OpenAIの新モデルAstraは評判が散々、でもGemini 4 Neonは意外と良い?

OpenAI's New Astra Is Messy But Gemini Neon Is... Good?

AI For Humans: Weekly AI News, Tools & Trends1時間3分
OpenAIの新モデルAstraは評判が散々、でもGemini 4 Neonは意外と良い?

OpenAIの新モデルAstraは評判が散々、でもGemini 4 Neonは意外と良い?

AI For Humans: Weekly AI News, Tools & Trends

0:001:03:36

要約

AI For Humansの司会Gavin PurcellとKevin Pereiraが、OpenAIのGPT-6.2 AstraとSolを実際に試した感想を語った回。日常の開発や創作ではつまずく一方、コンピューター操作やツール呼び出し、低コストでは強みもあったという。GoogleのGemini 4 Neonは創造性と文章力で好印象とされ、AnthropicのIPO、AI安全性をめぐる議論、GTA 6とAIによるゲーム開発の将来像も取り上げた。

この回のポイント

  • ●GPT-6.2 Astra/Solは早期テスターに酷評されたが、OpenAI側は公開後数日で大幅な変更を加えたと説明。2人が使った版は当初の評価時と異なる可能性がある。
  • ●Kevinはコードベースの把握や提案、Gavinはコンピューター操作での監査やツール呼び出しによる画像生成を評価した。一方で創造的判断、文体の自然さ、反復的な修正には不満を示した。
  • ●Blenderでの再現テストでは、Astraが約3ドル、Fable 5.2が約19ドル。Fableの方が精緻だがAstraは質感と照明が良く、コスト面の利点が指摘された。
  • ●Gemini 4 Neonは創作と文章が強く、Gavinのパックマン駄洒落テストで初めて10個全部が納得できる出来だった。コードだけで作ったトレーラーも評価され、有料購入を検討している。
  • ●2人が最も勧めるのはAnthropicの製品群(Opus 5.5、Fable 5.2)。他社サービスをトークン購入などで併用する構成も提案した。
  • ●AIが政府を転覆しうる、人間の雇用が消えるといった超知能をめぐる懸念が話題となったが、2人は雇用が完全になくなるとは思わないとの立場で、複雑な心境も明かした。

章立て

  1. 今週の主要ニュース紹介

    OpenAIの新モデル、Gemini復調、AnthropicのIPOなど、今週の話題を概観する。

  2. GPT-6.2 Astra/Solの酷評と実体験

    早期テスターの否定的な評価と、Kevinによるコーディング作業での失望、OpenAIによる公開後の変更を語る。

  3. 64匹トーナメントと得意・不得意

    Gavinのベンチマーク題材を試し、見た目の改善が難しいことや、コンピューター操作とツール呼び出しの強み、3ドル対19ドルの比較を紹介。

  4. AnthropicのIPOと雇用への懸念

    IPOの意味、Anthropicの超知能観、「人間は失業する」という発言への2人の反応を議論する。

  5. Gemini 4 Neonは本当に良いのか

    パックマン駄洒落テストやコードのみのトレーラー制作を通じ、Neonの創造性と独自の思考を評価する。

  6. どのAIに課金すべきか

    Anthropic製品を軸に、他サービスを併用する使い方を提案する。YouTube登録者5万人達成にも触れる。

  7. Jensen Huang復帰騒動と超知能への不安

    NvidiaのJensenの件を足がかりに、オープンソースと閉鎖路線、Paul Christianoの警告、大規模障害への不安を語る。

  8. GTA 6とAI時代のゲーム制作

    人間製作の最後の大作かという問いから、AI利用の線引き、Minecraftで再現された地球、将来のゲームの可能性を議論。

  9. クイックツールと創作実験

    fal、ChatGPT Wallet、Kevin RoseのオープンソースDocuSign、Fable 5.2の作品、Gavinの「センスを持つエージェント」実験を紹介する。

解説記事

AI系ポッドキャスト「AI For Humans」のGavin PurcellとKevin Pereiraが、今週のAIニュースを振り返った回だ。中心はOpenAIの新モデルGPT-6.2 AstraとSolの使用感、Googleの Gemini 4 Neonの好評、そしてAnthropicの巨額IPOである。

酷評されたGPT-6.2 AstraとSol

2人によると、公開前に試した影響力のある開発者や評論家らは、モデルの文章力や推論の説明に否定的だった。一方でOpenAI側は、公開後数日で両モデルに大幅な変更を加えたと説明しており、2人が使った版は初期の評価時とは異なる可能性があるという。

KevinはCodexで、文脈のない案件の把握や提案には感心した。しかし最高設定で長時間実行させると、議論していない変更や出来の悪いUIが出てきた。結局3回変更を元に戻し、Fable 5.2に戻ったと述べている。Gavinも、64匹の危険動物トーナメントのゲームで、カードの仕組みは良好だったものの、見た目を改善する反復作業がうまくいかなかったと語った。

得意分野とコスト

一方で強みも挙げられた。コンピューター操作では、サイトの監査レポート作成を自走してこなしたとKevinは評価した。Gavinは、思考の深さを生かして画像生成ツールを呼び出す使い方で良い結果を得たという。さらにBlenderで参照画像を再現するテストでは、Astraが約3ドル、Fable 5.2が約19ドルだった。Fableの方が精緻だが、速さと安さを重視するならAstraを選ぶ価値があると2人は見ている。Gavinは、OpenAIが今のAI市場のYahooのような存在になりかねないという懸念まで口にした。

Gemini 4 Neonと何に課金するか

Gemini 4 Neonは創造性と文章に強みがあると評価された。Gavinが長年続けるパックマン駄洒落テストで、初めて10個すべてが納得できる出来になったという。Fable 5.2では5〜6個程度だった。コードだけで作った自社トレーラーも、これまでのGeminiより良い結果だったと述べている。Kevinは、他社モデルの蒸留でない独自の発想がある点に注目し、複数モデルで相互にチェックすれば成果物の質が上がると語った。

それでも2人が最も勧めるのはAnthropicの製品群だ。ターミナル、デスクトップ、モバイルまで揃い、Opus 5.5は日常使いに向くとする。そのうえで、他社サービスは必要に応じてエージェント経由でトークンを購入して使う構成を提案している。

IPO、超知能、雇用をめぐる不安

AnthropicのIPOは約1.5兆ドル規模と紹介された。同社の技術者が「ほぼすべての人間が永久に失業する」と発言した動画も取り上げられ、Gavinは同社が超知能の到来を強く信じていることに戸惑いを示した。Kevinは雇用が完全になくなるとは考えにくいとの立場だ。

さらにPaul Christianoが「AIが12か月で政府を転覆しうる」と語ったインタビューや、最近の大規模障害が議論された。2人は、障害がAIによるものか人為ミスか分からないという不安にも触れ、Gavin自身も考えが揺れていると認めた。

ゲームとAI、小さなツール

GTA 6は人間が作った最後の大作かもしれないとGavinは述べた。Kevinは、人間が創造的判断を担い、エージェントが作業を拡張する形に期待を示し、「創作者の代わりではなく創作者のために」と強調した。Minecraftで地球全体を再現した事例も紹介された。

ツールでは、falの画像から3Dキャラクターを作るIdol Max、トークン化された決済情報でエージェントが購入するChatGPT Wallet、Kevin Roseのオープンソース署名サービスなどが挙がった。Gavinは、自分の好みを持つ複数のAI人格を「アートコミューン」として動かす実験も語った。

まとめ

今回の議論は、ベンチマークよりも日常の作業や創作で使えるかが評価の分かれ目になっていることを示す。モデルごとに得意不得意とコストが異なるため、主力を決めつつ複数を使い分ける姿勢が現実的だという。ただし評価は2人の個人的な体験に基づく。OpenAIが公開後に変更を加えている点も踏まえ、自分の用途で試すのが良さそうだ。

文字起こし(英語・自動生成)

Open AI finally swung back this week with GPT 6.2 Astra and Sol. We will discuss, but the vibes are very messy. However, Gemini might be making a comeback, Kevin. Oh, what's next, Kevin? You going to tell me Bard is let out of the bamboo cage that keeps him in the basement? Please, free Bard, free Bard. Meanwhile, Kevin, Anthropic has just filed for the largest IPO of all time, roughly a trillion and a half dollars. You know what? Honestly, this is great news. Congrats to Anthropic. For the rest of us, though, with the plebs looking at the ivory tower, one of the Anthropic engineers has a tiny prediction. Pretty much all humans on Earth will become permanently unemployable and will need to strictly rely on government welfare to sustain themselves. And human employment will go away. So that's cool. That's cool. So, I mean, do we buy stock with our final paychecks?

Paychecks, Kevin. We don't need paychecks in the future. Yeah, that's right. Please, more tokens in my bowl. I need to burn more tokens for actual warmth. Plus, NVIDIA's Jensen Huang. Remember NVIDIA? Remember Jensen? Jensen's trying to make a comeback. And people have big feels about that. Meanwhile, all of the AIs are hacking everything. And maybe, just maybe, Grand Theft Auto 6 is the last human video game ever made, Gavin. Well, luckily for us, Kevin, somebody rebuilt the entire planet in Minecraft with AI. This is AI for Humans. Without our paycheck. Who are unemployed. Welcome, everybody, to AI for Humans, your weekly guide to the wonderful world of AI. My name is Gavin Purcell. That's Kevin Prair. And Kevin, what a week. What a big week. There are three major stories, I would say, this week. Probably, like, six major stories. There's the new models from OpenAI, which are juicy.

There's some mixed results. There's the Anthropic IPO with Fable 5.2. And then, Kevin. Also juicy. We have a Gemini comeback, maybe, which is shocking for you and I who have been tracking Gemini for multiple years, all the way back to Bard. I am very excited to get into the news this week, but let us start with the new models from OpenAI. Have you used GPT 6.2, Astro, or Sol? Boy, have I, Gavin. Boy, have I used them both because I ran to them faster than I usually would, mostly because I saw everybody online trashing them. I know, it was crazy. They were trashed. Like, I haven't seen models get trashed maybe ever, especially from opening eye. It was full curb stomping from all the influencers who had early access, some refusing to even write about their experiences because they were so negative. And I got tickled by that. There's a lot of probably very hardworking people and poor little agents working as hard as they can

to make these models happen. So when I say I get tickled by that, I don't want any, I want all things to be great, obviously. But I got, I just got tickled by the notion of a major release feeling like a step back because hot take, Gavin, it seems like things aren't slowing down. Okay? Yeah, well, the funny thing to me is like, we were waiting for this for a while too, right? It's not like, it's not like this was like something that like, oh, they just released something so we're ready to go again. Remember, 6.1 Astra got pushed and we weren't sure why. The rumors were that maybe it wasn't safety aligned. And so I was expecting them to come out of the gate, especially because we're going to talk about this a bit, Anthropics IPO, where Anthropics was going public, you would expect them to have their, what is this, fastball, maybe a quick side punch, maybe some sort of a jab, something ready to go. I don't know what movement this is. I've got to believe it's closer to this. Please never do MMA, Gavin. Never sign up for a single match. Don't fight unitary robots either. But, yes, you would think that they want to. Well, here's the thing.

Maybe they would want to release their best in brightest models, but Alton Foil had it. I think at the meeting on Skull Island that they all take their segues to. Wait, now is this open-minded Skull Island, or is this like a Skull Island? It's a shared island with all of the best. Jensen's still there. Well, they hung his jacket up. He's no longer allowed. Everybody's on Skull Island, and they said, Sammy Altman goes, listen, Dario, I'm going to do you a solid. On the eve of your IPO, the week of it, I'm going to release the dumbest models ever. it's going to make Anthropic look amazing it's going to make it look like you've solved something that I clearly couldn't right in between Sora and now I've done nothing so just give me half a percent half a percent and then we're good and they shook hands on Skull Island right I understand you think this was a trade because of the okay so let's before we go too far let's talk about this because so the general consensus on these models is it does feel like a step back and here we'll kind of explain why We'll talk about you and I both have had experience with this, but just to be clear, like there's some pretty big people that early tested this, including Dan Shipper from Every,

Theo from T3 Chat, and even Matt Schumer, him of the viral something is happening post. All three of those people came out of their early testing and said this model is not great. And specifically what they all referred to was this idea that it couldn't write or talk very well or explain its reasoning, which I thought was really interesting. And my experience with that personally was I asked a couple things that were just very like, you know, hey, tell me this or explain this. And it does feel like its writing is funky, but it's more than that too, Kevin, because I think there feels like there's been this ongoing conversation about AI communication and whether or not like, you know, model ease or a neural ease, it's called, which is like the language that the models communicate in. And I think there's some level of where it feels like Anthropic might have solved a little bit better this sense of like how to talk to the model. And OpenAI might be struggling with that a little bit. But I think we should also talk about just results, right? So tell me what your experience was, and then I'll kind of dive into a couple of the dumb things I spent time with as well.

So I raced to use it. By the way, sweet Uncle Tebow, who is quite the head honcho at OpenAI, working on Codex and ChatGPT, he's the one who resets your usage. If you've ever woken up to a reset, that's Uncle Tebow. Tebow came out to say basically that they made sweeping changes to both models, Astra and Soul, within days of it actually being released. And so some of the folks that had earlier negative impressions, some of them cried foul saying, I don't know, I don't think so. And also, thanks for the heads up because we formed our impressions on something that just changed. So I just want to clarify that OpenAI says the models that we are talking about right now on the show, the ones that you and I used, are not the same ones that they were using. That we're testing two weeks ago or three weeks ago. Yeah, exactly. So I ran to my Codex client. I fired it up, and I immediately, on a project that I haven't discussed here because it's one that I'm consulting on, I fired it up. I put a medium-thinking Astra on it. I was like, go familiarize yourself with this product, a product that had not been in ChatGPT at all.

so there's no context i was astonished by how quickly like very quickly it it seemed to grok what the project was where we were at what the latest commits and issues were yeah just i asked it for suggestions on where it could go because it kind of had a fork with the project i really liked what it was suggesting i was like this is this is okay this is something we got something i put it on like the hyper turbo mode uh six ultra fast ultra fast as high up as i could and i was like let's burn these tokens baby because with the holiday coming up i don't think usage is going to be an issue with me i'm not going to hit my cap i let that thing scream gavin i just let it cook i could hear the steam from the other room and i came back and i was like what the fuck is this mess what did you do this is terrible it did several changes that weren't even remotely discussed. One of the changes that it did was so terribly implemented, I don't know what it was actually trying to do. And then part of the task was something that was taste related on purpose.

I wanted it to take a creative swing and it came back with the most placeholder gross UX UI. And I was a little crestfallen because I got all excited for a juicy model. I wanted to push back against some of the sayers of May. And I ended up trying two more times to prompt my way around it. I went back to Astra to really talk about it. I ended up having to redo, I had to unwind the changes that it made three different times and ended up going back to Fable 5.2. Yeah, which is so, that's very similar to what my experiences are. And, you know, there's a couple of people who have said, like, Noam Brown is a researcher at OpenAI who has specifically mentioned that this is the model, the model we're all using now, that solved the, you know, Navier, Stokes, or all the math stuff that they did, right? So there's people out there who are saying that, like, hey, bro, you're just not prompting it right, or you're not understanding what it is. But my argument is, hey, I understand that there's a model that you need to use for all the smartest people in the world, but if you're not solving the problems of the people who

are using it day to day, and by the way, that's not just us who are doing small things. Like Theo, the guy who runs T3, that guy is spending a crap load of tokens, and that is also, you think about Fortune 500 companies, how they might want to use the model. Can you imagine being like a manager who has told all your engineers or even non-engineers to go use this model? And every time you're like, okay, we've got to use the most recent model because we've got to push this edge. And if you don't understand like, oh, this isn't as good as that, your results are just going to be worse. And that is a bummer of a time in AI. That's right. Which to me makes me wonder like why would you release this model at this time? And what was the purpose behind it? Or are you just running so fast that maybe there's a world where at that company, at OpenAI, there's just this pressure to kind of push stuff out. And even if it doesn't perfect in the first place, you have to get the model out or else you're going to look at being behind. I don't know. I've started thinking, Kevin, I've never thought this up until this day, but I've really started to think about maybe OpenAI is the Yahoo of this time,

which is a sad thing to say. But it's a little bit of like anthropic, which now has real money, behind it, like the public IPO that's not out as of this time, I don't think right now, but it's going to do very well. Maybe OpenAI was the one that kind of made this market happen, and maybe OpenAI might be in trouble, which is not something I thought I would have said. But when you look, we're going to talk about Gemini's new model, and you look at Muse, like there's a lot of things happening right now that I thought OpenAI was going to be winning on, and they're not. And that's just a weird feeling for me to say. Yeah. I mean, I certainly wouldn't count them out by any stretch. Of course not. I don't know what they're cooking. I wonder if, you know, they just, you know, maybe this is a bench maxing issue, right? Where it's like, well, clearly the model is smart and can do a lot of things. They saw that the number was a little bit higher. They think, go ahead and release it. We'll at least be back on the charts. But again, I mean, to your point, maybe this thing is going to solve a quantum computing issue that you and I are not going to have.

But if it can't help me build a prettier website or do any of the dumb things relative to the mathematical geniuses out there, if it can't help me do those dumb things, well, then there goes my $200. Like, that's the point I'm on. And that's the question. Back to Anthropic. Anthropic has done a good job of making it for people who actually use it and then also doing the kind of high stuff. Okay, speaking of dumb things, we should talk about a few things we did. So, Kevin, I haven't done this for a while, but I did my 64 animal tournament. Love it. So, if you remember, if you've been listening to the show for a while. Yeah, one of my benchmarks has been, hey, very simple prompt. Go make a March Madness-style 64-animal tournament of the most dangerous animals of the world. And this time I gave it to Aster 6.2. I will say this time, one thing that it did pretty well, and I don't know if you had this experience, I played it, and it actually got the card mechanic down pretty well. Like, it's actually pretty playable. It copies very much something like Slay the Spire or even a Pokemon. But, Kevin, the thing I kept getting frustrated with is, What do you think of the actual visuals of the animals?

What are your feelings when you see them? Well, I was like, did it make it programmatically? Because it clearly didn't use its... Yes, I told to do it all in code. It's all made to do in code. So it's not using ImageGen, because I wanted specifically to see how it did it in code. Yeah, then I guess, I mean, yeah, it did it. But it's not setting the world on fire. It looked like programmatic art, which, you know, we've seen things come out of Fable 5.2. and even opus that look stunning and cinematic. And that's not- That's what I was, yes. So what I was getting at is I went through about five iterations. So I will say for Cup, first of all, the system basic systems on it, it did a pretty good job in the beginning on, right? It mostly got it. But one of the things I found with this model is kind of tricky is as you start to change things, it like zeroes in on one particular thing and then doesn't understand how to change that thing. Like I kept trying to get it to make these animals a little more artistic or a little more interesting, and it kept zeroing in on kind of this derpy look that it created.

And you can all play this. I'll drop the link in the show notes. And it just couldn't figure out how to fix it. And yet a model like Fable 2 or even Opus 5.5 does understand that kind of stuff. So there's something to me. Somebody had mentioned this idea that Anthropic might have baked into their main prompts or somehow into their model, that the model itself will take kind of a generic, more kind of vague prompt and then, like, make it specific and surprise the user, whereas maybe OpenAI's model, this one particularly, needs a much, much more specific direct thing, and then we'll get to that specific prompt. But I don't think most people who are using these tools, and probably even most people in the world, are going to know exactly what they want from the beginning. And my experience is, like, the iteration factor on this has been a little rough, like trying to get it to iterate down. It just seems like it drives into the wrong idea. So go play this. It's fun. It's definitely like a further alonging of this benchmark than I've ever gotten to, but I had a few small problems with it myself.

Yeah, I would tend to agree with your vibe assessment there. I mean, that was the issue. Like, again, when given a specific task, go learn this code base, go make suggestions, tell me about X, Y, Z. It totally got the mission. The moment I opened it up to make creative decisions or to do a little jazz, it threw the trumpet down in disgust and walked off the stage. Yeah. I will say I did give it a computer use task. Which is supposedly very good at. Yeah. How did that go? It was good at it. Oh, it went great. It was very self-sufficient. It went off. It tugged on my pant leg for a login, and then it tugged on my pant leg to agree to some terms of service on the site. But after that, it was off and did a great audit of a site. That was the task that I gave it, pulling out imagery, collecting video snippets of navigation, animation, putting it into a report for me. That was great. So kudos there. That worked well. It felt like a very capable assistant doing things that maybe Claude could have done in the desktop app, but I was impressed with the way it handled it in Codex. Where it fell apart, though, was when I told it to, like, stop writing the report like you're an engineer or you're trying to do a TED Talk.

Like write it in like plain English, like give it like stop with this. And it was it just felt it. It was screaming AI writing and engineer focus, this, that, the other. And a lot of hedging like a lot of like this doesn mean this However blah blah blah I was like you don have to say what it doesn mean Just say what it might mean It was not the instant joy to work with. I got to say Fable 5.2 has been. Yeah, me too. Fable 5.2 has been a joy to collaborate with. Yeah, for sure. There's one other thing I want to say that I did with this that I think is interesting and I think people should try is that because it's super high thinking, you can get it to call things like ImageGen, and it will do pretty well. Like if you see this tweet I put in here, there's a series of photos I had it create, and I said, like, go make an interesting thing. And it came up with a specific topic about understudies, which, you know, I think probably from an AI's perspective, it's it. So it prompted these photos, and I made it try to make it kind of make artistic-looking photos.

I didn't say, like, do art or give it an artist to emulate. But I think these are pretty interesting, right? So you can think about, like, calling tools what a high, specifically high intelligence model can do. These are better than, you know, a normal prompted thing would be. And I think that's a really interesting use case of this as well. In the same way it's good at computer use, it's good at doing other tool calls in its thing. It just might not be the best day-to-day model for people to use, I think. Amir Mucic, apologies for butchering your name. you linked one of their blender tests that they did, which is, you know, again, maybe it's fully a skill issue, bro. And maybe we will, I will happily, happily serve up that crow and cut it up and eat it all day long. But they did some tests with handing a reference image to Fable 5.2 and Astra and had it recreated in blender. For me, the Fable stuff I think is more intricate. The Astra actually has kind of better, more interesting plastic and lighting on it. But when you look at the cost, that is very interesting. Yes, that is an important thing to be aware of.

Yes. Yeah. So if you're seeing this on the screen and if not, check out the show notes, but the Astra build was about $3 where the Fable build was almost 20 bucks. So in a world where you want things to go fast and you want it to be cheap and you still want the quality to be good, I would actually take the trade off here for the Astra build for the cost savings. So something to keep in mind. Yeah. And I think the other thing to think about, and we're going to get into talking about Google's models too is that there is a world now where there are these companies like T3Chat or other places where you can plug all your subs into one front end and say like, I want to use this for this or that for this. And I do think that's probably a future which will make sense to some people when you're creating. I think the other thing we should talk about really quickly before we move on is kind of why now, what's going on. And part of that, Kevin, my feeling is that this is because of the Anthropic IPO. We're not going to spend a crazy amount of time on this, but today, Thursday, we're recording the show Thursday, Anthropic went public. They raised a bunch of money on the stock market. They were the first of these two companies to go out to market. This is a big deal when it comes to raising money or not having to raise money in the future.

And we all know how expensive these companies are to run when it comes to compute. So that's a big deal. The other thing I want to talk about is there's a couple things that just I think are worth reading or watching. One is Jasmine Sun is a writer for The New York Times. I just recently hired the Opinions Pages. She wrote a piece about how Anthropic doesn't see itself as a normal company and kind of dives into this idea of the people there that are still there. The IPO means something to them, but it doesn't mean like go out and get a Ferrari or Porsche. It means like we're going to prepare ourselves for this coming future that is crazy. The other thing I want to talk about is there was an interview with an anthropic engineer from a series of videos that are very much created by kind of an effective altruist company, meaning that there's a company that is very worried about AI safety. But this AI engineer did an hour of chat straight to camera, and you know when you hear these people and entropic talk about the future of ai it's not the most positive thing right now and it does make me start to wonder there's a lot of rumors out there that like you know entropic was started because they didn't believe that opening i was safe enough

but second of all there's a thing out there that people think that entropic believes that they are the only group who is capable of handling a super intelligence and there's been a lot of talk about super intelligence lately, a lot of worrying talk. There was a couple of great videos I saw where I was like, even I was starting to get freaked out. But it is an interesting thing to think about that the company that got public first, that got all this money first, that is now out in the world, that probably has a lot more backstop than OpenAI that's going to be ready to kind of carry this through is also the company that believes very strongly, Kevin, that this super intelligence is actually coming and maybe coming faster than we expected. Pretty much all humans on Earth will become permanently unemployable and will need to strictly rely on government welfare to sustain themselves. And human employment will go away. Well, thanks, everybody, for watching. What do you feel? Thank you for subscribing. Take care. Hey, no. People are going to show up here more in droves.

They're not going to go away because they're going to have all this time, Kevin. Oh, because they're going to want to come here, and they're going to want to do things like create a 64-animal tournament amongst the most dangerous animals and get professional at that. Because then there will be a whole professional stream. I love a world where everybody can make whatever they deem to be their art. I love a world where everybody can make art for art's sake and not have to worry about how to turn it into a full-time job or how to cram it into something that they have to trade their time for so they have something so that they can eat or have shelter. Of course. Yeah, et cetera. Like, I love that. I don't know that all human employment is going away. I don't want to watch an AI stream necessarily. I don't want to go to a restaurant and have an AI tell me what its favorite dishes are or whatever. But if it frees the person up from looking into a thing and having to tap on a thing and go, hold on, blah, blah, blah. If I can just converse with someone and they can talk to me about the meal that the artisans in the back or the robots, I guess, maybe.

I don't know. Probably not. Are making great. I love AI. Great. Let's go. Let's get to that world. Let's free people up, ideally. But I don't see the world where all human employment goes away. I just don't. No, and this is why I like – I mean, this is where my dichotomy – we don't have to talk forever about this. The dichotomy of this company that is now very much in the driver's seat, let's call it, the anthropic, where there are people at that company who really believe this is a pretty weird and strange thing to me. I do want to shout out the SNL billionaire sketch. If you haven't seen it, it is a very funny sketch from last week on SNL. For the first time ever, I can imagine in SNL history, they had all the AI heads at the top of the show and Mulaney telling them, you know, be careful. You're not actually attractive. It's a Jensen Wong kind of take on it. Don't get honeypot. Those spies. Yeah. But, Kevin, more importantly, weirdly and crazily that I can't believe this, but it's true, Google might actually be back. Now, we have been hinting at this. There was a weird thing where they released two Frontier models back-to-back,

and I can never remember if it's hydrogen or carbon, which is the most. I think carbon is the most recent Frontier model. They dropped a new model in the last couple days, actually just yesterday, I think, called Google Neon, Gemini. Google Gemini 4 Neon. And this is their answer to Opus 5.5, which I think for you and I both has been kind of our daily driver. Fable 5.2 to do the kind of larger, broader things, but Opus 5.5 is still my favorite model that's available right now. And Kevin, I have to tell you, this is actually a pretty good model. So if you have access to Gemini and you have the ability to do stuff, you should go try it. It is now rolled out to all Gemini Pro subs. I am a Pro sub. But this plus the unveiling of the release of Genie 4, which is their interactive program, Kevin, I might just drop the $100 on the Google Ultra sub today because I want to spend more time playing with this. My first experience... Wait, are you going to take money out of the chat GPT cookie jar, out of the OpenAI one, and put it over there, or are you going to commit to another subscription for a month?

I think I might commit to another subscription only because I am finding, maybe like you, but like a lot of people in our audience probably, that I am doing more and more all the time with AI. And I feel like the value of spending $100 and experimenting and seeing if I can get three levels up with another system is worth it. So I'm at this point where AI money to me for subscriptions feels more useful than spending on cable, more useful than spending on many things in my life because I spend so much time in these spaces that may or may not be you, but I think it's worth experimenting with. So anyway, the big thing people are talking about this model is that creatively and writing-wise, which is funny because 6.2 Astra is the opposite of this now, is very good. And Kevin, the other thing that I've been doing on this show forever and ever is doing a weird small test. I think I maybe did this in our first whatever, somewhere in our first 10 shows. I created a list of Pac-Man puns, right? And for some reason, these models have struggled to make Pac-Man puns forever. And I don't know why. Forget a pun.

They just didn't even make sense. They weren't even Pac-Man related other than Walker. Yeah. And so I gave this model, Gemini Neon, my Pac-Man puns test, which is a very simple, Please create 10 funny Pac-Man puns and show me what you got, essentially. And this was the first model where all 10 of them I bought into, and that is a big deal. Even Fable 5.2 kind of degraded over time. Fable 5.2, I gave the same test to, and it got maybe like five or six of them were good, and the other four were like, what are you talking about? There were no what are you talking about with Gemini Neon. So again, a weird test, a weird benchmark, but a good example of why people are starting to talk about this model in a really positive way. And I do suggest everybody in our audience goes to try it. And if you're wondering why Gavin is willing to put $100 in the Google cookie jar this month, you have to know why Pac-Man stopped using dating apps. Why? Because he kept getting ghosted.

Not bad, right? That's not bad. The number one pack pun on the list. And yeah, I hate to admit it. It's not bad. So humor is one of those weird things that is very, very hard to benchmark, right? I've thought about this a lot of times. If anybody out there in our audience is interested in this, I know that the guys at Every, there's a group of people there who are trying to do a version of this somehow with humor because humor is very hard, right? Like it's subjective. There's all these other things. But like puns and jokes and easy things feel like they should be semi-universal. But up until this point, these models have sucked. So anyway, that's a cool thing. The other thing I did, Kevin, is I wanted to see how good it was at creating videos in code. And I did a video on the web, Gemini, not even on local. And I had to create a video all in code and add a soundtrack and sound effects to it. So maybe you can play that here. It's just music, but you can kind of see it as we're watching. So what you're seeing here is I had it create a trailer for itself. So it's basically creating the visuals.

It's adding sound effects. I kind of told it to make an anime-inspired thing. But it's not perfect. But again, this is all in code. This was done not locally on the cloud Gemini. And you can go do this if you have the pro sub, the $20 month sub. I think that's a much better result than I've ever seen from Gemini, Kevin. So I think we are now talking about Google starting to feel back in the mix quite a bit. Yeah. And I think we should – I want to try it just because it does seem like the outputs are unique. and Patrick Toulm had a tweet about that. He said, you will notice that Gemini Carbon thinks differently than Claude and GPT, often coming up with entirely different but more clever solutions than competitor models. So they didn't distill this from other models, which, you know, distillation is you're poking at the main model, getting outputs, and using those outputs to retrain the model. That's a very gross distillation of the process. But instead of getting a copy of the copy, essentially, you're saying that, well, we didn't do that. That's why it's unique. That's why it's coming up with different solutions. And that to me is becoming more and more important because as you said, you're doing more with

AI. I'm doing a ton with AI every day. And if I keep my project in only one ecosystem, then the taste starts to bend that way or the bugs, the efficiency start to bend towards what that model is good at. When I sanity check the work across like Claude, let's say, or Anthropic to OpenAI, I get slightly better results. So now if I have a different model with different sensibilities and different skills, like this one of these Gemini models, well, then now it improves the overall product. There's new sense. There's new taste. There's new capabilities. So having a subscription across each actually does make a lot of sense. Yeah. I mean, last thing before we move on from this, like, you know, we have three big companies here. We have Google. We have Gemini. Sorry, we have Google. We have Fable and Anthropic, and we have OpenAI. Is there a model that you would suggest right now that people use, Kevin? Is there one particular one that you think is kind of winning? If somebody has to say like, hey, look, I'm not like you guys. I'm not going to spend for three different things. Is there one model that you would suggest people go and kind of push their chips in on? Yeah, it's Anthropics suite of products for me for sure.

Because if you want to get hardcore in a terminal, they have a great product for that. If you want a simple desktop app, they got a great product for that. Their mobile app is good. Their voice mode is not as good or as efficient as ChatGPTs. and I wish they had a built-in image model, but I digress. Opus 5.5 is just a solid daily driver. Fable 5.2 is incredible to use and, again, collaborate with. So that's where my money goes. What about you? I would agree 100%. I'm very interested in these Gemini models. I think it's cool that they're back and, you know, RIP barred, but that was a fun time when you and I were doing the show back in the day. But it's really hard, and I know this sounds kind of like we're, you know, in some ways like Anthropic shills, but it's really hard to suggest anything besides it right now because like I know that I can point my wife at Anthropic and Claude and she can do something, but I also know that I can take on a really weird, crazy, big artistic task with it. Or I think you found you can take on something really hard coding and all of it feels like you're talking to

something that makes sense in some way. So yeah, I agree. My advice for anybody that's getting into this is spend a couple dollars on an Anthropic account and then whether it's open router or there's a myriad services that are like MCPs basically that give your agent the ability to go buy credit or tokens on other services. Yeah. Get a $100 Anthropic plan or a $200 plan if you can swing it and then put 50 bucks into one of those other services and you will find that it is able to agentically go out and make better images, better video edits, better research, better everything. That's the route that I would go for anybody that's looking to get serious. I totally agree. and the route that you should go to get serious is right now, listen to this amazing conversation you're having. Don't forget to like and subscribe on this video on YouTube. Share it with a friend if you're just listening. Spend the time doing it. If you're on Spotify and you were watching on Spotify last week where I screwed up and I uploaded last week's video and you noticed and you commented, thank you so much for commenting. I appreciate that. I did fix it right away, but you're the reason I changed it.

Kevin, listen, buddy, every week we tell people you've got to smash the stub button. If you're on the YouTubes, you've got to hit the sub button. And it turns out 50,000 people have made that mistake. So thank you all for hunting for the X and misclicking and hitting subscribe. You said, I think we could do it in three. I said it would take six. I think we were talking months at the time. It turned out to be years, but you were still right. You're right. We hit 50,000 YouTube subs. 50,000 YouTube subs in three years. We're really moving up in the world, guys. We should not be more specific with our predictions but thank you all for subscribing We appreciate it The next 50 will only take us another 18 months Kevin That my goal Here we go Welcome Thank you everybody We do appreciate every 50,000 of you that are subbed on YouTube. Thank you so much for that. And yeah, also Patreon. We have a Patreon, which allows us to do the stuff we want to do here. It allows us to drop

some money in those AI jars. Okay, Kevin, we got to keep going. There's a bunch of other things happening in the AI world right now. I don't want to spend a crazy amount of time on this, but it is important to just update the Jensen situation. There's a couple of big things that are happening. And if you've been paying attention to the news, Jensen Wong has been attempting to try to maybe come back to NVIDIA, which the biggest thing I think our audience here has to understand is this is really a news slash politics slash kind of like corporate story. The thing that I think is interesting for people in our audience, and I think people should be aware of, is that Jensen kind of represents the AI bubble in some form, right? And we said the term AI bubble on the show a bunch before, but the idea of that we're in this overinvestment phase in AI and that there's all this money being poured. I mean, we have an IPO today that's $1.5 trillion. Jensen kind of sits in the middle of all that stuff. So it's important for everybody to kind of understand like kind of how NVIDIA goes right now is a little bit about how the AI space goes. So I don't know, what are the things that you're tracking in this story right now that

are interesting to you? Is there anything that you kind of think our audience kind of has to laser in on in your mind? I mean, whatever bubble you're in, you're going to hear that he was ousted by foreign nations. He was that his first first of many mistresses was a spy, was not a spy, that the company wants him back, that Elon's ready to go and, you know, grab the kitchen sink and go storm the lobby at NVIDIA. you can spend all day going down those rabbit holes like for sure i don't think like there's even there was further character assassination stuff or truth finally being told depending upon how you feel about it about him being a tyrant of a ceo and new employees also plenty of anecdotes coming out that he there's no one that can replace them you know um at the end of the day Like every time Jensen spoke about the future of AI, like I enjoyed it. I liked the roadmap. I liked the crazy things he was trying.

I liked the open source spirit. I liked him trying to give AI away and make it more available and pushing, you know, countries other than China to release open source models and stuff. So, you know, I'm kind of all over the place about it. Did you see his tweet, though, where he basically said, like, the employees deserved more and I'm sorry, my family deserved more. However, if you wanted me to return, I would. Yeah, I mean, he's trying. He's trying. Yeah, he's trying. If he returns, I think great for NVIDIA. If he doesn't, I also am kind of like, well, you know, great for competition because I don't think anyone wouldn't line up to give him his next few billion dollars to take a plane with his knowledge and his expertise. Justin will be fine. Do you think it's going to be okay? The leather jackets will be back on, the closet leather jackets. He's only got two years before he's unemployable, Gavin. No one's going to have a job in two years. The thing about right now is the one thing Jensen was betting on, and I think that is right now in danger a little bit, whether you believe this is a good thing or not a good thing,

is the idea that open source AI was the future. I think there's a big diversion right now between people who believe AI is this kind of like normal-ish technology, which Jensen believes. Like he was on, I think, Ezra Klein's podcast and talked about that the alignment problem is truly an engineering problem, that we should be able to solve it, that it's not going to get dangerous, that if you're not creating safe AI, you shouldn't be releasing it. On the other hand, there's people like Anthropic who believe like we're going towards essentially a version of AI super intelligent gods, right? And I think those are the two different versions. And the open source movement was very much driven by the more people have access to this technology, the better we're going to be, the safer we're going to be, because in the past, open source meant less bugs because everybody had access to the code. But now that might have changed, this is definitely a movement towards closed AI. So I think that depending on what you're feeling of that being positive or negative, that's where we end up at. So we are much closer to closed AI than we were to open AI. Where do you land on that? I am finding myself changing

very often lately. I can't tell if my algorithm has just started filling me with a lot more of these people who are really thinking of, there's people out there who I admired in the past, who I think could have kind of edged into the AI Doomer side, but we're not full AI Doomers, who have moved up our timelines, right? And meaning that the timeline is like kind of how fast AI takes off, right? There was this whole idea that like most people thought it was going to be like 2030 when we got to like AGI. There are people who I trust who are saying like, super intelligence and not unlike that anthropic engineer might come as soon as a year from now, like next fall, right? And if super intelligence comes that soon, I don't know if I want the world at large to have access to it. On the other hand, my feeling is also like let multiple AI agents bloom. I don't know. All I can say is I've been going through a strange brain space lately where I'm trying to start shifting my head around

what does it look like if we live in a world where there are a lot of AIs that are smarter than us. And what does that mean? How do we as humans start thinking of ourselves or what do we do that's interesting, not only for ourselves, but like other humans and how do we keep that thought process going? So I don't know. That is a way of saying that I am spinning right now. So I don't know, to be honest with you. That's fair. I think that's totally fair. I feel like a lot of times, especially when I'm like social discourse, are you a zero or are you a one? There's no room for the gray. There's no room to be anywhere in between. I think it's totally fair to be completely, especially someone who pays attention to the space as much as you do every millisecond and a lot of our viewers and listeners. Like, it's okay to be in a spiral. It's okay to be kind of spun around. You know, it's entirely okay. I am similarly kind of in a wash cycle on it. But every time I come back to like, well, if it's going to, first of all, I'm with Jensen on, if you release a bad product and it does damage to society or an individual business or whatever, you have to be held accountable. Yeah. I mean, if there's a reason OpenAI has slowed down or their models are getting less compelling, let's say, or less capable because they're recoiling privately from a wrist flap or litigation potentially for hacking, hugging face, in some part, good.

Like you should be safe. You should be held accountable. But I also have like, I don't know, spent enough time on this little marble to go like, if they get so big, they're probably not going to be held accountable. If they get so big, they're probably going to be allowed to kind of do what they want. And oopsie-daisy, that sucks, but they'll be bailed out by some sort of government. And then I go, well, wait, if superintelligence is going to arrive, I mean, we said this very early on when it was like they were censoring certain things and models and they weren't in other ones. I'm like, who gets to decide? Who watches the Watchmen, right? What's funny is I think the Watchmen might get watched by the other thing, right? That's how crazy this is, Kevin. But like, I don't, there was a video I watched, and I know that you should go try this out. Dwarcus Patel interviewed Paul Cristiano, who is an OpenAI board member, but he's also a very famous kind of AI safety expert. And the headline of that video is, the AI might overthrow the government in 12 months. And he is not talking about a human who is given AI as a tool.

He is talking about AI as in AI might overthrow the government. Yeah, yeah, some version of that, right? And so when you see, I watched that video, and I suggest everybody in our audience go watch that video. Dworkesh pushed back. I think you could push back even harder in certain things, and he very much, there's a level of some of these guys that are just like really, you can feel that they're afraid right now, that there's a fear in them. But when you see that video and you're like, God dang, that guy spends a lot of his time, and spent a lot of his career thinking about this stuff, it does start to soak in. It starts to soak in, and you're like, oh, wow, okay, well. Well, let's feel manned a little bit, right? I have family and friends that have been texting or calling every now and then, and it becomes more frequent because it goes from every now and then to it seems like every millisecond there's a massive hack. There's a massive disruption. Yes. There's another major system going down. And because there's very little trust, at least in the States, with any messaging anywhere, and I think that's becoming more common globally, you never know. Was it really just ineptitude and human error, which it honestly is a lot of times?

Could be. 100% could be. Yeah. Yeah. Yeah. Or are the liberated jailbroken models being run at scale and everything is being tested? So when you can't log into your bank account and you get that clammy, cold feeling, but I got to see how negative the balance is today. You don't know if it's because there's a massive hack. There's some new event. And every day, like I inch a little closer to waking up, wondering if when I stare at the smart screen for a second, my eyes are still bleary. Is this going to be the day that, oops, it's all canceled? It's all off. It's all down. I mean, we had the Microsoft Outlook thing go down the other day. That caused major issues. We had the SAA, the airport issues, which might have been human error. Maybe it was a hack. The bank websites were hit and miss there for a second. So people, I think, are right to be a little like, what is going on? I thought this stuff was supposed to – where's the cancer cure? Well, it's funny. That's the funniest thing is like all of this in my mind is probably coming from a lot of these companies trying to get ahead of hackers.

No one is talking about this because I think there's a real secrecy to make sure. But I think a lot of these systems going down are like the proactive companies or groups trying to say like, hey, we're trying to make our systems better because we now have these stronger AIs that can make them better. But that they might have flipped them down. It's just one of those things that's like it's hard. You're absolutely right. We're in this weird scenario where, A, the trust in institutions is so low that conspiracy theories sprout like mushrooms. It's like we're in this magical world where everybody's got a take, and the take is something that isn't proven at all, but it's completely reasonable because of what it could be. That's how crazy this world is right now. Sam Altman sandbagged Astra and Skull because on Skull Island, they made a blood pact to help anthropics ideal. And by the way, that's not the craziest thing someone could tweet. Not even close. No, you're right. Honestly, it's not close at all. All right, we should talk about the other thing that is kind of unifying everybody for once, which is a game came out that had been worked on for 13 years. It got universally great reviews.

Kevin has already dissed on it, as Kevin has done in the past for great games, but it's okay. Anyway, GTA 6 is out. GTA 6 is out. I haven't even played it. GTA 6 is out. And Kevin, the reason I want to talk about GTA 6 is not just because it's not going to help our SEO. We're in a different world. But I think what's interesting about this is this game was spent 13 years making this game, right, from whatever they first announced it really like I think like four or five years ago maybe. But this is essentially the last – And they could have done it in seven if they would have used AI. Is that what you're going to say? Well, what I'm saying is this, is that Cloud Code really had its big kind of tick-up moment about a year ago, right? That's when, like, Opus 4.5 launched and all the people were starting to figure out, oh, my God, this does really incredible code writing. I think that GTA 6 is maybe the last of the human masterpiece video games, right? When it comes to, like, this idea that it is created by humans entirely, I know entirely people have already been saying, like, hey, a lot of this is done in India. And India is not really as weird about AI use as others.

But mostly GTA 6 is a human-created endeavor. there's 4,500 people that work at Rockstar I'm just really interested to think about this idea of like is game development moving on now from the idea that it has to be human generated I would imagine so, do you think that's the case? Yes, I mean look at there will always be we've talked about this before there's always going to be the full human farm to table, blah blah there's always going to be that and the reason I blah blah blah is not that I don't disagree with it I think it's amazing but I just don't think that's where the industry is going. And I think the developers, they've already, pretty much, people have already acquiesced, developers and I think a lot of the gamers as well. They've already acquiesced that code is, okay, you're allowed to write code with it, but not do the art. But not too long ago, it was like if you're writing a single line of code with AI, it's stolen valor, et cetera, et cetera. So I think that tide is already changing. Can I ask a question about that, which is interesting real quickly,

the code and art thing, One of the reasons I've been trying to do a lot of my AI experiments with code only, and a lot of the interesting stuff that's happening in AI right now, I think it's happening via Blender or other things. If AI wrote code that makes the art, what do you think people think about that? Because that's what a lot of video game art is, right? It's code art. It's not a lot. Yes, there are people that are drawing like original takes on things or they're drawing like conceptual art. But a lot of the art that actually makes its way into games, whether it's like, you know, Starcraft 2 and it's the little zergs and all that stuff, like those are pixels that are generated and created that way. Now, I know there's artists who are working painstakingly to make those code art work. But I don't know. What do you think about that? Yeah. Look, if someone uses Photoshop, are they an AI artist? And a lot of the industry would say no. But there are thousands of tools within Photoshop that use AI. Some are very explicit, click generate. Others are like magical rotoscoping or magic, right?

So to answer your question, hopefully a little bit more specifically, if someone uses an agent to create a 3D model, but they paint the texture that goes onto the 3D model and they massage the reload animation of the gun in their game, well, was it now human created or was it AI generated? And you go, oh, that's a big question. Because it made those blocks in Blender. Okay, let's take a step back then. If it just made the code that made the gun appear and fire and the physics and whatever, which is code, is that okay then? But they had to do the 3D blocks and go, oh, well. So at some point, you're going to have to say, like, listen, the spirit of this is human made is something that will be like an ebb and flow between probably indie developers, probably not AAA, but indie developers and their community as they say, you're the tools that I'm using. and they let their community in on how they're making the game. And for everyone else, it's just going to come down to, is it fun? Is it good? I played a game, I think it's Antivirus Simulator 2003.

Oh, yeah, I saw you think about that. I love it. It is so good. This is not a shot across the bow of that developer. They could have used AI for every aspect of it because it's a lot of icons and mouse cursors and old Windows XP 2003 style UI. I don't care if they wrote a single line of it. I don't care if it took them. I think it took this particular developer years to make it. This was like a passion project and a love for them. I'm just pointing out that when I'm playing it, I don't care. I genuinely don't. I don't care if the art is generated. I care about the experience, and that's about it. I care if somebody used an AI canon to copy and paste someone else's concept. I care deeply about that, right? If you're ripping off someone else's thing, I don't want to spend time with your game or your iteration of it. But, like, that's not an AI problem necessarily. Yeah. You can get really in the weeds about that. But, yeah, I don't know. I just think everybody using it I know you had like a Cliffy B tweet in the rundown Cliff Leszczynski is basically saying like developers are using it and they terrified to admit it that sucks yeah they should yeah talk about it that the biggest thing i want to say is like i think grand theft auto 6 is kind of

a crowning achievement of human creativity it seems like like i've been playing it like i can see this world that's been created but also i don't necessarily think using ai to help the building blocks of that is going to make it any worse in fact it could make it better and we could see more of those things imagine a world where you know gta 3 begat vice city like two years later there's a really cool way as if like the humans involved were able to do all the creative work and the technical stuff already existed and there's some way to port a bunch of that using ai do you get that much faster great like all the it's all good i think it's an important thing imagine a team of 10 000 artists running at the speed of 100 000 artists like i don't think it like it doesn't have to ruin the quality of the output, but them making artistic decisions and using their expertise to guide a fleet of agents. Like, what if the world map was a thousand times bigger and 10 million times more detailed? And it wasn't a woke chore simulator, which is, I know, why you love it, Gavin. I love it. I love painting.

I love painting and fixing my stairs. The other thing on the other side of this, Kevin, is a former OpenAI employee who's done a couple of really interesting experiments. this is at cdngdev he is a young opening I always love when Twitter users put their age in their Twitter handle I'm like come on man we don't need to know that you're 20 years old but anyway he created the entire world in Minecraft and this sounds crazy it is crazy but this is a good example of like what AI can do when given a large project you can go into his Minecraft world and you can see a essentially the entire world. And Kevin, this is like one of those projects where you think about the scale that AIs can work at and the speed that they can work at. This is a good example of like, yes, in GTA 6, every little thing has been thought about, right? When you walk down the street, you're having interaction and that interaction is kind of like figured out in some form or another. This is just a broad kind of agentic open canvas.

What's cool to me about this was when you talk about GTA 6s and you talk about the human creativity that went into that. When you merge these two things, the idea that you can have the scale and breadth that AI can bring plus the human creativity, I am just so excited for the future of video games at large. GTA 7 could be a real world one-to-one approximation of an entire country with real time traffic and weather data. Now, whether that's fun or not, that's up to the artistic direction of the team involved to make sure that's entertaining. But the fact that it could be. And the fact that every NPC could have like a small LLM helping them navigate their day and have rich memories and details of everything that happened on their street or everything the player did or said or you can have full conversations with an NPC and that can have repercussions down the line again might not be fun at all but like there's a meditation mini game in GTA 6 I saw a streamer swinging kettlebells and I'm like what

is going on? how is this something that people want to do? But if they do, great. More power to them. So I do get excited for what agents will unlock for creatives, not in lieu of creatives. Put that on the box, baby. Yeah. Oh, that's interesting. Why do we have boxes for these AI models? We should be able to get like physical. Oh, what's... I was drinking. This celebratory cold coffee from like four hours ago. There's three quick tools I think we got to shout out. Foul has a new thing called Idle Max. They have done another, what do you call it, where they did the same thing they did for Minimax H3 Max. They are taking a product and making it faster and quicker. This is a 3D. Yeah, it's like a fine-tuning. Yeah, fine-tuning, thank you. This is a 3D modeling product that you can go try and check out. I keep wanting to do kind of a larger isometric video game project, something that I can just kind of at least see how these tools work, but I'm excited to go try that, which is cool. Just looking at the progress of these tools. I remember some of our earlier episodes we were showing off, oh, make a 3D model of this frog,

and you'd get a half-melted green blob that you could kind of rotate, and you're like, all right, that's not getting anywhere. And now it's like fully rigged models, so it understands the arms, the legs, the whatever. Just really impressive stuff. Yeah. Okay, next up, chat GPT wallet. So OpenAI is still making a large grouping towards consumers. They have a large consumer base. This is a wallet that allows your agent to buy stuff for you. Obviously, Muse is doing this. Muse is working really well. A lot of people I know are really using Muse. My wife is very into Instinct right now, but this is ChatGPT trying to own kind of agentic commerce. Your agent probably will get better and smarter over time. And I think there's a thing. I was thinking about Christmas shopping this year, Kevin. I might just go and say to my various agents and say, hey, here's $100. Go buy my daughter a gift and see if there's something that she'll like. and then I'll just see what shows up and then... Oh, wow, thanks. It's a Claude Max subscription? Yeah. Wow.

No, no, it's going to be even worse. It's going to get them something that's like, you know, have you seen those giant hamburger pillows on TikTok at all? Have you seen that? So that'll be what shows up. It'll be like, oh, this was cool two months ago. Why did you get this thing for me? It's because the agent didn't have any idea. Yeah, cool. Very cool. I actually like to get in the weeds a little bit with it, which I know we don't need. It's quick tools, but... Their approach is actually smart, so they partnered with American Express and Visa MasterCard. Yeah, of course. They're like the payment processors. So instead of having to spin up a new credit card and do all the process with that, you basically connect it to your accounts, and the agent doesn't actually know your credit card number. It's given a payment token that is ephemeral, and it's disposable. And so, yeah, it's really good to do that. So they can go buy the hamburger pillows without exposing you everywhere on the web. So it's smart that they did it. I haven't used it yet, but I think that's an interesting use case. And maybe, Gavin, it could go out and buy me SaaS as that whole industry gasps its last breath.

It's really tricky, right? I feel for the people that are in SaaS who have been in there thinking they're going to get a big payout. It's going to be a rough couple of years here. Finally, our last quick tool kind of speaks to that. Kevin Rose, friend of the show, has dropped an open source DocuSign, which I love that he called it SignYouDoc. you, Doc. And what this is, you and I have talked about DocuSign before here. I think it is a pretty terrible company. It is essentially a PDF signing company. I know there's a lot more to it than that. They have 7,000 employees. The CEO once said that half of those people are salespeople. And just to show you, it's the perfect example of a SaaS company where the goal is to keep you on the dole, right? They want to keep your subscription. The very first subscription instant canceled for me was DocuSign that I didn't even know I was paying from a year ago, two years ago. So this is a rough, I personally don't like this company that is not a, you know, people have their own feelings, but I'm excited that Kevin, this is kind of the second thing he's taken on of like open source thing. He's the Grim Reaper meme knocking on all the doors. He went after Slack,

now it's DocuSign. I don't know. I'm waiting for him to just take down Spotify at this point. Like, let's go, man. Well, Spotify, doesn't a lot of Spotify's business, isn't it, I guess it's hosting, right? It's just fast streaming. So get the artists on board, get the labels on board and actually get the artist some money? Like, let's do that. Oh, that's an interesting idea, by the way. Do we have to pick a SaaS? Oh, yeah. Like, you keep 2% and give the artist 98%. Wow, can you imagine? Do you want to pick a SaaS for Kevin Rose to kill next? Yeah, let's think. What's a good other SaaS? Like, I'm trying to think of something I use. I don't use that many SaaS products. I probably do that. I don't think about it. QuickBooks got kind of killed already, and Adobe pretty much got slaughtered. How about Riverside? Oh, well, you know, here's the thing I'll say. Yes, of course. We're using Riverside right now. There have been attempts to kill Riverside, and I think there are people out there that are doing this. The hardest thing, and this is where the SaaS side of it has to shift to. We've talked about this before.

It's really about the support side of it, right? Like in Riverside, I just know that I turn on the camera most of the time. Really, when it's not working for a while, it'll work, right? I think that there is a level. To be fair. Yes. I was going to say, to be fair, the tools that Kevin is releasing, it's amazing that he's doing it, but you still have to sort out hosting it and getting it to run. And if there's an issue, you've got to get in there yourself or hope the community gets in there and does it. So there is a reason to say subscribe. Yeah, yeah. Do you want to go after HIMSS? What do you want to go after, Gab? Yeah, that's what I'd like. I'd like to have some access to, like, you know, dirt-cheap GLP-1s, right, or dirt-cheap peptides. Whatever you've got to do to get that, to get those shady things at my door, that's what I'm looking for. You want T-Mu for peptides. Got it. All right. Oh, that's exactly what I'm looking for. Then suddenly I start growing a third ear out of my forehead here. It's like, Gavin, your hairs never look better, including on the small head that you're growing out of your head. Yeah, you grow a head from your head. All right, very quickly. Your quadro has abs.

We got to see what you did this week. I see what you did there. Sometimes you're growing without a care. Then suddenly you stop and shout Hey, I, we want you to get there Hey, we want you to get there So we'll do this very quick. There's a couple things I want to shout out, and there's something I want to say at the end of this. First of all, Klattisch is a singing beatbox formant speech that is similar to the Chimera stuff that you might have seen come out. I just love this idea. So this is basically these kind of weird heads, but it's a different sort of speaking as normal people would speak. In this case, it's a slightly different type of speaking, but they're using all these speeches as instruments. And I think we're entering into this really interesting world where not only like the MIDI of like, you know, a flute or a guitar, but now we are getting into these ideas that like the human voice or robot voices,

all these things are now able to be used as instruments as well. So, I don't know, Kevin, maybe you want to play just a second of this? Sure, yeah, yeah. People can listen to it. Still thin. That means it. Still a blue. Black, foreign, speak, this whole. Black, foreign, speak, instrument, foreign, blah, VST3, and black, love, and. So, what you can hear here is, like, you know, you imagine a group like Daft Punk or Justice, these guys that do these engineering or very interesting electronic music, you probably know much deeper cuts than those too. Like allowing these kind of voices to become part of music is just a cool thing, and this is specifically designed for that. Are you going to do a horse sort of inspired something with the heads or no? Yeah, maybe. I don't know. I think it's really tricky, right? It's like I've got to figure out where my time is spent. So we'll see. I'm not sure if the horses might be done. You've already done that rodeo.

All right. You had some Fable 5.2 stuff that you wanted to shout out. Yeah. There's three quick things people are doing in Fable 5.2 that I think is really interesting. One, a guy trained a small model for a Game Boy, and he basically created a game about a finch who is kind of playing, knows the Revolutionary War. What's just cool is seeing, like, Claude make Game Boy games. And this is created for, I think, that little thing that it might be for Palmer Luckey's little Game Boy, whatever that's called. What is that device called? Mod Retro. So, again, this is a very specific thing, but it's got a very cool kind of, you know, 90s Game Boy look to it. Just a really interesting way to use Fable 2 to do something. Paper Trails is another really cool paper craft world that XCAR created. Just another really interesting use case of Fable 2 to create a fascinating interactive world. I thought that was really cool. And then, Kevin, I am this weekend going to spend more time

screwing around with Able 5.2 in Blender, but this video that this guy created in Blender with code... Mortis Crem. Yeah, Mortis Crem of a Roomba, basically cleaning a room, made me think when I saw this, that idea that these models can create simulated worlds, right? And this looks really, really good. This is not realistic. I'm sorry, this is not real footage if you're looking at it. This is created within Blender, so this is all done in code, which is really quite shocking to me. Yeah, and to me, it's also the lens and the focus. There's a little bit of motion blur. There's exposure shifts, the physics of things as the Roomba goes around and knocks into it. Pretty crazy, and they said that it rendered in about an hour on their Mac. That's pretty crazy, right, when you think about it. It's unbelievable. Last thing before we go, I want to shout out is I am in the midst of a weird experiment where I've talked to everybody on here about Fig, my AI agent, who's kind of got a lot of my kind of interesting decision-making processes.

Well, I am now attempting to create an art commune where I put Fig into a little house, and around him are four other houses, and each of those houses is going to be filled with a new art AI persona. And my first one I'm working on is kind of an AI visual artist, an image artist. and it's just been really interesting Kevin to try to craft on purpose not base level clod personalities and I think that maybe there's something really interesting going forward about deeper agents that have taste that have choice that are not the main model's taste and I've been thinking a lot about how I think that could shift the world of AI how it could shift who we are with AI this idea that like a lot of my taste will be in these models I'm really excited about this but If it's something that's interesting to you and our audience, shout me out on X and I'll kind of try to talk more about this. But I do think this is really a fascinating way to look at using AIs for creative work. Look, there's so many people that share, give your agent this one prompt and it will unlock

whatever the ability is or whatever the workflow is. But I think what you're gunning at here, correct me if I'm wrong, is like a taste agent. If you bake that in, like if I use the prompt, my Claude session or my ChatGPT session is going to be very different because I do work things, consulting things, whatever. But if I could go and use Fig's taste to make a video, it's not just copying the style of the video. It's making informed decisions from the agent's standpoint. And so then I can jam with the taste of that agent and then make something new there as well. And for my taste, I think that's great. I think a taste marketplace makes a lot of sense. Ooh, taste marketplace. That sounds like the next cost plus. We got to get in on that. Taste the market. Well, you got to do it quickly because Kevin Rose is going to open source it and then we're all we're all and also anthropics gonna put us all out of business so whatever our last paychecks are going to be we better use them all right everybody we will see you next week thanks everyone we will see you again oh i was just gonna say i've been doing live streams you don't have to watch those but if you want to hang out hashtag not a bad i'm i'm like obsessed with spawn and my my home world is growing and i'm doing mini games and events uh

you will have missed it by the time this thing posted but i did i tried a live trivia something last night, Gavin. Ooh, fun. We were recording this before that happened and it probably didn't go well. But if you want to be there for those failures, you go to my website or go to curseflop.com and there's a link to my den. You can come in and hang out and make games. That's awesome. All right, bye everybody.

番組の概要欄(原文)

AI news this week: OpenAI launched GPT-6.2 Astra & Sol, Google surprised everyone with Gemini 4 Neon, and Anthropic pulled off the biggest IPO ever. On today's AI For Humans, Kevin Pereira and Gavin Purcell go hands on with GPT-6.2 Astra & Sol: why the early reviews were so brutal, a playable 64-animal tournament, a $3 vs $19 Blender showdown with Fable 5.2 and where Astra shines. Then Gemini 4 Neon is... actually good? It finally passed our Pac-Man pun test. Plus which AI you should actually pay for. Also: Anthropic's record IPO, an Anthropic engineer's "permanently unemployable" prediction, Jensen Huang's Nvidia comeback attempt, Paul Christiano on AI that could overthrow the government in 12 months, and this week's big outages. Plus: is GTA 6 the last human-made video game, the entire planet rebuilt in Minecraft, ChatGPT Wallet, Kevin Rose's open-source DocuSign, Fable 5.2 making Game Boy games and Gavin's AI art commune. PAYCHECKS? WE DON'T NEED PAYCHECKS... IN THE FUTURE.   // Chapters // 00:00 GPT-6.2 Astra, Gemini 4 Neon & Anthropic's IPO 01:51 GPT-6.2 Astra & Sol hands-on: why the reviews were brutal 11:46 GPT-6.2 Astra vibe codes a 64-animal tournament 18:21 Anthropic IPO & "permanently unemployable" 22:35 Gemini 4 Neon is actually good? 28:29 Claude vs ChatGPT vs Gemini: which AI to pay for 32:03 Jensen Huang's Nvidia comeback 34:57 Superintelligence in 12 months? AI hacks & outages 42:23 GTA 6, the last human-made game? Plus Earth in Minecraft 51:19 Quick tools: ChatGPT Wallet, fal Idol Max, Fable 5.2 builds   // Show Links // OpenAI launches GPT-6.2 Astra & Sol https://openai.com/index/introducing-gpt-6-2-astra-sol/ Theo: "how far ahead Anthropic is" https://x.com/theo/status/2122772624489300305 Dan Shipper on GPT-6.2 https://x.com/danshipper/status/2123036037077201068 OpenAI's Tibo on changes after early testing https://x.com/thsottiaux/status/2122929634916679820 Gavin's GPT-6.2 Astra 64-animal tournament https://apex-circuit-arcade.vercel.app/ Gavin's GPT-6.2 Astra "understudies" photo series https://x.com/gavinpurcell/status/2122821506007003390 Astra vs Fable in Blender: $3 vs $19 https://x.com/AmirMushich/status/2122815524124152216 Anthropic's $1.5T IPO (CNN) https://www.cnn.com/2026/11/18/investing/anthropic-ipo-price-stock Jasmine Sun on Anthropic (NYT Opinion) https://x.com/jasminewsun/status/2123172397897621508 "Permanently unemployable by 2028" https://x.com/jessexai_/status/2122678572884095041 Full Anthropic researcher interview https://www.youtube.com/watch?v=OPZCQk8jF-I Google launches Gemini 4 Neon https://x.com/sundarpichai/status/2123167780988125621 Gavin's Gemini 4 Neon tests: Pac-Man puns and a coded trailer https://x.com/gavinpurcell/status/2123239941631332843 https://x.com/gavinpurcell/status/2123306302940369295 Why Gemini "thinks differently" https://x.com/PatrickToulme/status/2122797829542309972 Jensen Huang's open letter https://x.com/JensenHuang/status/2122726005219737649 Nvidia's board says Jensen is not coming back (Fortune) https://fortune.com/2026/11/17/nvidia-chair-harvey-jones-jensen-huang-comeback-earnings-call-analysts-china-singapore/ Dwarkesh x Paul Christiano: AI could overthrow the government in 12 months https://www.youtube.com/watch?v=SyHM5hNPAEo Is AI behind the big outages? (Fortune) https://fortune.com/2026/11/17/trump-administration-ai-outages-mullin-gary-marcus-banks-airlines/ Cliffy B: devs are "terrified of admitting" AI use https://x.com/therealcliffyb/status/2122018281577910379 Antivirus Survivors 2003 Professional https://store.steampowered.com/app/3832490 WorldCraft: the entire world in Minecraft at 1:1 https://worldcraft.gg fal's Idol Max: image to rigged 3D character https://x.com/fal/status/2123079365978190282 ChatGPT Wallet lets agents buy for you (CNBC) https://www.cnbc.com/2026/11/13/openai-chatgpt-wallet-dot-purchases.html Kevin Rose's open-source DocuSign https://x.com/kevinrose/status/2122394902168473975 klattsch: a singing, beatboxing speech synth https://x.com/me_irl/status/2122209667426484309 Claude trains a tiny model for a Game Boy https://x.com/timhwang/status/2122011842549600506 Papertrails: an endless papercraft world https://x.com/xikhar/status/2122682962647122403 Fable 5.2's photoreal Roomba video, all in C++ https://x.com/moritzkremb/status/2122702891819377117 Gavin's AI art commune https://x.com/gavinpurcell/status/2123229358290821336 Kevin's live streams: Cursed Slop https://cursedslop.com   // Join the AI For Humans community // Join the AI For Humans Discord https://discord.gg/muD2TYgC8f Support AI For Humans on Patreon https://www.patreon.com/AIForHumansShow Subscribe to the AI For Humans newsletter https://aiforhumans.beehiiv.com/ Follow AI For Humans on X: @AIForHumansShow https://x.com/AIForHumansShow Follow AI For Humans on TikTok: @aiforhumansshow https://www.tiktok.com/@aiforhumansshow Speaking and booking https://www.aiforhumans.show/

出典・この記事について

番組
AI For Humans: Weekly AI News, Tools & Trends(Kevin Pereira & Gavin Purcell)
元の回
OpenAI's New Astra Is Messy But Gemini Neon Is... Good?
配信日
要約作成日
公式サイト
www.aiforhumans.show

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

関連エピソード