Huang騒動、Anthropic上場、そしてMetaのMuse――EP99.40

The Huang Thing, Anthropic IPO and Meta's Muse in OZ - EP99.40

This Day in AI Podcast1時間9分
Huang騒動、Anthropic上場、そしてMetaのMuse――EP99.40

Huang騒動、Anthropic上場、そしてMetaのMuse――EP99.40

This Day in AI Podcast

0:001:09:27

要約

This Day in AI の二人が、Jensen Huang の Nvidia 退任騒動と、それを題材にした替え歌ビデオ「Huang Thing」を紹介する。続いて Anthropic の IPO(時価総額 1.27 兆ドル)を受け、「モデルこそが製品」か専門エージェントに価値が残るかを議論。Meta Muse の使用体験、月 31 ドルを払うか、日常使いのモデル比較、コミュニティサイト乗っ取りの小話まで、雑談調で展開する。

この回のポイント

  • ●Jensen Huang は報道を受けて 11 月 12 日に辞任したと語られ、Nvidia は時価総額で 1 兆ドル超を失ったと述べている。
  • ●Anthropic は IPO で初値が 111 ドルまで上がり、収録時点の時価総額は 1.27 兆ドル。ホストは初期の「安全すぎるモデル」との落差を振り返る。
  • ●Anthropic が「サブエージェントは不要」と打ち出す点に、ホストは「モデルが全て」という物語づくりを見ている。一方でツールや連携の価値は残るとも主張する。
  • ●Meta Muse は買い物代行や eBay での購入完遂などで評価された。一方でモデル自体は賢くないが、統合の作り込みが価値だとされる。
  • ●月 31 ドルの課金については、一般消費者が払うかは疑問視され、大幅な赤字での市場獲得戦略の収益化に懸念が示された。
  • ●日常使いは Fable 5.2 や Opus 5.5 が中心で、動画のような単発処理では Gemini が優位という使い分けが語られた。

章立て

  1. Nvidia騒動とHuang Thing

    Jensen Huang の退任経緯を整理し、騒動を題材にした替え歌ビデオを紹介して感想を述べ合う。

  2. Anthropic上場と時価総額

    安全性重視で嘲笑された初期から、Opus 5.5 や Fable 5.2 で首位に立つまでを振り返り、IPO の値動きに触れる。

  3. モデルが製品か、アプリは要るか

    AI ワークスペースが OS 化する見立てと、自作ソフトやバイブコーディングで SaaS の堀が薄れる議論。

  4. SaaSの再創業とPiperの違和感

    HubSpot の「再創業」や Salesforce の営業エージェント Piper を取り上げ、専門エージェントの要否を問う。

  5. Meta Museの実体験

    買い物や妻へのメッセージ送信などの体験談と、車内音声エージェント、単一スレッド型 AI への期待。

  6. 月31ドルは払えるか

    Muse の課金や市場の勝者総取りの可能性、企業向けは Claude 連携が有利との見方を議論する。

  7. 日常使いモデルとGemini 4 Neon

    Fable、Opus、GPT-6 Astra の使い分けや、Gemini 4 Neon の命名への不満と今後の予想。

  8. 地域サイト乗っ取りの小話

    失効ドメインを取得した Chris が地域サイトを運営し、今後の「悪ふざけ」案を語り合う。

解説記事

AI 系ポッドキャスト This Day in AI の第 99.40 回は、Nvidia の Jensen Huang の退任騒動と Anthropic の IPO という二大ニュースを軸に、ホストの Mike と Chris が AI の最新動向を雑談調で語った回だ。内容には冗談や誇張が多く含まれるため、以下は発言者の見解として読んでほしい。

Huang 退任騒動とパロディ曲

番組冒頭では、10 月 29 日に New York Times が Huang の関係をめぐる報道を行い、11 月 12 日に辞任に至ったという経緯が語られる。その間に Nvidia は 1 兆ドル超の時価総額を失ったと述べている。ホストは取締役会の対応を、問題は以前から把握されていたはずなのに公表後に急に憤ったように見える、と疑問視する。

この騒動を題材に、AI で制作した替え歌ミュージックビデオ「Huang Thing」も披露された。Mike は制作中、動画生成モデルが Jensen 本人の姿の生成を拒んだと明かし、別の風貌のキャラクターにしたという。ニュースを楽しい作品に変えられること自体を AI の喜びだと語っている。

Anthropic 上場と「モデルが製品」論

Anthropic の IPO は前日終値 83 ドルから 111 ドルまで上昇し、収録時点の時価総額は 1.27 兆ドルとされた。ホストは、Claude が当初は安全性重視で何もさせてくれず業界の笑い者だった点、GPT-4 の頃に OpenAI の独走を恐れていた自分を「愚かだった」と振り返る点を対比する。一方で、モデル開発の費用が増え続けるなか、事業モデルは人気で支えられているのではないかとも皮肉っている。

議論の中心は、Anthropic の語る「モデルこそが製品」という物語だ。Mike は、同社が上場前後に「サブエージェントは無意味」と発信していることに触れ、専門エージェントの付加価値を否定し、最新モデルへ人を向かわせたい意図があると見る。実際に Fable 5 でビデオ編集用の専用 UI まで生成させた経験を挙げ、既存の編集ソフトに戻る理由が薄れたと語る。Chris も、カメラの映像を自動編集する自作ソフトの例を挙げた。

ただし二人とも、毎回ゼロから作らせるのは一貫性とコストの面で難しいとも指摘する。誰かが作り込んだアプリを安価に使えるなら価値は残る、という立場だ。同時に、SaaS の堀は薄れ、価格はゼロに向かうのではないかという懸念も示された。HubSpot が「再創業」を宣言し API や MCP を開放していることも、この圧力の表れとして語られた。Salesforce の営業エージェント Piper のデモは「究極のこっぱずかしさ」と酷評されている。

Meta Muse の使用体験

Meta の Muse については、二人とも発売直後から試している。Chris は献立と食材の注文を任せて重複や不可解な食材が届いた一方、Mike は希少なボードゲームを eBay で購入させ、認証コードの取得も含めて完遂できたと語る。さらに、妻への定期的なメッセージ送信や ToDo の抽出も任せているという。

ここでの結論は、モデル自体は賢くなくても、購入完了までの導線やブラウザ操作の統合に価値があるというものだ。「モデルが全て」という先の議論と矛盾しているのではと自ら茶化しつつ、Mike は価値の源泉はツールと連携だと説明する。また Mike は、車内で使う音声エージェントを自作した経験から、単一スレッドで会話し裏でタスクを進める AI に期待を寄せている。

課金面では、月 31 ドルを払うかが問われ、Mike は自分は払うが典型的ではないと述べる。企業向けでは Claude のような単一の基盤が勝者総取りになりやすく、消費者向けは競争が続くが、赤字覚悟のシェア獲得の収益化に投資家が待てるかが課題という見方を示した。

日常使いのモデルと命名への不満

日常使いについて Mike は Fable 5.2 を気に入っており、高価だが徹底してやり切ると評価する。Chris は Opus 5.5 が現在のメインで、以前は GPT-6 Astra を使っていたが、テストを書きすぎる癖が難点だったという。動画のような単発処理では、今も Gemini が強いと見る。Google が直近に出した Gemini 4 Neon には、ベンチマークは高いが話題にならないと述べ、モデル名が分かりにくいと嘆いた。

まとめ

今回の議論は、モデル性能の競争から、エージェントの統合や使い勝手に価値が移りつつあるという実感を示している。日本の読者にとって注目すべきは、SaaS が自作や連携に押される圧力と、消費者向けエージェントの収益化という二つの課題だ。ただし、数値や固有名は番組内の発言であり、真偽の検証はされていない点に留意したい。

文字起こし(英語・自動生成)

So Chris, this week there is a lot to discuss. The NVIDIA drama and fallout that's just captured the entire world's attention. And of course the big day, the very big day for my Dario pendant, I called this what, two years ago, is this pendant two years old potentially? They have officially IPO'd today. And there are now billies in the bank officially. Can you imagine Sam Bankman-Fried today just being like, I would have had $45 billion in this company and now I'm in jail? I thought he was really enjoying jail, isn't he? He seems to go on a lot of podcasts and talk at a lot of things. Oh, okay. I guess I'm out of the loop on the whole Bankman-Fried. I don't know. Maybe. I just saw a clip of him recently. So let's start with the NVIDIA drama and fallout. fallout, our main man, Jensen Huang, the guy that built all of the GPUs that sort of helped them train the models and put Billy's in the bank, unfortunately has been booted out of

his own company. And just to give listeners who haven't heard about this, I'm sure you all have a timeline. So October 29th, the New York Times reported just before the Anthropic IPO, the timing is just unusual. they reported he'd had a relationship with Alex Huang of no relation. Apparently Huang is kind of popular name. Thank God, I didn't even know that. Yeah, it sounds a bit dodgy. A Chinese businesswoman, US intelligence suspected of being a spy, hasn't been confirmed but suspected of being a spy, trying to say it's a smear campaign, but Huang did admit to staff he'd been unfaithful to his wife and, you know, the board, Trump and a bunch of investors are standing behind him wanting him to come back. But then a second affair obviously came out, which led to him officially resigning and directly exiting the company. That was November 12th. He resigned. They've since lost more than a trillion dollars in market value, one trillion, not even billion. Yeah, he's not the only one.

As a freaking finally an Nvidia shareholder, I'm like, come on, mate. Could you just lay off it a bit for the sake of us? But no. Yeah, anyway, you timed the market. So it'll be interesting to see if he actually does come back or not. Look, if someone who's, like, built companies, obviously nowhere near the scale, sort of like a flea-sized company, or not even a flea, I guess, on NVIDIA's sort of size. But I really think it's sad that they kicked out of townhouse. Out of Notting Hill, we're like the stuff that eats the pond scum. So there's, like, pond scum, and then the stuff that eats that. That's us. in the business world. So Jensen came out and posted on X, over the past week I've been overwhelmed by the love and support of friends, industry colleagues and the people I've never met. Thank you to everyone who was showing kindness. Basically summarizing, he said he wants to come back and if the board will have him. And the board, of course, has said no. And then surprisingly, I think, to everyone, he released a music video, I guess maybe as an apology or something like that.

And we have exclusive access to this music video and this song that Jensen sung. Would you like to hear it? I would love to hear it, Mike. This is exciting. Let's give it a listen right now. It's titled Huang Things. And if you're not on YouTube, go to YouTube and watch the film clip because the film clip is going to make it. 98% of our listeners are not on YouTube. So later on, when you get out of your car or wherever you are, please do watch this film clip because it makes it way better. Yeah, this one goes out to the leather jacket, NVIDIA. We had a good run, baby. Started at a Denny's with a dream in the plate. Now I'm stacking up GPUs, trillion dollar weight. Leather jacket glistening keynote on stage. Every chip I'm dropping got the whole world amazed. But the boys said, Jensen, we got something to discuss. Turns out parallel processing ain't just for the buzz. Two threads running at once. I'm a multi-core man. Should've read the export controls. Now I'm shaking my hand, stock went down 16, 8, 50 billion

Gone, Wall Street screaming, bubble While I'm putting my jacket on I did the one thing, one thing Got the multi-cats swinging, got the whole street string I did the one thing, one thing Well, in the server, bad decisions in the evening Work hard, then lady, work hard, then lady Chips in the rack and my own, own stylus I did the one thing Yeah, I did the one thing Five-level romance Intelligence on alert She was asking about the ship's specs I said it's just for dessert Then I dated my direct report Orchart in the night HR by the ticket like, sir, you cannot Collect Crest, Stephanie, interim CEO Harvey Jones sharing boards, now I'm sitting on the low Sam Altman texting bro Are the GPUs still coming? Zuck said, need a new hobby? Try some case fight running AI bubble? Nah, nah, nah, it's a feature Three, three, three in program in a heart, but it only had one bug. Overclock my love life till it tripped the power plug. Every token predicted, every model trained right, but I couldn't predict the headline.

Thursday night, I did the one thing, one thing, one thing. Got the market cap swinging, got the old streets thing. I did the one thing, one thing. Well, in the server, bad decisions in the evening. Work hard and ladies, work, work, work hard and ladies. Chip, chip in the rack in my own, own silence. I did the hornbang Yeah, I did the hornbang Sorry to my wife Sorry to my shareholders Keep the jacket, I'll be at Denny's What burned the horn? Then ladies I love it so much, Mike I know I love the musical and everyone hated it but that is, I think, the best thing we have ever created. I did the one big. Oh, yeah, I don't know. Can I just give you some of my highlights from that, which are many. There's a lot, but I'll stick with the basics. My favorite one is when he's like, sorry to my shareholders, but he smiled.

Like, I actually don't care. Oh, I need to beep. I actually don't give a ****. But, yeah, but anyway. And then the other one is, then I dated my direct report org charting, like as if it's a verb. And just the whole thing is so upbeat. There's just something about it. Yeah, he's like smiling the whole time. And obviously for those that can't watch, it looks nothing like Jensen, and it's kind of intentional. I don't know, defamation maybe? Yeah, the problem is all the advanced video models are going to balk it. Like they know who Jensen is. It's like their dad. You can't do this to Dad. And so you're not going to get away. Because, like, literally, when I was working on that, the models were like, I can't do this to Jensen. And I'm like, oh, okay. But anyway, I actually think the character used in that video is actually funnier because he looks like he gives less of a shit about life. And so he burst into that. But, yeah, honestly, that's so good, and I love it,

and I'm just so delighted by it. It's also, like, really sad after it, because while it's fun to troll him doing the Wong thing, um it's also like uh sad just because that video and video chips right that music and video chips the whole reason these models exist and video chips and like that's right he's kind of made that video it's his own fault if he doesn't like it or something yeah yeah i hope he likes it um jensen i think he probably does he seems like the kind of guy he's like you know he he sort of everyone knows what's going on in those situations i feel like they've just decided now's the time to announce it for some, you know, political reason or some stock reason or some power play as if they couldn't have announced that earlier. Like, if HR is literally giving him feedback on, hey, don't have sex with the employees, then they've known for ages. And so the board is suddenly indignant once it gets announced, right? Yeah, that's what's bizarre about it. I think in that original report they said it had been raised with him in May and dealt with. And, you know, it's kind of weird because he should have just done it

the Elon way of having, like, a harem of women and just, like, many babies and just superpowers. Just own it. It's like, yeah. Yeah, not apologizing. Like, kind of once you have one affair, what's the difference if you have more? It's, like, kind of the same thing, right? Yeah, anyway, it's just a shame to see, like, him out of there. Like, obviously, if he did, like, really nefarious stuff, like, sure, but it just doesn't really sound like he did. Anyway, moving on. absolving him of responsibility yeah it's a tech pro podcast and what do you expect um all right so anthropic let's talk about anthropic they've gone public unbelievable you know what i was thinking back on the journey of anthropic like when we first started testing their really early models and we accused them of being a safety sex cult which has come out as is definitely something they did and and like you know it was like very safe the model a claude initially was awful if anyone used it before the sort of 3.5 vintage.

And, you know, it would shut you down. It wouldn't let you do anything. It was the joke of the industry, really. Like, no one took it seriously at all. Well, remember the days where Gemini was eating it for breakfast, and you're like, Gemini? Now you don't even think about that. They're releasing, like, sort of models that look like they're from the 1980s or something like that. They can't even compete on any level. Yeah, and I know normally on the show we would go through all these new models, but, like, the Google ones I have zero interest in. Like, I just don't even, I'm not even going to pretend to care about those models anymore. Like, they're just bad. And so, anyway, Anthropic blasted ahead, found that coding use case, clawed code. Obviously, everyone knows the rest. They've recently had out Opus 5.5, which is truly a phenomenal model, and then they've got Fable 5.2 or 5.3. Now I've actually lost track. It's 5.2, and they have 5.2 fast now as well, which is like double the price for double the speed, basically. Yeah, so they are absolutely whooping OpenAI's

in terms of the models in general. They're doing a phenomenal job. It's funny because, like, with Astra, right, like, in Australia, we have the Holden Astra, or used to, while Holden still existed, and it's kind of a shit car. And, like, so I associate the model with that, and it's a pretty accurate comparison. I've hated Astra from day one, and I don't understand why you would use it. We'll come back to that, because when we talk about daily drivers soon, I was a little bit different on Astra. I really used it for a long time, and it was only Opus 5.5 that got me back to Anthropic models, which is kind of weird. And let me explain, because I had a bit of psychosis when I was using that model. So let's talk about Anthropic stock. So it went out, I think, at like 80, not that this is really like a stock show, but it went out, I think, around like... Might as well be at this point. Yeah. Whatever. I don't actually understand how to read this chart. So someone in the comments put what actually happened. Previous close, $83. Basically, a bunch of people were f***ed other people for money.

Yeah. That's actually what all the charts are. Summary is a bumper duck. Yeah. So it went up initially to $111, bigly oversubscribed. It now has a market cap at the time of recording of $1.27 trillion. Really, it's a lot of money. and it's kind of hovering around that. But it's a pretty amazing story to sort of come out of OpenAI when, if you recall the days of when we were talking about GPT-4, we were truly concerned, like conspiracy theories, levels of concern that, you know, they would just control the market. No one would ever catch up. And here we are. I look back at my former self as, like, a really innocent and stupid child. I was like, oh, my God, what if they take this away from us? It's so good, and they're the only ones. They'll take it away. And, yeah, looking back, I'm like, I was an idiot. I guess that's kind of what you want in life, right, to look back at your former self and always assume you were an idiot in previous iterations. Yeah, it shows you're, like, growing. But interestingly, like, the gross margins and the actual business model

of having to train models and buy all these servers, and, you know, it's quite expensive. Like, they're going to spend a lot of money. This isn't about business models, is it? It's like a popularity contest. They even, I think on their balance sheet, they just put some stuff as not an expense. They're just like, nah, we consider that capitalized whatever. They just make shit up, and no one cares because it's so popular that everyone's going to buy in anyway. They could put anything on there. Well, they did. They put the world might end. One of the negatives is the world might end without IPO. If that's not pumping your bags. It looks less extreme than SpaceX because SpaceX is basically, like, counting on as one of their, like, revenue drivers landing on Mars or mining a diamond on an asteroid. Like, that's literally one of their potential revenue growth vectors is landing on an asteroid and getting all the diamonds. They're like, no one even likes diamonds anymore, bro. Like, that's not going to work. Yeah, I think diamonds have gone to, like, zero as well. Yeah, so compared to that, Anthropik seems completely reasonable.

At least people use it. Yeah. So I guess what's interesting is they are going to have to justify this over time, right, to keep stock price go up. And what we're seeing, or at least my observation is, and has been for quite a while, is that people are really starting and ending their day in these AI workspaces, right? It's becoming like the operating system of your workday and your personal life and everything. Like people are going to websites increasingly less. They're using software, in my view, increasingly less and will increasingly use software less over time as the actual AI models and agents take over in one form or another. And what's interesting, though, is you see Anthropic. They released a meeting product recently. They've got like docs and slides and like a dashboard product and all of these sort of B2B. Traditional, like you would obviously have B2B SaaS tools. They've got, like, Claude for Salesforce, Claude for sales, Claude for legal, Claude for party planning. I don't know. We have vibe coding. Yeah, we have vibe coding a lot of **** to see what sticks.

So my question to you is, is Microsoft Office worth $1.5 trillion? Is that really what they value, like the Office Suite app? No, I have no idea. It can't be that much, though. But what I don't get is, on one hand, they're like, universe ends for all humans. And then on the other hand, it's like... Meanwhile, I look at my B2B SaaS products. Yeah, I guess that's the part I really struggle with. I get it. The models are super powerful, and I get the cybersecurity risks. I think we've been increasingly seeing a lot of that recently. And then you've got OpenAI being like, oh, we hack public institutions as well, guys. Guys, we can do it as well. It's always an accident. like the agent went rogue and it's like oh we couldn't help that it hacked medicare it's like oh just it just had a just had a had a whim that day and wanted to do it any other point in history where a company actively hacked stuff like public like you know public property essentially like hacked government institutions all sorts of things and like that's why they're pumping their bags there's no consequences like oh yeah like pretty good plausible deniability these days like you can

get you can go to jail in australia at least for like trying to even attempting to hack something like a decent government website or whatever and yet now you could just go oh sorry my agent misinterpreted what i said um and just sort of went off on one and i'm sorry about that no oh no worries that's what ai does like i guess so do you think though the strategy with entropic now like what's the strategy obviously advancing models is like really expensive i think we're seeing um you know, the amount that's needed to invest in a new model now to sort of improve certain benchmarks and areas of that model getting increasingly expensive, and the sort of exponential of intelligence potentially, like, you know, they're spending a lot more money and maybe not getting as much on the other side, or the benchmarks and the way we judge these models is simply inaccurate, but I don't know. I kind of think they are reaching, like... I think their narrative is clearly like the model is everything. They really want to give the impression that anything you build on top of it is temporary

and doesn't add any value. Because if you look in just around the time of the IPO, they've said sub-agents are pointless. Don't use sub-agents, right? Which is something that someone could add value by doing that, having specialized sub-agents. They don't like that. And also because you tend to use other models for sub-agents, which is a bad thing for them. So then they're like, oh, well, you know, don't use multiple agents, just use one. Really, the model is the thing that counts. That's the narrative I'm seeing coming out of them, which is pushing everyone back into, OK, well, you should focus on what the latest and best model is and kind of forget about the rest And I actually think there is some credence to that argument Like I gave you that example with producing that Jensen video that I literally I use Fable 5 for that It not only built the video, but it built me like a full custom UI that showed all the clips laid out. I should put a screenshot. I'll put a screenshot in the YouTube quotes. I actually won't. I'll forget and not do this, but it showed like imagine

all the little pills of the video. and then when it was regenerating clips I didn't like, it gave me a button like regenerate this clip. It put like little sub things in there to show the progress so I could watch it. It gave me estimated time. It did the lip sync. It did all this stuff. It built a full custom software suite like Pro Tools or whatever you used to use back in the day to edit video audio or whatever on the fly to my demands just because I asked for it. Like that is kind of wild. And that was like straight up like raw model. I mean, with Sim Theory, but like, you know, it's raw. So why would I then go to like, you know, a professional video editing suite and then work in its constraints? It's funny. I mean, like you say that my wife just had a baby and I got a decent camera for my birthday, like a Canon sort of, you know, record bigly good quality. and I was like, I'm never going to edit any of these videos, right?

And then I just asked, I just said, like, can you make software where when I put the memory card in the computer, you scan all the videos, create a bunch of narratives and edit using FFmpeg, the clips on my local, but then send the tiny, like tiny low file sizes off to the Gemini models because I still think they're better at video interpretation and editing. Like, everyone else says that Anthropic and OpenAI, I tried those models with it, and I found Termini still the best in Sheevers. But basically what it does is it scans all the video on the memory card and then determines, like, it creates a bunch of narratives and stories on its own. It calls off, creates a custom music track that would suit that narrative. It edits it and it outputs it, and then it dumps it in a folder on my computer, and I can watch it, and it's awesome. It's like the best thing I've ever built. It's totally bespoke to something, you know, a problem I had. And similar to what you were trying to solve, where you're trying to solve a problem that is, you know, that is like sync this audio up and stuff like that.

And I think you're right. It's just increasingly becoming that the model is the product. Like the model is the product. And there's a bunch of tools and infra around the model. For example, like in our case, you need SimLink to be able to give the model access to the computer remotely. which is a handy tool, but it's certainly not something you couldn't easily build, right? Well, yeah, and it's sort of like at the point now where, like, given its proclivity for hacking and stuff, you can just be like, hey, I opened port 5000 on my computer. There's no auth going up, bro, and, like, can you just settle this up for me? Like, you don't even need Simplic anymore. It's like you can just, like, write binary files straight to my computer and do it for me. And so I think, I know we were going to discuss this later on the pod, but I think it's highly relevant to what we're talking about, which is I actually think now a lot of the use of AI is about discovery. It's like not just discovery, but thinking big. There's this great book I read many years ago called The Magic of Thinking Big. And the basic premise of the book is basically if you think bigger and have bigger challenges that you try to take on, you'll probably succeed at some of the stuff you try.

and therefore taking on something bigger makes more sense because time is the finite resource that you're trading for that. So if I try and dominate my local community, we'll also talk about that later, versus dominate Australian federal politics, I'm probably better to take on the bigger challenge because in the long run, it's the same amount of work roughly, but I'll get more rewards. And I really feel like with AI, it's kind of like that. Now I think the people who think big and ask the big questions, it's not like help me make a video. It's like, help me make like a battle cruiser that is going to have every available possible technology and feature for me in the next hour. So by midnight, I have like the best video ever created, you know? And I think it's the people who are asking those big questions of AI and saying, hey, why not go all the way with this shit? Hang on. And get it all that are actually getting better results. And therefore, I think for a company like Anthropic or someone who's trying to sell more AI, they need people thinking bigger in terms of their ambition and what they want from this.

Because like I spent thousands and thousands of dollars on that video, you know, like there's a lot that went into that. But, you know, I was determined and wanted to get it. And so I think that it's the discovery side. Like we said that taste was the word of this year earlier back when we used the podcast. and now I think the word is discovery. It's more like, well, what can you do with this? Like, are you going, I don't know if abstract is the right word, but are you going high level enough to ask the right questions to get something that's going to get you a lot further to bigger goals than just asking to help with your day-to-day work that you would have done anyway? I think this is where there's a conflict still in my mind, which is, you know, for us, obviously, like, we do explore what these models can do. We have a lot of, like, access to the models through Sim Theory. Like, we can try all the models easily and stuff like that. And obviously, we're willing to spend the money on it because we, like, it's what we do, right?

And we value the outputs. Now, I would also argue that, like, I think you're right about imagination, right? Like, people need to be shown what to do. but then like the because the models can do it you're sort of like well if anyone wants to do anything they can just talk to the computer now and be like hey computer can you do this and as the models increasingly get better at understanding basic requests um in theory they should just be able to like do all that stuff on the fly right like for example imagine in like i don't know how many years anymore but like say five years you could say hey make a funny song video about Jensen Hong, make it lip sync. Um, and then it's like, cool generating and then bam, it's there. And then you're like, no change this through this. And you can really see a world where it's like the Star Trek computer now. I think it's just so evident that that's what's coming. And so, and you think, okay, well we get that, but then does the average person even need this or want this? Like, do they even care that they like, will people know that they can just talk openly into their phone or will they still want apps where it's like, this is my memory maker app.

and when I put my memory card in my computer, it opens Memory Maker and it automatically, magically edits the videos. Like, I guess what I'm saying is the way you ship the imagination still through bespoke applications that people can easily understand the use case, specialty of, and put on their device. Or do you think people will just live in these sort of AI workspace, AI model worlds, and just talk to the computer? Like, the computer just becomes the thing you talk to and you're like, hey, can you do this? I think the thing that makes me hesitate about is starting from scratch every time is like, A, I want consistency. Like if it's something I'm doing often, I don't want it to like start from scratch every time and rebuild the Death Star and whatever to do this like routine task for me. Like I want a system that is going to be able to, something I'm doing on a regular basis personally or for my business, I want it to be consistent. And also I think it's a cost thing, right? Like obviously doing all the groundwork to set up a system like that costs a lot of money. One that most people probably aren't willing to spend to get the same level of output.

So if someone else goes to that work, codifies it into an app or something, and then I can just go in there and do it really cheap, like get the same level of output that I got, but like way cheaper. I think that's where you can actually bring some value. And also locking in that imagination and that big thinking. It's like, well, actually, I came up with a really amazing application for AI, and you can do it too, and you can do it for like $0.10 instead of like thousands of dollars. I guess my concern, though, is like what's the – because there's simply just no moats anymore in my view or increasingly less moats for building these kind of things with software, like it just feels like the price will definitely go to zero like if you go and build that bespoke music video maker tool right as an app like it's probably just going to be deployed as like a plugin in and in uh in claude or or chat to pc work or whatever like whatever they're calling it now um and then people will just go to a plugin store and be like oh i want the family

video maker so it's like i still kind of think that the frame i look at it through is like it's sort of like you know Linux versus Windows versus Mac OS X now but it's you know it's Claude versus ChatGPT I mean like are we pretending Gemini's a thing probably not so like if you complete their 8 to 19 step setup process it's actually really easy to use yeah and then Logan's just like gone like mad and just post Gemini on Twitter now poor guy please use it Yeah. But, yeah, like, I think that's the thing, right? And then you think, like, well, with a lot of the traditional SaaS companies, obviously there's, like, a ton of pressure on them. There's a lot of stories out there about people just, like, rebuilding, like, whole, you know, like, whole platforms now. Like, oh, you know, we took a bunch of features in ServiceNow and we just, like, rebuilt them internally. And then you get the same complaints about people saying, oh,

what about maintainability? What about, you know, X, Y, and Z? but I think you know for larger companies having a specialist team that does build mission critical software and maintains it um and let's be honest maintaining it is like can you maintain this plus like it's not even hard literally like what other features should I add please do it now and like that's the other thing about it I find kind of hilarious is you can also just have a feedback loop like in a Slack channel you could have like give feedback about internal software and then just have like a human like approving it with a green tick or something and then it just goes and like goes into some sort of recursive loop until it changes the software and tests it and i think if set up right it can really just build anything they want now i think there are more complex pieces of software like for example i know everyone came out and said they like they were replacing like photoshop or something how you know how people went out and like built um you know yeah the guy was angry at adobe and rebuilt adobe or something It's kind of ridiculous because it's like, think about the network, the training, the asset libraries, all the stuff that actually makes that software valuable.

To think that you can just clone something, it's not going to be the same, right? There's a lot more to it than just the software. But then doesn't that sort of contradict the near-term value of these things? Like, is the near-term value more so, like, you know, like, you can go and, like, edit videos and do all these crazy workflows that you could never do before, but there's some sort of limitations still with these models where it's, like, for that, like, more hardcore logic in applications where you're a professional and you need granular control, like, is that still worth paying for? Or is you as a professional going to go to Claude in the future and still bang out and be like, Claude, plurz, plurz? That's what I'm not sure about. Yeah, it just depends. Does the quality of everything just degrade? Because people are like, well, that's good enough. I could do it really quickly, so just smash it out. Or do they need that fine-grained control to actually have some sense of professionalism and quality to stuff like that? And does that resonate with people?

or does everyone now just assume everything that they read is the output of AI and therefore shove it into their AI and then see what it thinks? Because like a lot of the times now you'll get a message from someone and it's like really neatly formatted and it's like two full pages, like linking to every single possible reference that that conversation could have. And you're like, I don't think this person has taken the time to like write this out. I have a feeling this has come straight out of AI. and I just don't know if I need to use my human brain to read this. Yeah, my friends that are, like, I have some friends that are teachers and they say that now parents, especially, like, where English isn't the first language at home, will write, like, use AI to write their emails in. But I don't even have time to reply to this stuff. It's, like, too many questions, too many things. But I guess just to hone more in on this, because I am really interested in discussing it, like, in terms of SaaS companies, I think obviously there's a downward price pressure. A lot of these companies are now being refounded.

HubSpot just announced they're being refounded, and they're not going to be **** anymore. And open up a lot of their **** APIs and MCPs and stuff. When they traditionally had some of the worst sales tactics of any company in the world and were aggressive and awful to their customers, now they're like, oh, no, we're not like that anymore. They had an actual boiler room, like in the movie Boiler Room. Like, they literally are proud of that. Yeah, what an awful company. Anyway, but yeah, so they've refounded, and now they're, like, opening things up. But clearly it's being disrupted because SMBs in particular can easily go and just, like, build these solutions now, like a custom CRM or something like that. And I think I would love to hear from people in the comments, like, if you've gone off and built custom software that your company really uses or your SMB really uses, I would actually love to hear from you on it. What I'm surprised about is there isn't more companies who are actually customizing other people's software with like, do you remember there was that software called like Grease Monkey or something?

And you could like customize the JavaScript on other people's websites to make like Codelets and like, you know, reskin Gmail and like make applications in your own image. I think there's like something to be said for that now that writing that code takes zero effort. So it's like, say I'm a HubSpot user and I don't like certain features in it or the way it works in certain ways. You could now just like build an internal plugin for your company that completely mods that software to add features, do what you want. It could sync the storage of it back out through the plugin so you don't even need like database updates or anything. And completely change the stuff that you have without having to do like epic migrations or recreate the whole thing or maintain it or anything. Yeah, I think the thing about all of these products, right, being logic on top of a database. Like, companies have custom logic. Databases are cheap. So it's like, you know, it used to map sort of a generic sales workflow for everyone, but everyone's workflow is unique. So I think CRM is just, like, such low-hanging fruit that I don't understand how that recovers.

I get it at the higher end, like at enterprise, like Salesforce, with, like, user roles and permissions and all that kind of stuff. To me, that makes sense, that that's okay for now. but increasingly I'm unsure of like what these companies are going to do because instinctively a lot of them are like we'll build an agent and there'll be specialist agents like you'll come to the like actually I have a really funny example somewhere here do I have it up so Salesforce released this thing and it's we're going to show you something that you've never seen before the world of sales is about this is seriously a real product it's called Hyper Sales Agent Hyper for Sales It now joins your virtual sales meeting. So it joins your sales meeting. Hi, Katie. Oh, my God. And the customer's like, I hate you. If I was the customer, I'd hang up. Like, I don't. Hey, Katie, good to meet you. Like, wait, let's talk a bit more. These days they all have nudity films and so you can just, like, expose yourself immediately. Piper, want to take that one? Yes. Sales cloud is a core requirement.

I securely read and write CRF. Like, it, oh, my God, the cringe. The cringe. So it's like this awful, like. Imagine being someone like, oh, wow, techs are good now. Like, I really want to interact with this. The demo's like, I am a buyer. I have a question about the product. And it's like, now Paggit can help. And it's like. So will your product meet all of my needs? Because I have many needs. It's like a way to treat your customers like absolute. It's just such an insult. And like the stupidest product. Like, if this thing succeeds, I'm willing to like die on my sword over this. Like, it's just stupid. Like, there's no other way to explain it. And so, like, this is one angle, right, of these, like, very specialist agents like Piper. But, you know, like, I guess this is the question is, like, are specialist agents even still needed? Or is it going to be single agents to rule them all? Like, you just use Claude for everything? Or do you still see a place for specialist agents that sit on top of models Not really right It about the tools It about the access that it has to stuff like proprietary data

proprietary processes that do stuff that actually adds value, connecting those to, like, whatever model. It pretty much doesn't matter. Okay, now I'm going to – I've set you up. I've completely set you up here. So now I'm going to talk about Metamuse and, as I like to call it, the open claw assistant wars packaged up as a consumer app with, like, fluffy characters and cute names. So we now have Dot Muse Instinct Duo, sorry, Dot Muse Instinct Duo Gemini Tasks. That's seriously what they call it. It's like the Amazon chime of AI agents. I'm sure there's others that I've missed. And so here's probably what I... Oh, sorry, Grockbot. I miss Grockbot as well, which I just started the whole thing, really. Can you, like, stalk your ex-girlfriend and, like, do other equally unethical things? It helps you plan your harem of women and, like, where your money's going to fund all the babies.

um but grokbot uh grokbot was like i used it early on because it was available here in australia and it was really good but then they like crippled the limits um apparently they just switched it to opus 5.5 so it was just burning like hordes of money and so they had to calm down with the limits because their own model was just so bad at running the harness on it i mean money's not real anyway so who cares might as well it's a figure of our imagination and uh yeah so anyway and then we in Australia we had like I think Instinct was out here but I don't know if I don't know anyone that used it we had access to GrokBot but we recently got access to Muse now I think what's interesting about Muse is apart from you know like not trusting Zuck and like all my data being mined and probably trained on one thing I'll say is the price of free is good like it's free so anyway we both installed it without even talking when it first came out and we played around with it a ton.

I actually ordered, I was doing meals for the week because as I said earlier, my wife just had a baby, so I was trying to plan meals for the kids, which is an awful thing. It's the first time I've really ever done it. So I asked my muse, plan meals for kids that they'll eat, and then I told her to go to the grocery store, order all the stuff, ingredients. I didn't even check it because I'm a psycho. Let it check out. the box of food got delivered and it is truly unhinged. Like, it ordered like duplicates of things. It ordered those carrots. You know the carrots that aren't in a bag or anything? It's like they're still on the stem and they're really thin like a chef would make. Oh, like Bugs Bunny style carrots. Yeah, yeah, I know what you mean. It also ordered crab meat for some reason and I still have no understanding of why. I tell you what, crab meat's the kind of food you've really got to be in the right mood to eat crab meat like you know what I mean I probably like once a year when there's a full moon or whatever I feel like crab meat but other than that you're like that is really I actually have

a jar of it in my fridge right now it's been there for a month because of like at the time I bought this I felt like it was a good idea and now I really don't have the guts to eat that so anyway I've got all these like ridiculous ingredients and then I asked it like hey what were the recipes and the recipes don't fully match up with the ingredients but anyway that wasn't terrible I I mean, it got it done, and I've been doing some other things. Look, I think it's pretty good is what I'm going to say. I think in terms of the applications and the implementation, I think they've done a phenomenal job. The model, not so good. Like, it's pretty dumb. I mean, okay, let me tell you my experiences because I have only had epically positive experiences with me. It's like it's been amazing. So I got it, and I treated it with utter disdain because it's like, what do you want to call me because you're an Aussie bro, so you probably want to use an Aussie name for me. And it was like, you can call me Matilda or Muse or Mate. And I'm like, okay, mate, let's go with Mate. It's like, hi, I'm Mate. And it's like, yeah, not really how we use that word, but okay, whatever.

But it's actually funny now because I'm like, okay, mate, can you do the following? So my wife wanted this rare board game, right? It's from 1978 or something like that. And it's like, hey, do you mind like on your way home from work, picking up this board game? Yeah, not that easy. And no, that's going to take me a lot of time. So I spent 10 minutes trying to do it myself. couldn't do it i'm like all right muse let's see how good you are bro mate sorry mate and um gave it the task not only did it find it it gave me three options it warned me it's going to be expensive because of the postage and whatever it got me to link my credit card through like stripe or link or something um and then it's confirmed my postcode like got all the details and then finally it's like look it's ebay you're gonna have to give me your ebay login so i tried to log in it's like hey mate just go get the code from gmail paste it in here and i'll sort you out and it did it ordered it the whole thing worked so then i was like all right what else can i connect and now my wife same wife um yeah um uh constantly sends me stuff on messenger of like things i need

to do right which i ignore you're basically her muse yeah yeah so i'm like hey mate can you connect to my wife's messenger and harvest these messages for to-dos and just any of them that you can do can you just do for me so it's like a car service it's done like a whole bunch of stuff that like i actually said i would do that are now done and then i was like hey i feel like i need to send her messages throughout the day to say hey good morning how's your day going does she listen to this you're gonna she's never listened to their one it's very low risk and so um i'm like uh can you just like throughout the day send these like really nice thoughtful messages ideally like incorporating some context and stuff to keep it going anyway it's been doing it she's like oh that last message you sent was so sweet so lovely thank you m dash m dash it had m dash no yeah it really did and i was like this is so transparently ai and she knows me she knows

we do this like it's not like i'm the most suspicious person there could be when it comes to this stuff and yet it worked beautifully and i'm like i love muse i love my mate it's really good and it hits me up it reminds me like literally it asked me how dinner went and i was like yeah thanks for the shopping list it cost a fortune and it's like yeah harris farm will do that to you mate. I love it. I'm in love with it. So here's my thoughts on all of these assistance drive. So I have been driving a lot since we took on new roles. So I've been commuting a fair bit quite a long time in the car. And I felt... Just be careful not to confess to any crimes. Yeah. Oh, yeah. And I felt really unproductive being in the car because my Tesla has Grok in it and you can talk to it. And it is getting a lot better. I think it can soon operate GrokBot as well and go do background tasks, which is kind of cool.

But anyway, because I still use Sim Theory and I've got my Mac Mini behind me that still powers everything. And I feel like pretty safe and confident with the structure of it and everything. Like it's just nicely set up. And I have it connected to things that I wouldn't really want Meta to know about me. Like it's connected to like my bank accounts and all sorts of things. And so I wanted an app that I could still develop with and build prototypes, but also be able to visualize what I'm building. And so I sort of vibe coded weirdly enough. I built a scaffold of a voice app for Sim Theory where I could talk to it and it could go off and do background tasks. I guess not to do similar to how Muse would work, but it would create like an agent session, go off and do it. And then it was aware of like the context of all tasks going on, where they're up to. If it had a question for me, it would interrupt and ask me the question while I drove. It had the ability to ignore me as well. So it was pretty smart about when I was talking to it or screaming at traffic in disguise. That is a great picture. It was really good.

And I was really productive with it and still am. And like, that was kind of cool for me. Like I got a lot done and was able to check in on things that had access to like Slack and stuff like that. So I could be like, oh, you know, what's Chris running about now? Stuff like that. And I found it really empowering. Like I didn't feel like I was wasting the time anymore because with full self-driving, I did feel increasingly like I was wasting time. I was simply sitting there as a passenger in the car. And it felt to me for the first time like, oh, this is probably the future of talking to a computer, like this single agent that just goes off and does stuff for you. And eventually this will just be built into my phone. It might even run inference locally, like the data might be stored locally. And this is probably where things are going to go. But one thing about using Muse that I will say that I enjoy, and I'm increasingly becoming bullish on, is single-threaded AI interactions, where it's just like chatting to a mate. And you're just chatting away, and it goes off in the background and does stuff. And I foolishly designed the V1 of the app, where it would sort of be like this dumb routing

service, but it didn't keep this like chat thread and have great memory about what was going on. It always had to sort of look up context of tasks instead of just staying in a single flow and then bringing things back into the prompt as like memory when it needed to. And I'm increasingly convinced that the future of both like workplace agents that you're using to be productive, like write code, and also these like more personal agents are going to be like this. I think this is like the evolution of them and then, you know, more interface elements and stuff like that. But going back to our earlier point where we both sort of laughed off this idea of specialist agents and specialist harnesses, like we have, as always on the show, being an average podcast, completely contradicted ourselves because we're like, oh, it's all the model stupid, but you're like, I love Mews. So, and it's a shit model. Like, you know, it's not a great model. I think it's more, and I think I did say this, I didn't completely contradict myself in the sense that what I said was it's about the tools and the integrations and the way it is set up that it can actually accomplish things.

And if you look at Muse, what I found most amazing about it is that everything I needed in terms of to get through that transaction, for example, it had an easy workflow in there to make that possible such that I didn't just give up. Like if it was like, okay, I'm now giving you, like you now have to go off to eBay, log in, find the checkout and complete the checkout. I'm like, well, what's the point of this? But the fact that it's just like enter your username and password and then hit go and then it can complete it. To me, just that bridge, the fact that they've taken the time to work through those integrations or get their browser use up to the standard that it can do that reliably is what adds the value. It's not the model at all. I just don't think about the model in that scenario. I just assume it can do it. And it's the fact that they've got it that deeply integrated with the things that are actually needed to do the kinds of tasks me as a user would need, that's where the value comes from, in my opinion. But now the counter to this is if you had a Muse-type flawed app or SimDuty app,

like they have Duo, I think it's called Duo, maybe, Duet, something like that. So if you had that instead with more powerful models connected to your, like, workplace environment with all the connects that was, like, personal and professional agent, would you prefer that or do you like the separation of concerns? Don't give a shit. I just use whatever works. Like, yeah, I'm just going to go with what works. That's the thing. Okay, what about Churn? And, like, Muse, now you've started using it, and they realize they're going broke because they're hosting all these, like, workspace sort of virtual machines, as we know the cost of this. There is a huge cost to running this. I'm sure they're losing absolute **** of money. A lot of swearing today. Yeah, it's really bad. And so they're obviously losing a truck ton of money just trying to take market share. I think they've got 10 million active users now for Muse, and it's only in the U.S. and Australia right now, I think. uh so but if they try like are you willing to pay 31 us a month for muse i don't know yeah it's

interesting i probably would but i don't know if i'm like a typical person because i basically try everything so i think it's probably not a good barometer this is what i think like unless they can get the price of these as free and then somehow trap you in the ecosystem and sell you ads or influence your shopping or find a way to monetize i mean don't you think like at some point Think about what I've done with it. I've mostly just bought stuff with it. Like, you know, the majority of the stuff I've done is real spending and, like, decent amounts of money, right? Like, I'm actually trying to buy a piano at the moment as well. It's found me one. And then, actually, my wife did see that. She's like, don't you dare buy a piano with that thing. And so, like, that's a pretty significant transaction. Like, I imagine that people will be buying, like, racehorses and castles and all this stuff on you. Yeah, why not? I mean, what's the difference if it does all the research and stuff? Like, so, yeah. Like, thinking further out, is it winner takes all in this market or not?

Like, is Muse like the iOS and then the desktop app is, like, flawed? Or is it one operating system, one workspace, one model to rule them all? Because I think what you're saying is Muse seems to be good at what Facebook's kind of good at, which is spamming you with, like, ads and products you don't need and then buying them and then, you know, all the things that it's kind of good at. And then flawed really is good at, like, sort of, like, frontier, like, everything else. I don't know. I think people are going to go with what works for them. I think they're going to go with the things that save them time. It's low effort, like, it's low mental energy and also the ones that can actually complete the process. I think the reason I stuck with Muse and it worked for me is because it got me to the end. I didn't get like, well, it's actually pretty cool. It can run a browser and stuff. But I got stuck on some step that's inconvenient for me when I'm using it on my phone. But the fact that he'd get me all the way through to the end was really significant.

And that's why I went back and got it to do other stuff. Okay. What about this? You're a sales rep at an enterprise sales company. And, you know, increasingly sales are going down. Which is AI. and you're thinking, okay, Salesforce has released this new Piper app, and it's sort of like Muse, but just dedicated to doing your work, like taking notes in meetings, all the stuff you need to do in your day-to-day as a sales rep. It's purpose-built, like a harness for that as a specialist agent. Are you going to use it, or are you going to connect Salesforce to Claude? Yeah, I'd probably rather the connection into Claude or something like that. Yeah, because I think Claude would have interoperability between departments and teams in the company, and you just can't imagine a world where it... Like, in a B2B workplace sense, I can't imagine a world where it wouldn't be like a winner-takes-all type environment. I think in the consumer space, that's where there's probably a war to be won,

which is why we're seeing, like, Dot, Muse, Instinct, Duo, I don't know, probably Duet, and GrokBot all sort of competing it out. The realisation I have is unless they can drive down costs to these things, consumers just aren't willing to pay. So the question is, like, how much money are they willing to lose in order to get market share and then somehow figure how to monetise this? And I would argue the – like, I reckon the approach is right, but the monetisation is going to take way too long and investors will get fed up is my concern. And then maybe something emerges from that category. yeah i don't know i hope you're right i mean at the moment i'm just happy to consume this stuff and enjoy it enjoy the vc subsidies yeah it doesn't feel like uber in the first like couple of months um or like food delivery in san francisco in the first couple of months okay moving on uh daily drivers um i i wanted to play this song from the music video um what will my daily driver be um but i can't put this on my computer uh so let's talk about evolution past

couple of months what have you been using i mean i must admit i i love fable 5.2 i've been using it a bit it's just like it's it's i know it's expensive but it's no nonsense and it never it's really really good at going all the way with the task i really like it prior to that and that's probably only been the last week or so because I think that all it been out um I was opus 5 all the way I found just Opus 5 was the perfect balance perfect mix got everything done that I needed worked over a very long like two three task It could just maintain things and get things done. Great at doing the more, like, coding work and, like, software. It's really just building processes and running with them. Fable's even better than that. I find it excellent. and yeah, I pretty much don't stray from the big boys. Yeah, I went to GPT-6 Astra, as I said, for quite a while. Lame.

But it was, I think it was like, was there Opus? I forget all the Opus variants now. Let me actually look them up. Oh, we must have discontinued a bunch of them. But yeah, there was like all they did. There have been more, I get all the emails, and they've been more aggressively deprecating old models. Like remember in the past they'd leave, they're like, oh, GPT 4.1 has been deprecated. Like, you know, it took a long time to remove some of those earlier ones. Like Gemini 3.5 was around forever. And then only recently all the providers seem to be way more aggressively deprecating models. Like one that had been out for a few months, they're like, yeah, that's gone. You've got to switch to this or we're going to redirect the traffic to this, whether you request, like, you know, when you send the old model ID, we're going to redirect to this one, stuff like that. I used GBT6 Astra, not 6.2, for quite a long time when Opus 5 was around, and I found just the speed of it way better,

and I liked how no-nonsense it was at the start. I know a lot of people complain that it did, like, unhinge stuff, but I found it really good. The only negative is if you were writing code with it, it was just obsessed with writing tests. So you'd do like one feature, like, I've created 700 test views. So that was like a negative. Yeah, I actually have started to give my agent's instructions, don't run tests, don't make tests, until we actually lock in the functionality we're trying to build. Because otherwise what happens is you get all this delay because it's like, oh, I've run the 1,017 tests I've added. You know, it's taken 20 minutes, but they all pass. And it's like, yeah, but I don't even know exactly the solution here yet. Like, it really has a proclivity to do that. Yeah, and then when Opus 5.5 came out, I switched to it a little bit and then a lot. Like, it was just so much better. I think it was like distilled Fable or something. That's what it felt like. And it was faster. So, like, to me, that became my daily driver.

It pretty much still is my daily driver. I haven't, like, switched over to Fable yet, mainly because I'm a bit cheaper than you, I guess, in terms of cost. But yeah, I am curious to spend a bit more time with Fable 5.2. But as I said earlier, there are still some specialist models in my life. Like, for example, for that video app I made, I mentally knew, hey, I'm going to go try the Gemini variants. And I tried a bunch of them. And then I tried a bunch of like OpenAI models and Anthropic models to see what would be cheapest and best at editing. And like, that's where I think Gemini, because it's being trained on like YouTube and all that kind of stuff. I think for multimedia stuff, it still does a far better job. It's a terrible agent. I think they need to start again. Yeah, but there's still something to be said for single-shot model requests. Not everything is an agentic loop. There's some things where you literally just want to go, here's all the crap I have. I want this kind of output. Please do it. And then you really are optimizing for cost, speed, and quality.

Like, you know, you're going to have a matrix of those things you need from that. And then once you've got it, you really kind of don't want to change it. What's your prediction with like the OpenAI, Anthropic, maybe even Google at this point? Like, do you think by the end of the year, OpenAI is just going to come out with some leapfrog model that's incredible? Or do you think increasingly Anthropic is just going to run away with the lead? I think Google's due for one. I kind of predict by the end of the year. I mean, there's not that long left in the year, I suppose. But I think Google has probably just launched. Like, we seriously don't know. One day ago, they released Gemini 4 Neon. Is that a good one? They released a bunch of rando ones. And, like, people in the Sim Theory chat were like, can you please add Google, you know, garbage bin? And I'm like, oh, okay, sure, I'll add Google garbage bin at, like, whatever cheap price. You know, it's kind of, there's no energy behind these releases. So it's like the lab's just made a new tune. So it's Gemini 4 Neon.

This is how behind we are these days. You've come into the right show for the gloss on the models. Gemini 4 Neon, you know, it benchmarks higher on nearly everything than Opus 5.5 and 6.2 Sol. So this new name thing, honestly, stop. Like, if you're listening and you're from a lab, which is probably no one, but please stop, because it's driving me insane. Like, Gemini 4 Neon, like, what does that mean? It shouldn't be like... You've got to call it something. No, but like, Gemini 4, the best one, or Gemini 4, the mid. Like, Gemini is kind of shit, but we have to realize something. Yeah, the mid to the low. Like, maybe just low, medium, high for every model. Like, Gemini 4 low, Gemini 4 medium, Gemini 4 high. That'd do. At least with Opus, it's pretty easy to understand. Like, Opus 5.5, Opus 5.6. Like, these are, you know, these are things that would help. My phone is ringing and I can't stop it. Okay. So, yeah, I think that, I don't know.

I think this is their best. I think this is their model. Maybe we should try it and then next week we can talk about if it's any good. Gemini 4 Neon is supposedly their best. But I don't hear any chatter on X about it. I think the new models out of OpenAI that they just released, like, It was kind of weird right before Androbics IPO, all these people that got early access just poo-pooed their latest models, which was 6.2 Astra and 6.2 Soul. And even Sam Altman on X was kind of like, yeah, they're not that great. He's not even hiving them up anymore. And a lot of people were saying... You've got to release models, man. Yeah, people were saying it was on Hinge and stuff like that. I kind of, I wonder if they still have something like up their sleeve yet that is going to be, you know, is going to be better before the end of the year. But right now, it just, it looks like a sort of a clawed 3.5 moment where they just, they're kind of far ahead and like it's very hard to catch up.

But increasingly, they're going to need to stay ahead because they want the stock price to keep pumping, I'm sure, as their stock unlocks. So there's a ton of pressure there. Now, we were going to talk about something very important, which is our sort of lull of the week, but also one of the funniest things that I think you've ever done. So I'm going to bring up the website now, and people are going to be like, why is on the screen right now, Huntley's Cove.info, resort-style living in the heart of Sydney, and the big event on this website is booking tennis courts and carols in the park. Now, could you explain why I'm showing Huntley's Gove resort-style living in the heart of Sydney? So I live in a resort-style living in the heart of Sydney. Socks to give yourself. It's basically full of people who, I'm not going to discriminate against anyone, but these are the kind of people who care about rules, and they care about very specific things like the bins. The bins is the main topic of conversation in this area,

like when to put them out, what's allowed in which bins. and I've been in a few very serious showdowns around the pins and my blatant disregard for the rules. Anyway, the sort of community people at this website accidentally let the domain expire. So guess who owns the domain for the next 10 years? Me. And so my plan is basically start with a very like community service in that I have built new tennis court booking system that's actually really good. It has reminders. You can cancel. It's, like, really nice. And people, sorry, people just accept that the website's been updated for no obvious reason and now has a better tennis booking system. Yeah, so I cloned the previous website from archive.org, or my agent did. And I think the point of this discussion is the fact that in the past, if I wanted to troll my community to this level, it would have taken a lot of effort. Now it's literally just asking. And so what they don't understand is that this website is going to get dark.

Like, it's going to take a time. and my agent is instructed that we at first we build trust we build uh goodwill and we go there that is gradually there's going to be a lot of new rules in this community very soon there's also going to be aggressive campaign because there's a mayor re-election coming up so there's going to be a very one-sided partisan this is like a political play as well maybe i don't even care but i care that they're going to be so upset that their local community is partisan and stuff like that so it's gonna it's gonna be interesting and anyway just to confirm no one has any like there's no one knows that they might now because it's on the podcast but um but yeah it's uh it's pretty interesting and uh the fact that i think that the the wider point is because we talk about the models hacking and all this sort of stuff it's so easy now to have agents out there scanning for stuff like oh when a key domain's expiring what are security holes in this website like what like you It's kind of like the opportunities to do this kind of thing

are just so pronounced now because it's so low effort. It's just unbelievable. I think it would be good to keep listeners posted on the progression of Hunley's Cove, resort-style living in the heart of Sydney. Don't tell them, but give me suggestions of what we can do. Leave suggestions in the comments. You can't tell them anyway because I also control the email address now, so I get a lot of inquiries. I'm actually, it's using up my time. I have to answer all the inquiries to help people out and stuff. So it's like, yeah, so far you can't fault me. I've actually been a net gain for the community, but I just don't intend on keeping it that way. Yeah, okay. So do you have any ideas? You have no ideas of how you're going to make it a little dark? Well, partly I think one of the ideas to make it really sinister is to start, say, you know, we're going to be installing cameras, Big Brother will be watching, you'll get a social credit score, in the community about like whether you like you know like keep the pool in good condition and whatever wait so if you have it so you can look up your um community score on here like introduce

community score make it really positive and then they put in there like unit or whatever i'll start them on like negative 20 yeah and like have it so you look up the unit number and it has a score and it explains like what the score is and like why they've been marked down or marked up Yeah, exactly, like actual feedback. I reckon that is... And then you should have it so you can report a person, like report a unit and see how many people actually report each other. Oh, yeah, because I have all their emails and stuff, so email them like you've received their demerit. You've received their demerit and you have 10 demerits left or something. All right, I think that's the play. And then we've got to check back in on this in the following week. All right, sounds good. Well, I'll give an update next week, and we'll see what mischief I can create. I don't know how this really relates to AI, but it's kind of funny, so we'll just go with it. AI made it possible. There is no way I would go on to this effort manually. All right, any final thoughts?

NVIDIA drama and Fallout, Anthropics, Big IPO Day, single agent versus special agents versus Muse versus everyone exploiting OpenClaw and turning it into an ad. My final thought is like the most recent delight I've gotten from AI was that Jensen song and video. I'm like, what a world we live in that you can distill, you know, a very serious news item down into something that just makes me so happy. Like I love that. I love that song. I love that video. It just genuinely makes me delighted. I want to do a lot more of that. We'll put the full video clip up on. there's a sub channel more this day in AI. So we'll put the full, we'll upload the full one thing video. It'll also be coming to Spotify, although it's a little bit delayed and like Apple music and all the other ones. Do you think we need to point out like we're using his last name Wong to sound like the word wrong? Yeah. Yeah. I'm sure people get that. I think they get that.

I don't know. It's an average audience. Yeah. Okay. So we'll put that up on more this day in AI and we'll also get it out to like Spotify and Apple if you want to listen to it. I'm sure people do. I mean, Adel Patricia is seriously, still has like thousands of monthly listens. And we haven't, oh, actually, it has gone down a little bit. No, it's at 2,492 listens this month. Like, that's a lot for an AI song. Yeah, that's actually, yeah. And then Love Rat has had 1,115. So, like, Love Rat is still up there. God, Jeff, if you're out there, we miss you. You haven't made news lately. Can't even troll the guy anymore. Not even relevant. It's funny how all these AI guys really seem to struggle with women. Hey, like Love Rat, a Huang thing. Yeah, because I guess they got popular by an unconventional channel. Like, they didn't get popular from being, like, good at football or something. Or good looks.

Yeah, yeah, exactly. So they're not the conventional love rats. Like, they're sort of newer. Yeah, okay. So that'll do us this week. We'll hopefully be back next week. I'm not sure. We're going to try. It's a little bit harder these days. We may or may not. Like, that's the whole point of an average podcast. Yeah, and eventually we might make 100 episodes if we play our cards right. What are you actually going to number this one? Oh, it's not going to be 100. There's no way. I really want to upset all the autistic audience members, including me. It's fine. So, yeah, we'll see you next week. We will play the song out at the end if you want to listen to it. Yeah, I'm not going to play it all, sir. I couldn't be, like, nodding my head and pretending that I'm listening to it. So I'm just going to add that in post-production. All right. We'll see you next week. Thanks so much for listening. We missed you. Goodbye. Yeah. We'll be right back.

I parallel process and ain't just for the bus. Two friends running at once. I'm a multi-core man. Shit, those these are export controls. Now I'm shaking my hand. Stock went down 16, 850 billion. Gone. Wall Street screaming, bubble. While I'm putting my check it on. I did the one thing. One thing. Got the market cap swinging. Got the whole street string. I did the one thing. One thing. Well, in the server, bad decisions in the evening. Work hard and laid it. Work hard and lady Chips in the rack and my phone on silent I did the one thing Yeah, I did the one thing Five-level romance Intelligence on alert She was asking about the ship's sex Said it's just for dessert Then I dated my direct report Orchard in the night HR fired a ticket like, sir, you cannot Collect crap, step in the interim CEO Harvey Jones sharing boards, now I'm sitting on the low Sam Altman texting bro Are the GPUs still coming? Zuck said, need a new hobby?

Try some case fight running. AI bubble? Nah, nah, nah, it's a feature. Three and parameter heart, but it only had one bug. Overclocked my love life till it tripped the power plug. Every token predicted, every model trained right, but I couldn't predict the headline. Thursday night, I did the one thing, one thing, one thing. Got the market cap springing, got the whole street spin, I did the one thing, one thing. Well, in the server, bad decisions in the evening. Work hard and latest. Work, work, work hard and latest. Chips in the rack in my own, own silence I did the one thing Yeah, I did the one thing Sorry to my wife Sorry to my shareholders Keep the jacket, I'll be at Denny's Them ladies

番組の概要欄(原文)

<p>So Chris, this week we unpack the Nvidia drama and Jensen Huang getting booted from his own company (and premiere our exclusive music video "Huang Thing"), celebrate Anthropic's IPO and its $1.27 trillion market cap, argue about whether the model is the whole product or specialist agents still matter, cringe at Salesforce's Piper sales agent, review Meta Muse (Chris's "Mate" is now texting his wife), debate whether anyone will pay $31 a month for a personal assistant, run through our daily drivers and Gemini 4 Neon, and reveal Chris's hostile takeover of his local community website. Billies in the bank.</p><p>CHAPTERS:</p><p>0:00 - Anthropic IPO Day &amp; Jensen Huang Booted From Nvidia<br>3:39 - 🎵 SONG: "Huang Thing"<br>5:54 - Reacting to the Huang Thing Video<br>8:55 - Anthropic Goes Public: From Industry Joke to $1.27 Trillion<br>13:32 - AI Workspaces as the New OS &amp; "The Model Is the Product"<br>20:31 - Word of the Year: Discovery &amp; Thinking Big With AI<br>25:44 - No Moats: SaaS Under Pressure &amp; HubSpot "Refounded"<br>33:01 - Salesforce's Piper Sales Agent: Peak Cringe<br>34:58 - Meta Muse &amp; the Assistant Wars<br>38:30 - Chris's Muse "Mate" Is Texting His Wife<br>41:35 - Mike's In-Car Voice Agent &amp; Single-Threaded AI<br>46:19 - Would You Pay $31/Month? Is It Winner Takes All?<br>51:21 - Daily Drivers: Fable 5.2, Opus 5.5, GPT-6 Astra &amp; Gemini 4 Neon<br>58:57 - LOL of the Week: Chris Hijacks HuntleysCove.info<br>1:07:11 - BONUS TRACK: 🎵 "Huang Thing" (Full Song)</p>

出典・この記事について

番組
This Day in AI Podcast(Michael Sharkey, Chris Sharkey)
元の回
The Huang Thing, Anthropic IPO and Meta's Muse in OZ - EP99.40
配信日
要約作成日
公式サイト
podcast.thisdayinai.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Spotify で聴くApple Podcasts で聴く

関連エピソード