument, and then we can go back and forth. So, you know, when we spoke last time, I think you mentioned that, like, a new paradigm shift is needed. Yeah. You know, I think... And there actually has been. Yeah, there has been. Right. I think that the point is that they have so creatively anthropic open eye they have so creatively not been religious about their their pure techniques the pure LLMs they have creatively incorporated other techniques they didn't say which was my whole argument you're going to need to do that and they never said we're just going to keep scaling. Well, they might have thought, let's keep scaling LLMs until they get it to AGI. They did say that. Sam said that a lot. They said that. Sam and Darren both made that argument. They're not doing it anymore. They realized, I think what happened was it's pretty amazing. They realized that didn't work. Yeah. Or that wasn't going to

work the way they wanted to. Which we could say they realized that Gary was right. They would never admit that publicly, but that is what they realized and that's why they're doing something different now. And what's happened? That we were, you know, the data wall was blown right through. Reasoning really worked. Reasoning still has some problems. Yeah, I know you called reasoning hype, but this idea... I mean, I didn't... That's not a quote from me, but I have said that the thing they're calling reasoning isn't really reasoning. It's not going to be fully robust. They're going to have problems with it. Yeah, calling it reasoning is hype is the direct quote. Calling it reasoning is hype. Yes, right, right. I can look at that. So that's worked well. They're starting to, you know. It's worked well-ish. I mean, you know, all the best results right now are in math, right? Yeah. And there's a reason for that. The reason is in math, you can have an LLM create whatever it creates, sometimes right, sometimes wrong, and then you send it to an actual reasoner, which is usually lean,

is a formal reasoning problem solver that works with known axioms. He can do verification. So the reasoning there is not done by an LLM. The reasoning to decide whether your proof is good is not done by an LLM. It's not done with an LLM with extra compute time and so forth, the kind of stuff that I was saying calling reasoning. It's done with an actual reasoner. There's a symbolic system called Lean, right? And so they are not doing the reasoning with the stuff that I said would not work. They are doing it with the stuff that I said would work. But is it the LLM? Help me understand this. Is it the LLM that attaches to that? That's because the rough way to think about it is like a generate and filter. The LLM generates a bunch of possibilities, and then the external reasoning system, which is lean, at least as far as I understand what's going on, goes through it and decides which is right. That is the whole essence of being a neurosymbolic system. Neural network generates. I mean, there are different ways to do neurosymbolics, but this is a completely neurosymbolic solution.

The neural network generates something. The LLM generates something. And then something else says, does this make sense? So you're doing external verification with something that is not an LIM. And it's a version of a symbolic AI idea that goes back to the 1950s, right? Newell and Simon had general problem solver in the 1950s, which, by the way, actually solved some novel theorems. Nobody remembers that. But the notion that an AI system could solve a novel theorem goes back to long before neural networks were doing anything of substance. In the 1950s, Newell and Simon did this. So you have an upgraded version of what Newell and Simon did, pure symbolic, working together with the neural network. So it's a partnership, like I always said, you need. So you would say, you might say, well, Gary, then why aren't you satisfied now, right? Gary, why aren't you satisfied now? And the reason is because the way that they're doing it is still kind of narrow AI, right? It is focused on a theorem-proving symbolic system that does that and nothing else,

and they're building a hybrid that is really focused on that. The problem with AI for 70 years is that we know how to build narrow solutions to narrow problems. The classic version of that is like a chess computer could play chess, but it couldn't do anything else. You can do this particular trick where you have a whole bunch of axioms like you get in math, and you can verify them. You can do this trick to some extent in coding. So you can have verification of a set of benchmarks. They're not really benchmarks. I'm blanking on the word people use. typically used, but you have a bunch of tests to see whether, unit tests and so forth, to see whether the system performs as expected. And so you can verify, not 100%, but you can do a lot towards verifying that your code works. And so that's why code works, and it's why the math stuff works. But if you want to, say, make a plan to invade Taiwan or to defend Taiwan or whatever, you can't do the formal verification techniques.

And then this technique that is super cool in these very verifiable domains doesn't really work for you anymore. And that is really where AI is currently stuck. I mean, you know, AI keeps evolving. It's been evolving for 70, 80 years. The place where we are right now is we know how to make this stuff work in domains that we can formally verify that the answer is correct or at least come pretty close to that. And most of the world isn't like that. And that's where we still have work to do. For those who are, let's say, on Wall Street and they've heard neurosymbolic systems a couple of times, just give a one-sentence explanation of what that is in a simple way. Neurosymbolic AI is basically the marriage of the two dominant approaches of AI for the last 70 years. One of them is called a neural network, and it's very loosely modeled on the human brain. You train it on a lot of data. It does some statistical approximation. the best known example of this is the large language model. Symbolic AI looks like classical computer programming.

It's verifying theorems. It's having functions. It's writing loops, do this over and over again until this condition is met. It's running if-then statements. It's not probabilistic the way that... That's complicated because there are ways of doing it probabilistic, but it's much more about writing a function to do something in an abstract way, You're using those parts, building modules. You know, it's the stuff that you would learn in a classical computer science class or a classical AI class. And so these two fields or approaches, I should say, were really hostile to one another for a really long time. And it goes back to the 1950s with fights over resources, who would get the grants, who would get the best graduate students and so forth. These people really, really have come to hate each other for decades. And my whole career had been about saying you actually need parts of both. So my dissertation in 1993 with Steven Pinker was looking at child language and neural networks and symbols and saying you need both, right?

And my most important book was my, although not most accessible book, was called The Algebraic Mind in 2001. And the subtitle was Integrating Connectionism, which is another name for neural networks, and Cognitive Science, which is really pointing towards symbol manipulation, the classical AI stuff. And although the book was really about human intelligence, it was more broadly about intelligence in general. And it made the argument, again, that you need both. You cannot do just one. The symbols alone, the old stuff isn't going to work. The new stuff isn't going to work. You need to bring them together. And my famous piece, Deep Learning, a Critical Appraisal, said, I'm not saying we should throw away deep learning, but we need to have symbolic stuff. We need neurosymbolic stuff. My famous and very notorious paper, Deep Learning is Hitting a Wall, did not say we should throw away deep learning, which is what everybody who read the title and not the paper believes that the paper said. It actually said, but you need to add symbolic stuff to that, too. You need to have neurosymbolic systems that use elements of both. And that is, in fact, what has happened.

You must be impressed with these companies that they actually went out and did that. Well, I think they did it out of desperation, but it worked. And I think Cloud Code really showed the way. Like I give Churney a lot of credit for getting rid of the dogma. Or is Churney the founder of Cloud Code? Yeah, I don't know. He's been on the show. Right, yeah. I mean, he's really the driving force, as far as I know. I'm on the outside. True, true. I believe he's – and so the other piece of it as part of the war was a lot of dogma. and people like Hinton said you don't need any of the symbolic stuff it's old fashioned, you're wasting your time and he told them meeting G7 like you in Europe, don't spend any money on the symbolic stuff it was super dogmatic and a lot of the hostility to me over the years was because I kept saying yeah you're going to need the symbol stuff too and so it is to Cherney's credit that he said you know I don't care about the dogma I want to make this thing work and that was a huge advance and even still you're not going to buy any anthropic shares I mean, not at this price. I just don't see the route to profitability.

And so it feels to me like it's going to be a long, slow decline. I think they have a lot of smart people at Anthropoc. There's plenty of reasons to like the company. It's the valuation that I really think is absurd. I just want to end here. What is your AGI timeline? Because this is, honestly, Gary, it's the most hopeful I've heard you about AI since we started. speaking. I mean, hopefully in some ways or in others. I guess my timeline is probably sometime next decade. Anybody who gives you a specific year already has no credibility, right? You're not a scientist if you don't have boundaries, what we call confidence intervals around it. I still don't see it happening this decade. I still think there are core discoveries especially around world models to be made and not world models the way Lacoon does, but that's a story for another day. I think there are core things we have not yet figured out and that it typically takes a while once you've figured something out to make it work in practice. So deep learning is like 50 years old or 70 years old, and it took a while even after the 2012 discovery that you could use GPUs

in order to make your neural networks work better. You know, that was 2012. We're, you know, finally reducing it to practice kind of now, like 14 years later. So let's say somebody actually has, like, the right idea now about world models. It's still probably going to take several years until it's rolled out in a way that really works and so forth. I just don't see us, and that's not the only thing that we need to solve. Like, that's like a prerequisite to getting to continuous learning, but continuous learning isn't solved, et cetera, et cetera. So, like, you know, people like Rich Sutton are on my side now. You know, he made all these arguments about the bitter lesson, and now he basically sounds like me. Like, we need some more discoveries here, right? I think a lot of people who thought about these problems for a long time realize we need more discoveries. So if you need more discoveries, then that's just not going to happen, not be reduced to practice this decade. We might make, in the most optimistic case, we might make those discoveries in what remains of this decade, but getting them into practice will take a while. And so, you know, real Asia, I just, I don't see how it happens this decade.

But if you're talking about, like, what could happen 13 years from now, that's a really long time. There's a ton of money being spent on it. We're getting a little bit smarter about what we don't know. We're getting much less dogmatic about the, you know, it used to be neural networks or the highway. We're not dogmatic about that. Actually, if you actually read the 2022 paper, which nobody ever does, they just read the title, Deep Learning Sitting Wall. Headlines matter, I guess. Yeah, I mean, that was like too good a headline in a way. If you actually read the paper, I express the same note of optimism as I am now. I sort of mistimed it, but I said, I think people are finally getting a little open to neurosymbolic AI, and that would be a great thing. And it took like two years past that before they really became open to it and maybe four before it just became a standard thing everywhere in the industry. But that was a note of optimism then, and it's a note of optimism now. It is the dogma more than anything else that has screwed us. The idea that you just need a pure LLM was a really stupid idea,

and the idea that all you need is data centers is kind of an extension of that, and that's a stupid idea. The idea that, hey, maybe we should use old AI and find new ways to put them together is a great idea. It's much less dogmatic than these other ideas. And so, yeah, to the extent there's less dogma, I am more optimistic. I will, however, put in another pessimism, which is what we haven't made progress on at all in any real substantive way is AI safety. And I don't think we want to get to AGI if we don't make progress on AI safety first. Yep. All right. The website's garymarcus.substack.com. It's actually, there's three headlines at the top are amazing. A report from the Vatican with Gary shaking the Pope's hand. Just say no to Anthropics IPO. And Jensen is cooked. That's just the world we live in today. So, Gary, so good to see you. Thank you for coming back on the show. And let's do it again soon. Awesome. All right. Thanks, everybody, for listening and watching. And we'll see you next time on Big Technology Podcast. Thank you.

番組の概要欄(原文)

Gary Marcus is a professor emeritus at NYU and the author of Marcus on AI. Marcus joins Big Technology to discuss his bear case for Anthropic’s IPO and whether its business can justify its enormous valuation. Tune in to hear us debate the company’s explosive revenue growth, massive compute commitments, and prospects for turning AI into a profitable business. We also cover competition from Chinese models, Anthropic’s push into software, and why Gary is becoming more optimistic about the technology itself. Hit play for a spirited discussion about AI’s business model and what it would take to prove one of its biggest skeptics wrong. --- Enjoying Big Technology Podcast? Please rate us five stars ⭐⭐⭐⭐⭐ in your podcast app of choice. Want a discount for Big Technology on Substack + Discord? Here’s 25% off for the first year: https://www.bigtechnology.com/subscribe?coupon=0843016b Learn more about your ad choices. Visit megaphone.fm/adchoices

出典・この記事について

番組
Big Technology Podcast(Alex Kantrowitz)
元の回
Why The Anthropic IPO Will Fail — With Gary Marcus
配信日
要約作成日
公式サイト
www.bigtechnology.com

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

X でシェアApple Podcasts で聴く

関連エピソード