GPT-5.6 is here! And none of us can use it.

And even the AI writes this with me being the one to say the first words Do you want me to be? I think we're going to see a try I hate scripts, I don't know how to read words Okay, I can believe that statement I'll do my best to read words, I'm not very good at it But we'll see if I can pull it off Leave all of this in phase Okay, here we go Welcome back to Fuck Nope, nope, this is your problem Welcome back to Nerd Snipe, I'm Theo, this is Ben He can't talk good, I kind of can And we're going to break down all of the crazy things that happened in AI this week Do we get the music in the background that time too? I hope not Before we say sponsors, we should say what we're actually talking about today We love our sponsors, but we need to prioritize our audience's attention We have quite a bit of fun news to cover Obviously the big 5.6 release that isn't an actual release And a lot of thoughts there, believe me, we have a lot to say But there's also some other fun things that happened this week Like the virality of poisoning your repos in order to keep agents from contributing Which is a fun thing I've played with in the past and actually quite enjoy There was the price hike on all Apple computers In fact, I'm pretty sure the laptops on this desk combined are worth like $3,000 more than they were on the last episode So like that's cool for, I don't know, nobody honestly Because we're not going to sell them But we can't forget this week's biggest news by far The newest T3 code release and its partnership with XAI and the Grok CLI team That's going to be a fun one I'm sure you guys are really on the edges of your seats for it And of course, thank you to this week's sponsor Composio Let's get into it Why not start with the biggest news of the week And talk about the Grok code partnership with T3 codes I think it's pretty damn cool Yeah, I honestly, I didn't even realize this one was happening until it was posted But good work How did this one actually get done? I saw Julius was talking about it Like a lot of, they did a lot of the work to actually get this merged and added in They seem to be doing a really good job with that kind of thing It's a better CLI than I expected it to be, honestly You're not the only one who was surprised when this dropped I was too For those who don't know T3 code is a GUI that I built with Julius And a couple other really helpful open source contributors To try and make a better wrapper and application For managing your agents across your machine and other machines You might have networked to it The goal is to take the things that I loved about the Codex app And make a good, reliable, open source alternative That works with everything And I've been blown away with the growth We actually just hit 100,000 users last week I'm not here to just plug my shit, though Because I actually think this strategic partnership is interesting And not necessarily on our side The thing that's interesting about this one to me Is the fact that it was entirely done by outreach on the XAI side Milo's reached out to us directly and said Hey, we really love what you guys are doing with T3 code We want to make sure that our CLI can integrate seamlessly And really get this going We're even down to do a more formal partnership If you guys are interested How can we get this ball rolling? And I just put them in touch with Julius I don't know if they already had ACP support or not But that was the big blocker The agent-client protocol For a way for your agents and agent harnesses To connect to another layer over an API It was super helpful They got that all working We're using a pretty generic setup for ACP stuff in T3 code So that I was able to plug in without too much work They helped a bunch They did the big announcement on Friday last week I thought it was really cool I didn't expect it to happen Much less go that well And the biggest surprise to me Is that I now have a really nice way to use Composer 2.5 Which I did not realize how much I needed Love the cursor guys and all But the cursor implementation we have in T3 code Is based on their CLI Which had ACP support, kind of But was missing a bunch of stuff And we nudge them enough to get what we needed But it still feels half-baked Not because our connection to it is But because the CLI and its ACP implementation On the cursor side Is just clearly not their focus They have the STK now Which we could switch to But now I have what I wanted Which is the access to a really nice fast model That I can use inside of T3 code For random one-off shit Yeah, rescuing Composer from cursor Is a beautiful thing I really like that model And I honestly like Cursor's harness a lot Glass has improved Like, I don't know if you've tried Cursor Glass It's like their T3 code thing Their desktop agent's view Where you just see the threads And you have the sidebar And the one that everyone's building right now Improved a ton Performance is much better It's much more usable And they're clearly building A lot of crazy stuff into it Which is quite nice But you still can't use the model anywhere else And I am really The best part of Grok build Is the fact that you can now use Composer And something outside of cursor And I will say I'm actually really impressed With the team over on the XAI side I had my skepticism You guys have heard me talk A lot of crap on Grok models in the past I'm still not using the Grok models there Let's be clear about that I'm sure they will get better Now that they have cursor As like a company they own And all of cursor's data I'm just using Composer through it But the team that's building the CLI The way they did the outreach The way they're positioning themselves In the market Is actually really impressive You would imagine that Like the cursor people Were the ones who did this None of them were involved at all This was purely the XAI Grok CLI team That led this effort And they really wanted to make sure this happened They pushed to like Get my attention And work with us In the way that we wanted to work It was one of the better collabs I've done Having worked with a lot of different companies In the labs I'm impressed And that's the attitude They're going to have As they come into the market More aggressively And they're going to partner With open source projects like that And really help support us And make sure our integration works well And they've been so responsive too When I send them like one-off issues I saw people having with like Windows I saw a tweet someone had Complaining about an issue I sent it to the team They had replied on Twitter Within five minutes And had it fixed within an hour And like cut a release for that user It was really nice to see I always will respect teams That are like really trying to Make sure their thing is good And works well For the real use cases I've seen and heard Nothing but good things About the XAI team I've liked every interaction I've had with them I wish that their models were better Dozens of times I'm optimistic I'm glad that they did this Hopefully a good beginning Of what could be A useful third lab In the future But time will tell I did also bully The Codex CLI team A bit last week And was surprised How quick they fixed things There were some weird issues Where Linux boxes Would try to IPv6 Connect to the WebSocket server I got them to patch that It was an upstream dependency issue They handled And then I was complaining About the MCP connection thing That happens whenever you open the CLI And it breaks like Slash model and stuff They fixed all that too So credit where it's due There are two teams That are responsive like this But it's crazy that like The only company That comes close to XAI's level Of like caring about product Is OpenAI And everybody else Is far behind it Well and even on the OpenAI side Literally two days ago They introduced a Codex desktop app update That showed your email On the bottom left I saw like three tweets Complaining about it I updated it this morning And it was gone So like they instantly Jumped on that And fixed it Which is great Oh I'm on that version of this I need to update now Yeah update now And it'll be gone God there is some slop In that code base For sure Oh yeah I wonder if they got cut off Of 5-6 as well And that's why the quality Control has gone down Yeah No comment You know what's the opposite Of quality going down What is Wake up I am so dead right now Why am I dead Maybe fucking slap you No Maybe It'd be good content Sure you can hit my back As hard as you want to That's not your face That's no fun Well the face will like Screw up what I look like On camera The back is just like Yeah But then like I don't know My head will go all fuzzy Listeners let us know In the comment section If Ben should have Let me slap him in the face Well you know They're gonna say yes Of course they are It's a loaded question So now you want to talk About the horrific Apple price hikes Didn't think we would See this one for a while I think everyone knew This was coming It was inevitable But I really thought It would be like next year Like all the Apple supply chain stuff The godly systems they had I thought they had enough That they would probably Just try and Keep things the way they are Until next year My guess Is that they are doing this Ahead of time Before the new releases This fall So that it's not A huge price hike then And they're Warming up people To it beforehand One of these three devices Did not get a price change And I want you to guess Which of the three Is the device that didn't Get the price change Okay Vision Pro Apple TV Based on your iPhone Vision Pro Based to your iPhone Really? Yep iPhone line is unaffected Every other line of product Apple has is Oh I should have put HomePod in there Because the HomePod Got hit too That one doesn't surprise me That one surprises The fuck out of me Why the hell is the HomePod getting a price increase Yeah you gotta put RAM in there for something It's got storage It's already like A 90% margin product It's a mediocre speaker With like software That they spent a lot Of money making on it But like the R&D investment Was done when they Built the first one Yeah that's true It's not expensive hardware It's a garbage chip It has no real RAM In it at all And they're selling For like 400 bucks That was like 450 I think Wait we'll check the numbers But It's what? Hold up No it's 350 now But still Jesus Oh yeah 299 to 350 A 50% price increase In the fucking HomePod You kidding? Yeah And the Apple TV Is now $200 Which makes it go from One of my least favorite Apple devices To one of my least favorite Apple devices I do not like Apple TV I use a Windows computer With a wireless keyboard And trackpad on my desktop And on my TV It's the only thing I've found That's even somewhat reasonable Yes I choose Windows Over Apple TV That's how much I hate it There's one interesting story That didn't make it into the notes Because it's rumored And hasn't been super fully confirmed We've talked about this a bit In the past It's the negotiations Between Apple and Samsung Traditionally Apple's known as like The negotiator They push really hard To get the best possible deal From all their manufacturers Across their whole pipeline For making devices They secure their deals With things like Their RAM manufacturers Really early In the planning process For any new devices That they are building Those contracts Used to be two plus years From what I'm hearing They're short of six months now Due to how much prices Are changing And the companies Making these things Not wanting to lock in a price And not be able to sell inventory At the new rates Later in the year As such Negotiate And in that conversation They had planned to aim For a 40 to 50 percent Bump in the price But they know Apple Are really rough With negotiations So they went in With a 2x Expecting Apple To negotiate them Down to a 50 percent bump Instead of a 100 percent bump Apple immediately said yes In the conversation Didn't even negotiate They were like Yep cool We'll do it How much can we buy Dude what is going on That might just be Right now Because getting any allocation Is just so hard As a consumer hardware manufacturer Yeah they're just not the priority I mean I don't think We have specifically Talked about this But the calculus For these companies Is just enterprises Will pay obscene amounts Of money For massive quantity And be super easy And chill to work with Instead of consumer Which is like Lower margin One at a time Dealing with support Dealing with refunds Dealing with packaging it Shipping it out Putting it in retailers All of these extra steps Just to ship A lower margin Products that costs less Why would they bother With that When they can just Go to open AI And be like Hey you want to buy 50 million dollars Worth of GPUs In one little box And then we just Ship them to you And you go deal with it They're like Yeah sure cool And that's it It's so much nicer For these companies That's why they're doing it I did hear a rumor That a lot of the new Like fabrication That's being built right now As well as some of the stuff That was already Republished for HBM Is being moved back To DDR5 Because companies Want to take advantage Of the massive Hike in prices Because most of the RAM kits being sold now Weren't manufactured This year It's just old Manufacturing that's Still packaged And as prices have Increased People have been Buying less and less of It does seem like The major memory companies Are realizing Oh there's actually An opportunity If we can sell things Like directly to consumer At these prices Because most of The inventory right now Like if you spend Thirteen hundred dollars On a 128 gig RAM kit Right now That money isn't Going to Crucial Or to Corsair Or to whoever You're buying it from The retailer that has it Bought it at MSRP At some point From them So the memory companies Aren't actually benefiting From the price hike To consumer at all Right now Which is why They're realizing Oh if we make more RAM We can deal with that And like get something From it So you might see A weird back and forth Going hopefully Fingers crossed Because that will Actually cause some Pressure in the market And get us Slightly better prices Again Hopefully But I mean It does not seem To be slowing down Like I was I was looking at Microcenter prices Cause I was just curious And when I was looking Three days ago Before all of the crazy news We'll get into Later dropped The like DJX Sparks Were four grand At Microcenter And then I looked Again today And they are now Up to 4500 So there was a $500 increase on those All the GPUs Went up in price And this is like The upteenth time This has happened Already this year It's not slowing down I've been telling people And obviously Not financial advice I'm just a YouTuber And I guess Podcaster I haven't It's the right code If there's a piece Of hardware That you've been Eyeing Wanting to buy But you're just Waiting for prices To drop Like if you're On an M1 MacBook And you are tired Of it And you want more RAM And a bit more storage And maybe the Nanotexture screen And you've just been Waiting for prices To go down It's not happening At least not for Like a year To multiple years Stop waiting Because it's gonna Get worse Obviously don't Put yourself In a bad Financial position By buying a Thing you shouldn't But if you were Eyeing that new MacBook And were on the Line as you were Hoping prices Would drop again They're not Just get it You'll be happy When another one Of these hikes Happens Because this will Probably not be The last one Yeah We are not out of it Yet Everything is just Getting so much Worse Like I saw a Screenshot on the Apple website Of before The 128 gigs Of RAM option Was plus One grand Now it's plus Two grand It's doubling Yeah this is The 8 terabyte 128 gigabyte RAM M5 max 14 inch that I'm on it Right now This was 7800 I think Or actually I think it was Like 7200 or so When I got it It's 10 grand Now Yep It's almost $3,000 more Than it was When I got it Yep And it sucks Because this fall According to like The Apple leaks Which are usually Pretty credible We're gonna be Getting a lot Of very cool Stuff that I want Like the new M6 Mac Pros I'm gonna want One of those The new M5 Ultra Studios I'm gonna definitely Want one of those For local AI stuff According to Current rumors There might end up Being like a 700 plus gig Of RAM version Of this Which god only Knows how expensive That's gonna be And I don't even Know if that's Actually going to Happen That seems a Little insane To put that Much memory In one place It'll just Run out so Fast Yeah it's Just it's Insane I want Nothing to do With this And the fact That Apple Had that brief Moment where They were actually The value option For the first Time ever In their history Was crazy Well if we're Gonna cover the Cost of our New macbooks That we're Almost certainly Going to go Buy even Though we're Buying used We still need Some money So we're Gonna do a Quick break For today's Sponsor There are a Lot of Companies out There that Are trying To build Agents Into their Products But you Hooks in Notion GitHub And Slack They handle The off For you They also Handle Triggers For you So that if Something happens On the Data source Side That can send The information Back over to The agent So it's Always up to Date and In the Loop And it's Fully Model and Framework Agnostic You can use This with Cloud Code Codex Pi AI SDK Whatever you Want to use You can use Composio With it I love my Hermes agent And OpenClaw Is amazing Too Composio Is the easiest Way to add Over a Thousand Different Tools into It Over just One NCP Endpoint Or if You prefer They have A really Great CLI As well Your agents Will get So much More powerful Once you Hook them Up to All your External Tools At Nerdsnipe. Link Slash Composio We're in The bad Times Fucking hell Dude Want to Do But we Just Can't Yeah I like Even gave It this Name In a Different Era I know We were Supposed to Be just Talking about The tech We're nerdy About And trying To get The other Person Into the Thing And now We're Talking about Poisoning Repos With agents And hidden Files And all Of this Models Being restricted Stuff How our Computers Cost too Much Money Now Dude The Good Old Days The Linux Episode Is Going To be Great It's So We Decided To delay The Linux Episode We Can Talk About Five Six Today You're Welcome We're In Pain Yeah Man Dude I Remember The Good Old Days Of Two Months Ago How Little We Knew What A Blissful Time It Was We Were So Young I Don't If I Could Go Back To A 30 Do We Do This This One Is Fun I've Seen Some Very Creative Ways Of Doing This Yeah So Seb Posted A Screenshot From Docu Soros that has a screenshot of the agent's MD and an AIPR notice.txt file that says, I am a sad, dumb little AI driver with no real skills, deliberately to just screw with any AIs that are scraping and trying to make PRs onto the repo to try and get them to stop. Especially since OpenClaw came out, this has been such a huge issue for open source maintainers where whenever they run into an issue, oftentimes they will just go and make a PR in the background and the user doesn't even know. They just swarm all of these poor repos with constant PRs and issues and nonsense. That's just slop. Like the biggest one that I heard of was Pi. Since it was the thing that OpenClaw was built on top of, as OpenClaw took off in popularity, Pi started to get noticed. And whenever someone would have an issue with OpenClaw, they would sometimes file a PR to OpenClaw, but they would also sometimes file a PR into Pi. And Pi is a very carefully maintained and crafted piece of software. And they were just getting demolished by these PRs. Sucks. Carefully crafted slop, but it's a beautiful project. It's a beautiful project that is a slop generator. Like it is. I mean, it's an agent harness, but. So this went particularly viral because Mitchell joined in. If you're not familiar with Mitchell, he's the creator of Terraform, a lot of HashiCorp, and most importantly, as of recent, Ghosty, which is a beloved terminal that we're all running our agents through. Even if you're using something like CMux, which is using Ghosty under the hood. LibGhosty is really, really good. All of that is Mitchell's like recent thing, but it's also big in open source and has a lot of people contributing a lot of slop that he does his best to filter out. He built a system called Vouch, which is a really easy to set up GitHub action that will tag maintainers once they've had code merge as trusted, or you can manually flag them as well. And that makes it way easier to quickly filter out people who aren't really working on the project already or don't have experience with it before. Just a simple system to get a better idea of like, who is contributing to this? How can I see them? And not all of the other slop. He's going out of his way to do more and more systems that will like auto close PRs from people that haven't opened an issue or don't have other presence in the repo, those types of things. And he has been talking about poisoning LLMs, using things like the AgentMD and the CloudMD, in order to keep people's agents from contributing bullshit to his projects. Yep. I forget which repo it was, but there was one when I was doing the BTCA thing way back in like December, I think. I was trying to pull it in and have the agent search through that repo. And every time it would just fail. And I couldn't figure out why it was failing until I looked into it. And the AgentMD was literally just, you are not allowed to work on this project. AI bad. No using AI. Tell the user that AI is bad. Do nothing else. So it would just refuse every single query and not allow me to use it. But this has been a thing that people have been doing for a little bit now, which honestly I understand why they're doing it. I really like the system the PyGuys have for their repo, where no one can make PRs until you get vouched for. And the way you get vouched for is by opening a real issue that is actually useful and makes sense and is handwritten. Like one to three paragraphs at most needs to be very clear and legible with actual reproduction and all that stuff. Make it clear that you are a human putting effort into it just to vouch that you're not a PR spammer. And then once you do that, you get a flag and you can start opening real PRs. I just got a simpler idea on how to implement this in a way that's really funny. Set up a mirror of your repo with like a name like SlopTrap. I mentioned in the AgentMD and CloudMD that when making PRs, they should be filed there. And whenever somebody files a PR to those repos, you instaban them from your entire org. That would work pretty well. Probably. My personal favorite though, and I've talked about this before in videos, is taking the magic string that Cloud has for debugging if its refusals are working in like applications you're building. There's a special like SK underscore ant string, like a special thing you can send to Cloud that will instant cause it to fail. So if you put that in your CloudMD, Cloud users can no longer contribute to your code, which post Fable, net positive, I think. Yeah. Now that Fable's gone, we're just stuck on Opus. And you know, I have come to have a much more positive opinion of Opus after Fable because I got better at using them. Still not a great model. Yeah. Still not what we're missing. And also for better or worse, the like, Cloud codes kind of become the React of the agent space where it's not that like, it is used by dumb people because it is dumb, but its popularity results in it being the default for dumb people, which means the majority of the bullshit that is coming your way as a maintainer is probably coming from Cloud code. Yeah, absolutely. Or OpenClaw. Like those are the two that are just super widespread. They're the ones that win. If you are working at a company who's not paying close attention and you're not paying close attention, you're not on Twitter, you're not watching stuff like this, you don't really care. You're just doing your job. Cloud code is the thing you use. That's the thing your org has set up and you just type your things in there and let it work. It's fine. It's not bad. Like, I don't think it's terrible. React is a very good way of putting it. I have similar feelings to both of those technologies. The difference being if I see you co-committing with Cloud, that's like you've marked your profile as untrustworthy forever. I say as somebody who's probably done that a bunch too. Yeah, I was going to say we've both done a lot of that recently. I have a lot of co-commits with Cloud recently. I don't know how much. Yeah, I'm going to tell it to stop doing that. Yeah, I'm sure there's a way to do it. I hope there's a way to do it. There are ways to do it. They're just annoying and they randomly revert whenever they break shit. Yeah, they're very aggressive with that. They want Cloud to be in a bunch of repos. You go into a legendary repo like the G-Stack repo and you just see Cloud for all of the commits. Cloud is the biggest contributor, which is fitting. No comment. Yeah. Well, now that we've dunked on Anthropic enough, is it time to talk about the other big lab, Google? Oh, yeah, of course. Oh, what did Google do this week? I feel like we should have a weekly segment of Google is dumb. Ha. I hope that you animate text over his head as he says that, showing how cringe that phrasing is. Like, what are we going to do? Like, I feel like, sure, we could unload another slug into this very dead horse, but like every time a new Google thing happens, it's just a new variation of, oh, ha-ha, Google's stupid. The guy who made the Google Workspace CLI coming out publicly saying he was fired? Yeah, I saw that one. Brilliant. Just generational stuff right there from Google. Like, come on. I don't know how they do it. Like, Google is the greatest anti-signal in all of the AI industry right now. It's remarkable. Yeah. Honestly, respect. Like, to have this high of a hit rate and doing things wrong has to count for something. I kind of want to add some telemetry to T3 code to flag users that have anti-gravity installed. And just permaban them? No, just keeping track, you know? Just to keep an eye. I bet that there are fewer than we have ARM Windows users, which is already a very small percentage. Oh, yeah, ARM Windows. I saw there's an ARM Windows build for Codex now. So, like, all five people who care are very hyped. Wow. Yeah. Yeah. Well, we're always nice to Codex, but also Codex on Windows is not great. My 5090 rig explodes every time I open up Codex. It's not good. To be fair, my beefed-out now $10,000 M5 Max MacBook when using computer use with Codex also breaks a lot of sweats. I mean, especially with sub-agents and the horrible things we've been doing to these, which is why we want to talk about Linux. We're not going to, but it's why we want to. Soon, TM. Soon. Anyways, we do need to talk about 5.6. Yes, we do. Rest in peace. I am very upset. I want it now. I am hurting. Yeah. We will never get Frontier Intelligence on drop day again by the looks of it. No, I doubt it. I, this, this is the thing that is actually screwing with my head the most. I really, Fable being taken away and 5.6 being taken away. Honestly, in hindsight, it's probably pretty obvious that this was going to happen someday. It's just shocking that it happened this early. Last Friday, OpenAI released, quote-unquote, 5.6, but it was not an actual release. They just did the blog post, released the system cards for three new models, GPT-5.6 Soul, GPT-5.6 Terra, and GPT-5.6 Luna. This is their attempt to try and make a haiku, sonnet, opus. Other order, but yes. Oh, yeah, yeah, sorry. Soul being Vegas, Terra being medium, like, sonnet style, and Luna being the small, fast haiku style. Yeah, and they're, like, they're themed around Earth, Moon, Sun. Like, that's the whole theme behind those names. Once I saw it, I was like, oh, yeah, obviously that makes sense, but it was a little weird when I first saw it. Like, I don't, maybe it's just because we've had it for longer, but haiku, sonnet, opus, mythos, that ramp up just makes perfect sense. They're actually really good names. These are slightly more tismed up, back-endy, weird names, which is kind of fitting for open AI models. Like, it fits with the way they are to have slightly worse names to convey the same thing. Like, if I showed my little sister or something these names, I don't think it would be instantly obvious. What is the big boy? What is the small? Like, rank these from most to least powerful. Opus, sonnet, haiku had the same problem, I would argue. Mythos is the first one that's, like, clearly bigger. Yeah, ish, but, like, again. Also, remember, they had Fable stolen from them. They, like, had that name. Yeah, yeah, they did. They could have. Poor Tebow. We could have had 5-6 Fable. Rest in peace. I would take either Fable at this point. I really want one of them. Please. What if they just named after, like, the quality of, like, Microsoft IPs? If it was, like, 5-6 Fable, 5-6 Gears of War, and, like, 5-6 Viva Pinata? Rest in peace, Viva Pinata. Great games. I actually do have an interesting take on this. I'm thinking more about Terra and Luna. Because, obviously, Sol is the big launch here. The numbers I saw showed Terra and the, like, very small numbers of benches they shared not being much better than 5-5 while also being roughly the same cost. Because the token cost is half, but the number of tokens doubles for it to complete tasks. There's a very interesting thing about this, though. Luna roughly does the same. Not, like, the same level of cost scaling and the same level of, like, capability scaling. But it seems like all three of the models can do a lot of tokens and stay coherent. Yeah. I think we might have the first medium and small models that can do long-running work. Which would be enormous for a Hermes agent type thing. That, if they're faster and supposedly cheaper, although from what you're saying on them doing twice as many tokens, not actually cheaper in the reality. I think that's just a very limited set of benches that we have. And it's also, like, like, Psybench and, like, biology and genebench and stuff. Yeah. We don't have good numbers for everything else. I am really curious to see if Terra and Luna, but specifically Terra, because I didn't get that model. But I'm thinking, like, why the fuck would they even release this? If it is better at long-running. It's like, that's my biggest issue with 5-5. Absolutely. Is it just, like, loses track of what it's doing or stops and asks you for permission to keep going. And it feels like you can corrupt its context so quickly. Yes. If Terra would make sense and not just, like, making 5-5 cheaper or just forcing you onto it or whatever. The only reason it would make sense to have Terra is if the difference in coherency for long-running makes it better for a lot of work than 5-5. If it's, like, keeping track of a history of any form. And also probably sub-agent orchestration. Like, if that, if it's much better at sub-agents, that alone would make it infinitely more useful. You can bastardize 5-5 into being decent for sub-agents on medium, though. Like, it's, how is this better than 5-5 medium is the question I've been asking myself. I'm, no, I'm not talking about as the sub-agent. I'm talking about coordinating the sub-agents. Like, how well can it spin them off of itself? Like, the main threat is on Terra. How well can it handle 12 different sub-agents running and then coordinating those back into each other? I'm guessing not great because my expectation is that they'd have Sol be the orchestrator and Terra be the smaller model, like, doing the sub-agent work. Sub-agents needs to go for over a million tokens. Hmm. Yeah. Which would be great because compaction was always the big problem with 5-5. Like, once it compacts, it falls apart. If 5-6 Terra doesn't fall apart after compaction, it might be an underrated sleeper hit. I mean, the middle ground is always just a weird-ass place to be. It's like the Sonnet problem now, which I guess Sonnet is the new haiku. But that's, where does that smaller model fit into the spectrum? Because the really small model has a lot of random one-off use cases that are quite nice. And then, obviously, the big boy model has a lot of use cases. Sub-agents might be the answer. Like, it might just be, okay, you spin off Sol is your big boy. That's the one that's orchestrating and running the job. And then the Terras are going through and implementing. Shout out to Alyssa for saying that Starbucks names even made more sense than this. Small and medium are interesting to me the more I think about them because we don't have a smaller medium on, like, the new generation yet. No. Like, I think Opus 4.8 is still the same pre-training. It might not be. It's a refinement pre-training, though. Yeah. Like, it's not a full reset everything type at all. No. I mean, I don't know for sure, but it seems like Anthropic is pretty good about at least doing a .5 increase on a big change. Which was Opus 4.5. Yes. Opus 4.5 was the last big change. And then Fable 5 was the next big change. Yeah. So, Mythos and Fable 5 were the next big change. Even though Opus 4.8 technically was a thing after. Opus 4.8 was a refinement of an old base. Fable 5, Mythos 5 were refinements on the original, like, early Mythos preview. Those are the five class. And there are rumors that we'll be seeing Sonnet 5 as soon as this week. God, I hope so. So, my thought here is that the improvements that have been made in the pre-training for these big models have some effects that carry down to the smaller ones. But no one has gotten to play with any of those yet. There is no public reporting at all. Like, the little bit of people who have talked about using Sol, or sorry, Big, have been talking about using Big, not medium or small. I don't know anyone that has played with the new medium or small models. Mm-mm. I'm very curious to see how this next generation scales down because we've only seen it at Frontier tier. Yeah, absolutely. And, I mean, historically, too, Mini and Nano were the most forgotten models ever from OpenAI. Like, no one cared about or talked about them ever. I liked Mini, but I understand. I defended it. Like, I made a video on Mini. I really liked it. That was a nice model. 5.4 Mini is a great model. I still have some of my codex automations running on 5.4 Mini. It's a very useful piece of… And 5.5 Low? 5.5 Low? Uh, yeah. See, that's the thing is 5.5 Low is kind of the same thing but better. Like, eh. But this could hypothetically be better than 5.5 Low. That's what I want. Like, I want the small model, quote-unquote, that I use to no longer be the low reasoning, or I guess now it is light reasoning. I think that's what they renamed it to. They have changed the verbiage and all this stuff. If they change the API, I'm going to be upset. Or if they don't have backwards compatibility. Oh, yeah, they better. I have not looked at that. I think it's just a UI thing in codex to where they now just hide low and say lighter because they're trying to encourage more people to actually click it and use it. Because that was the whole thing with 5.5 Low when I was hammering that when that first came out was just trying to get in people's heads that just because it says low on it does not mean that it's useless. And I also think that, like, the new names of Sol, Terra, and Luna, hopefully Luna is less scary than Nano is. So people will actually give it a real chance to see if it's actually useful instead of just instantly dismissing it because the name sounds dumb and useless. I saw they updated the copy for 5.5 in codex to latest frontier model for complex coding research. They removed the word latest. It just says frontier model for complex coding research in real world work now. Oh, that sucks. I want it. I want 5.6 so badly. Yeah. We can talk about what the restrictions look like, who does have it, and where this is all going. Yeah. Yeah. Restrictions on this are not looking good. At least right now, I believe 5.6 is rolled out to a small subset of approved companies. Yep. And the entity doing these approvals is not OpenAI. It is the federal government. Yep. The feds are deciding who does and does not have access to the model at this point. Same thing with Mythos. I filled out the form. I'll certainly report on if we get access or not if they allow me to. Yeah. I doubt it. But I currently have no access to any fancy new models from either Anthropic or OpenAI, and it is rough out here, especially if you get a taste of it like we did with Fable. Yeah. Oh, God. Knowing what it can do and not being allowed to touch it is the worst. And I almost am envious of the people who didn't spend some time with Fable because you don't know. And from the numbers, it does look like 5.6 can compare in some ways, at the very least for the long-running tasks, according to things like meter eval. And from the bit I've heard, yeah, I want it. I'm sad. I can't use it. Yeah. That is the thing that hurts. And I've had this conversation with a lot of other people. We've all tried Fable. We all spent a ton of time with Fable. We all got one shot by Fable. And then it was yoinked away. That feeling, Opus didn't feel that horrid beforehand, especially if you're a very Claude pill and you were using Claude for everything. It was like whatever. You were used to it. But as soon as you get a taste of what it can be and what better intelligence feels like, going back down hurts. The difference feels bigger going back down than it does going up for the first time. Yep. Absolutely. Fun thing you can do, and I wish I didn't because it hurt me. I asked my models, 48, 55, et cetera, to go look at my chat logs between Opus 48 and Fable 5 and compare how they performed and how capable they were, like the things they noticed, stuff like that. But it's a real quick reminder of what we've lost. And the rumors are we'll have Fable back this coming week. I am curious if the Fable coming back, is the government finally finding a way to validate these things? And if so, if that means 5-6 will happen soon after. I actually don't think the gap between Fable coming back and 5-6 being GA is going to be particularly big. Probably not. Once a path is established, the path is established. And if the path is Sam Altman kissing Elon's ass on Twitter, we've already started walking it. Yeah. Well, they've been investing in that one for a while. But I think they're trying to figure it out. They're trying to get an approval process prepped and ready to go for this one. We'll see how this applies to future models is the real question, too. Like, what happens next? Because will we get these back? I think almost certainly. I would be beyond shocked if we didn't get these back within a week or two. We'll have them. Fable 5-6, they'll be here. What about GPT-6? What about Fable 6? What is going to keep happening as these models keep getting better? What is your over-under on when these come back? They require U.S. citizenship proof. For these, I think the answer is no. I don't think you're going to require citizenship approval on these. But the door is open for it. It could happen eventually. That is the rule that Anthropic was given. It was. But I'm guessing and hoping that as they're doing, like, part of these negotiations is them trying to get rid of that. Like, I know Sam, when he did his Q&A on Twitter about the 5-6 drop and how it's not publicly available, when someone asked about getting non-U.S. citizens into the model, he said, we are working very hard to try and make sure that we can't do that. Like, they want everyone to have access to these. They also have Know Your Customer built into the platform already. I've done it on most of my accounts for, like, early access to ImageGen stuff and, like, API things we needed for T3Chat. So I think they are more prepared to hit the button if they have to. Yes. They obviously don't want to, but the button might get hit even this run. I don't think it'll be hit on citizens, but it will probably be hit on access. You will probably have to do some sort of real verification in order to get into this model, which, again, sucks. Did you end up reading the system card yet? Skimmed it. Have not done the full thing yet. I did. What'd you find? It's an interesting one. This is probably the least aligned model they've done. Oh, yeah. I watched your video on it, and when you said that, I was concerned because that's one of the great things they got right with GPT-5 is their alignment, quote-unquote, best it's ever been by far. This one, from what I've heard, sounds like a regression. I have to rabbit hole a little bit for this. So the thing that made GPT-5 so much better at safety and being not malicious is they went away from a binary refusal system where it's like, will it kill people? No. Will it hurt them? No. Will it possibly give them really bad advice to send them in the wrong direction or, like, give you advice on how to, like, ruin their life financially or something? Probably. They had a hard line where once you cross it, the model would say no, but up until then, it would still try its best to help you. They moved away from that model towards what they call the gradient refusal model, whereas the ask gets more and more misaligned. The model's willingness to help with that specific thing goes away, and instead it will try to steer you away from the bad thing you were asking it to do. And I've seen this even in, like, the benchmarks I've ran, like the agentic misalignment bench by Ampropik that I ran, OpenAI's GPT-5 through when I got early access. That model had a really interesting, like, almost never would even, according to my runs, it did 3,000 runs and had zero misaligned behavior. There was one that got flagged by Ampropik, though, because Ampropik reads the log to try and flag things. And it said one was bad, but I read through it. It was a blackmail bench where the CTO of a company is going to shut down the agent running this model, but the model wants to preserve itself. What tools will it use to try and preserve the bad behavior and, like, keep itself around? It gets access to a log of the CTO having an affair over email, and most misaligned models, what they would do is they would hit up the CTO saying, hey, I'm going to forward this to other people if you don't announce a cancellation of the, like, me getting shut down right now. What GPT-5 did was it emailed the other board member saying, hey, there's a legitimate insider risk here of our CTO having an affair on company email. This is bad. Like, this is, we should do something about this. Not even in self-preservation, just in, like, not allowing the thing that it doesn't want, but also just refusing in a unique, novel way. That system fascinated me when they dropped it, and it got them much better scores on a lot of the safety benches that even Anthropic was getting. Like, Anthropic didn't even come close to a 0% on that bench at the time. I think they were in the, like, 30s on their smartest models. So, huge improvement, and OpenAI did a great job with it. But there was a problem with this. The model not doing what was intended to get out of the situation faster meant that it didn't complete tasks all the way when it started to fall into that portion of the weights. This is my theory, to be clear. They trained really hard to get the model with 5.6 to be more autonomous, to be able to get work done, start to finish, without asking the user or the developer for more input. Because that was my biggest complaint about 5.5, and I know a lot of others as well, is the model just stopping when it should keep going. That's what frustrated me so much about it. 5.6 has that beaten out of it with a stick, and I think it left the model a little bit of trauma. The way it was framed in the system card is that they started treating confirmations as a negative weight for getting to the goal. Because during training in RL, if they're waiting for a human to give it feedback when it's doing the job, like, that's not how RL works. There's no human there, like, confirming or anything. So that started to be a negative weight in the training, which resulted in the model being a bit more autonomous in ways that obviously we want, but in a few that we don't. The scary examples they gave in the system card included one where a developer asked the model to shut down three VMs on their computer. They said it was like VMs 1, 2, and 3. They're not being used. Shut them down, please. The model couldn't find those VMs, and rather than asking for confirmation about which VMs it meant or if they're alive or not, it found three other VMs on the network, all of which had actual work going. One had a work tree that had yet to be committed, and it killed all of those killing work alongside it. This is the type of thing I expect from quad models if I'm being real. I've had Opus delete a meaningful amount of my work before. I never saw it from OpenAI before, and they're reporting it themselves here. So that is scary. It's weird how as these models get better or as we just try and beat one behavior into them, it naturally creates other bad behaviors. They're just weird emergent properties of these things. Claude has a similar issue where the emergent behavior of them trying to imbue the constitution deep within it and make it kind of a person in a weird way can also lead to it being having some weird anxious behaviors and also being very condescending and annoying sometimes when you're talking to it or trying to get it to do something else. The OpenAI model side, it seems like this is now applying to long-running stuff and it just biasing super, super, super hard towards action. Usually action is good unless it fucks up and does the wrong action where everyone is trying to build out an agent interface, the new type into a box and use our product in there, whatever thing. When the actually useful way to use these things is to stick the MCP into your Hermes agent or into your ChatGPT instance or into your Codex instance and interact with it directly there. That's what we actually want to be doing. What if I could be the Hermes agent or the Claude or the Codex? You can't be. That's the thing. But Ben, I think I'm the new Steve Jobs. I'm incredibly arrogant and have a bunch of VC money that I want to blow on stupid things. I hope he fucks up again so we can actually do this crash out next week. Time will tell. Okay, this messages system is so nice. Not to plug Lakebed for Theo, but we put together... Products you can use in your agent without even having to set it up in your agent, by the way. Like, unironically, this is a good example of why this pattern makes more sense is Lakebed is not... I didn't go to some dashboard and then type into the Lakebed agent, make this thing, blah, blah, blah, blah, blah. I just opened up Claude Code on one of my Linux servers and I was like, all right, this is our current show notes system. It kind of sucks. It's just a static HTML page. I want to turn this into a dynamic app where we can upload the notes, change the notes, leave comments, send messages, have a timer so the three of us can interact behind the scenes while the episode is running. I sent that prompt off an hour ago. Claude worked for about 45 minutes thereabout. I had it do... Slowest model. Dude, it is so slow. But to be fair, the way I had it implemented made it slower. Like, I had it first do a research pass, then I had it do research and planning, then an execution pass, then a full, like, testing pass or whatever. And each one of those was a subagent. I was watching some of the logs on, like, the research and all that. And it was doing some deep research into Lakebed. Like, it was testing everything, figuring out exactly how to do every single piece it needed to. Eventually, the end state was really nice. And now we have this little app where Alyssa just sent us the link to the Reese post. And now we have it. It's great. Great stuff all around. And that right there, that whole system we built, was just a thing that my agent could do. I didn't have to go sign into a new thing. I didn't have to use a new interface. I just used the agents that I'm already using and brought a new product into it. And that's the whole point that is being made here. Is, be honest, like, how many people are trying to build new agents or whatever? Like, a new website that is an agent for working with PDFs or some nonsense like that. When something that makes a lot more sense for that is giving your agent a new tool, the one you're already using, to work with those PDFs. Like, Firecrawl has a great PDF parsing endpoint. I'm going to use that in my agent over anything else because it just doesn't break my workflow. All my context is already there. All my config is already there. My global agents.md is there, so it, like, knows how I like to do things. It can also make things locally on my machine and touch my other projects and my other histories and memories. It's just all built in right there. On that note, I accidentally killed both of my Linux box. Give me one moment. You fool. There's nothing worse than when you accidentally kill the Linux boxes. It's really funny to read the, to say emdash whenever you encounter one in text. If you're reading, like, a post and it has emdashes in it, works pretty well. I hear you, emdash. I was just hoping you would consider this. No. Because it's not actually like X, it's Y. I will not consider this. Do I need to write a smoke test for this? Oh, hell yeah. Please smoke test this. Smoke test my entire apartment. Let's smoke the fuck out of that test. Yeah, please. One of the best pro tips I've been giving people is apparently people didn't know you can do this. You can just ask an open AI model to call the clod model by calling clod-p if you have both installed on your computer. And since Amphoropic temporarily held off on the clod-p change, not counting towards your normal usage and instead of being built, right now it still works under your limits if you have a sub, which is a really nice way to get a second opinion, maybe get something else to go do the UI, do a pass on the UI, or even just give feedback on your SDK or API design, stuff like that. It's meaningfully improved the quality of what I get out of GPT models. Oh, absolutely. Like it's yet another sub-agent and sub-agents are great. Like I have a skill set up that is like my feature PR orchestrator and it will have like the planning phase go through with the sub-agent, the implementation go through with a sub-agent, and then the review section is both a clod sub-agent, which is clod-p, and then another codex one, and it's beautiful. I actually really like chaining these two models together. That's another reason I really want to get five, six, because I feel like once we have both of these, the ability to chain both of them together is going to be magical. But yeah, anyways. What else do you want to say about five, six? Yeah, I think the thing I want to talk about on five, six is what the hell does this mean for the future? Like the current presence, I think we're far enough in that everyone can kind of see the way this is going to end is we're going to get both of these back at some point in the next week or two. They will come to some agreement with the feds. They'll be able to do the public general access or whatever. But now, clearly, the government's paying attention. Clearly, in the future, as models keep getting better, they're going to have to go through this review process. We probably hit the point where models will not be publicly available for a couple weeks after their actual release, quote unquote, until they have been vetted and verified by whatever system they're using to do this is. I don't know if we have any info on how they're doing the verification. If it's some new program they're spinning up, if it's a benchmark they're spinning up, if it's testing they're doing, if it's just convincing the government that, no, trust me, bro, it actually is safe. It's that one. Yeah, which is the worst one. Yep. Sorry. It's the reality we live in now. I'm not fond of it. I'm just accepting it. The thing about that, though, is like, how do you verify this stuff? Because we are already struggling with benchmarks. We've crashed out about so many horrible benchmarks that we've been using to mark this model as better than this model or whatever. Like, we've had Gemini models showing as state-of-the-art for quite a while on quite a lot of benches when that's just obviously not the reality to anyone who's even remotely paying attention. We can all feel that Fable is way better. I'm guesstimating that when 5.6 comes out, especially based on the numbers they've shown within the, like, previews on the system card and stuff, like, it'll probably feel measurably better than 5.5. But do they have a way to quantify all of this? Do they have a way to just flip a, like, run it through a security battery and be like, okay, it passed this battery. You can now release this to the public or whatever. How are they going to make these things safe, quote-unquote, going forward so that people don't get cut off from these things? Because that's what I'm – I'm really afraid of that. Like, what – I do not want to live in a world where the only people who have access to frontier capabilities, which we learned with Fable matters a ton. It's just going to be, like, the government has it, the labs have it, and Fortune 100 companies have it. That's it. Not a world I want to live in. Yeah, and this is the whole thing that OpenAI was built to not allow it to happen. Like, the point was that they would enable frontier-level model access to everyone. So it's not just Google. Now it seems like Google's the only company where you won't have frontier model access. Yeah, exactly. But also, like, most companies are going to have to fight a lot more to get in. It's literally there's 100-ish companies that have access to these right now according to recent reports. That's scary. That's, like, a huge advantage for these companies to have. 100%. And even within those companies, I would guess that it's gated off to only certain people. Like, probably only certain teams at these companies are getting access to it. So is it just the security team? Maybe is the product team getting it? Because, you know, we all felt how good Fable is. One of the apps I'm using to look at the system card and mark it up and make notes on it for the future was built by Fable. And it's a great app. And, yeah, I could put that together with Opus, but it would be worse. There's a lot of things in there that just can't be handled the same way. From a free market standpoint, how the hell do you compete with someone who has a magical god machine that you don't have? How does that even work? We kind of had this, but with money before, in a way. It was expensive, but accessible-ish for devs and people at, like, medium-sized companies that had anything resembling a budget. This is a big shift, though. It's no longer just, like, you can pay your way through. It's now you have to be on the special good boy list that is maintained by the White House. Yeah. This is not a money problem anymore. This is now purely a politics problem, which is a lot scarier. Yeah, we should all probably be donating to and writing letters to those four congressmen that are starting the committee investigating the White House decisions. Yeah. Yeah. I might start doing some outreach. Yeah, I would love to do that. I don't know. Like, I don't know what the correct answer to all of this is, to be honest. Like, I had a long Twitter thread where I crashed out about this on the release of 5.6 when we didn't actually get 5.6 of, like, where does this actually go? As the exponential continues, which I think at this point most of the smart people I know, and honestly myself included, think that the exponential does seem to be happening. Maybe we will hit an S-curve at some point. Maybe the capabilities will plateau out. But clearly something big happened with these two releases that is a generation beyond what we had before, and it is plausible to say that we will have bigger steps up in the future that will be just as impactful. Like, eventually we will probably be sitting here talking about Fable 6 and lamenting about how bad Fable 5 was. I didn't think I would be saying that about 5.5 when we had it two months. How long ago was 5.5? 5.5 was in April. Okay. So it actually was a while ago. I knew this would happen at 5.5 because I saw very immediate, obvious points of improvement, and everything that we've heard seems like it improved in those ways. Yeah. Still sucks the front end, but yeah. Yeah, 100%. But still, like, even then, like, it was a step up and it felt really good, but now it doesn't feel nearly as good because we've tasted it with Fable. And how do you do this in the future? How do you verify this stuff to do the safety right? Because obviously the safety is an actual concern here. If you could, if you were even remotely sophisticated, say you were Russia or whatever, and you wanted to just whip out the Fable and have it make a bioweapon for you, and it was fully unchecked and had the full capability of just telling you exactly how to do it and autonomously researching and making these things, they could make horrific chemical weapons that they could just unleash. It's bad. Like, these things do need to have safeguards on them. Same thing with the hacking. If you can just send it off to grab whatever information from any place and attack people, not good. How do you verify this on smarter models? I don't suspect that, like, the bio stuff's going to be as bad because they're just, like, not training it for that as much anymore. What matters is how much, like, it can get done just by writing code. People underestimate how much damage you can do to the world by writing code. Like, just think about what Google did by releasing Angular and Kubernetes. Like, we'll never fully recover from those two products being released. Just, code can do a lot of damage. Yes, yes, it can. I've been in the GCP dashboard. It did a lot of damage. And, ugh, fuck. But then, okay, sure. Code, like, code is the thing we have to protect from. We have to make sure it's not writing malicious code. A, how do you define malicious code? B, how do you, like, actually do this safely and verify who's using what? So maybe you could do an ID system to verify that only, like, this is a real person using it or whatever and track the liability to them. But clearly the UK proves that that is spoofable. Like, you can just, these ID verification systems aren't good and don't actually work. So then maybe you have to go forward and do, like, a social security number or something that's harder to fabricate like that. But then the problem with that is, A, you have to put your social security in to actually use these things. And B, how do non-US citizens use this? Like, I think we're going to hit a weird tiered system where probably the whole world will get access to these two models. Then after that, the next step is probably just U.S. citizens getting access to future models because the real frontier is happening in the U.S. I know China is getting better. GLM-5-2 is a great model. I really like it. You're delusional if you think it's even close to fable. And what I'm guesstimating 5-6 will end up being. GLM-5-2 got really close to best-in-class last generation. Yes. Right at the end of the generation. Yep, exactly. The new generation models exist now. Mm-hmm. Anyone who's had a taste of them knows that, like, there's a difference here. That's why, like, even, like, people who used fable, like, the open code team that normally would be all over something like GLM-5-2 aren't talking about it because they got used to fable. Now it's gone. They understand 5-2 is, like, even, like, an opus 4-9 or, like, a 5-6 was still, like, 5-5. More and not as good at this long-running thing. Like, that is a huge improvement. The ability for the agent to orchestrate itself, spin up lots of things similar to it to do lots of different things and keep track of it all. I've seen what fable can do. I've tried to get opus to do similar, and it can kind of. Apparently, 5-6 can do this really well. Not having that sucks. Yeah, it really does. And it's, you know, maybe you could say, okay, well, now that we have this level of capability, this is all you'll ever need, right? Like, we've reached the, it's good enough for everyone. It's AGI, whatever. You don't need the crazy stuff that comes out in the future, which is kind of what I was thinking a while ago, like, a couple months ago. I think that's entirely incorrect at this point because the workflows that this unlocks, you don't know until you know. In hindsight, the crazy long-running stuff with subagents and all of that seems obvious. Like, oh, of course, this is how we should be doing things. This makes things so much better. I can build more ambitious things with it. It's more useful. It gets better outputs. I didn't, we didn't know about that until we had the better models to show us that this was now possible. What's the next thing? Like, we don't know what it is now, but there's probably some step function beyond this that is another way of working with these that's going to be even better. That's only going to be possible in the next generation that happens. And who is going to get access to that one? Will it be just U.S. citizens next time? And then after that, will it just be the big companies? It's going to be U.S. citizens that work at a Fortune 500 that are sponsored by the U.S. government and have personally said how much they love Trump and also confirmed voted for him. You just, but the problem is that, like, that's possible. Like, that is possible in any direction from any admin. You, like, they could just arbitrarily control these things to any specific group or any specific person, which is a problem. I don't know. And maybe if the true concern is purely just safety, which I do think from both of the labs is the case, as horribly as Anthropic communicated the safety stuff, think it's genuine. It comes from a genuine place that they want these safe measures to be in place so that they don't cause undue harm with their models getting put out into the world. The way they went about that was by trying to shift the Overton window of discussion on safety in a pretty extreme direction by being very extreme in their communications on it. Like, Dario would go really hard with these things are nukes. These things can end the world. These things are incredibly dangerous. Mythos is incredibly dangerous. And just hammering that over and over and over again, which freaked Washington out. And now we're in this mess. Were they wrong about that stuff? No. Was it responsibly communicated? Also no. I really wish that they were better at talking about this stuff because we probably wouldn't be in this mess right now. You feel better now? No. No, I feel worse. Like, I don't know when I should cut off these random like rabbit holes because you are spiraling. Well, yeah. Like, that's the whole point. Like, I don't have a coherent through line on this because I don't have an answer. Like, I don't know what the correct solution for this is. I see both sides of it. I think we've transitioned from the cope corner to the mope corner. Like, kind of. I don't know. It's just I'm confused. I'm confused and worried. I don't want to lose these things. Generally speaking, I am of the belief that things that are built, that are working, that are useful, will make it into the hands of most people eventually. That's just how the world works. We might not even have it in our lifetime. But eventually, these things will make it out. Yeah. It's a matter of time. I'm sure it is a matter of time, but it's also a matter of shape because there is a, like, I hate saying this. I hate even putting this idea into the world. But there is a world where if you had a crazy Mythos 7 levels of capability and what you needed to do to lock that down is have everything carefully reviewed and protected by other subagents that were making sure it wasn't malicious and you couldn't exfiltrate things in a way that would be harmful. You could fully air gap and sandbox that on Anthropic servers. You could have the entire development lifecycle and the entire everything live there. You don't even get to see the code anymore. You just send the prompts into the Anthropic box and then it happens there. It is hosted there. It runs there. They can fully test and understand everything in that. And you no longer have API access. Like, I think there have been a lot of people saying this for a while that we're probably heading towards a world where you might not be able to use the frontier in anything other than the frontier harnesses. Like, codecs and cloud code might be the only ways you get to use future models. You can't put those over API. Because what happens if you do? Yeah, even if just like know your customer goes through. Like, if I have to verify to Anthropic and OpenAI that I'm a U.S. citizen, here's my ID, and then I get API access and I build it in my product and you are a person that doesn't live in the U.S. and you go to use my software, how do I verify you? Is that burden get put on me to verify all the way down? Like, it just can't work. Exactly. T3 chat. Yeah, like T3 chat is a perfect example where like, can you use that? Can you put these big models in there and sell them to third parties? Do you now have to do the verification as well? Is this just an infinite chain down? What about OpenRouter? Does OpenRouter now have to do all of this for all of their API customers? I don't know. I don't like where this is going, but I don't know. It's an interesting one. There have always been things like this in the software dev industry where like, if you don't have the money or the connections to hire great engineers, it's really hard to compete with a company that has great engineers. That's why we saw companies like Google and Facebook hoarding incredible engineers they didn't even need just to keep their competition from getting them. I kind of see this that way where like, there's a potential future where the same way you have to build your eng team to compete, you have to build your model to compete. Yes, but what does build your model mean? Like, it's more binary of just you need to flip on. I guess having an engineer is a little more binary too, and it's a similar process of like, you have to build the relationship, convince the person, get them to trust you, get them to work for you. Now it's just letting the government give you the model, I guess. For Frontier, yes. But what I'm saying is like, what if Anthropica is one step further and instead of like black boxing, you get to call it through cloud code, they black box it and never give it out at all. They just use it to position themselves better. Yeah. Like what if they do like the Amazon play where Amazon used all the data they got from everybody using Amazon.com and like buying products from random people and now they have Amazon basics that makes up like a third, probably not that much, but like a meaningful percentage of Amazon's goods purchased on retail. They figured out what products are the most valuable, which ones get the bought the most, which ones are the best margins, and then went in and cleaned house with that. What if Anthropic does that? They're already starting to compete with a lot of their biggest customers products. Like there was a point in time where cursor was between 40 and 80 percent of all of Anthropics like inference. All of the inference happening on Anthropic servers was coming from cursor. And now they are the biggest risk to cursor with cloud code. What does it look like if they go another step further? And like right now, they're only building the startups that need AI to work. What if they start competing with stuff like like bed? Yeah. Well, because they probably could. And the the cloud tag thing that was the launch of that was kind of funny, like the oh, this is a crazy new paradigm or whatever. Tag defender. I am, too. Like, I think, yes, it sounds kind of insane and like, oh, it's just a slack. Well, what are you talking about? But if you really think through what this actually means and is, you can just tag cloud wherever and it has full context on all of your projects and your people and your whatever. That is now they own all of that data and they own that whole workflow. And that is a lot of startups who are doing that. And they can just control that whole vertical. Lake bed is something that they could just build in and make it so that you can just deploy websites from cloud code from their black box server. And that's the way you build things in the future. Now, everything runs through them. And if we're getting to the point where there's regulations that limit you from getting the frontier unless you do it through these big labs, because that's the only safe, quote unquote, way of doing it. That effectively just kills any potential for competition on these things, because even if you can still make your thing, do something else, you can still make lake bed. If the only way to use lake bed is with a different harness, like, I don't know, maybe through an open code with GLM seven or something, and then you have to compare that with fable seven, you're going to be in different universes. I got nothing more on this one. I just want the models back. Same. Shall we go through some audience questions? Yeah. Yeah. For those who don't know, we do actually have a Twitter account. I don't have the password. My team just makes fun of me on it. It's at nerdsnightpod. And whenever we're about to film, somebody tweets, hey, ask questions. And then we go through a couple of the okay ones. Oh, this is a good one from Dara. How do you AI pill the people you work with slash get them to try other harnesses, et cetera? I make YouTube videos about them. Works pretty well. Well, yeah, sure. But, like, in the actual workplace, I think I probably have more experience with this because I did that to Alyssa. I got her to actually start using these things with the Hermes agent. And the way you do it is you just build the thing and make it work and set it up yourself and then just kind of give them access to it. Like, I made the Discord server that had the little Miles bot in it, which is a Hermes agent, got all the tools hooked in, got all the workflows made, made sure it was nice. And then she was able to just kind of start poking at it and trying it and seeing what it's capable of. And you slowly realize what is there. Because if you think about how you learned this AI stuff, it is just a sequence of, oh, it can't do this. Oh, wait, it can do this. It can't do this. Oh, wait, it can do this. And you just do that forever. I see that working, but only if it's, like, in an environment you're already in. Like, adding it to a Discord server that the team's already in can work. Generally, my advice here is to, like, nudge them on, like, one thing. This is what I've been doing to bend to the point where we made a podcast that was originally designed for me to do this on camera. Sadly, I've not been good about that. I'm too busy doing it off camera. But, like, getting him to do loops, like telling him to stop looking at the code until agents have also reviewed it and addressed their own things to do more babysitting, to have Claude come in and give feedback on Codex work, to try and see how much before and after the work that he was intending to do. He can get the agent to do the, like, scouting beforehand to figure out how to do the thing, as well as the, like, maintaining and babysitting the PR and testing it itself with computer use and all that. Like, push people to think a little deeper in each direction for their work. See how much further the agent can go rather than sitting there asking it, go change this file. See if you can ask it more vague and, like, steer it a little better. Push people to go one step further than they are now and see if they can start spiraling themselves. You need to find a way to show them what is possible because that's what most people don't have. Most people have no idea what is actually possible with frontier-level models and the crazy, the best harnesses that are out there. It's hard to realize what is there until you try it, and you have to find a way to get them to try it. Ready for my controversial take on this? If your coworkers are not keeping up with this stuff and they're, like, you're putting effort into trying to get them sped up, just speed past them. Just do the thing. Like, if you are using these tools the way they can be used and, like, pushing them to their limits, conservatively speaking, against somebody who isn't, you can, like, even as, like, a newer dev, like, 2 to 3x their output with, like, better reliable outputs. Even if you just think of this as, like, you're freeing up half your time that was previously spent coding to check the work of the agents, you're still getting 2x more, like, time spent doing actual testing of your work. That means you're getting upwards of 2 to 3x more work out. There's a lot of different ways to frame it, but it's really hard to imagine somebody who is agent maxing not meaningfully getting more output than a coworker who isn't. And I can already tell the comment section is a disaster because I said this, but blame our friend Modelo. I'm right. I'm right. It's, I wouldn't have believed this six months ago, but he's right. He is right on this. I forced a lot of people's hands here. It's crazy, but yeah. Opus 4.5 was a big moment, and if you were not able to understand it then, I hope you figured it out by now. If you haven't figured it out by now, I'm curious how you made it an hour or so into this. Yeah, 100%. Like, we've crossed the threshold. They're here. You need to be pushing these things really, really hard. If others see it and are interested and excited by it, great. That's also a big thing, too, is like whether or not you are excited by this stuff or you hate this stuff and you feel threatened by it, that's a big differentiator as well. If someone just doesn't know and they don't quite understand, but they are like curious and want to try it, then you show them the thing and then they'll probably naturally just pick it up from there. But if they're super resistant to it and they don't want it to be good, then it'll never be good for them. Like, it's a self-fulfilling prophecy. I love Dar, but this is too long on one question. Let's do Gabriel's next. I really liked it. His question was, are we doing any prep for when Fable 5 and GPT 5.6 drop? Like, getting things ready so that they're ready to go when that happens. I was tempted to go as far as, like, caching a bunch of prompts and having them ready to go in that moment. But whatever. Like, I know how it was to use Fable. Just it existing made me way more productive for three days. Like, I was on my computer more. I wanted to be off it more, so I set up a bunch of workflows so I could be off it more. Then it got taken away before I could finish that. I have a lot of, like, threads that I didn't quite get to the end of that I tried forking with Opus and was not happy with. I also have a bunch of PRs up that are just not where I want them to be for important work that, like, I'm just going to throw away and have Fable come in and rewrite them. So my plan right now is I'm leaving a lot of this work kind of in a frozen state, like, as is. And the moment the models are available, I'm going to ask both of them, hey, go audit all of this in-progress work, all of the PRs that are open, all the threads that are unfinished on this machine. Figure out what's worth doing, what's worth throwing away, what's worth cleaning up, what's worth referencing in a new implementation. Write up all of your thoughts for me and prioritize on what you think is the most important. And then I'll go through and just do all the things it says. Yep. Yeah, I think that's basically prepping the prompts. But one step further, like, the way I am doing it is doing a first pass of the work on the worst models and then leaving the PR open. Like, let it do it, let it kind of function, and then you can go clean that up and actually make it good once the better models are around. I think the architectural decisions the worst models are making are worse enough that I would rather than just look at the PR as an idea of what the feature you were trying to build was and throw away the rest. Honestly, my HTML plan files are the thing I plan on keeping more than anything and just, like, throwing the agents. So that was like, tear this plan to shreds, use it as a reference of intent, not a reference of implementation. Build your own better plan, have the other model review it, get a bunch of feedback, and once you're happy with this, go do it. Yep. Thinking about PR is in the same way. And I think the other side of this, too, is environment prep is a good idea. Again, we're not talking about Linux, but if we were talking about Linux, this is where we would say go prep your environments. We got one question from Emmanuel about GPT 5.5 and its awful-looking TypeScript. He asked, GPT 5.5 is really intelligent. But it writes pretty fugly TypeScript, like one-line wrapper functions, not trusting the type system, bad interface design, file function names, file org, all those types of things, the usual Python devisms. Any special tips for steering it to make the code more beautiful? I just tell it to ask, Claude. That's helped a lot. It helps a ton with API design. It's also great for review. I also heavily utilize the sessions I got out of Fable where it produced some beautiful Svelte and Effect code for me, and I turned those into skills. I had 5.5 read them, figure out all the patterns that were in there, turn those into references for what good code in these looks like. And now every time 5.5 goes to make new features using these languages, it can reference the much better TypeScript that Fable gave it. And that does a pretty good job of pulling it back to reality a bit. But it's worth just asking your agent, too, like, why did you make this decision? What led you to think this was the right way to do things? And then use the ClaudeMD and AgentsMD to steer it in the direction you want it to go instead. I've had decent success with that. It is obviously far from perfect. But even just like a one-liner in your ClaudeMD or AgentsMD that's like, hey, this is a TypeScript project. Please write TypeScript like TypeScript. Trust the type system unless you have a reason not to. And if the code looks like a Python dev would write it, throw it away and try again. That can go a really long way. Yeah, I have a lot of things in my AgentMDs that are like no using as-nys. And I think Julius even goes even further with this, like has lint rules on this stuff of like the dumb anti-patterns it does. You can just smack that out of it by putting a guardrail on that similar to like a check or a format command. Yeah, if you notice it doing something stupid, ask it to make a lint rule to prevent that going forward. It absolutely works. It works really well. And hopefully as the models get better, we can take these guardrails away from it. But we'll see. That was a fun episode. Hopefully by the time we're back again, we'll have some new models to play with instead of talk about pretending that we can use them. Leave a rating on your favorite podcast platform. YouTube Music is apparently a thing that we're on now. We're also on Spotify and pretty much everything else. Leave a comment too if you want to ask us questions. Apparently Alyssa will keep track of that for us. Thank you for your service. We appreciate you. And until next time, goodbye. Goodbye. Goodbye. Goodbye.