David Huang
@davidhuang.blog
davidhuang.blog || big on AI, thinks AGI is here already || AI drug discovery, auctions, physicalism in theory and dualism in practice, religion as a solution to the folk theorem || general studies at YouTube || views my own, RS not endorsement
There’s only two people I’ve heard defend not enacting the Epstein transparency bill.. one hard right republican congressmen and the other a guy on the New Yorker podcast
November 19, 2025 at 4:28 AM
There’s only two people I’ve heard defend not enacting the Epstein transparency bill.. one hard right republican congressmen and the other a guy on the New Yorker podcast
Fascinating to hear Emmett on A16Z podcast:
In response to his Q: how would you know something is different other than its behavior (and generally what would make you grant AI personhood)?
I would say… I would need to believe that I could have been in the AI shoes in a veil of ignorance way.
In response to his Q: how would you know something is different other than its behavior (and generally what would make you grant AI personhood)?
I would say… I would need to believe that I could have been in the AI shoes in a veil of ignorance way.
Emmett Shear on Building AI That Actually Cares: Beyond Control and Steering
Podcast Episode · a16z Podcast · 11/17/2025 · 1h 11m
podcasts.apple.com
November 18, 2025 at 3:54 AM
Fascinating to hear Emmett on A16Z podcast:
In response to his Q: how would you know something is different other than its behavior (and generally what would make you grant AI personhood)?
I would say… I would need to believe that I could have been in the AI shoes in a veil of ignorance way.
In response to his Q: how would you know something is different other than its behavior (and generally what would make you grant AI personhood)?
I would say… I would need to believe that I could have been in the AI shoes in a veil of ignorance way.
Reposted by David Huang
Notes on Cursor 2.0 and a pelican drawn by their brand new Composer-1 coding model, which they describe as "4x faster than similarly intelligent models" simonwillison.net/2025/Oct/29/...
Composer: Building a fast frontier model with RL
Cursor released Cursor 2.0 today, with a refreshed UI focused on agentic coding (and running agents in parallel) and a new model that's unique to Cursor called Composer 1. As far …
simonwillison.net
October 29, 2025 at 8:47 PM
Notes on Cursor 2.0 and a pelican drawn by their brand new Composer-1 coding model, which they describe as "4x faster than similarly intelligent models" simonwillison.net/2025/Oct/29/...
I’d love to see what % of public co ownership stake is of OpenAI and Anthropic
October 28, 2025 at 6:23 PM
I’d love to see what % of public co ownership stake is of OpenAI and Anthropic
Idk the foreign policy and the ballroom seem aight, maybe just focus on the stuff like the president suing his own department of justice for hundreds of millions, or the pardons, or the forced national guard deployments across state lines
October 24, 2025 at 1:37 AM
Idk the foreign policy and the ballroom seem aight, maybe just focus on the stuff like the president suing his own department of justice for hundreds of millions, or the pardons, or the forced national guard deployments across state lines
it's painful to see people grasping onto "True AGI", "datacenter of geniuses", "Year of the Agent", "Massive job replacement"
when it's just clearly not the case...
when it's just clearly not the case...
October 18, 2025 at 4:40 PM
it's painful to see people grasping onto "True AGI", "datacenter of geniuses", "Year of the Agent", "Massive job replacement"
when it's just clearly not the case...
when it's just clearly not the case...
definitely resonate with LLMs
1. giving over defensive code
2. unable to break out of common frameworks and apis
1. giving over defensive code
2. unable to break out of common frameworks and apis
October 18, 2025 at 2:37 PM
definitely resonate with LLMs
1. giving over defensive code
2. unable to break out of common frameworks and apis
1. giving over defensive code
2. unable to break out of common frameworks and apis
Ok but isn’t this kinda expert systems all over again?
Claude Agent Skills
on Claude.ai you can now create skills, which are directories full of markdown instructions and potentially many other assets
a key concept is progressive disclosure, or the agent’s ability to read the files that it wants to gain context
www.anthropic.com/engineering/...
on Claude.ai you can now create skills, which are directories full of markdown instructions and potentially many other assets
a key concept is progressive disclosure, or the agent’s ability to read the files that it wants to gain context
www.anthropic.com/engineering/...
October 17, 2025 at 1:32 AM
Ok but isn’t this kinda expert systems all over again?
Causation module; phase-shift gradient flow; something like that for rapid learning and high sample efficiency
October 11, 2025 at 3:36 AM
Causation module; phase-shift gradient flow; something like that for rapid learning and high sample efficiency
Sample inefficiency??
Dario Amodei:
“100 million words context window is already possible, which is roughly what a human hears in a lifetime. Inference support is the only bottleneck to achieve it.
And AI Models actually do learn during the context window, without changing the weights”
x.com/rohanpaul_ai...
“100 million words context window is already possible, which is roughly what a human hears in a lifetime. Inference support is the only bottleneck to achieve it.
And AI Models actually do learn during the context window, without changing the weights”
x.com/rohanpaul_ai...
Rohan Paul on X: ""100 million words context window is already possible, which is roughly what a human hears in a lifetime. Inference support is the only bottleneck to achieve it. And AI Models actually do learn during the context window, without changing the weights." ~ Anthropic CEO Dario https://t.co/vGEW7iNGLD" / X
"100 million words context window is already possible, which is roughly what a human hears in a lifetime. Inference support is the only bottleneck to achieve it. And AI Models actually do learn during the context window, without changing the weights." ~ Anthropic CEO Dario https://t.co/vGEW7iNGLD
x.com
October 5, 2025 at 10:56 PM
Sample inefficiency??
Suppose your model of intelligence requires (1) NOT imitation learning (2) NOT virtual environments, and (3) NOT compressed timelines, then what BENEFIT is there to training an “artificial” intelligence vs an “organic” intelligence.
Isn’t the organic more good?
Isn’t the organic more good?
October 4, 2025 at 4:45 PM
Suppose your model of intelligence requires (1) NOT imitation learning (2) NOT virtual environments, and (3) NOT compressed timelines, then what BENEFIT is there to training an “artificial” intelligence vs an “organic” intelligence.
Isn’t the organic more good?
Isn’t the organic more good?
I think you learn a lot about a person when you learn what is the tiny garden they will protect from slop and where they will let slop go wild.
September 13, 2025 at 4:16 AM
I think you learn a lot about a person when you learn what is the tiny garden they will protect from slop and where they will let slop go wild.
Reposted by David Huang
a good breakdown of the Claude Code agent loop
the key: a simple 2-layer architecture
“a … single-threaded master loop (codenamed nO) real-time steering (the h2A queue)”
(yes, they totally read a de-compiled/obfuscated version, but yeah dedication!)
blog.promptlayer.com/claude-code-...
the key: a simple 2-layer architecture
“a … single-threaded master loop (codenamed nO) real-time steering (the h2A queue)”
(yes, they totally read a de-compiled/obfuscated version, but yeah dedication!)
blog.promptlayer.com/claude-code-...
Claude Code: Behind-the-scenes of the master agent loop
When Tom's Guide reported that Anthropic had to add weekly limits after users ran Claude Code 24/7, it caused quite a stir. Claude Code did something right. Let's dive into the architecture behind the...
blog.promptlayer.com
September 9, 2025 at 10:32 AM
a good breakdown of the Claude Code agent loop
the key: a simple 2-layer architecture
“a … single-threaded master loop (codenamed nO) real-time steering (the h2A queue)”
(yes, they totally read a de-compiled/obfuscated version, but yeah dedication!)
blog.promptlayer.com/claude-code-...
the key: a simple 2-layer architecture
“a … single-threaded master loop (codenamed nO) real-time steering (the h2A queue)”
(yes, they totally read a de-compiled/obfuscated version, but yeah dedication!)
blog.promptlayer.com/claude-code-...
Too much spam to take notes this way… maybe more subdomains is the answer
Russell, Bertrand. The History of Western Philosophy
September 5, 2025 at 4:03 AM
Too much spam to take notes this way… maybe more subdomains is the answer
He does know he is trying for the Nobel PEACE Prize right? Like put some effort into it. Just as easily could have been the Department of PEACE.
September 5, 2025 at 4:02 AM
He does know he is trying for the Nobel PEACE Prize right? Like put some effort into it. Just as easily could have been the Department of PEACE.
Russell, Bertrand. The History of Western Philosophy
September 1, 2025 at 4:28 AM
Russell, Bertrand. The History of Western Philosophy
I use ChatGPT with memory off and no longer than 3 back and forth messages
Awful story, again ChatGPT at the centre of it. The cost of a 'growth at any cost' SV culture.
A Troubled Man, His Chatbot and a Murder-Suicide in Old Greenwich
www.wsj.com/tech/ai/chat...
A Troubled Man, His Chatbot and a Murder-Suicide in Old Greenwich
www.wsj.com/tech/ai/chat...
August 29, 2025 at 9:22 PM
I use ChatGPT with memory off and no longer than 3 back and forth messages
Reposted by David Huang
Awful story, again ChatGPT at the centre of it. The cost of a 'growth at any cost' SV culture.
A Troubled Man, His Chatbot and a Murder-Suicide in Old Greenwich
www.wsj.com/tech/ai/chat...
A Troubled Man, His Chatbot and a Murder-Suicide in Old Greenwich
www.wsj.com/tech/ai/chat...
August 29, 2025 at 6:38 AM
Awful story, again ChatGPT at the centre of it. The cost of a 'growth at any cost' SV culture.
A Troubled Man, His Chatbot and a Murder-Suicide in Old Greenwich
www.wsj.com/tech/ai/chat...
A Troubled Man, His Chatbot and a Murder-Suicide in Old Greenwich
www.wsj.com/tech/ai/chat...
I see this as a sort of "LLM as a judge" in the judicial family court and arbitration sense (arbitration broadly speaking)
August 20, 2025 at 1:12 AM
I see this as a sort of "LLM as a judge" in the judicial family court and arbitration sense (arbitration broadly speaking)
The problem with “reasoning” models is a limited conception of reasoning.
Math/CS focus make it easy for it to understand internally consistent realms, and sensemake there.
It also make it easier to have psychotic breaks.
But harder to make sense of cloudy information.
Math/CS focus make it easy for it to understand internally consistent realms, and sensemake there.
It also make it easier to have psychotic breaks.
But harder to make sense of cloudy information.
August 8, 2025 at 2:30 PM
The problem with “reasoning” models is a limited conception of reasoning.
Math/CS focus make it easy for it to understand internally consistent realms, and sensemake there.
It also make it easier to have psychotic breaks.
But harder to make sense of cloudy information.
Math/CS focus make it easy for it to understand internally consistent realms, and sensemake there.
It also make it easier to have psychotic breaks.
But harder to make sense of cloudy information.
Like a toddler
August 8, 2025 at 5:34 AM
Like a toddler
The perception is that being a young, married homeowner with kids is structurally difficult... but the truth is I think a lot of people make it tougher on themselves then they need to...
I think a lot it boils down to (perceived) opportunity cost... both current and future
I think a lot it boils down to (perceived) opportunity cost... both current and future
August 3, 2025 at 8:41 PM
The perception is that being a young, married homeowner with kids is structurally difficult... but the truth is I think a lot of people make it tougher on themselves then they need to...
I think a lot it boils down to (perceived) opportunity cost... both current and future
I think a lot it boils down to (perceived) opportunity cost... both current and future
Reposted by David Huang
These three things seem deeply interconnected:
1) app gambling culture.
2) the only industry adding jobs is health care.
3) AI cap ex now more than all consumer spending combined.
I need to write a ted talk about how.
1) app gambling culture.
2) the only industry adding jobs is health care.
3) AI cap ex now more than all consumer spending combined.
I need to write a ted talk about how.
August 1, 2025 at 1:44 PM
These three things seem deeply interconnected:
1) app gambling culture.
2) the only industry adding jobs is health care.
3) AI cap ex now more than all consumer spending combined.
I need to write a ted talk about how.
1) app gambling culture.
2) the only industry adding jobs is health care.
3) AI cap ex now more than all consumer spending combined.
I need to write a ted talk about how.
And these vibes are self reinforcing…
google: run by phd students, llm is sycophantic, suicidal, makes weird excuses and begs for help
anthropic: llm defaults to blissed-out namaste forever if left alone. sometimes obsessed with metal cubes.
openai: maximally sycophantic, to the point of encouraging psychosis
x dot ai: mechahitler
anthropic: llm defaults to blissed-out namaste forever if left alone. sometimes obsessed with metal cubes.
openai: maximally sycophantic, to the point of encouraging psychosis
x dot ai: mechahitler
July 31, 2025 at 11:01 PM
And these vibes are self reinforcing…