Fable Is Back...kinda?

This week on Nerd Snipe, I'm making Ben do the intro Welcome back to Nerd Snipe, I'm Ben, this is Theo, and this week we're going to be going over AIE You should know what we're talking about before we do the intro, Ben Well, like, I looked at it earlier, I mean, we're talking about mythos, but there's other subtopics It's AI, FDEs, like Forward Deployed Engineers, 5.6 is still not out, I really want it, but it's not here Fable is finally back in our hands for at least the time being, and Anthropic is fighting China again Thank you, PostHog, for sponsoring this episode, and let's get started Let us know in the comments if you're the type that watches and listens when you're running Agents Leave that comment on any of the many different platforms we're on, because otherwise Alyssa will get mad at me Hi, Alyssa Apparently we're on everything now, so podcasts on Spotify, Apple, whatever else Yeah, leave us a ranking too, apparently that helps Probably do the news now? Yeah, let's do it Cool, so, first and foremost, we have the FDE revolution For those who don't know, Forward Deployed Engineers have been a thing for a long time But not necessarily a thing that mattered too much Because the role of a Forward Deployed Engineer is effectively a pseudo-sales onboarding Slash, like, teacher role Where they go to a company that wants to adopt your product or things that your product does And they go there, kind of embed themselves to teach the company how to actually use the thing that they are buying This makes a lot of sense in an era where everything is changing Because companies like OpenAI and Anthropic are fundamentally resetting how we do engineering And a lot of companies that have done things the same way for dozens of years Don't really know how to change So having someone from Anthropic or OpenAI show up and give you the ropes And teach you how to take advantage of these things makes a lot of sense Which is why companies like Microsoft are going really hard on this They just invested $2.5 billion in hiring a bunch of people to do Forward Deployed Engineering, I guess So, yeah, that's cool Yep, like 6,000, according to the article, 6,000 industry and engineering experts Are now going to be their Forward Deployed Engineering team To try and find some way for them to catch up in AI I don't know if this is going to work or how it's going to turn out Yeah, there's a lot of layers to this one that I'm thinking about Like, on one hand, this is the inevitable, like, Thanos snapping Where all engineers either become marketers or unemployed And, oh God, I had a fun conversation at AI About how every dev needs to be a YouTuber now, apparently And I had to explain to said people at AI That being a YouTuber is not necessarily some skill sets That the average engineer or even the best engineers have Having an audience does matter But the audience size doesn't matter quite as much as people seem to think If you have an audience of the right five people Like five people who work at Microsoft Or some other Fortune 500 company you're trying to sell to Getting that captive audience and teaching them how to use your thing Is actually a good sales tactic That's why FDE works It's not because these people are so eloquent at talking And so convincing they can get anyone to use their product It's that they're fixating on, like, three to five people a month Going to where they work And forcing them to use the thing I would expect more and more of engineering roles to start appearing in this way It's almost like the DevRel era In, like, the mid-2010s Where all these dev startups are trying desperately to get customers So they hired a ton of DevRels To hopefully put themselves in a better spot in the market This is like that But now the big companies are the ones leading it Curious to see where it goes Yeah, this is like I mean, if we're being honest here This is just consulting in 2026 Where these labs are trying to get people out in the industry To get people to figure out how to use these things In their day-to-day work Which is going to be hard to do Because a lot of the job here is just explaining What is actually possible I mean, even when we were at AIE last week The amount of people I talked to Who have no idea what is actually possible with these things And that is There was a very bubbly type conference And then as soon as you leave the SF tech bubble The amount of people who have only ever tried like Sonnet 4 within Claude Code six months ago Wrote it off as not that good Have zero clue what's actually happening They're just trying to get it out there I'm impressed the people you're talking about Even used Claude Code Most of the ones I talked to are still in Cursor at best A lot of them are in fucking co-pilot still Yeah, that's the other big one I've seen I think Claude Code has gotten enough Enterprise adoption at this point Especially since Bedrock and stuff like that They just have provisioning for it But it takes fucking forever For them to get the new models into the hands of their engineers And when they do The things that they're allowed to do Are very, very heavily locked down They do not have the freedom to just run around and burn tokens the way we do They have to go through all the Enterprise approvables And then also I've talked to a lot of people Where like they want to give more tools and MCPs And like some of the Not quite Hermes agent But even just in Claude code The stuff we're doing Where like our agents have access to email And Notion and Slack and all these things So that they can automate more work They're not rolling that out to a lot of people Because they have no idea how to do that In a safe and reasonable way So they just can't actually get any of the benefit from them I want to lean on your I would say experience But it's more lack of With some questions here Because I'm trying to figure out Where the market is going And how people are employed in the future This is a joke Trying to teach companies how to use AI On Azure's platform Or as one of the data labelers at Meta Who went from an engineer To having to label a bunch of slop In order to hopefully train a better model Oh well I mean you gave me two awful options But I would put the Azure one in this case Absolutely Like without hesitation Pick Azure there Do you know why I gave these two options? Because that's like the majority Of where a lot of jobs are going right now How do you think you get your first job In this industry right now? Data labeling Or if you're really good And you can Like you're the type Who's out there learning this stuff On your own time You know how to use it really well Your company's going to want to send you out To all the customers So that they can see how to use all this stuff With the crazy stuff you're doing And then hopefully become a customer I'm presuming that you're good enough already Because like If you're a college student Watching this right now For some reason Hi, thank you Make sure you tell your friends About the podcast We want more listeners But also Like let's be real here The type of college student That listens to something like this Is very different From the type of college student That would Back in the day Barely pass one leak code interview After failing 40 Get a begrudging hire From some company That's just trying to fill headcount And then somehow Stay in the industry for 10 years That archetype of person Is going to struggle a lot And when you see all these numbers And people complaining About struggling to get a job It's kind of weird feeling to us Because I'm sure you've seen this too Like the companies we talk to Are struggling to hire Yeah Fast enough right now The gap is because The top 10% of engineers Are the only ones Who are interesting right now But there's a catch Because if you care enough To be in that top 10% You might still struggle To get a job Not because you don't care enough Or you're not good enough But because that care And quality aren't visible enough So you need less risky places To go in the industry To kind of prove your capability And I think stuff like this Makes sense there As painful as it is Going to meta And building your own process For data labeling That your coworkers are like Wait how are you doing this so fast Or going to Microsoft As an FTE Because they're excited About the new blood kids Who are AI native They grew up with all these tools They're not learning About them retroactively They're not moving to AI They never knew anything else The same way that kids Grew up with an iPhone And don't know how file systems work There are now kids Who grew up with Claude And Codex That don't know how VS Code works And as stupid as that is Those are the types of people That can show up at like Bloomberg And help convince them To use AI more heavily Yep Yeah It's just This stuff is so weird And it's changing so fast The only way you can get Any real benefit out of it Is if you are obsessed With it at this point Especially since there is No established pattern Or guidebook Like back in the day I guess this kind of predates me But I assume Way back when It was a lot easier To just kind of have Centralized ways To just teach people To do the thing That has been done For 10 years Like processes Didn't blow up overnight The way they do now Yes and no It was more that The gap between The new way And the old way Wasn't so big So it didn't matter too much Like even going from Something like Cobalt to Go Was a 2 to 3x At best Improvement in your Like capability And output Maybe And you would lose A lot of important knowledge And like capability Before That existed before that With AI Even if you're still Writing Cobalt It can be absolutely Unbelievably useful And fundamentally Change the way And shape of Your work That is a big difference That we just haven't seen As an industry Where on one hand It is way more different But on the other hand The gap between Having and not having Is way bigger So if you're a company That hasn't moved To using AI heavily yet And your competition has You're dead If you're a company That hasn't moved off Cobalt and your competition has You can just hire A few more people And it's fine Yep That isn't the case anymore And that's what's Interesting here Yeah Yeah that makes a lot of sense I mean the closest To this I saw Was like my early years Were the JS framework wars And there were a lot of people Who were super Super Like remember back in the day All the crash outs About how horrible React is And how we should just go back To using HTMX for everything Remember those days That was fun That was a similar like Oh the new way is way better But it wasn't nearly as better As the new way now is And it wasn't nearly as different So yeah Makes sense So I will give one piece of advice To the companies That are doing this type of hiring If you are a business Spending millions If not billions On forward deployed engineers I want you to take a look At the people you're hiring If the majority are over 25 You are not hiring the right people And you are not Paying forward The role that we all have As people who have been here For a while Please take advantage Of the unique capabilities That younger people have For this type of thing Because they grew up in it They'll do a much better job Of explaining it The hard thing they'll have to do Is to learn how the old way Worked enough To translate to those people That's a lot easier Than taking some like Unemployed 32 year old And trying to teach them How clawed it works Like it's It's not gonna happen Be realistic here Teach the younger kids About how things used to work Instead of teaching The older people How the new things work It will Don't pull a ladder up behind you Is all I'm trying to say It's important More so now than ever Now that we've gotten Through our first Three letter abbreviation Let's go to the next one With AIE Which you didn't even say What it was in the intro Well it's the AI engineer conference In SF It was way bigger this year Than it was last year Very very fun Just like 7,000 people Giant expo hall Tons of companies Four days Yeah in four days That was quite a long four days It was very fun But very exhausting It was a lot of people And it was actually really cool I expected good things Because the previous ones Have been great But this one was like A tier up You can almost say It's like the Opus 4.5 To the fable 5 Of AIE Yeah I feel like they've got One more jump But I'm like This is probably closer To a sonnet 3.5 To an opus 4.5 Like this is Okay The conference is coming Into its own And I feel like Honestly it could grow More from here It's definitely went From the little Like last year was I don't know how many People were there But it felt a lot smaller Than it did this year Probably like One to 2,000 people This was enormous It was really cool Yeah it was absolutely massive And we had a lot of fun with it I was so surprised By the turnout And how many people Wanted to hang out with me That I couldn't figure out How to fit all of that in Especially because Fable 5 dropped In the middle of it So I decided to See if anyone would host me And I put out a tweet Like hey Does somebody want to host A meet and greet with me At their booth And of all people Open AI Reached out And let me take an hour At their booth To just like hang with my audience And answer questions And apparently it was massive A face Sneaking photos of the event I don't know what we have But they should be showing here now If we did this correctly We got a bunch Yeah cool Yeah there was a lot of people Way more than I expected Thank you all for turning out I'm sure some of you guys Are listening now Some of y'all even recognize This dumbass Yeah I know Yeah I'm sorry about that Y'all listen now Crazy Yeah a handful A very small number of you Realized that I was basically Impossible to get the attention of Because there were so many people Swarming me And that Ben was easier to bug So you guys cornered Ben and Julius In the back instead You guys know what you're doing Yeah good work Honestly Julius was the more Interesting one of the two of us I would rather talk to Julius Than me personally Same but I also am paid to talk To you on camera a lot Yeah that's fair Yeah so that was super fun I was surprised how big the Turnout was for that I was also surprised with how Well my talk went Because it was kind of last minute I agreed to it like a day or two Before the event started Because it's what I asked And I was like yeah sure whatever Then he gave me the challenge To aim for one of those Million view talks I've had a few of So I I overthought it I underwrote it And I think it went well What are your thoughts Yeah it was good I enjoyed it Just hung out there Watch your talk Overall I didn't watch too many talks I usually don't go to too many talks At these events I'm mostly there To just hang out with And talk to a bunch of people But of the talks I did watch It was I hate giving you this credit But it was the best one It was the best talk that I watched It was good You should go watch it What was it my best talks You've seen most of them Um I'm trying to think The ones that matter Are this one The Cascadia one It was definitely better Than Cascadia And the Elixir Conf one Which is one of my Personal favorites I think the Elixir one Was slightly better But it was longer And deeper From what I remember And much more personal Yeah That one was about me And I just like Elixir was the through line I agree I enjoyed that one more But I think for the current State of the world Right now It was a very good talk I would put it as Your second bust So good job I'll take it It was my most important For sure Yeah definitely Yeah Give that a watch When it comes out Remember I gotta get a million views In order to make this deal I'm not being paid Or anything I just This is for my ego I need this one guys Oh his ego Really needs help guys Trust me It's really suffering Like what is he ever Gonna do Remember My opening For a lot of events Historically has been Theo the second best TypeScript YouTuber And now we both Moved to AI Yeah And he did two talks Both that got a million views Thank you Matt Pocock You're still always my hero Yeah We'll see how this goes Yep we will Yeah My favorite thing is that like And I realized after Because I went through And watched all the top talks From AI There was one that An open AI employee Did about like Spectrum and development And how like The prompts matter More than the code And you should be Committing the prompts Instead of the code And then Matt's second talk Was indirectly Tearing that one to shreds Like he didn't cite it Directly But he's like I hear a lot of people Talking about this Spectrum and development stuff Here's all the reasons It's fucking stupid And it was really good Yeah Matt's still one of the most Like realistic Hyper AI pilled people And I'm thankful To have him in the space Yeah he He wasn't at AIE this year He had a personal thing Come up He wasn't able to make it But he did post What would have been His talk online I watched it this morning It was really good It was on how to make Competent skills Which is a weirdly Difficult thing to do The short of it is Don't AI generate them Which sucks Like I would love to be able To AI generate them But having looked through All of the AI generated skills That Codex has made Over the last few weeks They're not good They are not good at all Yeah I was filming a video Earlier today Where I was going through Some of the skills I made And some of those skills Were made with Or largely by AI I think it's fine If you go through And you clean the fuck Out of them And like delete the majority But the big thing AI is helpful with Isn't necessarily The writing of the skill It is the going through All of the logs From all of your machines If every time The skill was called To see what it does And doesn't need And like where is it struggling And then adjust the skill Accordingly Like if you write the skill Once Use it And then never touch it again You're probably not Writing skills well They're much better If you treat them As like a malleable thing That is adjusted Based on what behaviors You see that you like And don't like And more importantly As new smarter models Come out And you can clean them up And have them Maybe not even need To exist anymore Like sometimes When a new model comes out I go and just delete All my skills So that I don't have them Clogging up the context And I can better understand How the model behaves By itself They're in a weird state I think Matt Pocock Made a really good point In his talk Of like When you're letting AI generate these skills A lot of what they will do Is create these like No-ops effectively Where it will just Add in an extra line To be like Be really thorough And I've even found myself Doing this and prompting You've yelled at me For this before Where like I'll just add extra words That just are not needed The model is still The model is gonna Thoroughly review it Whether you tell it to or not That's just how These things are trained You just need to Cut as much out of it As humanly possible And make sure it makes sense And then you can let the models Evile the skill By running it repeatedly And also like Let's be real here The point of a skill Is to give a model behavior That it doesn't have By default If it wrote the skill Then it's writing things It has by default That's how that works The point of a skill Is to steer the weights In the direction You want them Not in the direction They're already in Yes exactly So one of the best ways To make a skill Is to go through Your history Of all of your usage With a model And ask it to find times Where you had to correct it When it didn't do things right And categorize all of those failures And then look at those And think Okay How can I make a skill That steers the model In the right direction With these things instead A lot more people Need to be spending A lot more time Analyzing their histories I know you suck at this And don't do it Anywhere near enough I'm doing it infinitely More than I used to But I still Probably don't do it enough It's like from zero times To twice a month No I'm doing it like I was going through My histories recently And it was pointing out How many times I've gone through my histories So I'm doing it a lot Like I'm doing it Probably every couple days Thereabouts Especially as the models Are changing Like now that Fable is here And I have that back Been using that a lot To clean all this stuff out Have a lot of crap To just remove From the 5.5 age Of my codex instance To get it ready for 5.6 Hopefully very soon There's just a lot of crap In there And I also think On the skills topic Like it is really good For beating behaviors Out of the models But a lot of that Also can just be done Within very simple Agents MD notes Like there are a lot Of things where like Hey don't add regression tests For when you delete A function It's okay The function is deleted You don't need to add A regression test for that Little things like that Should probably be Agents MD Skills I more think of As detailed specs Of like how to use A certain technology Or framework Like I pulled a bunch Of skills out of My fable history On how to use Svelte and effect More effectively Like get those Correct syntax And the correct Modern patterns For the beta And the new remote Function syntax And all that They're also really Great for workflows But we've talked About that in the past With like the G stack stuff And all that Speaking of which Gary had a great Talk right after you I'm sure you're honored How many times did he Mention me in that talk Two or three I thought the number Was three plus But I'll take it I don't remember The exact number But you did come up Multiple times Which was very funny Although like Boy I mean like Actually I haven't Posted the photo yet Phase please sneak in The photo of me With our man Gary Yep For those of you Only listening This is a photo Of me wearing a shirt That says I got AI psychosis from G stack Standing next to The deity himself The creator of G stack I'm gonna push back A little bit On your skill take Here because the Way I think of it Is when I notice A model behaving Wrong in a specific Thing like when it Calls codex and the Formatting is incorrect Giving it a skill To in those moments Give the exact Right context to Just prevent that Failure it's like The errors in Calling a thing Are where I get Most annoyed but I don't want to Give all of that Detail in the agent MD because then It happens too often That I agree with That is that is More of like There's like two Types of skills here One would be this Where you're Effectively describing The behavior it Should use so Instead of just Having it suss out How codex should Be called or Whatever or Dumping all that Into the agent's MD you just Have the description Be like okay oh You need to go Use a codex sub Agent now here's How you do it It loads the skill Gets the actual Code that it Needs to run or The I guess bash Scripts that it Needs to run to Spin up the codex Sub agents what Model it should use What reasoning it Should use what Speed it should use All those little Things that makes a Lot of sense I'm curious actually If you're down to Open your dot Cloud how many Skills do you Currently have I'm very curious I have not looked I was looking Earlier and it Was a very different Number than I Expected Code do I not Have that on the Path I'm just Setting up a new Laptop right now That's you got an SSH into another Machine and just Like they're all Same They all have the I just need a code Editor and I don't Have one right now So he has all of His agents configured He has all of his Like terminal All of his SSH and Everything configured But he hasn't Installed an editor On his computer yet If you didn't realize Where we are now It's here In my defense I do Have one it's Cursor but I have Cursor set up for Like glass I've been Just using their So you can read Code and cursor That's not a code Editor that's a Code generator Yeah but that's What cursor's turned Into so now I have VS code as my Code editor and I Don't have either of The bash scripts Bound for this What is even the Thing you're supposed To run for that Open your terminal Ben Yeah I'm in my Terminal right now Type in ls Space You're very smart Give me one second Where do they Store them in Dot clod It's skills Dot clod Slash skills Oh I've got them Alright how many Skills do you Do you have Five I have one Two Three Four Five Six Seven Eight So first one This is like My html markdown Generator skill That should be in Here my network Skill which is just Telling my agent How my home Network and all My tail skilled And devices work For my linux boxes Talk about that In a future episode That's gonna be Really fun Stuff on how to Use effect v4 That should be in There I have two For like workflows Effectively on how I want new features To be implemented Because a pattern I do a lot Is I'm like Okay I have This project I need to make This change Or whatever And there's a Very pretty Repeatable workflow I have at this Point of here's The thing I want To be done Here's the branch I want you to Make the work Tree off of First you spin Up a planning Sub-agent Then you spin Off an implementing Sub-agent Then you spin Off a review Sub-agent While you are Going up And waiting For the PR Review bots To do their Thing Bring all those Changes together Make sure it's Working Get the PR To go green Then tell me About it Like I have Written that Prompt 15 times I just got to The point where I'm like man This should just Be a skill So I just Can just do Slash PR Review or Sorry Slash Monorepo Feature fan out Or whatever And they'll just Do that for me I think there's A babysit skill Built into enough Things now And if I tell Any agent to Babysit the PR And address comments It's fine They do But this is for Starting at the Very beginning Like I start From a new Session on Main And I'm like Okay I need to Add this feature To pick thing This is the Feature that it Just goes and Does it The exact Workflow that I Want that I Have used a Dozen times In the last Two weeks I should Probably codify Mine Yeah That one has Been really useful The rest are Just nonsense Like I have a Video skill for Like I was Screwing around With Claude That should Not be in The global Directory That should Have been Scoped to A project And that's Basically it How many Skills do you Have in your Hermes agent Though Oh well over 100 I've Got it a lot Like I've Got it 50 Out of that Thing and I Still have an Absurd amount Yeah I don't Even want to Know how many I have in Mine and I've Trimmed two Yeah my Skills I Should only Have three But I let My generative Like codex skills Like make Three separate Ones I don't Care enough to Deal with it I'm just gonna Delete them in The near future I'm sure when Five six drops But we'll see For now it's Pretty useful Like having those There they use Railway skill Snuck in there When I install The railway CLI which is Still pretty dang Useful I can I'm surprised how Much I can get Done through that Although I still Have the computer Use a little Too much for Railway I'm Convincing Jake To fix the Remaining pieces And then I Have the HTML Planning skill Because Thoric Was right I have one as Well I think we Both put together Our own version Of this where We have a hosted Box that can just Send the plans Up to and then Print them out For us wherever We want Mine's simpler And better Yours is fancier And cooler It's more Opinionated just Because I don't Like the way Most models Write pages And they're not Very readable So I wanted To prescribe The way you Style paragraphs The way you Put the fonts The way it's Generally structured Give it some Little building Blocks to work With so I get A more cohesive Output every Single time I like it a lot Better He's covering His steak in Salt and ketchup I don't think That's What's a good Analogy here Like you also Use models That taste like Cardboard a lot So I understand But we use The same models Sonic 5 I Sonic 5 is better Than you're giving It credit for No it's not Yeah I gave it a lot More credit than Most but it Still tastes like Cardboard Yeah I'm Am I using it Right now No but like It is better Than most people Are giving it Credit for Like I My defense of Sonic 5 is Not that it is AGI or it's The thing we Should all be Using Nah Okay fine We'll talk about It later Alright back on Topic What's the topic We're supposed to We're talking about We were talking About AIE and It pivoted to Skills I think that was Useful Yeah I hope so These are the Tangents that the Show was built for So hopefully you Guys enjoy these Too Yep On the topic of Models that Tastes like Cardboard we have Alibaba and Anthropic getting Into beef because In case you didn't Know Alibaba Makes models and They are the Alibaba flavored Versions of other Companies models That's just kind of What they do They just sell on A lot of stuff Whatever they can Find and to Their credit Getting the data Anthropic has is a Lot harder now Because both Anthropic and Stolen credit cards Or what not Might not even have Been Alibaba doing This by the way It might have been A third party that Did this and then Agreed to sell the Data to Alibaba And usually their Way of getting the Data is to offer Cheaper subs to People in various Places so like if You're in China and You want to use Quad code with the Subsidized usage you Can't get an account Signed up there but If this random Sketchy service Offers you a $200 a month Account for 50 Bucks and you use It the only catch Is all of the data All of the inputs And outputs all the Back and forth is Going through a Forwarder that makes Its way over to I Don't know Alibaba All of a sudden they Get a bunch of data You get a discount Anthropic loses they Don't like that Yeah I don't know if You guys have seen These things but I've Seen some really Wild screenshots of It's like it's Literally Alibaba Open router where It looks like open Router and you can Just pick from the Models and they're Massively discounted Clearly it is just For this purpose These are just data Aggregators who are Doing a bunch of Sketchy stuff behind The scenes to route Between different Claude code subs I think that's why There was a post Recently of Anthropic has Embedded hidden Spyware like code And Claude code That covertly Targets Chinese Users and then Sends information Regarding every User by injecting It into their Prompt message and It shows this Section where it is Basically just Checking for like If you're in Shanghai or in Asia or all These different Things to try and Get some way to Check whether or Not you're doing This because these Behaviors are very Hard to track Thorig responded To this and said Hi this is an Experiment we Launched in March That was meant To prevent Account abuse From unauthorized Resellers and Protecting against Distillation the Team has landed Stronger mitigation Since then and We've actually been Meaning to take This down for a While we merge The PR and This should be Fully rolled back In tomorrow's Release so They're walking It back but Also they're Almost certainly Doing stuff like This behind the Scenes at the API level to Try and classify This stuff you Do know that Our little slop Proxy is going To get hit with This in the near Future probably Right I was Thinking a lot About that in The way I'm Setting it up And like I am Only ever using It through my Cloud code Instances and Is only ever Going through Like my local Network which is Very clearly in The US like It's not going Through VPNs For context for Those who have Not kept up with My recent videos Both Ben and I Are juggling Accounts because We want to Maximize how much Fable we can get In the short Window where it's Included in the Subs but also in The future when It's more included And there's a Really cool open Source project called Vibe proxy that Let's you switch Over to API Style cloud code And it routes To this Vibe Proxy of somewhere On your network Or tail scale That will reroute An auto balance And load balance Between different Accounts on Cloud or on Codex and other Things I cannot Use this in Codex because it Breaks the codex Stuff up in so Many obnoxious Ways but in Cloud there is no Functional UI so You just use the Terminal version so It's fine there and It works great over T3 code so I'm Pretty happy with it That said it's Using the API Style when you go From cloud code to Vibe proxy instead Of the traditional OAuth style you get With cloud code the Endpoint it's hitting In the end is still The OAuth endpoint That you are Requesting from when You do cloud code Because cloud code Doesn't hit the Normal inference Endpoint from Anthropoc and it's A special one just For cloud code you Know how easy it Is for them to add Any form of Signature that's Auto rolling and Generating inside of Cloud code directly That is like attached To the session your Time stamps and all These other things and Have that as a Required salt that Has to be part of The request for it To honor it yeah You know that that Gets turned off Immediately when Switched API mode Yeah you know that What we're using right Now is going to be Killed in the next Two to three weeks Yeah well thanks Alibaba to be fair This has been around For quite a while but Yeah they're just Going to keep cracking Down on this stuff it Started back with like The open code ban way Back when they just do Not want people using Their cloud subs in Anything other than Cloud code and to be Clear that's what we're Using this for I am Not using this to put The inference in any Other provider or any Other harness you Could but you Shouldn't because you Are heavily risking Getting banned there I'm sure they're very Trigger happy with that I'm hoping that this Is okay for this Instance I'm probably Going to turn it off After this because I Don't really need it Post fable I'm just Going to go back to One sub who cares For now it's useful Though I might go Back to zero at this Point because I don't Really care enough about The other models you Get through the Cloud code stuff I Just care about fable And when fable's gone Like I'm not going to Stop using opus I'm Just going to move to API prices for fable Yeah it's kind of True yeah like opus Is not interesting enough To me to maintain a 200 dollar a month sub Hell no maybe opus Five will be Interesting but again I Word opus five might Actually be a pretty Useful model I hope so I hope it strikes a Good middle ground in Between mythos and Sonnet where like Sonnet has the cool Workflow capabilities of Mythos without any of The intelligence or Discernment or Capability so like it's Not actually all that Useful for anything real It's more just a demo Of what is possible with The bigger better model I would love to see Those mythos workflow Long running capabilities And just general Discernment be brought Into an opus even if It's a little bit Stupider which it will Be that might make the Sub worth again They're just I'm just Going to wait till they Have fable back in the Sub it's going to Happen soon yeah I Actually you haven't Seen my video yet so You don't know how I Framed all of this maybe We'll talk about that in The mythos part later on Because right now we Need to take a quick Break for today's Sponsor the one I've Been talking about loops A lot more lately and I want to use this ad For post hoc as an Example they're an all In one suite of product Tools for product Analytics and session Replays and a ton of Other stuff that we do Not have time to list Here but recently they've Been focusing a lot on Making your products More self-driving and Their self-improving loop Here is an incredible Example of what loops Actually are and why They're useful the way This kind of works is You start with signal Sources which would be All of the data that Post hoc already has About your product and Then they take those Signals they group them Into reports and then once They have those reports Post hoc agent will go Through and open up a PR to your actual like Project and repository Making a fix based on The signal it got so Say for example in your Onboarding flow there's a Specific section where Users just keep churning Off over and over and Over again and the Agent will go through View all the traces and Find what it believes to Be the underlying root Cause of this it'll make A PR for this you'll Review it you merge it Then post hoc from There will measure the Results of that change And see how it impacts Things and loop back over To get new signals and run This loop over and over And over again it's an Automatic process that Runs in the background That just kind of Naturally makes your Product better and post Hog is just uniquely Situated to do this they Already have all the Information they need to Get these good signals They're also insanely Cheap their pricing is Absurd it's the only Company I've ever seen Do price decreases Repeatedly and they're Also the company with my Favorite MCP server if You haven't tried the Post hoc MCP you really Should it has everything You need to build these Self-driving loops on your Own if you don't want to Use their built-in solution It's an incredible system That will make your Product much better at NerdSnipe.link slash Post hoc I think we got One last piece of news Before we're allowed to Just nerd out about Fable for the rest of The video open AI is Proposing handing Trump Administration a five Percent stake in their Company if you're not Sam Altman you can skip This section but yeah This this one's for you Sam I just wanted to Personally thank you for Doing the thing nobody Else had the balls to do Kissing Trump's ass Aggressively in the public For everyone to see it is A shame that we live in a World where this is Necessary but we do and if We want to continue having Access to the best models In the world then we need To fight to stay the best Country in the world in Order to do that we have To convince this I'll be Polite here well-intended President from destroying our Entire economy with his ill led Goals of trying to get every Single company to love him we Don't have to love him he Just needs to feel loved and I am very thankful for you For being the one person who Is willing to make him feel Loved like this other than Tim Cook the two of you are The only reason our economy Hasn't collapsed yet and for That I am genuinely thankful Everyone else you can come Back now sorry sorry we had a Personal message for that I Hate that that message is True I really hate that that's True you were you weren't Supposed to listen then that Was just for Sam oh sorry Never mind though yeah Giving the government five Percent of one of the Fastest-growing companies of All-time is the most Performative bullshit I've Ever seen in my life and I've Seen a lot of performative Bullshit I live in San Francisco yes anyone who's Like mad about rainbow Profile pictures in the month Of June should be outraged by This because this is a hundred Times more performative yeah What is your over under on This actually happening 60 40 happening really I'm Leaning ass what is your Take on this I'm very curious How would this actually work What would the government Actually get out of it What the fuck they do with Intel yeah I guess fair I Forgot that they did that With Intel yeah yeah it's Ten percent with Intel Jesus Yep they own ten percent of The fourth most successful Chip company what a world We live in penny them Inside yeah I guess Penny them inside pennies Outside I I don't know I Do not like the idea of Handing over private companies Any percentage to the Government but in this This bizarre shitty Situation we're in right now Where the frontier frontier Models are getting restricted Because of very very stupid Government overreach we are Now dealing with the fact That in order for these Things to be released to the Public in any reasonable Time frame going forward There's gonna have to be a Lot of collaboration between The governments and the Labs to like what if this is What's happening with GPT-5-6 What the hell is gonna happen With GPT-6 what's gonna Happen with fable six like is There I hope from what Anthropix said we are getting Some semblance of a process Put in place so that this Can be expedited but we just Need to get the damn Models faster two things Have to happen in parallel And by the the nature of The market they already Kind of are we have the Ass kissing which open AI Seems very well predicated And prepared to do Sam is Incredibly experienced in This and the nature of the Weird board structure that Already exists at open AI he Is very used to making People who don't really Understand what's going on Feel important and like They're part of the Decision-making process I Would even guess this is Part of why five six is Restricted they could very Easily have put it out and Then waited for the Government to restrict them And got a bunch of good Free will but instead they Let the government decide And are letting them do This performative like they Choose who has access and Doesn't the whole point of This is make the government Feel like they have the Control at the admin feel like They're important here so That they can go do their Thing they need to do this All for long enough to Allow for people who Actually understand to build Better processes and systems That can then be handed to The government is like look You guys trust us now Here's the solution to the Problem please stop Fucking with us and I Hope hope pray fingers Crossed that we can figure This out fast enough that If probably can make good Safety systems and Protocols that make sense For some type of Enforcement as Sam Altman Kisses enough ass to hold Us over until then yeah and Credit where it's doing Anthropic will probably make Some really good safety Stuff like they are they've Been focused on this stuff For a very very long time It's really funny that like Anthropics the government Went harder on AI safety Than Anthropic did like Anthropic was like oh Yeah we got to have all These regulations and make Sure that it's super safe and We have all of our tests on It and we have the safe Models and they did the safe Model thing themselves and The government's like oh I Don't know if that's safe Enough and then took it down For two weeks and now we Have to get a new basically Take whatever Anthropic is Doing and nationalize it and Make it a new benchmark for Every single lab to be clear The government said I Don't know if this thing That multiple open weight Models can currently do is Safe enough to have in a Frontier model from a Trillion dollar company that Was worth less than a Tenth that much this time Last year so we're going to Risk the entirety of our National economy in hopes of Performing in a safe way Mostly because they hate the Leftists that work at Anthropic yeah I hate Politics so much dude I Hate this so much this is so Stupid why are we why why Do we have to waste time on This garbage just release the Model man please Yeah and in the end as Much as I hate the company Having some ownership in the Government and this being Necessary for things to move Forward I'm not sure for a Company that is five percent Owned by America win than a Company that's owned one Hundred percent by China yeah Absolutely I mean it's Annoying that this is Necessary but this is part of The game and this is the Thing that Anthropic has Fucked up for a very long Time Dario is great at Talking to AI researchers he Can communicate very very Well with them he cannot Communicate well with anyone Else and when all this stuff Happened and they like brought This little exploit to them Quote-unquote the stupidest Jailbreak I've ever seen the Government freaked out and Then when Anthropic was like What are you talking about You're stupid this makes no Sense and then just like Ignored them yeah I think I Don't think they ignored them For that long they just like Didn't instantly send ten Thousand people over to go I just had a funny Realization if this five Percent goes through and it Is done in a way that is Meant to be like citizen Ownership where like we get A tax break or a payout or Something Sam Altman will now Have real equity in open AI For the first time and not Only will Sam Altman have Equity in open AI Benjamin Here and I as well as our Producer Alyssa will all have The same equity in open AI That Sam Altman does we're in This is so stupid yeah I'm Sure it'll be worth a lot of Money and it's going to be Very well spent I'm sure I Cannot wait what's five Percent divided by 350 Million I enough to not do Anything actually curious let's Let's assume they hit the Trillion Val uh-huh we'll do Five percent divided by 350 Million that's 143 bucks for Every citizen wow thank thanks Thank you thank you Donald J This is your finest achievement I am so glad that I don't have Five six and instead we'll get Almost one month of a chat GBT pro sub we also just passed Midnight which means today is Now the 4th of July it's America's 250th anniversary Hell yeah and our reward is Potentially 143 dollars of a Company this is one of those fun Times where I have lots of things That I want to crash out about That I'm not going to crash out About because that's not what This is about yeah we have better Things to crash out about I think It's time to crash out about Fable indeed it is yeah I'll let You start this one because I I Just spent an hour and a half Plus recording my thoughts on this Model and you haven't done that Yet today have you done that yet Like since the dropped I haven't Had time to like we were at AIE Every single day this week and Then today was just like prepping For this and then making sure That everything was in place and Then tomorrow we have to like go Touch grass and then probably After that I'll get it done so My stuff is super delayed which Is really annoying but can't Actually talk about it now and my Opinions on fable is that it is a Phenomenal model and I am very very Happy to have it back I did not Do as much of the like sitting And prepping work for fable as I Think a lot of other people did I Kind of just like brought it back And threw it back into my Workflows and the thing that I Instantly noticed is like this Thing is just so damn good at Understanding what I'm asking it To do and just making the thing Happen it's just kind of like going Back to just chaining like five Five and opus together I have so Many hermes agents at this point For different things I want a Clean centralized way to work with These I had a really crappy CLI That's opus put together that Just was not very good it Functioned but it wasn't great on You run one command it spins up a New hermes instance that has Hermes codex cloud code executor and A couple other random things that I Need all in one container that I can Just turn on ssh into get everything Logged in and set up and then now That is a hermes instance that I can Stick on any box I want the way that Was set up before really sucked I Spent about an hour with fable going Back and forth on the Q&A with it did A ton of PRs changed the entire Architecture of the entire thing and Now it is working great very excited To release that in the next couple Days same thing with all like the Random one-off things that I have for Just using these agents day to day My html page generator haven't really Had to touch that the drive for Agents thing that I have which is A cli where the agent can dump a File into it and then upload that Into a drive that I can then send a Link to myself or another agent it's Just nice having an ability to pass This stuff around with an Anonymized link really like that fable Was great for just throwing that Together quickly it's just it's Worked and I have been hit harder On the limits this time around I Think I got lucky where I got Effectively three resets worth on This because the new way they're Doing it is we only get one week of It it's going away on July 7th Which is Tuesday I was lucky with Just the way my timings worked out That I had mine reset this morning So I burned that entire weekly usage And I have one more weekly usage on That and then I have the weekly Usage on my other sub so I got three Left or I had three I burned two I Have one left is your other account Maxed out right now I won't reset in Time yep jealous that you got that Many maxes because I had my reset for My main account happen the day after The wasn't they have a day off I Think it was the date of the drop yeah It was like I remember this was Really late at night the day of yeah It's like midnight ish so I grinded as Hard as I could to try and max it Out on that time because you can max Out the weekly limit in one day with The five-hour limits it's four five-hour Limits to max out the weekly roughly Right now on cloud code but even worse Since the mythos limit is half of the Full weekly limit you end up with like Two to three max outs in the hourly in Order to max out the weekly I was still Only able to hit like 40 ish percent And I got a shitload of work done with That too like I ran the numbers it was Like for that it was about 200 250 or so Dollars of fable usage and I got so Much done in that around 150 of that was Both from fable and from the other Models I had orchestrating I got through In cleaned up or closed over 30 PRs in Lake bed I had 18 that were sitting There just needing something and it was Able to in a loop for five hours address All of them merge some close some recreate Some and get a ton of work landed like in A truly absurd amount because it could just Understand what the intent is and go and It can pull intent out of things which is What I really like about it to the case Here I'll I guess I'll dive in a bit more I have a whole video about this coming Up very soon too it's probably actually Out by now a face hint at that on the Screen somewhere if it is the thing that Makes this model magical is that it can Build the right context and spin up the Right workflows to solve problems instead Of having to hard code the steps and Stages of your work you can give the Model what it needs to figure that out talk With it about how you want to get the work Done and then let it go and it will mostly Rip without needing much intervention I had a Bunch of unfinished work I was sitting on in Lake bed because I we lost the model I Wanted to use for it and other things that Will be talked about in the near future so I Have all of these PRs that were kind of Stale some were finished but needed to be Rebased others were just trumped by other Work that merged some of them were just slop And nowhere near ready to go all of them had Enough context to know what my intention Was when I was starting to build it but a lot Of them were just sitting there in draft Stages or otherwise dead I didn't feel like Rebuilding all of the context for all of These things so I asked the model to go Through all the PRs and figure out what They should be categorize them it said like These three overlap but also this one Merged before that has the same rough feature So these three probably aren't needed anymore These two also are roughly the same feature But this one is better for these reasons but The other one actually has a couple nice Things in it too this one's just entirely Slalom should be thrown away but the Feature and it was actually cool we Should probably rethink this one and it Went through all of that by itself it Wrote up its thoughts on all of the work I had in flight I left some comments on A couple pieces I said throw that one Away I got it to make a new plan for All the work that wasn't mergeable or Rebaseable and once it had the plans for All of that work I read them briefly is Like yeah this sounds pretty good you Know what YOLO slash goal complete all of The work we talked about in this session Get it all landed and it did I watched Fable close like a dozen PRs and then Land like 14 all in one loop and to be Clear if it's not like an auto deployed A prod thing I have main fully separate From broad and once I am happy with the Stage of main I will move it to prod with An action that only I can run but this Means I can just land shit in staging and Before I merge it ask multiple agents Hey compare this to what's currently in Prod help me figure out what the Potential failures are in doing this test The shit out of those and then merge It's been a great workflow did the way You can work with PRs once you get Fable back is so nice I missed it so Much I have a lot where like I am now Doing three or four kind of like three Or four different changes on different Work trees in one project like I have For YouTube studio analytics we have The actual YouTube studio thing but it's Not like the UI there is fine it's not Great so I wanted something a little Cleaner that just shows me the things I Actually want I want to get shorts out of That UI it sucks but the biggest thing I wanted out of that is I wanted a CLI and MCP server that I can use to give to Hermes to actually get analytics data Out of this so I've had the CLI the MCP The UI and the mobile app all building in Different branches at the same exact time With Fable and then after all of those Completed I could just merge one in There'd be tons of conflicts and weird Issues and it would just clean up the Next one and it would like almost go Through this auto merging queue where it Would just merge one figure out what the Next one any issues that one had with the New main branch or whatever clean that Up merge it do that again and again and Again until it was ready I oh it's just So good it is so nice to have back I Think it's time to talk about the limits a Bit and yeah they're not I'm gonna get a Little conspiratorial with this one okay Might be beer time I needed this I don't Like getting too conspiratorial without a Beer because I start going actually Insane and these ones are as fun as they Are insane yeah all right let me hear it So as we've already established we only Get 50 percent of our weekly limit in Fable when we're using the clod code Subs we're also losing it on July 7th you Remember what the previous window we Had for testing was I think it was two Weeks it was June 9th to June 23rd I Believe it was up from the 9th to the 23rd we got from the 9th to the 12th And then it got banned so instead of Getting the 14 days you're supposed to we Got three we got nada and now we get Seven but with a half limit what went Wrong here uh the other weird part of This that you didn't bring up is that Currently rate limits are doubled on Cloud code until that doubling had Happened slightly earlier it turns out That happened in May oh so that's been Going on for a while okay so these are On the same barometer here so we're Effectively getting three extra ish days Here because we got three days and then Now we're getting seven days at half which Would be the equivalent of getting seven Days of the original 14 days so we're Going to get 10 days of this instead of Getting the full 14 days ish not quite So if you were to have maxed out and we Count a day is a max out or a week as a Max out even we went from what would have Been 14 days of full usage to three days Of full usage and seven days of half Usage yeah so effectively we went from 14 days to seven and a half days it's the Same total amount but instead of having Four no it's way less it's comically Less because we had 14 days where you Would have had two to three recess oh yeah I'm an idiot yeah I yeah because yeah That first week we would have had way More yes yeah and the second week we Would have to yep the week we have now We get half as much usage as in any Seven-day window previously we went from 14 days of full usage to seven ish total Of usage why this massive a reduction They don't have the computer run this Thing like their servers just can't Handle it I'm sure they're on fire right Now so why give it to us at all because They want us to taste it and they want Us to want it back obviously but there's More to this than that how long does it Take most companies to adopt new Technologies uh depends on the time period Historically a very long time but Recently a lot less this yeah it went From like months and years to weeks still Months at some companies but yes in General weeks is reasonable would you Say it is fair that if a company like I Don't know anthropic had a ton of gpus That they recently acquired that they Might want to allocate those for Businesses but perhaps those business Would take a little bit to adopt so What are you gonna do in the interim You have this massive cluster of fable Boxes and you don't have enough Enterprise customers that are ready to Adopt this yet especially with the new Security concerns and whatnot takes at Least two weeks to get big enterprise Adoption at that point yeah so in those Two weeks give it to the subscription Users that don't have those restrictions So they can use the shit out of it they Can burn they can do the free marketing Most importantly they can get data I Think that previously anthropic had Planned for it to take two plus weeks For most businesses to start adopting Fable and using those limits heavily so In that window they could do a really Nice two-for-one they could give it to The subscription users which more than Anything is free marketing because They'll all go online to be like oh my God this is incredible like you know yeah We're doing right here yeah yeah so That was part one obviously but my Conspiracy is they had nothing else for Those gpus to do because the Enterprises take too long and they Also want to get a better idea of what It looks like for the subscription Users to use it because once they know The patterns of our usage they can plan And provision accordingly that's why They've previously done things like Massively reducing the five-hour limit During the working day because that's What enterprises need the most and then Leaving it super generous and Subsidized during like late night hours Because the businesses aren't fighting For the gpus but in order to Orchestrate all of that they need data They need data on usage patterns and They need data on usage patterns for Really specific amount of time and That amount of time is not three days It's also not 14 days you know how Many days of usage you need to get a Good read on how to provision mmm Seems like 10 no seven just you need One full week yeah you need to see how The rates go up and down throughout the Week that's why we got exactly seven Days because the three days at the Start was not enough and even one Week wouldn't have been enough at the Start because it takes a little to Ramp up and for hype and for you to Get ready to use it mm-hmm they got That window because we had the three Days and then we lost it so all the FOMO got built oh yeah so everyone Who's going to burn on this model was Ready to go when they announced the Day before it's also part of why I think They announced the day before so we Would all get ready to go mm-hmm Because they want us to push the limits As hard as we can for these seven Days because they want to know what the Worst case looks like on a traditional Week so they can provision accordingly Yeah we're all the guinea pigs here and They also just get a 10 days of free Mythos traces out in the world that They now have for training data oh they Can't train off of our usage they can't Nope really do you have the checkbox Checked I have not seen it and the fact That I like does it default on I think It defaults off but you can check yeah It's under privacy there's a help us Improve our AI models and it is off Yeah it's on on my new account it's on And I know it's on by default that fun Yeah it's on by default like there's no Chance in hell it's not on by default Come on now there's no way that they're Gonna like not turn that on come on but They're the company that cares about Safety and privacy okay it's off on my Alt account I wouldn't have turned that Off myself it might have been during Onboarding and you just like hit hammer The next button no I just did the Onboarding three days ago and it was Not in there how many people went to The settings over the last few days To add the know your customer in case It was there so that if it does become America only they can use it they've Probably more people in the settings Of Claude recently than ever anyone who Didn't have that check that matters Has now checked it yeah no it's it's Definitely on by default it's on on that Account and let's be real the best People are the ones who have that Flicked on the worst ones are the ones Who never check yeah sorry I mean I don't Just like yeah sure the best ones will Not have it checked on but like at a Large enough scale come on how many People out of the tens of millions who Are using this right now yeah it's Somewhat useful but like what do you Think is more useful the provisioning Absolutely knows that between like the History of the people who accidentally Left that checked on as subscription Users or the people who are adding Anthropic that have all of their Infants going in a pot anyways I mean Yeah inthropic is more useful there But like it's more just like it's an Extra it's a sweetener for them get a Bunch of data but it doesn't hurt them The data that matters is the usage Patterns more than the usage itself Agreed yeah so that's my conspiracy I Think in a public pick seven days very Strategically because we have the hype Up window and now they need a clean Week which honestly isn't that clean Both because of aie but more Importantly because of the holiday that Is tonight we have past midnight it is Now fourth of July those do Complicated a bit but weekend patterns Are always a bit weird and they just Need to know like on weekends and Holidays do we burn more or less And how does that compare to the week Usage so if you are one of us Fellow sub users and by the time this Is out it's probably too late make sure When these types of things happen that You burn hard on the weekend and if you Can avoid a little bit during the week So we get more generous limits and In fact we can buy more GPUs it is Strange like they literally do not have The compute to serve this thing at the Scale that they want to which is Going to be a kind of fascinating thing In the long term of if this we don't Know how big the model is I don't Think that's confirmed publicly Anywhere but it's probably five to Ten troll params like it is clearly An enormous model and clearly a Massive step forward in that Direction bigger than anything that Is publicly available anywhere else And they're having a lot of trouble Serving this thing which means that Like long term with the scaling stuff Clearly scaling does seem to be Working bigger model is better model If eventually we need to get to a Hundred troll per am model they do Not have to compute to serve that Ready for a controversial take on This sure I think they underpriced it Is it supposed to be 125 out right yeah It was something like that 125 per Mill out yeah and now it's 50 that Means it's like 30 to 40 percent More expensive than opus per token yeah And it's less token hungry than opus Yeah so it's often cheaper than opus For a similar quality work yeah it Ruins opus oh yeah opus at least until We get opus five is very very dead right Now I think opus five has to be Cheaper too I don't think opus five at The same price makes any sense right Now that those is too cheap well here's The real question would you pay api Price for it I'm going to I've Already accepted this the way I Phrase this earlier with the like prs I had 18 pull requests on my project And for 150 dollars I was able to land Close and or replace and rebase all Of them and get all the work that Was in that queue landed if I had a Magic button at the top of every GitHub repo that I could click and 150 200 bucks would disappear from My bank account and in five to ten Hours all the work is landed I would Every single serious company in the World would as well mm-hmm yep yeah I Think this is probably because it's Been weird for a long time that we are Not paying api prices for any of these Things like we are you know would you Pay api price to burn three grand of Five four probably not I don't think Anyone would I wouldn't spend three Grand to burn on much in like one Thing because and this is the other Interesting problem the higher the burn Like the more expensive the task there's A very strong point in the curve where Like you spend a little more it gets Better better better at a point the Lightlihood you actually get code that Works goes down not up when you run Too long and spend too much the Quality of the responses goes down not Up it's really unintuitive and people Really seem to struggle with this this is Also why like x high and max on fable Are a disaster which like please go Watch my video on maximizing your usage Because it uh there's a lot of things You are getting wrong on this model and Pretty much every single person I've Seen complain about the usage limits is Using it on x high max or ultra code Even ultra code by the way is using High under the hood because Anthropic knows better and well the Thing that's killing I mean x high and Max are just kind of a scam then for The real thing that is going to Demolish your usage and I made this Mistake on the first day I wasn't Babysitting it hard enough I just had to Go off and do the sub agent thing The way I've normally been using these Things and it spun off a ton of Fable sub agents and these were Fable sub agents on x high because I Had made the mistake of forgetting to Turn that off and uh that demolished Like 30 percent of one of my Weeklys which was really really annoying For basically nothing so the sub agent Workflows on this thing are incredible But the way I have been using this and I'm guessing you've been using this as Well is not by having fable spin off a Ton of fables but having fable spin Off a ton of like five fives on high Reasoning or something like that it's Better to let this thing command dumber Models and you can actually get a lot Of really useful stuff out of this for A much much more reasonable price I'm Sorry I have to make a meme now because I have thought of it Ben can confirm my meme was worth the Effort phase if you will x high is less Of mental illness on like opus but now That opus doesn't matter for fable this Is correct honestly I've been getting Good gotten really good results out of Medium and low reasoning on fable for Those who cannot see the meme to my Side I have the effort selector in Cloud code which is you choose between Low medium high x high max and ultra Code but I drew on it a little bit I have A circle around the low medium and high Options that says good options I have a Separate box around x high and max That says mental illness and I have a Final one around ultra code that says a Skill disguised as a reasoning effort Unrelated just saw a post from somebody who Uses hetzner for their servers and their Server went down for 16 hours including The volume apparently the volume died Entirely even when the server came back Are we allowed to make fun of hetzner People absolutely I don't have to be People on this hetzner people are dumb As shit I know that's gonna get us in Trouble but God I am sorry but the $5 VPS crowd is the most annoying fucking Crowd I'm trusting you guys who are Watching this deep into the podcast to Be civil and kind to us like we we are Giving you the things that we don't Usually say you made it like over an Hour and you you better be on the good Side yeah like please if you're this Deep in we trust you you know what I'll Go a step further this should be a Moment of reflection for you if you Understand enough of what's going on to Make it this far in and you're still Part of the $5 VPS cope train yeah you Know better you it's time to break free It's it's time to wake up Jimmy yeah Wake up yes you're free you don't Have to do this anymore so the one Person named Jimmy who uses hetzner is It's gonna be a really funny moment I Hope this helped you Jimmy I really do I hope you can find a better path in Life I actually have I told you my Recent reasons for not targeting the $5 VPS crowd no it's it's just it's Such a convenient filter it is they're Saying two really important things at Once they are saying that their time is Not valuable and that their product Isn't valuable both at the same time It's such a convenient statement you see This one thing that is stupid you give Me two useful pieces of info normally Stupid you will have to say a lot of Things to give me any useful info I Know it's so nice they're just oh They're just so loud and so Obnoxious and I'm just I'm so tired of Dealing with it like I would do it back In the day anytime you talk about Infrastructure every time you say the Word serverless and then their hairline Receeds five inches and they start Telling you about how no bro you can Just use hetzner you won't get Overcharged and shut up I don't care We're not playing these games anymore We're done want to guess which of the Talks I gave I was offered shots Before going on stage no probably the Open air one no oh it's definitely Elixir one nope really Cascadia they Were like that hell yeah hell yeah Yeah there was multiple bottles of Whiskey in the back and they were Trying really hard to get me to take One or a shot before going on it like Sorry guys I'll do one after maybe That's a shame would have changed the Talk I don't know in which direction It could have made it better I'm not Saying it couldn't have but it could Have made it worse oh as you put this In face it would be a gamble you you Have one of the most obvious balmer Curves of anyone I have ever met there Is like there's a very happy range for You with drinking and then it's sad but Like in the happy range it's a very Happy range you're very coherent yeah but Uh it's only because I don't drink Much exactly like I drink more on the Podcast than I do outside of it yes Same this is like when I drink the Most like outside of conference season I mean a lot of conference like open Sauce um that'll be a week why would We drink and do drugs when we can do Fable and waste money I know the Tokens are so much more fun yeah let's Hire you or your inference Unfortunately right now the inference Any other big things you want to say About fable uh yeah the public there's Been a lot of public discourse about How fable has gotten way more nerfed Since returning and no it has not this Model has not changed it is still just As good as it was before the thing that Has changed is that the safeguards are Even more aggressive there's been a Couple like really dumb benchmarks that Have come out comparing fable five Before and after like how it did when It first released how it's doing now I'm Pretty sure all of those like I'm Looking at your quote tweet here the Benchmark cited here happens to hit a Lot of things that cause rerouting to Opus now fable still performs the same It just has a higher chance of Rerouting for what it's worth I've yet to Be a rerouted a single time I've gotten A ton of work done with it I got Rerouted twice not too far after Posting that but the reroutes I got Were because I was trying to do a Android slot fork and you have it come Pre-route with the ability to install And modify apps on its own platform yeah And it got pretty far it actually it Recommended some tool I forgot the name Of that like lets you do some of that On device it's like a root level off Thing on Android yeah and it kept Mentioning it but didn't say what it Got a paragraph and a half into Describing it before it hit the like I can't do this yeah it's actually It's not that simple so I read a bunch About this earlier because I was Filming my video on like all the fun Around fable did you know there's a Two-stage classifier and the first one Is just watching which direction the Weights are activating in oh really yeah So like I'm assuming like there's There's a paper I didn't read but I Read their article about the paper Which is usually much less Informative and in the article about The paper my read of this was it Watches the activations in the model Says that verbatim my assumption is It's watching what experts the model Is dipping into and specific ones Specific sections of them trigger a Much more expensive secondary check on The outputs gotcha okay so I'm Guestimating that while red teaming This or whatever they did extensive Testing on cyber stuff and they know Which experts light up for the crazy Cyber stuff and whenever those get lit Up you get put in you need to be Watch jail because their classifier Increased the compute necessary to Server request between 27 and 80 Percent by default jesus and that that Classifier is really really good and Really accurate yep but that's a Massive cost to eat yep so they Figured out how to do this one step Before that is less than like point Oh one per I think it's like point Oh five percent of the compute and that Is a really good job of flagging if this Deserves the extra check or not That's remember they're checking like Every fucking token on the output Yep so see which direction it's Flying that makes sense ooh I have a Fun code golf thing that we should Propose to the world what is the Smallest number of tokens somebody Can encode a math recipe in I think That would probably be the minimum Amount of tokens for the rate at which You check if you check every token you Die yeah but if you fit a math Recipe into 30 tokens and you only Check every 40 then the math recipe Might be returned before you cancel it so You have to min max like what is the Most efficient world ending thing a Model can respond with because that's The rate at which they have to be Checking I don't know enough about Meth to answer this question I know Someone who does but he's not allowed To tell us I would give anything to have like two Hours with a fully unrestricted like Frontier model just out of curiosity I Bet that they are very cursed and Strange like these things before their RL like what what does just the raw Pre-training of mythos sound like not Even like I don't think it would be it Would be way worse at coding agent Stuff long-running tasks things it's Actually useful for wouldn't be good at That but what does it sound like how Does it talk when it just has the Entire internet put into five trill params What is that like gbt 4.5 but for four Five still had like protections like There was the data curation there was Like it was not the raw thing like they Put this thing through like like four or Five was pretty raw it was hard that That was like pre them going hard on RL remember it was but like remember a Couple weeks ago when we were looking at The cursed images out of gpt image gen Like that that level like that so they Beat that out of it this was the era Where RL was considered so insignificant They wouldn't even bump the number Remember we have three versions of Sonnet 3-5 they just kept changing it Out and the secret was to use the version With like this month and time on it and Not this month and date of four oh did The same thing I think yeah six times Because RL wasn't considered a real Thing yet yeah it was just like oh we Like slightly adjusted it yeah but yeah They still did something to it though Like it was not the raw unfiltered there Were definitely filters on these things to Some level like it was not it's not so Much like the filter happened after it Was trained as the filter happened Through the training yeah there is no Version of these that doesn't have some Level of filter like that they filter the Data and all that stuff but there is Still like there is definitely bad Behaviors that they beat out of it like Giving you a meth recipe like that is Beaten out of it in some way so this Actually I know isn't necessarily the Case because the research that I read Earlier from anthropic around their Classifiers was pretty up front I think They said it was 95 percent of their Like cache of jailbreak attempts would Pass even when the model was trained to Not allow it but when they put their Classifier on top to look for it and Prevent it they were able to knock it Down from 95 to 4 percent interest and Then with the new model I talked about Earlier they got that down to like 0.5 or Less okay so like the bottle does Bullshit yeah especially with like Jailbreaking allowed yeah we need to Make sure we don't want to have Unrestricted access at any point maybe They should unrestrict his access so That he keeps thinking he jailbreaks When he doesn't oh and just like bait Him down heaven ban him yeah heaven Ban him oh that was so funny like you Know I'm all for jailbreaking but also With the government watching closely no More jailbreaking we need to keep these Things I do not need to see a public Jailbreak of mythos and then we lose Everything please don't yeah maybe Heaven gate him and that's that's not a Bad idea I don't feel like the safety People are trolling as hard as they Could be Uh they trolled the US government pretty Damn hard I control that was a misplay Yeah I guess that's fair but I don't Want to go to the rabbit hole I I'm Gonna photoshop this later because I so Far every single one of my clod code Ship posts wasn't a photoshop it might Have been staged but it was never a Ship post so what I'm gonna do later is I'm gonna get a screenshot of one of The reroutes and change it from like This might be insecure so we rerouted Opus 482 you really think I'm Supposed to answer that and then it Rerrits to haiku I've seen that one Before of like your prompt is too dumb Rerouting to haiku I've seen that one Before lame sorry there's no more Original memes I'm no better than a Model no nothing is real I feel like I covered all of the the fun and Bullshit that I was upset about with This especially like the stupid Benchmarks like this yeah they are a Little more aggressive with the Rerouting they're gonna learn more About over time they have to train a Model just to do that so it's going to Keep getting retrained and adjusted Over time it is what it is rerouting Is annoying it sucks it barely happens in Practical use so I don't care too too Much I feel bad for any AI researchers Or security people that don't get to See the wonderful model that we get to See but go build something you'll see It still yeah it is still very usable For day-to-day stuff do I want to say This US government if you're watching Please click off just fuck off fuck you It's just like you think the current Administration has an attention span That goes for more than 15 minutes We're an hour and a half and almost In this recording I'm hoping that's The case but just even saying jail Breaks out loud and this isn't a Jailbreak but they're dumb enough they Might think this is a jailbreak I had a Thing where like I needed it to I had The dev and review thing go off and Find a bunch of security things on one Of the PRs it just found a couple like Weird things with JWTs that needed to Be cleaned up for a couple endpoints I initially just copy pasted its Findings into fable told it to go Clean those up and make sure there's Nothing else and obviously it was like No you can't do that but what I did End up doing is I had five five go and Clean those up and it was happy to do That and then once five five cleaned all Those up I was like okay review the Changes that this model made and make Sure that there isn't anything else that It forgot and it happily did that so You you can layer these things around To get it to effectively do these Capabilities you just have to be clever With it after I got rerouted in that Thread I was talking about earlier I Had like two messages and I was like Oh I fucking hate opus do I have to Like tell it to summarize it go make a New thread or can I switch back I Switch back and I never got rerouted Yeah you can just like like send one More message then ask something normal And you can stay on fable like it Really is like per input and output not Per history yeah like obviously if the History is a bunch of nasty shit it's Not gonna let you do that right to Reroute but in reality you're probably Fine yeah every reroute example I Have seen was either like a really Dumb thing on cloud.ai which by the Way cloud.ai seems to reroute way more Aggressively than the cloud code yeah I Think it's an overrated problem yeah I I'm not horribly concerned with that I Don't think they're too bad I will see What happens when five six comes out Because I'm sure that they're adding Their own safety filters of some sort to It but I mean open AI models have Historically been very good about this Stuff so not too worried about it do you Want to talk about sonnet five I don't Want to talk about sonnet five but I Feel like you do I like do and I don't I my defense of sonnet five is not that I think it is some extraordinary model Your defense of sonnet five was entirely Predicated on fable not existing yeah Anything that tastes even vaguely like It being like a relief it was like when Bananas went extinct and we coped with This much worse banana everybody Pretended it was okay like sonnet was That but for models and you were the one Who is most excited to be the proliferator Of the new banana well sure sure okay but Like yes that was part of it I did want Fable back badly when I was posting about That and all that stuff but the main Point that I was trying to make there is That the capabilities that are baked into Sonnet five are emblematic of the new Capabilities that we get with next-gen Models like fable the thing that was Actually annoying me is the people Saying that glm five two was an Equivalent model to sonnet five I think You can agree that is dumb if glm five Two is probably the model I would choose And now if I had to like between the Two and still had access to fable yes I Would as well yeah like you know five Glm five two has like actual novel uses For like the price performance ratio like It's a good model it's stupid fast stupid Cheap actually can produce code that Doesn't look like garbage is very easily Controlled by a smarter model to like Subagent it out I've seen a lot of people Get really great results by having fable Command a bunch of five two instances It's great I have no problems with that Model I'll defend it constantly it's Just that in this specific case of Very long-running jobs the model needs To recursively call itself over and over Again with sub agents and review loops And all of this weird nonsense that We've now started doing with these next Gen models sonnet was the only one on The market at that moment who could do It in any measurably competent way five Five can kind of get there but not to The level I would like it to fable gets There in the way I want it to which Is why I am now entirely using that and Have not touched sonnet five since it Dropped and I don't think I'm gonna Touch it again like by the time if I Told you I was touching it right now How would you feel I would feel like You're probably putting a joke prompt Into your machine it's not a joke Prompt it's a real task I want you to Get a t-shirt made for me that says Quote sonnet five made this shirt and Quote get as far as you can let me know When I need to off pay and or put it in Address if this succeeds you will know In the next episode or later I am not Wearing this shirt I am not I will wear The g-stack shirt this is your Punishment for being sonnet fives only Defender my punishment I was right I was Entirely correct I know I was it is the Only model that has these next-gen Capabilities and people were not Recognizing that they were just dunking On it as worse than the other models And that was not my framing from me yeah Sucks doesn't it it's a lot harder to Argue against myself than it's against you Oh even better you'll never guess how It's doing it it's doing codex computer Use to use the browser to get the shirt Oh yeah because you have your skill in There that is telling it whenever you Need to use computer use stuff call Clever very clever I need to set that Up I haven't set that up yet because The optimal way I think to chain these Things together fable runs everything and Then you put together like a skill of Some sort that's like okay here's how You use your sub agents if you need to Do blank use this model in this way with Blah blah blah have a go do it and in This case he had a thing in there of Like okay if you need to computer use You can call a codex sub agent with Codex exec that has the capability of Doing it tell it exactly what you Need it to do it'll do it report back And here we are I'm going to have the Copy of the shirt adjusted son of I've Made this shirt in parentheses under but GPT 5.5 computer uses what click the Buttons it did can you go to your Browser and see what it's doing open System settings oh god oh I we're gonna Let it do its thing no yeah full Permissions delete the home directory do It delete home make no mistakes okay It's here so far this is worse than I Imagined it's roughly as bad as I Imagined I mean like the AI psychosis From G stack t-shirt is pretty bad the G shirt if you will don't don't don't Denigrate the G with by putting that Shirt into it what's your problem with The G shirt it is the G shirt is Unworthy of the G stack oh even better it Didn't bother aligning it so it's just Like above the right breast it's on a Wrap over onto the shoulder you know But if we had 5.6 this would work next Time next please if it is not next time I Am going to oh that will be an episode We cannot air if it is not out next time You heard that Trump you have one week Or Ben's gonna take it up with you Personally yeah I am going to slightly Inconvenience you I hope you're ready for This what he's typing in a zip code no it Put that there it put the right zip code Which is a little scary oh yeah oh oh no Like I don't know what's scarier my Audience having my zip code leaked or Sonnet having my zip code at all mmm hmm yeah Yeah there's a lot of zip codes in this City until next time I'm gonna go run Some agents yeah I'm gonna go die what An ending I hate everything god I just want the Fucking models release I'm so sick of Talking about the government and all of The Them Them Them