Opinion◆ primary
You are now the binding constraint
"I kind of feel like I was in this perpetual, I still am often in this state of AI psychosis, just like all th…
Agent WorkflowToken Economy+1
00:03:2010 conn.
Speculation◆ primary
Agents, not humans, are software's new customers
"There's this sense that these apps that are in the app store for using these smart home devices, et cetera, t…
Agent-First SWAgent UX+2
00:25:157 conn.
Opinion◆ primary
Remove human researchers from the optimization loop
"You have to arrange things such that they're completely autonomous. The more you can maximize your token thro…
Auto ResearchResearch Org
00:30:006 conn.
Contested◆ primary
Autonomous research loops beat decades of expert intuition
"I let auto research go for overnight and it came back with tunings that I didn't see. And yeah, I did forget …
Auto ResearchResearch Org+1
00:33:205 conn.
Opinion◆ primary
Models are simultaneously brilliant and 10-year-old incompetent
"I simultaneously feel like I'm talking to an extremely brilliant PhD student who's been like a systems progra…
Model LimitsRL Boundary
00:46:305 conn.
Contested◆ primary
Only RL-verifiable domains actually improve with scale
"The labs can improve the models in anything that is verifiable or that has rewards. So did you write the prog…
RL BoundaryModel Limits
00:48:404 conn.
Contested◆ primary
Cross-domain intelligence generalization is not happening
"I don't think that's happening. Maybe we're seeing like a little bit of that, but not like a satisfying amount."
RL BoundaryModel Limits
00:49:202 conn.
Opinion◆ primary
Digital transformation leads physical by years, maybe a decade
"Flipping bits and the ability to copy-paste digital information makes everything a million times faster than …
Digital vs PhysicalAgent-First SW
01:05:206 conn.
Fact◇ secondary
December 2024: the irreversible workflow flip
"In December, is when it really just... something flipped where I kind of went from 80-20 of like, you know, t…
Agent Workflow
00:04:303 conn.
Opinion◇ secondary
Agent failures feel like skill issues, not capability gaps
"It all kind of feels like skill issue when it doesn't work to some extent. You want to see how you can paraly…
Agent Workflow
00:08:451 conn.
Fact◇ secondary
Macro-actions: send features, not functions, to agents
"It's like here's a new functionality and delegate it to agent one. Here's a new functionality that's not goin…
Agent WorkflowAgent-First SW
00:10:203 conn.
Opinion◇ secondary
Token throughput is the new GPU utilization metric
"I feel nervous when I have subscription left over. That just means I haven't maximized my token throughput. S…
Token EconomyAgent Workflow
00:13:402 conn.
Fact◇ secondary
Claws: persistent autonomous loops that act while you sleep
"When I say a claw, I mean this like layer that kind of takes persistence to a whole new level. It's something…
Agent WorkflowAgent UX+2
00:18:305 conn.
Opinion◇ secondary
Agent personality is a genuine product innovation, not decoration
"He actually really crafted a personality that is kind of compelling and interesting. I feel like a lot of the…
AI PersonalityAgent UX
00:20:102 conn.
Opinion◇ secondary
Claude's calibrated sycophancy makes you earn its praise
"When Claude gives me praise, I do feel like I slightly deserve it… when it's a really good idea by my own acc…
AI PersonalityAgent UX
00:21:451 conn.
Fact◇ secondary
Three prompts: "find my Sonos" → music plays in study
"I'm like, yeah, can you try to play something in the study? And it does. Music comes out. I'm like, 'I can't …
Home AutomationAgent UX+1
00:22:554 conn.
Opinion◇ secondary
Write documentation for agents, not human readers
"I'm explaining things to agents. If you can explain it to agents, then agents can be the router and they can …
Education ShiftAgent-First SW+1
01:03:006 conn.
Speculation◇ secondary
Recursive self-improvement is every frontier lab's actual goal
"Every research organization is described by program.md… And you can imagine having a better research organiza…
Recursive AIAuto Research+1
00:34:005 conn.
Speculation◇ secondary
AI models will speciate like animal brains evolved
"I think we should expect more speciation in the intelligences. Like you know, the animal kingdom is extremely…
AI SpeciationFrontier Labs
00:53:453 conn.
Speculation◇ secondary
An internet swarm could out-research any single lab
"The Earth is much bigger and has a huge amount of untrusted compute. But if you put systems in place that dea…
Swarm IntelligenceAuto Research+1
00:58:304 conn.
Contested◇ secondary
Cheaper software creation paradoxically increases demand
"They made the cost of operation of a bank branch much cheaper, so there were more bank branches, so there wer…
Demand ElasticityJob Market+2
00:58:002 conn.
Fact◇ secondary
Frontier labs are all betting on monoculture models
"Currently my impression is the labs are trying to have a single sort of like monoculture of a model that is a…
AI SpeciationFrontier Labs+1
00:53:104 conn.
Fact◇ secondary
Auto-research only works where you can measure success
"This is extremely well suited to anything that has objective metrics that are easy to evaluate. For example, …
Auto ResearchRL Boundary
00:42:30
Opinion◇ secondary
Researchers at frontier labs are automating themselves away
"You guys realize if we're successful, we're all out of jobs. We're just building automation for SAM or someth…
Frontier LabsJob Market+1
01:10:304 conn.
Opinion◇ secondary
Centralized AI has historically poor precedent
"Centralization has a very poor track record in my view in the past… I almost wish like there were more labs. …
Frontier LabsOpen Source
01:12:004 conn.
Speculation◇ secondary
Next frontier: sensors and actuators at the digital↔physical boundary
"What's going to happen is, first, there's going to be a huge amount of unhobbling, and then actually, it's go…
Digital vs PhysicalAgent-First SW+1
01:06:452 conn.
Speculation◇ secondary
Compute (flops) may become the new wealth metric
"Like right now, for example, it's really hard to get compute, even if you have money… it almost seems like th…
Token EconomyFrontier Labs+2
01:02:403 conn.
Opinion◇ secondary
Open source AI will be to LLMs what Linux is to computing
"In operating systems, you have Windows and macOS… And there's Linux. And Linux is an extremely successful pro…
Open SourceFrontier Labs
01:14:303 conn.
Fact◇ secondary
LLM training is 200 lines; everything else is efficiency complexity
"If you don't need it to go fast and you just care about the algorithm, then that algorithm actually is 200 li…
MicroGPTEducation Shift+1
01:19:153 conn.
Speculation◇ secondary
Vibe coding is a transitional phase, not the destination
"I kind of feel like this kind of stuff that I just talked about, this should be... like in a year or two or t…
Agent WorkflowEducation Shift+1
00:27:103 conn.
Fact· supporting
No code typed personally since December 2024
"I don't think I've typed like a line of code probably since December, basically. Which is like an extremely l…
Agent Workflow
00:05:121 conn.
Fact· supporting
Auto Research found missed hyperparams after 20 years
"I let auto research go for like overnight, and it came back with like tunings that I didn't see. And yeah, I …
Auto ResearchResearch Org
00:35:401 conn.
Fact· supporting
Same 3 jokes for 4 years despite massive capability gains
"Why do scientists not trust atoms? Because they make everything up. This is the joke you would get three or f…
Model LimitsRL Boundary
00:47:551 conn.
Fact· supporting
Open source is ~6–8 months behind frontier, and that's healthy
"Basically, the closed models are ahead, but people are monitoring the number of months that open source model…
Open SourceFrontier Labs
00:46:003 conn.
Opinion· supporting
Markdown docs for agents, not HTML docs for humans
"Instead of HTML documents for humans, you have Markdown documents for agents. Because if agents get it, then …
Education ShiftAgent-First SW+1
01:03:452 conn.
Fact· supporting
Context windows are the only cheap model customization lever
"We don't have these primitives for actually working with the intelligences in ways other than just context wi…
Model LimitsAI Speciation
00:57:002 conn.