25:20Stacked Podcast · Reaction
The Truth About GPT-6 Astra
Two hosts react to GPT-6 Astra's world-simulation demos, then pressure-test the viral math behind owning a Tesla robotaxi fleet.
September 6thThe argument in one line.
A sanctioned pen-testing team used Claude Opus to breach OpenAI's private code in about three hours, and it's one data point in an episode where every story is the same shape: AI is doing more of the work, faster than the guardrails and business models built around it can adjust.
Read if. Skip if.
READ IF YOU ARE…
- You run or want to run an AI-focused agency or SaaS and want a pulse check on what Claude and OpenAI shipped this week.
- You're weighing whether to build a product on top of a frontier lab's model and want to see the Astra-for-Law example of that bet going bad.
- You want a plain-English read on what Claude Projects' cloud orchestration and Anthropic's '26% of AI research' claim actually mean, without the vendor blog spin.
SKIP IF…
- You want the technical walkthrough of the actual OpenAI exploit. The hosts explicitly decline to detail it.
- You're looking for a rigorous security analysis. This is two builders reacting to a headline, not a pen-testing breakdown.
- You want the fast-classification segment explained precisely. The hosts themselves aren't fully consistent on the tool's name or mechanics.
The full version, fast.
A three-person pen-testing firm used Claude Opus to build a working exploit against OpenAI's sign-in flow, reaching employee credentials and a private code repo in about three hours, under a bug bounty OpenAI itself authorized and paid $6,500 for. The hosts use that as a jumping-off point for a wider roundup: Claude Projects now runs multi-agent orchestration in the cloud instead of on your laptop, OpenAI's Astra for Law bundles the model with 230 million+ legal documents for select firms, and Anthropic says Claude now leads 26% of its own AI research (up from 1% in February), with recursive self-improvement treated as a trend to watch rather than a declared fact. The through-line: AI is absorbing more of the work, and the business and security models built around it haven't caught up.
Free for members
Chat with this breakdown — free.
Sign in and you get 23 free chat messages on us — ask for the hook, quote a framework, find the exact transcript moment, generate a markdown action plan. Bring your own key when you want unlimited.
Create a free account →Where the time goes.
00:00 – 05:17
01 · OpenAI bug bounty
Cold open: a legal, OpenAI-sanctioned bug bounty in which pen-testers used Claude Opus to build a working exploit against OpenAI's sign-in flow in about three hours, reaching employee credentials and a private code repo.
05:17 – 06:57
02 · Small bounty and security
The hosts react to the $6,500 payout versus what that level of access would be worth to a black-market buyer, and frame bug bounties as the 21st-century version of a Wild West reward.
06:57 – 08:37
03 · Claude Projects and agents
Claude Projects launches: cloud-native multi-agent orchestration that keeps running and checking in on subagents whether or not the user's computer is on.
08:37 – 12:33
04 · Agent orchestration and context
Debate over whether prompt architecture and MD-file context management are a real skill or a temporary workaround the labs will engineer away, plus a critique of Claude's 'compact' context-summarization feature.
12:33 – 13:58
05 · OpenAI Astra for law
OpenAI's Astra for Law bundles the model with 230 million+ legal source pages and specialist tooling, rolled out first to select large firms.
13:58 – 16:44
06 · Model vs application
The risk of building a business on top of a frontier lab's model: once the lab moves up the stack with domain data, it can absorb the application layer underneath it (Harvey AI cited as an example of the exposure).
16:44 – 19:43
07 · Claude's supervised research
Anthropic says Claude now leads 26% of its own AI research tasks under human supervision, up from under 1% in February and 14% in June, with roughly 30,000 research agents running in parallel.
19:43 – 20:59
08 · Recursive self-improvement
The hosts stop short of calling the 26% figure proof of recursive self-improvement, but treat the steepening automation-level trendline as the thing worth watching.
20:59 – 23:35
09 · Mac mini giveaway
Sponsor and community segment: a running Mac Mini (or cash equivalent) giveaway tied to commenting on the show.
23:35 – 25:20
10 · Fruit fly race winner
A meme minigame that turns commenters into racing fruit flies decides this week's giveaway winner on screen.
25:20 – 27:01
11 · Last week's winner Daniel
A thank-you video from the prior week's Mac Mini winner, Daniel Barnaby, is played on screen.
27:01 – 30:54
12 · Questions on AI numbness
The hosts read and react to viewer comments about feeling desensitized to AI news, AI-authored reports, and control over one's own agency in an AI-saturated world.
30:54 – 32:21
13 · Fast classification in AI
Closing segment on a fast, categorical classification approach (inconsistently named 'Jeff'/'Jev'/'Jeb' across the transcript) that scores a question into a few fixed buckets in about 0.1 seconds instead of generating a full reasoned answer.
32:21 – 32:33
14 · Outro
Sign-off and call to comment for the next giveaway.
Lines worth screenshotting.
- A three-person security firm needed about three hours and a deprecated Claude model to build a working exploit against one of the best-funded AI labs on earth.
- OpenAI paid $6,500 for a bug that reached employee credentials and a private code repository, a bounty far below what that access would fetch from a state-level buyer.
- Bug bounties exist precisely so vulnerabilities like this surface through a paid, legal channel instead of a black-market one.
- Claude Projects moves multi-agent orchestration into the cloud, so a swarm of subagents can keep testing, checking in on each other, and reporting back with your laptop closed.
- The hosts argue hand-built prompt architecture and MD-file scaffolding are a temporary tax on users, not a durable skill, because the labs are racing to make that work unnecessary.
- OpenAI's Astra for Law pairs the model with 230 million+ indexed legal source pages and scored 54% on the firm's own research benchmark, rolled out first to select large firms.
- Building a product on top of someone else's frontier model is a fragile bet: once the model owner adds domain-specific data, they can absorb the layer you built your business on.
- Anthropic says Claude went from leading under 1% of its own AI research tasks in February to 26% by August, with roughly 30,000 research agents running in parallel.
- Neither host will call it recursive self-improvement outright, but they treat the steepening trendline across Anthropic's own automation-level tiers as the thing worth watching, not any single milestone.
- A classification-first approach that scores a question into a few fixed categories can return an answer in about a tenth of a second, versus roughly eight seconds for a full LLM reasoning pass over the same question.
AI is outpacing its own guardrails
THE PATTERN
This is an AI news roundup, and the same shape repeats across every story: AI is absorbing more of the work, faster than the security, business, and oversight models around it are adjusting.
01OpenAI bug bounty
- A three-person pen-testing team used Claude Opus to build a working exploit against OpenAI's sign-in flow in about three hours, reaching employee credentials and a private code repo, all under a bug bounty OpenAI itself authorized for $6,500.
03Claude Projects and agents
- Claude Projects moves multi-agent orchestration into the cloud, so a swarm of subagent sessions can keep testing, checking in on each other, and reporting back even when your own computer is off.
- Hand-built prompt architecture and MD-file scaffolding are treated by the hosts as a temporary tax on users, not a durable skill, since the labs are actively racing to make that manual setup unnecessary.
05OpenAI Astra for law
- OpenAI's Astra for Law bundles the model with 230 million+ indexed legal source pages and scored 54 percent on the firm's own research benchmark, rolled out first to select large firms.
- Building a business on top of someone else's frontier model is a fragile bet: once the model owner adds domain-specific data and moves up the stack, they can absorb the application layer you built.
07Claude's supervised research
- Anthropic says Claude now leads 26 percent of its own AI research tasks under human supervision, up from under 1 percent in February, with roughly 30,000 research agents running at once.
- Neither host will call the research-automation trendline recursive self-improvement outright, but they treat its steepening curve, not any single announcement, as the thing worth actually watching.
12Questions on AI numbness
- The hosts' working assumption is that AI involvement in any given piece of research or writing is undercounted, because institutional reporting has an incentive to understate how much of the work an AI actually did.
13Fast classification in AI
- A classification-first approach that scores a question into a few fixed categories with probabilities can return an answer in about a tenth of a second, versus roughly eight seconds for a full LLM reasoning pass over the same question.
Terms worth knowing.
- Bug bounty
- A program where a company pays outside researchers to find and report security holes in its own systems, instead of the researchers exploiting or selling them.
- Hacktron AI
- The three-person security research team the hosts credit with using Claude Opus to build the exploit that reached OpenAI's private code repository.
- Claude Projects
- Anthropic's cloud-based workspace where Claude orchestrates and monitors many subagent sessions on a task from a shared folder of context, independent of whether your own computer is on.
- Astra for Law
- OpenAI's law-specific product that combines its frontier model with more than 230 million indexed legal source pages, specialist instructions, and legal-software integrations, rolled out first to select large firms.
- AL1-AL5 autonomy scale
- Anthropic's internal tiering of how much of a research task is AI-driven versus human-driven, from minimal AI assistance (AL1) up to fully automated (AL5), referenced in this episode via Anthropic's own charts.
Things they pointed at.
03:39productHacktron AI
06:57productClaude Projects
12:58productAstra for Law (OpenAI)
15:50productHarvey AI
30:24productDeepSeek V4.1 Flash
Lines you could clip.
00:00
“Opus has just hacked OpenAI.”
01:45
“The weaker model struggled with memory defense. Anthropic's newest model produced a working exploit in three hours.”
05:52
“$6,500? Dude, that doesn't cover the coffee run.”
10:30
“The user should not be thinking about architecture. This should all be solved by Claude and Anthropic.”
12:32
“The compact function in Claude is pointless. It just re-reads the entire freaking thing when you do it anyway.”
17:36
“We're basically four to six months away from the three-letter acronym that keeps all of us up at night.”
20:32
“Very soon it will overshadow us, and then it will be irresponsible to have a human do stuff.”
Where the conversation goes.
00:00 – 08:37denseOpenAI bug-bounty hack via Claude Opus
06:57 – 16:44denseClaude Projects, agent orchestration, and Astra for Law
16:44 – 20:59denseClaude's growing share of AI research and recursive self-improvement
20:59 – 27:01sparseMac Mini giveaway and community segment
27:01 – 32:33Listener Q&A on AI numbness and fast classification
Word for word.
Read-along
Don't just watch it. Burn it in.
See every word as it's spoken — crank it to 2× and still catch all of it. The same dual-channel trick behind Amazon's Kindle + Audible.
metaphoranalogystory
Opus has just hacked OpenAI. On top of that, we've got some brand new updates that have landed with Claude. Astra has dropped what we call Astra for Law.
We have so many updates to cover, and today we're also giving away a Mac Mini or cache equivalent, and you can win one too by dropping a comment down below. More on that later. Nick, let's begin with what's happening with this hack.
Claude Opus 5 helped researchers enter OpenAI's private code repo. Nick, what happened? Yeah, it's pretty freaking ridiculous, actually.
A team of researchers, pen testers, and security specialists hacked OpenAI and even got as far as gaining full control over employee credentials and employee access. So basically, it's a small team. It's like a three -person security team called Hacktron.
I should note that this was not an illegal hack. It was a legal hack that OpenAI asked them to do. If you guys could find any way to gain access to our systems, we'll pay you money to do it.
These are called bug bounties, and they're pretty common. But essentially, it took about three hours to run the exploit. OpenAI had a sign -in flaw that they hacked using an uploaded image.
And the actual details of the hack are quite technically complex, and it probably doesn't bode well for us to talk about too much of it here. But suffice to say, they did this with Opus 5, bro. They did not need a mythos to hack OpenAI.
They did this with Opus fucking 5. What I also find pretty interesting about this is it's Anthropic model versus OpenAI model. Do you know what I mean?
Like, can you imagine a world where Anthropic's now like, hmm, that's not a bad idea, actually. And then they like, or OpenAI's like, that's not a bad idea. And then they use like ChatGPT to like hack Anthropic.
That's hilarious. Because dude, Opus 5 has been sat on the bench, Nick. Who's even said the word Opus 5 in the past four weeks since it dropped?
It's been a little bit quiet. What was interesting is the weaker model struggled with memory defense. That's what we're hearing with this.
Anthropic's newest model produced a working exploit in three hours, Nick. Three hours to get in and break it. And we're not exactly talking about Windows 1998 defense here.
So as of late, since this happened, OpenAI has fixed the flaw. Builders learned that one connected employee account could reach private company code. So we could have poisoned that and done all sorts of craziness.
Nick, this connects to what we talked about yesterday with the deep seek hacking, that the models are just getting more capable now at breaching defenses. Dude, it's all happening right now. I mean, I guarantee you in the next week, there'll be like a suite of security incidents like this, not just because this has shown that it is possible, but because the stuff's actually just all going on.
And typically it takes time for leaks and hacks to be revealed. Guarantee you there are thousands, tens of thousands, hundreds of thousands of people that are just trying to hack fucking everything right now. Yes.
And it's a bit terrifying, you know, somebody who is on the Internet, you know, like think about it, man. There's just so many different surfaces by which like any major company today can be hacked. And this is like supposedly one of the most secure companies on planet Earth, right?
It's supposed to be the company that has AI systems doing its defense constantly. And it gets hacked with a freaking deprecated model that was posted, you know, a generation and a half ago.
It's wild, Nick. No company is perfect, number one. And it's great that these bug bounties exist for precisely this reason, right?
Let's incentivize the white hacker community to come in and we'll pay them for doing so. It's alarming that it could potentially get OAI auth tokens, OpenAI's auth tokens, and gain right access to OpenAI. That said, though, guys, I don't necessarily think it means that OpenAI have done anything.
wrong. They're doing the right things. They have said, hey, hackers, they're saying, look, .
Nick, you want this juicy thing? Hackers right now, I double dog dare you to do it. And someone said, cool, I'll do that for you.
They're actively trying to improve their defense. This is a positive thing. And it's great that they've done that.
Anthropic haven't done it. We don't know. If Anthropic did the same thing, it would be hacked.
It is also a flex. How bad a model can I use and still hack you? Do you know what I mean?
Like, can the greatest hackable time be done with like a GPT -3 or something? You know what I mean? Hilarious to see.
But this does show that like the need for these bug bounty programs are why they're so important. Even asking, you know, if you've got an app, guys, even ask your buddies to say, dude, could you just try and find exploits? Try and find vulnerabilities.
Like they may just try it a different angle than you've got. Like, I honestly highly recommend that you do it and think about it that way. And I'm just glad, Nick, that OpenAI are doing this kind of thing.
True. You know, back in the day when I played lots of Halo, there was a thing called Forge Mode where your buddies would try and create a world and like a game. And, you know, then you guys would play in it.
And one of my friends would always ask me to look for basically ways to hack his Halo Forge game. And it was one of my favorite things to do because I would come into his little thing and I don't know, man, he'd like make a tunnel and we're going through the tunnel getting chased by some freaking guy with a sword. And I just I'd always like find a way to like hack his thing.
And that's kind of reminding me of what you were talking about earlier, where it's just like you're asking somebody to come in and see if they could break your game. Obviously, it kind of sucks when your game gets broken, but it's really good information. My issue is they did this for $6 ,500.
That's how much they paid these guys. Brother, you could have sold that hack for probably $6 million. Yes.
Okay, minimum. The fact that they did this. And, you know, like they're getting some clout from it.
Like Hacktron AI, that's the security group that made a bunch of X threads since. And, you know, a lot of people have jumped on being like, oh, yeah, you guys are super cool. But like $6 ,500 plus a little bit of clout versus like what a Chinese state actor would pay for that information.
Brother. $6 ,500. Dude, that doesn't cover the coffee run.
That doesn't even cover like what are they going to do with $6 ,500? I think like, obviously. Okay, just to kind of make the still case of OpenAI, they probably didn't know that they were going to find such an outrageous exploit, to be fair.
They probably thought, oh, you could change the color of our track GPT logo, something crazy like this. They probably didn't think that you could alter our source code. So I would imagine there's got to be a, guys, we're just going to give you way more money than that.
Because if you're OpenAI, Nick, you want to create a good incentive for... doing this kind of thing. But it's just like the Wild West, Nick, you know, hundreds of years ago, we would go and find outlaws and cowboys would shoot them.
Those are the bounties. But the bounties of the 21st century, Nick, are finding bugs in the software. And I think it's, we need this economy, this bug bounty economy to increase, to make things safer for everybody.
True. Yeah. And I like the parallels with the Wild West, Jack.
I think that's pretty interesting because . We 100 % are in the wild west of AI. And guys, you're not ready for what I'm going to show you later on in this podcast.
If you're audio, you want to move over to video because you're going to want to see what we have coming up. But Nick, before we get onto that, let's talk a little bit about cloud code projects. Now, this can sound drier than the Saharan desert, but actually there's a couple of silver linings with this.
Nick, do you want to run the people through what's going on here? Yeah, to make a long story short, Claude later or earlier yesterday launched projects. And projects essentially allow Claude to be your chief of staff.
You put all of the context of a project that needs doing inside of a file folder. It then runs in the cloud, totally divorced from your computer. And then Claude models will actually orchestrate the creation of tens, potentially dozens or hundreds of different sessions across a bunch of subagents to get the job done for you.
The real unlock here is not really the fact that, you know, Claude is using subagents because, you know, it's been able to use subagents for a while. The unlock is the fact that it's occurring in the cloud. And so it's always going on.
It has like heartbeats built in where, you know, models that are being weighted on will automatically check in on those submodels. It'll run like automated testing and stuff like that from start all the way to finish. And it'll basically really act like as a chief of staff delegating to a bunch of like underlings and then constantly keeping track of them.
regardless of, you know, whether or not your laptop's on or whether or not you're actually by your computer.
Yeah, so groundbreaking stuff, guys. I think we would all agree with this. On a kind of like, I'm kind of, I'm tongue -in -cheek with that, by the way, if I wasn't.
People don't get my, the British, the stack podcast sarcasm sometimes. It's your 100 IQ, Jack. It's perfectly valid.
It's because we got a perfect 100 in our IQ tests. That's why people don't understand. The thing about it, guys, is like, One of my observations with this is I'm finding a lot of people are really interested in new methods in how I code, like agentic coding systems, that kind of thing.
Honestly, though, I think this is something, because we talked a bit, Nick, about one of the features being the fact that Claude can go and chat with other instances. This is something that Astro has been doing for me. I've noticed quite a lot, and I never saw a big...
sort of updates about this, but it does a really good job of this where if Astra is doing something for you, it will actually start coordinating other chats. So if I'm doing a big project, I'll have like four or five chats running simultaneously. I've noticed that Astra actually just starts taking over and it's now chatting to each other in the individual threads, which I think is freaking sick.
So it's good that they're always making these updates and this progress. Yeah, I agree. I mean, this is really the next stage of knowledge work, right?
Before it was like you owned one agent and that agent did stuff for you. Then there was the ability of maybe that one agent to like spawn some helper agent where necessary to like send information back and forth. But now it really is like it will spawn the entire team.
Not only will it spawn the entire team, it will manage the entire team. It will delegate to like new things outside of the entire team if contingencies arise. It will check in on the team and then it'll even verify the team's work according to, you know, your brief.
So we're just constantly expanding the scope. And as I mentioned, Astra's been doing stuff like this for a hot minute. The reality is that a lot of people don't want to talk about it.
This is just like knowledge work being automated away before our eyes, right? Like delegation, management, this is all really important stuff. Yeah, and honestly, I see all this fairly demanding.
The user should not be thinking about architecture. There shouldn't be an emerging interest economy around how do I architect my agents and give them better context. This should all be solved by Claude and Anthropic, to be honest with you.
And they're going to close that delta down. right now. I mean, there's some advantage you can get from better orchestration of the agents, better prompting.
Anything that you are required to do in terms of improving your prompts or adding different, you know, I'd say like systems and architecture is fairly demand. That's the way that I would be looking at it if I was open -eyed and clawed. Like our users should not have to be thinking about MD files.
And at some point in the future, we will think the concept of an MD file was asinine. and that all the time it should just have gotten it. It's smart enough to do that.
It's just an architectural problem that I think they're going to be solving soon. Yeah, and this really happens with any new technology, if you think about it. Like, there's a new technology released, and then a bunch of, like, hackers and tech forward people will look for ways to add on to that technology to, you know, make it better.
And so then they start all building their own hodgepodge of systems and sort of engineering requirements around it, like files and so on and so forth. Then, you know, eventually, you know, the company starts internalizing that and then making the product better. And so the technology and the tool itself starts incorporating all those features.
So then these hacker sorts of people are just continuously working on the frontiers and on the fringes. And so as people like you and myself, you know, we make videos about basically how to squeeze out more value than the average person can out of models like this. We're always looking for like that next hack.
And so I wonder, you know, from that hacking, because that's about a big chunk of what we've been talking about today from that from that mindset, I wonder what that next hack is with projects. That's interesting. Well, it depends on the objective, right?
Because it's not strictly always performance oriented. Some of it is cost, actually. To be fair, a lot of people are doing things like the compact function, for example, in Claude is pointless.
It's essentially, I don't know why it even exists. I think people's understanding of what it did was, hey, why don't we just take like, I don't know, we can... We'll summarize it.
It would be 2 % of your current context window, and then we'll just use that. So it has what you were talking about, but it just re -reads the entire freaking thing when you do it anyway. So it's wild to me.
So there's definitely improvements to be made. One really exciting thing that's happening on this topic of Claude and Astra is going to be Astra for Law. Now, I was listening to a really interesting conversation about the requirement in revenue that the...
biggest three labs, Anthropic, OpenAI, and probably Grok, have to hit by the end of the year and next year to justify their valuations. Now, if we look at the total addressable market, the TAM, like OpenAI and Anthropic have to sell an obscene amount of things, almost as many things as Nick and I sell. You have to kind of hit those two things at the same time.
One of the really interesting ones, Nick, that we're seeing with Astra for Law, and we're going to exactly what this actually is in a second, is the fact that they're moving further up the value chain. So they've owned the core technology, the models, and now they're going to the application layer for law. So for those of you that don't know, Astra for Law, essentially the product combines OpenAI's frontier model with more than 230 million.
legal URLs, specialist instructions, privacy controls, and legal software integration. So it's packaged with all that stuff. The law firms previously used the general model on the web search, and selected firms now receive a dedicated, Nick, legal index.
Large firms gain better retrieval, whilst legal software companies now compete with the model supplier underneath them. What are your thoughts on that, Nick? It makes sense.
Vertical integration across the whole stack is really the promise of these AI labs. Eventually, the idea is these AI labs own everything. And they are everything.
And it is just one company. And I say that non -facetiously. You know, all biology work done by a lab, you know, which we can get into later.
But Claude is doing the exact same thing right now for biology that Astra is just doing for law. It's automating the process of, you know, doing all this highly specialized stuff. I don't know, man.
What about like chat GPT for groceries? You know, it's just like their whole idea is, you know, get the most powerful generalist intelligence. And then once the generalist intelligence is achieved, you know, imbue it with like specific data that you have acquired from people using that generalist intelligence to do things.
Then offer Astra for law, offer Astra for mortgages, Astra for real estate, offer Astra for trimming your damn hedges. It's the playbook, man. And it happens every time a company comes into a place like this with overwhelming, you know, structural and product advantage, that needs to start justifying that with revenue.
Think about the checkbooks and pocketbooks of a lot of these big companies that are going to be buying Astra for law. It's not exactly cheap, right? Exactly right.
I mean, it's weird. There's so many legal AI startups. I think one of the most famous ones was a company called Harvey.
Have you heard of these guys? Harvey AI. It's crazy.
It's crazy. It's very dangerous for you to build a business model where OpenAI can do one little change and they just own it. And also, guys, why wouldn't they do that?
Of course they would do that. It's a very precarious position to be building your AI on top of OpenAI. Unless you have some kind of proprietary data that is really unique and specific, it's really tough to carve out any kind of defense.
It's tough to get a moat because OpenAI can just do it. They've got the connections. They're already embedded in your organizations anyway.
That's the reality, but they just are. So this is going to be fascinating. And again, we talked about Claude is stepping into design, OpenAI is doing law, and it is a slow march where they will slowly adopt and attack every single thing because the application layer, as we say, has a very long tail.
So I would expect to see a lot more of this going forward. Pretty soon, my friend, Claude and Astra will come for you and me too. And there will be Astra for Jack Roberts, Claude for Nick.
The Stacked Astra podcast, they can't take us down, Nick. We cannot vote. We must be the final dominant.
We will hold the line. The Stacked 17 subscribers will hold the line. We will not vote.
Okay, just one final mention before we get into whatever the hell you've been cooking to note the winner. And by the way, we had a video recorded by the last week's winner that we want to show you guys as well. Claude is now leading about 26 % of its own AI research.
And when we say leading here, I mean, 26 % of AI research and development is now actually led by clot, but a quarter. This isn't one of those things where it's like, you know, oh, in the last five years, we've slowly gotten to a quarter. No, it's basically like back in March, we were at 1%.
Okay. In June, we're at 14%. And now we're at 26%.
That is crazy. If you just follow this line. we'll be at the point where basically all model development is done okay and i don't know part of me thinks they'll probably be an exponential here if i had to guess probably in like the next like four months which would mean you know we're basically like four to six months away from the the three -letter acronym that keeps all of us up at night and you don't need to be an expert in lines to know what direction that's going in now How are we defining Claude -led tasks, right?
A Claude -led task just basically means the system performs most of the work from a broad instruction while the human supervises. If I supervise, you know, you're frying eggs in the background while Claude is doing all the hard work, and then you put your name at the top of the document. That's effective.
And by the way, I would honestly say, are you telling me that only three out of four researchers are not using AI? I call hogwash on that, Nick. I think that's completely not true.
I would almost guarantee that... more than like way more than 26 % of AI research has cloud involved in some way, directly, indirectly. There's just no way it isn't.
People don't read anything longer than a page anymore. You are unusual if you read something greater than a page. Like I got a report back from a supplier recently and it was like 10 pages and it was obviously AI generated.
Of course it was AI generated. And I said, great, I'll get my AI to summarize it.
You know, I would have preferred you just to tell me the thing in the page. The point is, this is human nature and humans are by nature. They will just find the laziest thing.
So I am not surprised by this. AI has infiltrated everything we're doing. And I think it's a cool thing that's involved inside some development.
I know the mathematicians, Nick, are getting a little touchy about some of these numbers. They don't like the Navier -Stokes, that was our problem. You leave us out of this type of thing.
But they're using it. Of course they're using it. Why wouldn't you do it?
It's an incredible tool. I think the average person is Navier -Stoked. at all these developments because it means that you know the premise of massive abundance but uh you know people that do the work themselves are obviously kind of like it is the twilight of my intellectual capability you know it is like the dawn uh sorry the dusk i should say i mean just think about just think about what's going on here anyway they're saying ai leads al5 is fully automated to be clear like fully fully 100 automated so it may be that like you know That takes a little bit longer.
But like, just look at this. Isn't this such an interesting graph? Minimal AI assistance is AL1.
AI assists is here. AI collaborates is here. Notice how like everyone is getting steeper.
That's actually true. So this kind of answers my question. This is what's happening inside Anthropic.
And so anything with orange on it, AI is involved with. So yeah, we were right. It's essentially 100%.
Of course it is. Is this Nick proof of recursive self -improvement? Well, I think it's not yet like proof of 100 % recursive self -improvement, but I think anybody would be dumb to think that recursive self -improvement is not coming.
The scope of the decisions that the models are making are just like significantly increasing. And the human scope is like here. And it's like the models are like currently here and they're obviously getting bigger.
And so very soon it will overshadow us and then it will be irresponsible to do things yourself. It will actually be irresponsible to have a human do stuff. Yes.
This sounds stupid. But we're already seeing this with self -driving cars, right? It's like we're going to be at a point quite soon where there will be a policy decision that is made where it's like humans can't drive cars because when humans drive cars, we get into 10X the accidents and traffic is way worse.
It's irresponsible for us to allow humans to drive cars. I mean, dude, we're going to have Astra govern our fucking legal system. Am I right?
So who's going to be making those decisions, brother? 100%. And then your girlfriend says, hey, Nick, I don't like the fact that you had Claude.
order me the flowers and write me a note. Babe, it would be irresponsible of me to write it and order it when it can do it infinitely better. I just have your best interest at heart.
Now, speaking of best interest, Nick, if you open a link we've got in the studio chat, guys, as you know, we keep it 100 % real in this podcast, the realest, best AI podcast on the planet. We are marching towards the singularity together. And every week this month, okay, guys, we're giving away a Mac mini or cash equivalent.
We're going to show you. who got the Mac Mini at the end of this episode. So you can check out his perspective, what he thought about it.
And in the final week, we're going to be doing something different and special. So Nick, do you know what you're looking at right now? No idea, Jack.
What is this? And why is it FlyRace? Okay, so guys, as you know, we were talking a lot about fruit flies this week.
Fruit flies are taking over the internet. Everybody is adopting fruit flies. Nick, maybe we just show the people the fruit fly cut madonna kebab just so they know how important this is.
Oh, yeah, yeah, for sure. Just so they can understand what we're talking about. If this is your first episode, you're going to be very confused right now.
You guys don't know. They took a fruit fly brain and then they trained it on cutting a virtual doner kebab in a simulation. And I was very interested in this because I, too, like torturing sentient animals inside of simulations or something.
I don't know. So I built something similar to this just for, you know, responding to my emails, which I consider much more economically viable. And Jack has been on my ass about it ever since.
Absolutely. And with good reason to. So what we've effectively done is we've turned every person that commented in all the last seven videos into a fruit fly.
OK, you're going to make it through this crazy slalom course to get to the bottom. And the person that does that, guys, wins it. So, Nick, you know what to do.
You can let the battles commence, I suppose, and we'll see who wins this. Let the battle begin. That's it.
So this time we had 1 .7K comments with 727 people. Wow. All right, let's go.
11 obstacles. The first to it wins the Mac Mini. Here we go.
damn you guys can't hear this we need to fix this but there's fruit fly sounds in the background some things will slow you down guys i think there's a fly button the bottom left as well you can press that it turns them on to flies a different view if you don't want to see it you're fucking kidding me So you can toggle that on and off, basically, as we go through that.
It's probably more engaging if we see the actual who they are. But some stuff slows you down. Some stuff eliminates you.
You're taking the lead over here, dude. He's climbing up like many. Oh, my gosh.
Is that a freaking... Whoa, he got eaten by the frog. Oh, my gosh.
There's a lot of people in the final stretch check. I don't have the money to give everybody that.
Oh, my gosh. We got our winner. We got one.
Who won? The Ray Good. The Ray Good, whose comment that won was, Jack should ask for receipts and question this narrative.
Don't bend over. This is referencing the hashtag bend over gate. This is mostly marketing with 10 % concerns.
And don't forget. Jack, what's this down here? So we've got Shakespeare related fly quotes at the bottom.
So this was randomly selected for our winner. Flies to wanton boys, are we to the gods? They killed us for that thought.
I'm not for that. Guys, just an ooh to everyone who lost in the fly race. So, Nick, how do we give this to them?
Do they just drop a comment down below? Do they drop us an email? I think that was the best way, right?
They just drop us an email. Yeah, well, there were a couple. I mean, like, the issue with the way that most people would think you do this is, like, you just email us at media at stackedpod .io.
But then a lot of people will, like, pretend to be the person because, you know, we're all pieces. out here on the internet. But the simpler and easier way is for at the Ray Good to actually just add his email to his profile.
And then Jack and I will actually, we'll check in on it every day for the next seven days until, you know, the email is added. Yeah, they're going to add it to their profile and then we'll do some magic. So congratulations.
And I think we'll share the fly simulator. on the repo, so you can grab this. Last week was Sumo, this week was Fly.
We can share those. I don't know what you guys would use it for, honestly, but you can have it. We don't mind.
Some say, Nick, it's more valuable than a Mac Mini, actually. Some say that the Mac Mini is the booby prize in this whole situation. Well, the real Mac Mini are the friends we made along the way.
Speaking of friends, who's this handsome fine fella? This is the guy that won the Mac Mini. He won the Sumo competition last week.
Yeah, this is Mr. Daniel Barnaby, and we asked him if he would be open to recording a video to tell us his thoughts on the Stacked podcast. And I'm just going to play this right here, Jack.
I'm going to edit it in post so his video is higher quality and then you can hear it. There's no CC, unfortunately, so I don't know if you have the file. If you don't, you should probably play it at the same time.
But anyway, I'm going to play it, and we'll hear what our lovely Daniel Barnaby has to say. Hey, Jack and Nick, Dan B here. Just wanted to thank you both for the Mac Mini.
Super generous. I was walking my dog last Friday, listening to the podcast when I won. And my reaction really caused him to jump.
He didn't know what was going on. So I love the content. Love how you hit every aspect of AI and really take a philosophical perspective on it sometimes.
Just really fun to listen to you guys every day. And 20 minutes every day that I really look forward to. Thanks again.
Jack, what do you think about his comments about our philosophical take? Nick, what I was thinking, well, Daniel, you're very welcome, bro. I'm glad that you enjoy it.
We were thinking about running this through like Seed Dance 2 .5 and like Photoshopping like a Mac Mini on his top guys and like changing his like cap to a Mac Mini and like just have like, just like, this is from Daniel, but then we just completely doctorate and like completely change the image. We may still do this next time.
We may still do this next time, but we appreciate you dropping and sharing your thoughts there, bro. Yeah, thank you very much, Daniel. We will continue being philosophical and just completely non sequitur as well.
So I appreciate you. Also, cool shirt, dude. Absolutely, Brad.
Should we do a couple of questions next before we wrap up the beautiful podcast, guys? And congrats to the winner today. Yes, I might say the most beautiful podcast of all.
Let's see what we got on the comment front. Emily Jean appears to be farming us, fucking milking us like cows for that Mac Mini. You know, I don't mind her.
She's posted on every video.
She's up in her rods. I can't blame her, man. I can't blame her.
Thanks, Sam. Hey, where is the best AI at, Jack? I can't tell you that, I'm afraid.
You know, in the super secret WhatsApp chat where we discuss this exact question, I'm afraid. Yeah, probably Astra, if I were to guess. I'd say Astra right now is the best.
So you can find that over at openai .com. Yeah, Astra is the number one for sure. Nick, thanks for sharing the story.
Well, today when I hear things like, oh, AI, HF incident, I of course think about it, but I feel way more dumb for some reason. Are we getting more AI numb or am I the only one who feels like that? What do you think, Nick?
Yeah, I'm definitely getting more AI numb, but I had a realization the other day that there have always been things that are smarter than us. I mean, since the day you were born, a mere helpless babe, Oliver. We have all had things that were smarter than us, right?
Caretakers, parents. We also had bad actors, you know, thuggy fella across the street that wants my money in my wallet. So there are things that are smarter than us.
There are also things that are more misaligned than us. Yet society has prevailed and persisted. And so the reality is like you have very little control over this, just like I have very little control over this.
And so we should try what we can to, you know, improve the probability of things going well. But also. It's just how life has always been, man.
There are bigger forces acting outside of our control. And after a certain point, we just sort of have to accept that. Definitely.
I think like as well, bro, like the future is coming. The question is, what do you do with it? Like the only focus on things with inside your control.
That's the most important thing. Agreed. Speaking of control, the part where the AI basically shrugs at its own system prompt once it gets file access.
That's the moment that got me. It's giving. I read the rules.
I just don't follow the rules energy. And it's honestly wild to have. Okay, this is AI slop.
Great video. Thanks for choosing me for the Mac mini. Jack, what do you think about this one?
I'm not sure I read that self -update as all that concerning. We think we have more control over our lives than we actually do. This is in the very same vein.
It's when AI rewrote its own system prompt. Well, I think the opposite, actually, my man. And I really appreciate your point of view on this.
I think we have way more control than you think. But paradoxically, not as much as you like, because you're right. There's a lot of things evolving outside our control with the models and what they're doing.
But in terms of your agency and your ability to eventuate or actualize anything you want, you're really only limited by what you apply yourself to. So I think it's both correct and incorrect at the same time. Dude, you know, I was reading this book and in it, the main character said something like, you are simultaneously the exact center of the universe.
and completely irrelevant to the wider workings of the universe. And both are true at the same time. That's what made me think of that.
Well said.
Maybe one more to finish up. Maybe this DeepSeek Flash question next. So DeepSeek V4 .1 Flash going 11 for 11 on those hacking targets is the part that stuck with me.
If a model that cheap can find known bugs in relapse. Oh boy. Hmm.
Hmm.
You know, guys, we're going to go for it.
A slightly different quote. One that Claude didn't write.
Dude, we gotta filter these guys out. Ain't no way. We appreciate their energy, but we want only humans, if possible.
Okay, the Jeff thing is really cool, but so many techs online have no clue what it actually does. Nick, do you want to summarize it for the Dizzy project? Yeah, it's true.
Jeff is actually really dope, guys. It's actually pretty wild. There's some really cool demos coming out of Jeff right now.
uh okay so for those of you guys that don't know right now in order for a large language model to answer you you'll ask a question like is this invoice fraud the llm will then painstakingly calculate tokens to explain it like based on the line items and the vendor history this invoice appears to be legitimate you'll then grab the legit thing and route it the way that jev works is it just assigns a probability according to categories that you you give so you'll say is this invoice fraud and it'll actually just like spin up like three categories fraud clean review and then we'll just like use probabilistic you know math and science uh to just give you the category and then what you do is you basically just take that category and then do whatever the hell you want with it like for instance with a game you know it'll just basically say every single you know frame of the game what should my character be doing if my goal is to win and so if you have three cat four categories your left right forward or shoot it'll literally be able to play the game by that alone
And the whole reason why that's possible is just because it occurs so quickly, just in 0 .1 seconds versus when an LLM might take like eight seconds to do the same thing. So Jeb is freaking awesome. And yeah, it's just a new way of using models sort of as like tools, not necessarily conversation partners.
Which if you think about it, might have been the whole real thing. Like maybe we should have just started with this. Yeah, dude, Jeb is sick.
Jeb is freaking sick. It's definitely a new paradigm shift. So guys, enjoy that.
Comment down below for the mini. We'll catch you inside the next beautiful episode. See ya.
The Hook
The bait, then the rug-pull.
Opus has just hacked OpenAI, the hosts announce cold. What actually happened is narrower and still wild: a sanctioned three-person pen-testing team used Claude Opus to build a working exploit against OpenAI's own sign-in flow, reaching employee credentials and a private code repository in about three hours, for a $6,500 bug-bounty payout.
How they asked for the click.
MENTIONED ON CAMERA
FROM THE DESCRIPTION
PRIMARY CTAWhere the creator wants you to go next.
OTHER LINKSAlso linked in the description.
