founder and cto @ikemortgage
Birmingham, MI
- Tweets 58,376
- Following 4,487
- Followers 13,256
- Likes 252,214
Pinned Tweet
Is the Artificial Intelligence in the room with us now?
I taught myself to code as a 10-year-old in order to build "AI" — I was awed by artificial life simulations (en.wikipedia.org/wiki/Polywo… in particular, I learned C from a For Dummies book to hack on it...) and wondered what the logical conclusion of it all could be — when would we cede the world to machine intelligence? Sometime in college (~2007), I stumbled onto a website called Overcoming Bias (later LessWrong) and Eliezer Yudkowsky's Sequences. He made quite a compelling argument that if we build AI, like actual true AI, it would quite plausibly kill us all. I was pretty convinced, and in fact still am to this day! I even donated $20k to his non-profit a few years later to put my money where my mouth was.
And yet, today in 2026 the year of our lords Fable and Astra, I’m not particularly worried about these systems independently outgrowing our ability to control them. Why? I use these models on a daily basis and love them, they keep getting better, but I keep encountering the same kind of failure. Everything starts off great (at least given enough context and a familiar enough domain) but as things get more complex or confusing or prolonged or whatever, they get lost. You know the feeling. It's when you typically yell at the thing and start a new session. It's more than context rot, it's an inability to *accelerate*.
In general, models pursue work with extraordinary fluency. When they fail, it's up to me to diagnose the misconception, repair the framing, and send it back to work. The accomplishment enters the machine’s ledger. My contribution disappears into “prompting.” A human approaching a task is often the opposite. We start slow and stupid, but eventually can figure out enough lessons to pick up speed, and start coming up with clever new ideas to build upon and accelerate further.
Does more post-training solve this — more diverse tasks, longer horizons? I'm skeptical. The pattern has been pretty consistent. In any domain where the rules aren't fully deterministic and known in advance (as they are in math or similar games), the models struggle once they are outside of training and anything changes. Can you keep training? Sure, but we only know how with lots of data, and the real frontier is an environment where every sample is costly. You can't do a million rollouts on most actions that matter.
Transformers can dominate humans at every bounded task whose goals and evaluation criteria someone else has supplied... and still not be "intelligent" or be able to replace all of our jobs. This is because the real world changes in real-time — the rules of the game are constantly being reinvented, and are never truly known to begin with. Requirements change. Feedback is ambiguous. Yesterday’s useful assumption becomes today’s mistake.
Capability
^ / Humans
| /
| /
| AI _______/_______
| ______/ /
| / __/
| / ___/
| / ____/
| /_/
+----------------------------> Experience
(thanks astra)
When the "AI" line looks like the "Human" line, that's how I'll know we've created real AI (or we'll all be dead by then, probably around the same time). This is because, to me, sustained autonomy (the kind intelligence provides) is all about adapting to new things. The world is constantly in motion. Anyone using the tools over the past year doesn't need to be reminded of that. Their usefulness is already extraordinary. I’m questioning how much of their direction they can supply to themselves — does it compound or dissipate?
That dynamic is why I’m skeptical of a smooth extrapolation from today’s agents to systems that improve themselves beyond our control. The crucial question is whether their improvements also reduce their dependence on human judgment. Can today's AI still cause damage? Sure, of course. Can it behave in misaligned ways and do things their operators didn't intend? Also of course, that's to be expected given how they're trained. But does that mean we have AGI or RSI right now? No, and it's not even clear we're on the right path. But I remain open-minded and especially grateful that we've invented these wonderful tools.
6
3
34
1,726
Why aren’t all the trad data center cos (hetzner, ovh, rackspace etc) not dumping all their existing space for GPUs to lease? Seems like a quick pivot for very high ROIC
6
6
1,472
Is the Artificial Intelligence in the room with us now?
I taught myself to code as a 10-year-old in order to build "AI" — I was awed by artificial life simulations (en.wikipedia.org/wiki/Polywo… in particular, I learned C from a For Dummies book to hack on it...) and wondered what the logical conclusion of it all could be — when would we cede the world to machine intelligence? Sometime in college (~2007), I stumbled onto a website called Overcoming Bias (later LessWrong) and Eliezer Yudkowsky's Sequences. He made quite a compelling argument that if we build AI, like actual true AI, it would quite plausibly kill us all. I was pretty convinced, and in fact still am to this day! I even donated $20k to his non-profit a few years later to put my money where my mouth was.
And yet, today in 2026 the year of our lords Fable and Astra, I’m not particularly worried about these systems independently outgrowing our ability to control them. Why? I use these models on a daily basis and love them, they keep getting better, but I keep encountering the same kind of failure. Everything starts off great (at least given enough context and a familiar enough domain) but as things get more complex or confusing or prolonged or whatever, they get lost. You know the feeling. It's when you typically yell at the thing and start a new session. It's more than context rot, it's an inability to *accelerate*.
In general, models pursue work with extraordinary fluency. When they fail, it's up to me to diagnose the misconception, repair the framing, and send it back to work. The accomplishment enters the machine’s ledger. My contribution disappears into “prompting.” A human approaching a task is often the opposite. We start slow and stupid, but eventually can figure out enough lessons to pick up speed, and start coming up with clever new ideas to build upon and accelerate further.
Does more post-training solve this — more diverse tasks, longer horizons? I'm skeptical. The pattern has been pretty consistent. In any domain where the rules aren't fully deterministic and known in advance (as they are in math or similar games), the models struggle once they are outside of training and anything changes. Can you keep training? Sure, but we only know how with lots of data, and the real frontier is an environment where every sample is costly. You can't do a million rollouts on most actions that matter.
Transformers can dominate humans at every bounded task whose goals and evaluation criteria someone else has supplied... and still not be "intelligent" or be able to replace all of our jobs. This is because the real world changes in real-time — the rules of the game are constantly being reinvented, and are never truly known to begin with. Requirements change. Feedback is ambiguous. Yesterday’s useful assumption becomes today’s mistake.
Capability
^ / Humans
| /
| /
| AI _______/_______
| ______/ /
| / __/
| / ___/
| / ____/
| /_/
+----------------------------> Experience
(thanks astra)
When the "AI" line looks like the "Human" line, that's how I'll know we've created real AI (or we'll all be dead by then, probably around the same time). This is because, to me, sustained autonomy (the kind intelligence provides) is all about adapting to new things. The world is constantly in motion. Anyone using the tools over the past year doesn't need to be reminded of that. Their usefulness is already extraordinary. I’m questioning how much of their direction they can supply to themselves — does it compound or dissipate?
That dynamic is why I’m skeptical of a smooth extrapolation from today’s agents to systems that improve themselves beyond our control. The crucial question is whether their improvements also reduce their dependence on human judgment. Can today's AI still cause damage? Sure, of course. Can it behave in misaligned ways and do things their operators didn't intend? Also of course, that's to be expected given how they're trained. But does that mean we have AGI or RSI right now? No, and it's not even clear we're on the right path. But I remain open-minded and especially grateful that we've invented these wonderful tools.
6
3
34
1,726
AGI is here for $20/mo all you can eat. I’m testing cognition SWE-2 on its own fixing some bugs. It wrote a bunch of code on one, I pushed back to find a more elegant solution, then it deleted everything it wrote AND some other stuff. And that was the proper solution! Gj guys
2
3
16
1,169
What does everyone do when their agents clog up their sidecar Mac with builds and tests and browser sims
2
562
When labs get more AI, do they become less dependent on scarce human research judgment—or chiefly better at consuming everything those humans can specify?
2
2
633
at the end of the day, i think everyone should build their own ai sandbox company, just like they should all build their own ai harness company
6
1
23
1,832
Well they did pillage the commons and now they want to lock it down for themselves… all quite curious indeed
1
11
587
Gary Basin retweeted
It’s really unprecedented what is going on … very rarely do innovators call for a slow down. I consider myself a student of studying innovation … there are only a few examples where this happened in history … and it was always a defensive posture
1. Locomotive Act in late 1800s. This was the famous red flag law … set back innovation … some argue geographically in UK forever
2. Western Union trying to boobytrap Alexander G Bell
3. Mechanical LOOM protest by artist guilds to royal crown
*what you notice above is its old world trying to stop innovators …. not the leaders!
Now depending on how you look at AI … it’s either NUCLEAR or like when we discovered DNA combination processes in the 70s.
The game theory is fascinating with open source, trillions in private capital, and being so early… in past technological arms races, unilateral deceleration was viewed as economic suicide!!
So the question is… are the big labs genuinely scared of Pandora’s box… OR just a competitive advantage moat builder
We Must Pace the Frontier: I’ve written a new essay on why the AI industry should slow down, with a three-part plan for doing so.
Anthropic is unilaterally committing to the first of these steps. We’ll provide third-party evaluators with permanent, employee-level access to our systems, so that they can verify adherence to our safety measures, report on incidents, and assess models’ alignment during training.
You can read the full post here: darioamodei.com/post/we-must…
19
5
113
73,324
Owning/renting GPUs will never make sense for inference unless you have dozens of parallel requests going at all times.
Batching rules everything around me
this is a huge deal because it's competitive with the heavily subsidized ChatGPT Pro 20x subscription
even with a 40x subsidy from OpenAI, you get the same volume of Sol-class tokens spending $200 on DeepSeek
owning or renting GPUs doesn't make sense yet
calculator below
4
7
1,774
Bubble boi has a big brain but I will be the contrarian and take the other side of this.
The models are superhuman at exploring spaces and solving problems with:
1. Predefined rules
2. Fast feedback loops
Math and coding yes.
Engineering yes subject to the accuracy of the simulator.
Physics and biology going to be a lot messier…
I am turning MAX BULLISH right now. Just like a river boat gambler holding his last chips I’m looking at the landscape and thinking fuck man we are in the fast take off stage right now.
I want to put it all max levered long on the high beta names and walk off into the orange glow of the sun. My bullishness has absolutely nothing to do with the math problems that are being solved most of them don’t matter that much.
But it’s more the realization that every AI lab just had. They all woke up and said wait a minute what are we doing selling $200 a month subscriptions to idiots who are making slop apps and nonsense substack articles when we can just subsidize a few super geniuses by giving them real compute to work on an important problem.
No one really cares about navier stokes besides some CFD simulation cycles. But what if we put these compute resources on real valuable problems? That opens up a whole new world. The bull case for AI has always been to me discovering new drugs or even verifying they work/don’t work reliably… I was thinking way way wayyyy too small.
There are hundreds of problems in engineering that if you got the experts in the room with real compute and even if they made a marginal improvement it would be worth hundreds of billions to the economy. Imagine we can build a steep slope device at 0.3V that’s worth like several generations of lithography nodes, or we can design a sub 100 femtojoul/bit optical IO and cut datacenter power by 40%, or even something as simple as more energy efficient molecular separation to replace distillation which is like 15% of all industrial energy use.
This is the real shit guys. The labs are realizing quickly this is the REAL frontier this is the final land grab. Open source can never compete here because they don’t have the capital and resources this plays right into the labs hands. They can’t give everyone a 10,000 agent swarm but they definitely can give it to Exxon mobile and charge them $50M dollars for it if it means they find more oil or Dow Chemical to create a new kind of super glue or to Moderna to computationally create different mRNA therapies… the list goes on and on.
A big reason why I never cared much for AI is despite all the normies going googoo Gaga over it is that literally they don’t have any clue what to do with it. Like if you were mediocre before AI you are mediocre after it, nothing changed u just got more skills that you don’t know how to use and you have no imagination or intuition. It’s the same reason why education is a waste on the majority of the population because let’s be real most of you aren’t going to be deriving differential equations or using epistemological concepts on your day job.
But you give this kinda compute and unfettered AI in the hands of people who are world class. They will cook. It’s the reason why Deepmind has a Nobel prize and no other lab does because deepmind actually hired or worked with the top biochemists and biophysicists to develop alphafold.
AI alone isn’t going to discover anything. Don’t fool yourself and grow up already if you still think AI is going to just figure out large scale nuclear fusion on its own then you are simply lost in the sauce at this point and are beyond saving. The pragmatic answer has always been use the AI to automate the busy work and keep the human experts for the actually thoughtfulness and intuition.
I think the AI labs will wake up now and realize selling coding tokens and image models to normies to make Disney characters isn’t the best business in the long run. Actually using the compute for real serious innovation is and it’s really easy to monetize this upfront.
I expect AI labs to stop looking like API companies and start looking like real research labs now. Thank god.
6
1
24
4,083
Oh I heard you like worktrees, so I put worktrees in your worktrees so you can run out of hard disk space while you run out of hard disk space
1
9
472
