So obviously this is a joke and all, but, I do think there's actually something real in there too. One thing that happens in the transition from bullets -> paragraphs -> bullets is that the person writing the first bullets doesn't have to worry about how they phrase stuff ("Do I say, 'so sorry for the delay?' Do I describe why it was delayed? Do I not mention the delay?") and can instead just provide whatever facts they want the other person to know. And because the second person knows what they're reading isn't exactly what the other person said, they don't read it with the same sort of sensitivity to precise phrasing ("What does it mean that they said 'Thanks.' and not 'Thanks!'? Am I getting fired?") that they might otherwise.
Like, when we communicate, we do a lot of parsing of subtle language, not because either side necessarily wants to, but because we can't communicate without it. But this kind of transformation potentially means we don't do that? Which seems like it could make a lot of things easier, but is also very black mirror?
I agree. These tools seem more distopian that utpoian. Or more specifically enshitifying most things. Hell each day I believe the dead internet theory more and more
Mostly, though there is part of me that doesn't think it would be all bad. There is a lot of performative busywork (and communication) that we do that doesn't feel useful in any real way. It doesn't seem bad if people didn't have to do that. You could kinda argue that it might let people spend time caring about how they say stuff where they need to care. I'm not sure that we'd use it that way, and wouldn't just robots become our personality though.
It's happening on the inside. We don't have much visibility on this, but definitely 100% being worked on. Most code in the AI Labs is already being written by AI. Obviously the models themselves are the next step. It's happening faster than we think.
Our human minds cannot grasp exponential growth very well.
And I'm sure it's being worked on. So is cold fusion. That doesn't mean we're anywhere near it.
And, given the amount of money at stake, if some efforts of AGI research were starting to bear fruit, I'd expect we'd see snippets of it in commercial models before an all-at-once great reveal.
So I'll continue to say taking all bets against AGI appearing in early 2026.
We're not talking about AGI here (it's not even mentioned once); we're talking about self-improving AI. AI that improves itself independently instead of being entirely developed by researchers. AI models created by other AI models, rather than by humans. v2 created by v1, and v3 created by v2, etc. AGI will come later, and will most likely emerge from those very self-improvements (because humans cannot work fast enough, but AI can work 24/7).
That's the "takeoff" the post refers to: once model improvements become entirely independent and can continue autonomously 24/7 (and progress that much faster, exponentially). No other AI lab will be able to keep up unless they achieve takeoff also.
Ah, sorry, yes - totally misread the original - my brain saw AGI instead of AI.
I'm less skeptical about AI improving itself, but even if used this way, there are limits and barriers. This isn't automatically a gateway to a flawless AI and certainly not AGI.
Connect two computers together, and the next thing you know, we have the internet. And things are moving a lot faster since then. Exponentially faster.
AI can now use tools, and this will only accelerate with things like the Bun acquisition (they didn't acquire it just for fun). Tool use is the main differentiator between humans and other species, and the primary driver of our growth. Look at everything humanity has accomplished through the use & creation of tools (including creating better and better tools, and more powerful tools).
Today, AI can use tools. But soon enough, it will begin to create its own tools independently -- and then, everything will change (i.e. the cusp of AGI). Models are merely the foundation for what's to come; the versatile interface with which to interconnect everything.
And then fuse those tools directly into models (kinda like cyborgs), and watch what happens. Any "bug" will merely become a quick self-patch in the AI's internal toolkit (i.e. how it interacts with the outside world, or with its "body"). Having trouble catching that ball? Give it a few minutes (or seconds), and it will never drop it again. Give it any musical instrument, and it will master it in a few minutes (and play it better than any human ever could). In fact, give it just about anything (like a plastic cup), and it will figure out how to play music with it (just like we do, but much better). All it will need to do is to create & update an internal tool for that "instrument". It just needs to figure out how to play a note once, and it will be able to play it perfectly every time. Using it will become as easy as calling a function (like play_note("G#", 0.5s)). Remove a finger, and they will adapt automatically by updating their toolkit on the fly.
And then we just need to network AIs for them to be able to fork & customize each other's tools (if that is even necessary). They might end up creating their own version of GitHub. They'll be able to learn new skills just like in The Matrix.
And AI doesn't have to be *perfectly* intelligent to call AGI -- it just has to be *generally* intelligent (i.e. be able to figure out most things put in front of it, and able to use them if applicable). Like, at what age do humans become generally intelligent? Babies are useless, but a couple years later they start to figure things out. But they're definitely not perfectly intelligent or flawlessly able at that age.
Meanwhile, humans will be living in a perfect utopia! No problems ahead at all!
And then hopefully, we eventually realize that it's no fun at all to be using cheat codes (or at least, not for very long). It takes all the fun out of playing the game the way it was designed.
confederated work... indeed. btw. thanks for the Sagan reference (even if you made be dig to get to it) - I was thinking that was headed toward Douglas Adams...
Cowork was written in 10 days with Claude Code at Anthropic. It's basically CC with a little linux VM inside to perform tool calls and actions, so the thing that makes it convenient, makes it less powerful than CC because it doesn't have native tools built into the Mac. So it's kinda like Manus, which is a distill of Claude with Ubuntu, but encased in a browser window instead of a Mac app. Everyone is talking about ClawdBot/MoltBot though, this week.
I remember this thing a while back about how Google could've never been the first to build ChatGPT, because it was too risky for a big company like Google to launch a "research project" that mostly just ignored IP and safety concerns like ChatGPT initially did. And I think that's kind of the story about clawdbot; it's probably what Anthropic wanted to build, but unleashing a rogue AI on people's computers with aggressive access to everything is a risk they couldn't take.
The idea of compute becoming oil-like seems interesting. I'm not exactly sure what the implications of that would be, but I can see the outlines of that somehow.
More fluff than a bag of cotton balls. Cowork was built in two weeks, it's imperfect, sure so was CC 8 months ago when it released. Three years ago these models could barely write a poem. Why is the other work structurally unadressable? Benchmarks? arguments form cognitive science? All I see is prople are saying...
I think it's very plausible (and likely?) that the robots can do the work; what I think is much harder is that the robots do the work *as us,* where they represent us (which is in effect what we're asking them to do when they write our emails). But I think that gets "solved" by us eventually not writing that many emails (or reports, or whatever) as us, and we stop reading most emails we get as being from someone and not from their EA / AI / EAI.
And but so aping DFW in business communication would make writing emails more fun for the email writer, and because the incentives of AI apps lean towards the Entertainment of the operator, it's a plausible outcome
I mean, if all of my emails were fun to read, I'd probably do a better job of reading my emails. I'm not sure I'd ever finish them though, since they'd all be 10 pages long with 15 pages of footnotes.
What I've noticed in 4 years selling dev tools, is that outside of big tech and startups, most companies do NOT care about code quality. They care about perception management. Saying you have code quality feeds a narrative, not a reality most times. That's consistently been the game I see people play. There are literally dozens of companies I talked with in the past that literally do not test their data even with dbt (you'd be very surprised who). "It works" often is enough for a lot of people because to your point, end users rarely look at the SQL and python behind their reports.
Until technical debt catches up to them and they lose the best talents. But yeah, still not a concern for most companies. They'll manage with average (or even poor) talent.
No one really cares that most IKEA furniture is made with particleboard and MDF (covered by veneers) because it is affordable, and that's what they really care about. Similarly, most companies and customers only care about the veneer.
I looked at your blog and you are maybe 18 months behind the curve which, for someone trying to sell AI implementation services, is deeply embarrassing for you. This post is by someone 2 years ahead of where you are. Please pay attention
On a more serious note: a lot of the work I see AI eroding is of the “Bullshit Jobs” variety (à la David Graeber). Operational monstrosities that emerged from compliance requirements, poor app/org design, or operating models that shouldn’t exist in the first place (but somehow do - eg because of monopolistic markets, government subsidies, and other reasons the market often doesn’t solve)
A lot of emails + ppts + sending docs around fits into that description. You had to move information from A to B to C because the exec can’t walk the floor. Or move data from Salesforce to Finance’s weekly report because your company is too cheap to build a dashboard. Or nobody streamlined that horrible 45-step process that’s been growing steadily since 1989.
There’s a lot of pointless knowledge work that should have been redesigned long ago but survived for convenience or inertia - but it’s now becoming sufficiently convenient/expedient to automate, and so we do. Other such inefficiencies we dealt with when we got computers, or Excel, or the Internet
And so I think we agree on a lot of premises but less so the (full) conclusion?
Now, all this doesn’t add up to *Claude Cowork in particular succeeding* - e.g. if they stick to devs as their core persona, then Cowork will end up a slightly sad little experiment that never got the love and attention it deserved.
But I do think (or hope? Definitely hope) there’ll be a market for “Claude Code but make it less scary for those of us that dislike CLI UIs specifically”, lol. Maybe that’s an Opus Wrapper, or maybe Google will finally unfuck the way they’ve built Gemini (the app, not the models), or whatever.
In general we’re still very much in the “trying things out” stage of defining the right form factor + UX/UI paradigm for the age of LLMs (much like the iPhone wasn’t the first cellphone that could do more than call people, or why the PDA didn’t ever become a thing but tablets did). And I include form factor because hardware is as much fair game as software here.
Start following the AI Village experiment to see it doing much more than just BS jobs. It is still not good at it, but getting to a far wider field of action than people expect.
As an engineer who deeply cares about em-dashes... this burned through my soul.
According to a very crude search, I've used an 4,927 em-dashes in this blog, so...right there with you.
***thx claude. kek
https://marketoonist.com/wp-content/uploads/2023/03/230327.n.aiwritten.jpg
So obviously this is a joke and all, but, I do think there's actually something real in there too. One thing that happens in the transition from bullets -> paragraphs -> bullets is that the person writing the first bullets doesn't have to worry about how they phrase stuff ("Do I say, 'so sorry for the delay?' Do I describe why it was delayed? Do I not mention the delay?") and can instead just provide whatever facts they want the other person to know. And because the second person knows what they're reading isn't exactly what the other person said, they don't read it with the same sort of sensitivity to precise phrasing ("What does it mean that they said 'Thanks.' and not 'Thanks!'? Am I getting fired?") that they might otherwise.
Like, when we communicate, we do a lot of parsing of subtle language, not because either side necessarily wants to, but because we can't communicate without it. But this kind of transformation potentially means we don't do that? Which seems like it could make a lot of things easier, but is also very black mirror?
I agree. These tools seem more distopian that utpoian. Or more specifically enshitifying most things. Hell each day I believe the dead internet theory more and more
Wait, are you saying that replacing humans with machines is a bad idea?
Heretic!! Light the pyres!
Mostly, though there is part of me that doesn't think it would be all bad. There is a lot of performative busywork (and communication) that we do that doesn't feel useful in any real way. It doesn't seem bad if people didn't have to do that. You could kinda argue that it might let people spend time caring about how they say stuff where they need to care. I'm not sure that we'd use it that way, and wouldn't just robots become our personality though.
"The researchers said this could happen in early 2026"
Taking all bets against.
It's happening on the inside. We don't have much visibility on this, but definitely 100% being worked on. Most code in the AI Labs is already being written by AI. Obviously the models themselves are the next step. It's happening faster than we think.
Our human minds cannot grasp exponential growth very well.
AI writing AI code is not AGI.
And I'm sure it's being worked on. So is cold fusion. That doesn't mean we're anywhere near it.
And, given the amount of money at stake, if some efforts of AGI research were starting to bear fruit, I'd expect we'd see snippets of it in commercial models before an all-at-once great reveal.
So I'll continue to say taking all bets against AGI appearing in early 2026.
We're not talking about AGI here (it's not even mentioned once); we're talking about self-improving AI. AI that improves itself independently instead of being entirely developed by researchers. AI models created by other AI models, rather than by humans. v2 created by v1, and v3 created by v2, etc. AGI will come later, and will most likely emerge from those very self-improvements (because humans cannot work fast enough, but AI can work 24/7).
That's the "takeoff" the post refers to: once model improvements become entirely independent and can continue autonomously 24/7 (and progress that much faster, exponentially). No other AI lab will be able to keep up unless they achieve takeoff also.
Ah, sorry, yes - totally misread the original - my brain saw AGI instead of AI.
I'm less skeptical about AI improving itself, but even if used this way, there are limits and barriers. This isn't automatically a gateway to a flawless AI and certainly not AGI.
Connect two computers together, and the next thing you know, we have the internet. And things are moving a lot faster since then. Exponentially faster.
AI can now use tools, and this will only accelerate with things like the Bun acquisition (they didn't acquire it just for fun). Tool use is the main differentiator between humans and other species, and the primary driver of our growth. Look at everything humanity has accomplished through the use & creation of tools (including creating better and better tools, and more powerful tools).
Today, AI can use tools. But soon enough, it will begin to create its own tools independently -- and then, everything will change (i.e. the cusp of AGI). Models are merely the foundation for what's to come; the versatile interface with which to interconnect everything.
And then fuse those tools directly into models (kinda like cyborgs), and watch what happens. Any "bug" will merely become a quick self-patch in the AI's internal toolkit (i.e. how it interacts with the outside world, or with its "body"). Having trouble catching that ball? Give it a few minutes (or seconds), and it will never drop it again. Give it any musical instrument, and it will master it in a few minutes (and play it better than any human ever could). In fact, give it just about anything (like a plastic cup), and it will figure out how to play music with it (just like we do, but much better). All it will need to do is to create & update an internal tool for that "instrument". It just needs to figure out how to play a note once, and it will be able to play it perfectly every time. Using it will become as easy as calling a function (like play_note("G#", 0.5s)). Remove a finger, and they will adapt automatically by updating their toolkit on the fly.
And then we just need to network AIs for them to be able to fork & customize each other's tools (if that is even necessary). They might end up creating their own version of GitHub. They'll be able to learn new skills just like in The Matrix.
And AI doesn't have to be *perfectly* intelligent to call AGI -- it just has to be *generally* intelligent (i.e. be able to figure out most things put in front of it, and able to use them if applicable). Like, at what age do humans become generally intelligent? Babies are useless, but a couple years later they start to figure things out. But they're definitely not perfectly intelligent or flawlessly able at that age.
Meanwhile, humans will be living in a perfect utopia! No problems ahead at all!
And then hopefully, we eventually realize that it's no fun at all to be using cheat codes (or at least, not for very long). It takes all the fun out of playing the game the way it was designed.
That's certainly one way it could go. I don't think it will. Not, at least, with LLMS. So I guess we'll call me an AI-pessimist.
confederated work... indeed. btw. thanks for the Sagan reference (even if you made be dig to get to it) - I was thinking that was headed toward Douglas Adams...
I thought it was a Richard Feynman line before I wrote this, so neither of us knew where that one was headed, it seems.
Cowork was written in 10 days with Claude Code at Anthropic. It's basically CC with a little linux VM inside to perform tool calls and actions, so the thing that makes it convenient, makes it less powerful than CC because it doesn't have native tools built into the Mac. So it's kinda like Manus, which is a distill of Claude with Ubuntu, but encased in a browser window instead of a Mac app. Everyone is talking about ClawdBot/MoltBot though, this week.
I remember this thing a while back about how Google could've never been the first to build ChatGPT, because it was too risky for a big company like Google to launch a "research project" that mostly just ignored IP and safety concerns like ChatGPT initially did. And I think that's kind of the story about clawdbot; it's probably what Anthropic wanted to build, but unleashing a rogue AI on people's computers with aggressive access to everything is a risk they couldn't take.
Excellent points!
You might enjoy this one https://open.substack.com/pub/tatsuikeda/p/the-ai-adoption-paradox-why-95-of?r=az7fq&utm_medium=ios&shareImageVariant=overlay
Or this one! https://open.substack.com/pub/tatsuikeda/p/the-ai-dollar-part-16-ray-dalio-is?r=az7fq&utm_medium=ios&shareImageVariant=overlay
The idea of compute becoming oil-like seems interesting. I'm not exactly sure what the implications of that would be, but I can see the outlines of that somehow.
More fluff than a bag of cotton balls. Cowork was built in two weeks, it's imperfect, sure so was CC 8 months ago when it released. Three years ago these models could barely write a poem. Why is the other work structurally unadressable? Benchmarks? arguments form cognitive science? All I see is prople are saying...
I think it's very plausible (and likely?) that the robots can do the work; what I think is much harder is that the robots do the work *as us,* where they represent us (which is in effect what we're asking them to do when they write our emails). But I think that gets "solved" by us eventually not writing that many emails (or reports, or whatever) as us, and we stop reading most emails we get as being from someone and not from their EA / AI / EAI.
And but so aping DFW in business communication would make writing emails more fun for the email writer, and because the incentives of AI apps lean towards the Entertainment of the operator, it's a plausible outcome
I mean, if all of my emails were fun to read, I'd probably do a better job of reading my emails. I'm not sure I'd ever finish them though, since they'd all be 10 pages long with 15 pages of footnotes.
What I've noticed in 4 years selling dev tools, is that outside of big tech and startups, most companies do NOT care about code quality. They care about perception management. Saying you have code quality feeds a narrative, not a reality most times. That's consistently been the game I see people play. There are literally dozens of companies I talked with in the past that literally do not test their data even with dbt (you'd be very surprised who). "It works" often is enough for a lot of people because to your point, end users rarely look at the SQL and python behind their reports.
Until technical debt catches up to them and they lose the best talents. But yeah, still not a concern for most companies. They'll manage with average (or even poor) talent.
No one really cares that most IKEA furniture is made with particleboard and MDF (covered by veneers) because it is affordable, and that's what they really care about. Similarly, most companies and customers only care about the veneer.
I like that framing, "customers only care about the veneer" is more true than people (including me) like to admit.
I force myself to read claptrap like this to try and understand exactly how cluless humans think about ai.
Thank you for today’s example.
Teach us, Jedi Master.
I looked at your blog and you are maybe 18 months behind the curve which, for someone trying to sell AI implementation services, is deeply embarrassing for you. This post is by someone 2 years ahead of where you are. Please pay attention
You're totally full of nonsense.
18 months?
I doubt you could even accurately count to 18.
Have you even figured out 2 genders yet?
Another childish poster pretending to be something they aren't.
A condescending narcissistic BS artist.
Give it a rest.
Are you using an open model?
Frontier models can generate way better insults than this.
Are you stupid? I don't use models for insults. Another clueless boob who thinks they can tell the difference
Please go on. I think I'm learning.
simply type the comment into any chatbot and ask it
On a more serious note: a lot of the work I see AI eroding is of the “Bullshit Jobs” variety (à la David Graeber). Operational monstrosities that emerged from compliance requirements, poor app/org design, or operating models that shouldn’t exist in the first place (but somehow do - eg because of monopolistic markets, government subsidies, and other reasons the market often doesn’t solve)
A lot of emails + ppts + sending docs around fits into that description. You had to move information from A to B to C because the exec can’t walk the floor. Or move data from Salesforce to Finance’s weekly report because your company is too cheap to build a dashboard. Or nobody streamlined that horrible 45-step process that’s been growing steadily since 1989.
There’s a lot of pointless knowledge work that should have been redesigned long ago but survived for convenience or inertia - but it’s now becoming sufficiently convenient/expedient to automate, and so we do. Other such inefficiencies we dealt with when we got computers, or Excel, or the Internet
And so I think we agree on a lot of premises but less so the (full) conclusion?
Now, all this doesn’t add up to *Claude Cowork in particular succeeding* - e.g. if they stick to devs as their core persona, then Cowork will end up a slightly sad little experiment that never got the love and attention it deserved.
But I do think (or hope? Definitely hope) there’ll be a market for “Claude Code but make it less scary for those of us that dislike CLI UIs specifically”, lol. Maybe that’s an Opus Wrapper, or maybe Google will finally unfuck the way they’ve built Gemini (the app, not the models), or whatever.
In general we’re still very much in the “trying things out” stage of defining the right form factor + UX/UI paradigm for the age of LLMs (much like the iPhone wasn’t the first cellphone that could do more than call people, or why the PDA didn’t ever become a thing but tablets did). And I include form factor because hardware is as much fair game as software here.
Start following the AI Village experiment to see it doing much more than just BS jobs. It is still not good at it, but getting to a far wider field of action than people expect.
Or Gas Town: https://steve-yegge.medium.com/welcome-to-gas-town-4f25ee16dd04