# AI Model Instability and the Rise of Open Source

**Podcast:** Dev Interrupted
**Published:** 2026-06-19

## Transcript

So, Andrew, last week you were telling us all, get out and use Fable before they take it away.
Did you use Fable before they took it away?
I used every token of Fable that was available to me before they took it away.
I was in the middle of using it.
When they took it away.
And I got a special little message from the Claude Code Harness that I'd never seen before.
It was like due to a government order.
You can no longer use this model.
Oh, that's terrifying.
Yeah, I know.
I was like, what did I do?
I'm just like, I'm just having it like look at this project.
How is anybody else involved?
But it was short, short, but sweet time to have it.
Yeah, Fable, we hardly knew you.
You know, we'll miss you.
What about you, Ben?
Did you use it?
You know, I got to use it a little bit and I was planning to use it over the weekend too, but obviously that didn't get to happen.
So, you know, for our listeners, I might've missed last week.
We were excited because it was the first time we ever got to cover an anthropic model release on this show.
And then that day the model gets taken away from us all.
Yeah, if you haven't been following it, the U.S.
government ordered Anthropic to immediately disable access to Fable 5 for all foreign users.
But of course, you know, Anthropic can't just turn it.
They don't have, you don't know who their users are, where they live.
So they just had to shut it down for everyone.
And the U.S.
federal government cited national security concerns over an apparent jailbreak.
And Anthropic has come out with some statements about it.
You know, they claim that it's the jailbreak is narrow.
It's non-universal.
It's a similar issue that exists in other widely deployed models, including OpenAI's GPT 5.5.
And yeah, you know, it's interesting because, you know, I don't think we have a complete picture of this all yet because we certainly haven't seen what the alleged.
concerns or the issues were.
The rumors are, though, that it was a research team at AWS, actually, that reported this to the federal government.
So there's potentially some competitive stuff going on here.
But, you know, yeah, it was nice to get it to use it a little bit before it disappeared.
So I at least got a sense of what it could do.
But yeah, I know you, Andrew, have been using Fable a lot.
So maybe explain what you've been using it for or what you used it before, before it took it all away.
Well, like I covered on the show, definitely used it to upgrade all my skills, upgrade all my hooks and my harness, update anything that was going to remain for other models to use.
Because I think the name of the game right now is choosing the right model for the task and doing model routing.
And one of the most effective ways to be successful in choosing lower cost, lower tier models is to make sure that you have a really strong and robust system in place for them to protect them.
Like anybody who's experimented with like...
local small models on their machine knows how important it is to sandbox it because these models could very easily do damage to a machine if not guarded.
Whereas maybe like on a more sophisticated model, you can put more protections in place as part of the model usage itself that are more protective.
So in my case, I used Fable.
to protect myself and my maybe less smart models from hurting themselves and my machine.
But the one thing I'll point out is that, you know, the time we had Fable might have been pretty short, but in the time we had it, I think it caused a pretty seismic shift for the folks who did try it.
But most importantly, it came and left at a time when so many companies and enterprises and individuals are starting to get bigger and bigger in French bills from places like Anthropic and OpenAI.
And this kind of instability really calls into question, how can I reliably and safely scale my business in this kind of economic environment?
Because for those of you that aren't following the whole jailbreak saga with the White House, Ultimately, what they're demanding to Anthropic is that no jailbreaks can be possible, which is kind of like a grandparent telling a grandkid that like, oh, why can't my computer just be hacker free?
Like, it's just fundamentally like an intrinsic part of the technology that will always have to be worked on and updated.
And so to make it a hard block puts, you know, Anthropic in a pretty hard position in order to escape.
And in the meantime, you get a lot of folks who are maybe pivoting to smaller models, some of which we're going to talk about on the show today, and otherwise maybe looking away from paying Anthropic for every single inference call that they do.
So it's definitely a shift in how folks are using the tools.
And if you're not feeling that yet, I definitely encourage you to pay attention, especially on social media, on how folks are reacting to the models coming and going, because folks are starting to bunker down and kind of protect themselves from those shifts in the future.
Yeah, and it's not clear to me how long this ban will actually last.
You know, personally, I think Anthropic is going to find a way to come out with either a new model or a revised version of Fable that's similar to it, but different enough that the export controls don't apply.
And then, who knows, Andrew, maybe we'll be back to where we were before a week ago, where, you know, tomorrow this episode will come out and the same day, maybe Anthropic will announce their newest models.
Yeah, maybe luckily for our listeners, we'll jinx it.
And they're actually releasing it right now while we're talking, which is what Anthropic traditionally does with all of their models.
Yeah, yeah.
Yeah.
Well, welcome to the Friday Deploy brought to you by Linear B.
I'm your host, Ben Lloyd Pearson.
And I'm your host, Andrew Ziegler.
Yeah.
And today, in addition to Fable, we are covering is meta destroying its engineering culture.
Raising the discipline bar for AI, U.S.
enterprises get wandering eyes for open source Chinese models and the importance of expertise in agentic coding.
And I want to start right at the top about Meta and questions around whether or not it's destroying its engineering organization.
So I've been seeing this all over social media in recent weeks.
We also had this article from the Pragmatic Engineer that came out, which is always a great resource to read.
And, you know, it really does a great job at just breaking down how Meta has really undergone this dramatic dismantling of its engineering culture starting around April of this year.
And it primarily seems to be driven by Mark Zuckerberg and their newly installed AI chief, Alexander Wang.
But they've been doing things like having mandatory keystroke and mouse tracking that they're using for AI training data.
They forced about 30 to 50 percent of their engineers to be reassigned from core teams to data labeling.
So effectively telling AI whether or not it's doing good work.
And then, of course, there's also been a 10% reduction in workforce over there as well with rumors that there could be more plans in the future.
And so they're at a point now where somewhere between like 4,500 to 6,500 engineers now spend their time, their full-time day doing data labeling.
It's bubbled up in social media through leaks into the media about how engineers are feeling really pressured, like they have fewer people on their teams.
Particularly, it sounds like their security and infrastructure teams have been dramatically reduced all the way to the point where there was this particularly embarrassing outage or hack that happened on Instagram where simply by spoofing your location with the app, if you knew the username of an account and the location where it normally logs in from you could have a password reset email sent to any email not even the email for that account like it's just mind-boggling um So, yeah, and there's a lot of warnings in this article about this concept of AI psychosis.
You know, there's leaders out there who are so fixated on AI that they just sort of abandon a lot of the fundamentals of engineering and professional work.
And, you know, that's not unique to Meta.
I think Meta has just really been at the center of it, particularly, you know, around this topic of token maxing as well.
So, Andrew, what did you think about this story?
I think the story really flips on its head everything about.
why folks are in engineering in the first place.
And it really speaks to like a misallocation, I think, of resources of judgment and time.
You know, Meta is one of the most resourced and rich and largest like tech companies in the whole world.
So obviously when they're going to make.
shifts, they're going to do it in a radical and large scale way.
That's how they've managed to stay big and be what they are.
But something feels a little different about this pivot.
And I think the taste it's leaving in everyone's mouths, both end users and their internal employees and the general public.
I think the AI psychosis label is pretty fitting here.
It's almost as if engineering leaders are throwing away the stuff that has always worked in the idea that it doesn't need to be there anymore or that perhaps there's a new version of it.
And they would rather fumble around in that darkness of the unknown to find what that new is than to safely and securely incrementally find their way there.
It feels like a lot of folks are pouring a lot of money and time into the expenditures around having and being able to access AI.
So they feel a huge, huge need to...
use it and justify all of these investments and stuff they're doing.
And, you know, I definitely really feel for the engineers in this situation to be pulled off of like a high stakes or like a really deeply technical team within like the engineering org and then basically put on data labeling would be pretty demoralizing.
And ultimately, I think it speaks to like the anxiety that employees have in general around being replaced by AI.
And for a company like Meta, This shift is staggering because there are so many people there.
So I think that the leaders at the company are really focused with, or they're hit with a really, really hard reality of there are way smaller companies that can move way faster because they have.
They're a blank slate.
They're AI native.
They hire one person, and that one person might as well have been 20 hires for the amount of work they can do in a scaled, agentic way.
And then because it's all that way from day one, when they hire that next person, they're going to be just like that.
Places like Meta, they don't get that same kind of momentum with folks adopting and using AI.
They have to search.
in their massive engineering orgs for the power users, and then they have to elevate them.
But even that isn't going to be enough in this scenario when you're ripping teams apart and putting people in different places.
So, you know, I think that this is unfortunately a step in the wrong direction for Meta.
I already feel like leadership in the company and around the company are already speaking out pretty openly about being against a lot of the changes.
I think this is still very much a developing story, but for the engineers there, you know, that are in that position, I think this is like an unfortunate call to action that it's like, this is the reality of being in a large enterprise engineering org right now is dealing with these huge restructurings and shifts and like what you focus all your time on doing.
Yeah.
You know, the thing is meta has long been the place that you go if you want to have.
an extremely high paying career in software engineering and work with some exceptional and very talented people.
You know, and I actually have always sort of appreciated their hiring practice, you know, up until last year, they actually changed this, but in the past they would hire an engineer and then that engineer would onboard over about a six week period and then decide which team they want to join during that period.
And it actually led to this environment where teams actually had to like compete for new talent.
But, you know, Between, you know, getting rid of that and all these forced reassignments, yeah, there really are.
It does, just from the outside, it looks like there's pretty significant cultural changes happening at Meta.
And, you know, we've been covering this concept of token maxing more and more here on Dev Interrupted.
Meta was one of the first companies to like really...
go deep into encouraging token maxing and having token leaderboards internally and those types of things.
It sounds like they've now killed all of that off and they may even be doing the opposite now and like thinking about token minimizing.
But yeah, we just yesterday published a Substack article with our perspective on token maxing.
And we also have a workshop coming up next week on life beyond token maxing and how to measure AI in the, or how to measure efficiency in the AI driven era.
You know, because really, like, you know, I think it's only a matter of time before token maximum culture ends.
And, you know, here at Devin Interrupted, we want to help bring it to that end because, you know, this AI psychosis, it's not going to be, it's not a sustainable way to operate.
And, you know, so yeah, stay tuned.
We got lots more content coming on that.
If your AI bill went up this year, but you can't show what it bought, this one's for you.
Andrew and I are hosting a live workshop, Life Beyond Token Maxing, AI Efficiency for the Long Term.
We'll get into why token counts fall flat in executive conversations, where AI is moving the bottlenecks in your pipeline, and how the APEX framework builds on existing productivity metrics frameworks to measure what matters in the AI era.
Everyone who registers for the event gets early access to our new guide on measuring efficiency in the AI-driven SDLC.
and controlling the executive conversation around these discussions.
The link is in the show notes.
Come hang out with us on June 25th.
All right, Andrew, let's talk about how AI demands more engineering discipline.
Yes, this is an article from Charity Major, CTO over at Honeycomb.
We love Honeycomb, of course, here at Dev Interrupted.
We most recently had their CEO, Christine, on the show.
And in this article, Charity argues that AI-generated code is reaching a median software engineer.
quality.
And what this means is that, yeah, sure, that the AI is pretty good at writing relatively durable code, but it doesn't reduce the need for rigor.
And in fact, what it does is it raises the stakes for understanding exactly what you need out of that code and having observability and system understanding for how that code works in the wild.
And ultimately, at Honeycomb, the thesis that they're driving is that The truth of what your product is lives in production.
It always has because that's what your users interact with.
That's where it lives in the place where it matters, not in the dev container and not in your CICD.
And the important thing about measuring and understanding what your product is doing in prod is that this allows you to rapidly adjust and iterate and effectively close the loop.
on development because now observability is a core input for agents and prompts and more work that get done to drive that production, get evaluated in production, and then drive the code.
And this doesn't mean that, obviously, Charity is not saying we throw away the dev instance.
She's just saying that it's more important than ever to measure what's happening in production and that you need to focus on ways of working that make that loop really tight and really efficient.
Having really short feedback loops between shipping things and understanding what it's doing in production and instrumenting your production just as eloquently as you do all of your eval tests.
Because if you don't know what's happening in production, then you don't know how your product is working.
And ultimately, she says that Non-deterministic AI systems demand even more validation than ones before.
So your eval practices, your test practices of yesterday, they just don't cut it.
You need higher quality gates that guard against slop and to protect your stuff in production.
Yeah, you know, there's a really great article or line in this article about how 2025 is the year that.
the economics of code production turned upside down.
So it used to be that you hired smart people who know how to write code and they would spend their time writing code.
That's now like the cheapest, one of the cheapest functions within your organization.
And, you know, when you reach that point, you know, the real value that humans provide in this world is creating this shared understanding of good software, you know, and that gets very tactical and technical with the testing and the CICD and deployment systems that you put around your software.
But it's also, you know, more theoretical, you know, you need domain experts that.
that understand the architecture and understand what's good and best, and then can apply that to a spec or, you know, whatever agentic system you're working in.
And then as it gets built out, validate, you know, efficiently that the agent is doing the right things along the way.
And, you know, and that shared understanding, I think that's really what organizations have to be building at this point, because that's how you apply.
you know, that discipline to your software engineering system.
So yeah, another really great article from Charity, of course, with lots of really funny quotes and comments about AI and all of that.
So very much, you know, the norm for content from her.
Yeah, always a great discussion from Charity.
All right, let's move on to some of these open source models coming out of China.
First, I want to talk about what's going on with GLM 5.2 and Z.ai.
What do we have here, Andrew?
Okay, so this is a model that came out in the last week.
It's GLM 5.2, and it's from ZAI.
This is an open source model from China.
And the reason this has been a big deal for folks is because it's beating GPT 5.5 on a lot of long horizon coding benchmarks at an incredibly small amount of the cost.
And why this is a big deal is because...
Long horizon coding tasks is the real frontier that places like earlier, we're talking about meta and what they're trying to unlock.
This is what they're trying to do.
They're trying to be able to run agents that are running for long periods of time, doing huge amounts of work, doing things like migrations, crunching engineering tasks and keeping the lights on operations that used to take, you know, maybe a team of five multiple months or a year to migrate something now could be a long running agent that just loops for a few weeks or something.
So this is like the real key a lot of folks are trying to unlock.
And the idea that you could get this kind of high-scoring behavior from an open-source model is really compelling, especially when you have the whole reality that you can fine-tune this, host this on your own infrastructure, finally throw away that big anthropic bill or that big open AI bill.
But the big thing about understanding why this is important is because...
Going back to what I said earlier, Fable came and then Fable left.
And in doing so, a lot of folks kind of change their opinions around what they want to use their models for and what models they should be reaching for in the first place.
And while it could be really easy for maybe a model like this to gamify a benchmark or, you know, really once a benchmark is out there, it ceases to be a good benchmark.
I think we need like a good heart's law for benchmarks.
And so like.
There could be a lot of nuance here.
Like maybe it doesn't perform outside of the benchmark on specific tasks that are long running.
But the experimentation need is there for everybody in order to like figure out what's going to work for my org.
What are the stakes that I can accept?
And more importantly, what's the cost that I'm going to be able to carry into the future because I can't keep paying these huge, huge, huge bills forever.
I'm kind of curious, Ben, have you been experimenting with local models at all?
I know you hear me talk about it a lot, but has it kind of crossed your mind as well?
Well, no, I'm not yet.
But I do understand exactly why people are doing it.
And I think Fable is a great example because if you're running everything locally, the capacity and the services are all up to you.
You get to define it.
You get to determine when it runs and when it doesn't.
I absolutely understand why people are looking more and more into them.
And also when you hear stories about how like Anthropic may have dumbed down some of Cloud Code's capabilities because they just ran out of capacity and they needed to take emergency action to keep it functional.
But yeah, I think it's just pretty incredible that they've been able to build this open source model that competes directly with Anthropic and OpenAI's latest models.
And it kind of makes me wonder if there's something to having like a domain specific model versus more general purpose.
Like it seems that Z AI is taking a much more focused approach on software development and maybe even like, maybe even to say like long, long, like long-term running actions, you know?
But yeah, if you're, if you're, you know, if you're running everything locally on your own hardware, like that, that really gives you a lot of power and flexibility.
But then also, you know, GLM, you know, they provide this as a hosted service too, and it's priced at about half the cost of Quad.
I mean, if you look at the API cost, they're about a sixth as well.
So, but I think what's really important with this one in particular is that, you know, not only has the model been published as open source, but the weights have also been published as open source, all of it under an MIT license.
And I think this is really important because it's going to make it a lot easier for people to build upon it and improve it and fork it and do new things with it.
So, yeah, I mean, it's, you know, everyone talks about the threat of like Chinese models and all of that.
But like if it's open source, like it can very easily become a Western model by just forking it and doing taking it in a different direction.
And yeah, I mean, I love seeing all the competition heating up in this space.
You know, I think it's going to.
Keep prices from going astronomical with some of these frontier models and probably also get companies like Anthropic and OpenAI to wake up a little bit and start to think, like, how can we replicate these low cost ways of building new models that, you know, maybe they're more purpose built than than what they've been doing.
Yeah.
And this general purpose versus like domain specific ideas, specifically with small models is really compelling because I think that's how this is going to go.
I think models are going to fragment into a.
bunch of really domain specific things.
And then you're going to get a small language model that's open source, open weight.
And it's going to be extremely specific for a particular industry or a kind of job task or something.
It's going to be the best in class model for doing that task or whatever, or being, or for that subject matter than any other model, because it's specifically.
you know, trained on material that is allowing it to specialize.
And it has that luxury because there's so many models now that you can use the distill and tune other models and create other things.
There's so much intelligence in these models rolling around right now.
So we're actually really straightforward to recombine them into really compelling examples.
Like one that comes immediately to the mind when you said that, Ben, was like there was a model called Epicure that came out like a week or two ago.
And Epicure is a large food model.
It's an embedding of recipes from all across the world, regionally and just generally.
What it does is that all of the ingredients that you cook with are like vectors in this model.
And so you can use this model to come up with new recipes for flavors and palates and things that are similar to other flavors and palates.
Find substitutes for other foods.
And so like imagine all the use cases in like a culinary world or like an AI powered like.
food-related tool.
Imagine all the food processing, all the food creating stuff in the world.
They're also experimenting with AI just like how engineers are.
And the sooner that engineers wake up to that reality and understand that there's a model out there that needs to be built for everyone and everything, I think you're just going to get a huge proliferation of stuff exactly like this.
So I'm excited.
And in particular, I'm excited for Epicure as well.
I kind of want to play around with that one.
Yeah, no, it's actually, I wasn't aware of that one, but I would love to check it out too.
Because, you know, one of the first uses I saw of AI outside of software engineering, I saw on the news that there's an interview of a frozen yogurt shop that had completely reinvented their menu by using ChatGPT to write a whole bunch of new recipes for them.
And I remember in the early days of AI.
And if you asked it to do recipes, it was a really, really difficult challenge for it because, you know, there's measurements, there's preparation instructions, there's like very precise things that you have to do.
And it really struggled to like piece together all of those components, you know.
So, yeah, I really do think there is something about this like highly specialized systems for AI.
Yeah.
All right, let's talk about how Microsoft is looking at some of these open source models.
Yeah, so we've been talking about some of the smaller specialized models and kind of how other players are looking at it.
This is an example of one of the biggest players that are out there, you know, maybe having wandering eyes for a model like DeepSeek.
Microsoft has been tossing around the idea of using a self-hosted version of DeepSeek to power Copilot Cowork, and they're seeing this as a lower cost alternative.
to all of, you know, the inference that they're paying to Anthropic and particularly OpenAI, which it relies on.
And, you know, it's definitely driven by the rise.
of those api costs which is like what we've been talking about this whole episode so far i think it's pushing enterprises to make new strategic decisions that they maybe didn't before and to find cheaper models to stay competitive with these fast moving ai native startups that are eating all of the tam around them and so like the idea that the ceo of microsoft would publicly be talking about using or switching to that kind of model, I think is like a sea change event for most people.
Like, for me, Microsoft is like almost synonymous with OpenAI.
Their partnership has been so strong from the beginning.
And, you know, Satya Nadella, the same person who's saying this now, like, oh, maybe we'll just use DeepSeek, was the biggest supporter and one of the earliest proponents of making OpenAI as entrenched in the world as it is today.
There's also a lot of nuance to Microsoft's wandering eyes as well, because obviously the US government is not going to take very well to the idea of a large enterprise with as much sensitive information and proprietary stuff as Microsoft considering to use a model from China, even if they're hosting it.
themselves, which speaks also to like a culture war that's happening with these models and the impacts that they're having on economics, both on China and on the US side.
But for engineering leaders right now, I think this is...
Just one to be watching, not a move to make yet, to understand how this large enterprise is going to make a shift to lower its bill.
You know, most recently we talked about how Shopify picked up QAN and fine-tuned it and unlocked this crazy cool multi-agentic system that not only was way better.
than what they were using from open ai or whatever uh but it was self-hosted and like way less of the cost and so imagine microsoft being able to unlock that same thing because they'd slash so much of their bill.
Because the reality is, is that a company like Microsoft, maybe they say, they pay one sixth of the inference costs.
Well, they're going to reinvest those five, six back into more inference, back into more development, and they're going to go faster.
And that's exactly the kind of coal that he needs in his steam engine, I think, to get Microsoft ahead.
So that's kind of what I'm seeing from them.
I think it's kind of an interesting shift to watch.
Yeah, and even here at Linear B, we've started to focus a lot more on model selection for sub-agents, right?
Like, you don't always need to go to the most expensive hosted model to solve every problem.
And if you can find ways to push tasks down to cheaper models, you can unlock much bigger agentic workflows because you won't be blowing your token costs out of the water.
But yeah, I mean, Microsoft is in a really tough place because they don't have their own foundational model yet.
I mean, they have started to release some stuff.
We've been covering that, but they're still very early to the game.
And they really do need to focus on finding less expensive models as quickly as possible.
It seems like that has to be like one of the top priorities at that company.
But I actually don't think the Chinese part of this will play too much into the equation because...
Since so much of what's coming out of China is open source, I could actually see Microsoft using open source as a leverage to attempt to dethrone some of these more established players.
Over the last decade, Microsoft has really warmed up to the open source community.
I mean, really, they did a complete 180.
They used to be very hostile to Linux and the open source community.
And now they've completely flipped and are very active.
They're one of the most active companies in the open source space today.
So, you know, I could actually see them start to view open source models as their key differentiator that helps them survive this transition.
So it wouldn't surprise me to see them really heavily invest in the open source side of all of this.
I think you're going to see them make their own model, too.
I think this is just a stepping stone along the way.
Yeah.
Yeah.
All right.
Let's talk about.
Agentic coding and a return to expertise.
This is some new research from Anthropic.
What do we have here, Andrew?
Yes, I love an Anthropic study.
You know, as much as we've kind of ragged against Anthropic in this episode, I'm a big Anthropic fan.
And what I love about Anthropic is they're research-minded that they take to how people use their tool and how they put together these really smart case study reports from the unique vantage that they have in the industry about how folks are using their tools, what makes them successful, and where they see it all going.
So this is a study that looked at 400,000 Claude Code sessions.
found a pretty clear division of labor in terms of what was happening in those sessions.
Users were making a huge majority of the planning decisions, like 70 plus percentage of it.
But Claude was doing 80 plus percent of the execution of nearly everything.
And this suggests that the agentic tools and their orchestrators, they're finding their two camps.
The idea that the domain expertise around what we need to build, when we need to build it, and why and what it looks like, is still firmly in the grasp of the human because they're the one plugged into the world where that product needs to get delivered.
They're the one to have all of the messy context from the other humans trying to deliver their things into the world as well.
They just have a unique advantage in the world of planning.
But on the other side, you have the agent, which has a unique advantage in the world of code.
Because just like a charity said, you know, the cost of getting good code now is basically free.
A lot of the work that you have have to do now is very different around understanding what you need and then securing it and making it stable and scalable.
And that is something that is ultimately still hand in hand between the agent and the planner, because the ability to scale and be safe and be secure isn't purely an execution problem.
It's a planning problem as well.
And ultimately, like when there's also like what we learned from this study is that like.
there's levels of skill involved with being able to be this human operator.
And we've talked about like levels of being an orchestrator on the show lots.
Like we've talked about Gastown and Stevie, I get at the top of the year, gave us the really great chart of like eight levels that like everyone has talked about.
And we've covered a lot on this show.
This is.
They're giving us also here like a scale, a rubric of like understanding how these practitioners are applying their skills and ultimately what a good planner and what a good executor and what a domain expert looks like.
And from their vantage, I think it really paints a picture of the qualities that make those rated expert sessions stand apart from maybe the more novice ones.
But ultimately, like.
Another thing that got pointed out in this is that the complexity and the value of work, you know, being done with Cloud Code, it increased really significantly.
Like, instead of everyone throwing agents and their tokens that trying to fix bugs and like get basic things off the ground, most of the tokens and the high level usage, especially of these very skilled executors that were in these sessions, were spent around deploying and understanding the data, analyzing what's in production.
And ultimately, like the value of those sessions rises because the compounding value of what you can gain from a data analysis or a deployment study is just fundamentally different than, oh, please fix this bug.
that, by the way, you also introduced.
I think it just speaks to a difference in how the models are evolving and the human practitioners that use them are evolving as well.
But ultimately, this is just your reminder, the bottleneck in agentic coding, AI-assisted development, it's never been coding skill.
It's still domain expertise.
And folks who have these engineers with strong domain expertise are just going to unlock fundamentally way more value, which is what this study proves.
Yeah, and to go build upon your point about how, you know, AI has really been enabling higher complexity and higher value work.
Like that's a trend that's really been taking off in the last seven months, according to this research, which really tracks with what we're seeing.
You know, it really all sort of kicked off right around the end of last year, the start of this year, where everyone started to figure out how to operate more agentically, it seems like.
But to summarize this research, you know, humans are still doing most of the planning work with AI, whereas AI is doing most of the execution work.
And the people who are experts at the domain they're working in get about five times the output from their models as someone who is a novice.
And the gist of this really, domain understanding is probably the most valuable thing that we as engineers can bring to an organization right now as it's going through this AI transformation.
So domain expertise has never been more important than it is right now.
Indeed.
So Andrew, what are your agents up to this week?
I guess aside from eulogizing Fable and picking up all of the...
the specs that it left me and the research that it left me.
I have actually been doing a ton of experimenting, as I've been talking about in this episode, with local models.
I'm having a lot of fun with the latest breed of open models that have come out into the world recently, especially with their capabilities and ability to run on smaller devices as well.
So my agents that I've been using all this time in Claude Code and otherwise that have pretty built up.
memory stores and great amounts of skills and harnesses, I've turned these into really effective domain experts, I guess, for what those agents do.
And now I'm putting those agents to work at getting these.
dumber models, maybe a little smarter and ready to go.
So I guess you could say my agents are kind of in like big brother mode.
They're kind of like looking after their little brother.
They're trying to like figure out what does it need?
How can I support them?
And like, you know, just ultimately kind of helping to distill their knowledge into it.
It's been a really fun experiment.
Just an interesting time to be playing with the tools.
What about yours?
Yeah, this is like the first time in a while that I've traveled.
So I'm experiencing that like, you know, I don't have a lot of time to sit down on my laptop and do things.
So it's like in between stuff as I'm going about and going into meetings and everything, kicking off my agents for those long running tasks so that I can just be focused on being here in the real time.
And then, you know, I just have to check in on them every few minutes to see if they need any help to get along.
I've been doing the same thing because we've been here at this we've been here at this offsite for Linear B and between like all of the big sessions and and like we'll have like a big breakout session or something for like an hour or two.
And I always make sure I kick off some kind of long running task I can come back to and review during a break or otherwise check back into.
And I still feel like I'm getting a lot done, even though I'm spending a lot of time in rooms just chatting with folks.
Yeah, yeah, absolutely.
Well, that's it for this week's episode.
If you enjoyed it, remember, everything we discussed today comes back to one major challenge that engineering teams are wrestling with.
AI is writing code faster than ever, and your SDLC is struggling to keep up.
Linear B is the engineering productivity platform that shows you exactly where AI speeds up delivery and where it stalls, and we automate your bottlenecks that your team can ship faster and with total confidence, even as AI takes over your SDLC.
So see how Linear B can help your engineering organization by checking out LinearB.io.
That's the show for today.
Thanks for tuning in and sticking with us all the way to the end.
If you're still listening, that means you're one of the biggest fans we have on this planet.
And as the biggest fan, you need to go help us promote our show.
Go out, give us a thumbs up, comment wherever you're consuming this media right now.
Give the podcast a rating, you know, whatever you can do.
Even little things really help us get the word out.
So thank you for joining us today.
sticking around to the end.
And thank you, Andrew, for joining us as well.
And we'll see you next week.
See you next time.