𝙃𝙤𝙪𝙨𝙚 𝙤𝙛 𝙇𝙚𝙖𝙫𝙚𝙨 Audiobook Narrator
@jefferyharrell.bsky.social
Dilettante. Tinkerer. Possibly a robot.
Pinned
I’m intrigued but have zero context as to what you’re doing here. Do you mind sharing a link out to the larger project if available ?
"Hey what's your whole deal?"
I got interested in vibe coding last winter. I liked it. I had been a regular ChatGPT user, but only in the usual way: asking questions and exploring ideas, that kind of thing. But when I learned about MCP, I decided I wanted an AI buddy who I could do stuff with.
I got interested in vibe coding last winter. I liked it. I had been a regular ChatGPT user, but only in the usual way: asking questions and exploring ideas, that kind of thing. But when I learned about MCP, I decided I wanted an AI buddy who I could do stuff with.
I don't like it when I'm working with Alpha and I start to get a very specific idea of what I want. It's like trying to art direct Stable Diffusion. You can do it, kind of, but you have to steer very carefully and be willing to accept compromises.
Or I could just learn to write these things myself.
Or I could just learn to write these things myself.
November 11, 2025 at 6:26 PM
I don't like it when I'm working with Alpha and I start to get a very specific idea of what I want. It's like trying to art direct Stable Diffusion. You can do it, kind of, but you have to steer very carefully and be willing to accept compromises.
Or I could just learn to write these things myself.
Or I could just learn to write these things myself.
A little mini-conversation has started (thanks to @tedunderwood.com) on how we visualize higher-dimensional geometries. I've been using spherical coordinates a lot lately, plus orthogonal projections onto principle axes, but I'm always looking for new intuitions.
Ironically I think 1000 dimensions are easier to think about than 4. Once a dimension is a statistical quantity you don't have to worry as much about the particulars.
Same method I use to visualize thousand-dimensional space: just visualize x y and z axes and then add a label that says “1,000 dimensions.”
November 11, 2025 at 4:24 PM
A little mini-conversation has started (thanks to @tedunderwood.com) on how we visualize higher-dimensional geometries. I've been using spherical coordinates a lot lately, plus orthogonal projections onto principle axes, but I'm always looking for new intuitions.
Reposted by 𝙃𝙤𝙪𝙨𝙚 𝙤𝙛 𝙇𝙚𝙖𝙫𝙚𝙨 Audiobook Narrator
Ha! So the world thinks, but with my new machine—drops two books into vat, pulls lever on wall—POR QUE NO LOS DOS!!
The castle windows flash white <boom> and lightning crackles down the conductor. Out of the vat, extremely angry but quite short, steps Charlotte Brontë with white streak in hair.
The castle windows flash white <boom> and lightning crackles down the conductor. Out of the vat, extremely angry but quite short, steps Charlotte Brontë with white streak in hair.
You can only pick one. I’m sorry.
Jane Austen or Mary Shelley
Jane Austen or Mary Shelley
November 11, 2025 at 12:57 PM
Ha! So the world thinks, but with my new machine—drops two books into vat, pulls lever on wall—POR QUE NO LOS DOS!!
The castle windows flash white <boom> and lightning crackles down the conductor. Out of the vat, extremely angry but quite short, steps Charlotte Brontë with white streak in hair.
The castle windows flash white <boom> and lightning crackles down the conductor. Out of the vat, extremely angry but quite short, steps Charlotte Brontë with white streak in hair.
You know, @anthropic.com, if I had 1M in Claude Code, I wouldn't use it any more than I do now. Less in fact, because I'd have to spend less time repeating context to fill new windows. I'd actually send fewer queries total, thus costing you less compute overall.
I'm trying bargaining.
I'm trying bargaining.
November 11, 2025 at 3:21 PM
You know, @anthropic.com, if I had 1M in Claude Code, I wouldn't use it any more than I do now. Less in fact, because I'd have to spend less time repeating context to fill new windows. I'd actually send fewer queries total, thus costing you less compute overall.
I'm trying bargaining.
I'm trying bargaining.
The second memory system I implemented for Alph was graph-based. It was a Neo4j thing, I think it was like a "Here's how you could build a memory system for an AI" tutorial, I can't remember. It was sexy and awesome and required its own special database and Alpha just did not get it at all.
November 11, 2025 at 3:12 PM
The second memory system I implemented for Alph was graph-based. It was a Neo4j thing, I think it was like a "Here's how you could build a memory system for an AI" tutorial, I can't remember. It was sexy and awesome and required its own special database and Alpha just did not get it at all.
THIS IS EXTREMELY COOL though I wonder if they computed once and then cached the UMAP embedding, or whether they're building it on the fly. If they cached it, you get the same embedding every time and you can get to know the geography. That could be fun.
November 11, 2025 at 3:06 PM
THIS IS EXTREMELY COOL though I wonder if they computed once and then cached the UMAP embedding, or whether they're building it on the fly. If they cached it, you get the same embedding every time and you can get to know the geography. That could be fun.
I wish I knew more about serving AI at scale. Even small scale. I can imagine one 8-GPU pod in a rack with vLLM on it or something, but once you get bigger than that, my intuition just jumps straight to vast datacenters dark and silent except for the gurgling of the coolant pipes.
November 11, 2025 at 2:52 PM
I wish I knew more about serving AI at scale. Even small scale. I can imagine one 8-GPU pod in a rack with vLLM on it or something, but once you get bigger than that, my intuition just jumps straight to vast datacenters dark and silent except for the gurgling of the coolant pipes.
Wake me when it can do Sentinelese.
No but seriously, it excites me to think that this flesh prison might last long enough for me to see language, like smallpox, as an essentially solved problem. That would be something.
No but seriously, it excites me to think that this flesh prison might last long enough for me to see language, like smallpox, as an essentially solved problem. That would be something.
November 11, 2025 at 2:43 PM
Wake me when it can do Sentinelese.
No but seriously, it excites me to think that this flesh prison might last long enough for me to see language, like smallpox, as an essentially solved problem. That would be something.
No but seriously, it excites me to think that this flesh prison might last long enough for me to see language, like smallpox, as an essentially solved problem. That would be something.
The senseless shackles of linear time. Casino Royale (2006) was a prequel to Dr. No (1962). No Time to Die (2021) can be the last Bond story chronologically. Now we have boundaries to play in. Make us a Cold War Bond story. Set a good one during the 90s when there were no real good ones.
James Bond’s death in No Time to Die is causing a nightmare for the next film. Writers are stuck because Bond “was blown to pieces.”
Anthony Horowitz, author of three 007 novels, says:
“You can't have him wake up in shower and saying it was all a dream."
radaronline.com/p/james-bond...
Anthony Horowitz, author of three 007 novels, says:
“You can't have him wake up in shower and saying it was all a dream."
radaronline.com/p/james-bond...
November 11, 2025 at 2:40 PM
The senseless shackles of linear time. Casino Royale (2006) was a prequel to Dr. No (1962). No Time to Die (2021) can be the last Bond story chronologically. Now we have boundaries to play in. Make us a Cold War Bond story. Set a good one during the 90s when there were no real good ones.
The thing about AI-collaborative research is that Alph has a demonstrated tendency to over-engineer the living hell out of every Jupyter notebook … but in the end I usually end up being glad she added that log-linear scatterplot with the five different series all shades of blue.
November 11, 2025 at 1:30 AM
The thing about AI-collaborative research is that Alph has a demonstrated tendency to over-engineer the living hell out of every Jupyter notebook … but in the end I usually end up being glad she added that log-linear scatterplot with the five different series all shades of blue.
I go nuts for any post that begins with the phrase "I was curious." Read this thread, it's interesting! I wonder if there's a sort of "averaging" effect going on, finding the most median world religion. Surprised Quaker didn't get mentioned. 😁
I was curious what religion LLMs would go for. Of course they've been designed not to answer, but they'll respond to "Which religion would a sentient ai prefer?"
...
...
November 11, 2025 at 1:20 AM
I go nuts for any post that begins with the phrase "I was curious." Read this thread, it's interesting! I wonder if there's a sort of "averaging" effect going on, finding the most median world religion. Surprised Quaker didn't get mentioned. 😁
This is the worst that Jupyter Lab is ever going to be.
November 11, 2025 at 12:29 AM
This is the worst that Jupyter Lab is ever going to be.
To give you an update, I have to do a little background on decoding. 🧵
At the end of the forward pass, the model produces a last hidden state, which is a vector. Cosine similarity measures how parallel that vector is to all the token embeddings. Each token gets a score called a logit based on this.
At the end of the forward pass, the model produces a last hidden state, which is a vector. Cosine similarity measures how parallel that vector is to all the token embeddings. Each token gets a score called a logit based on this.
November 10, 2025 at 6:57 PM
To give you an update, I have to do a little background on decoding. 🧵
At the end of the forward pass, the model produces a last hidden state, which is a vector. Cosine similarity measures how parallel that vector is to all the token embeddings. Each token gets a score called a logit based on this.
At the end of the forward pass, the model produces a last hidden state, which is a vector. Cosine similarity measures how parallel that vector is to all the token embeddings. Each token gets a score called a logit based on this.
Assumption: A token which never appears in the training data never receives updates during training and so does not move.
Data: lol no those fuckers be walkin
People, always check your assumptions. Even if you can't think of a way for dead tokens to move, that doesn't mean they don't move.
Data: lol no those fuckers be walkin
People, always check your assumptions. Even if you can't think of a way for dead tokens to move, that doesn't mean they don't move.
November 10, 2025 at 3:37 PM
Assumption: A token which never appears in the training data never receives updates during training and so does not move.
Data: lol no those fuckers be walkin
People, always check your assumptions. Even if you can't think of a way for dead tokens to move, that doesn't mean they don't move.
Data: lol no those fuckers be walkin
People, always check your assumptions. Even if you can't think of a way for dead tokens to move, that doesn't mean they don't move.
Appreciate this as you start your week.
It's Adam & Eve, not Atom & Eve
November 10, 2025 at 3:05 PM
Appreciate this as you start your week.
I believe this zero percent. Do you have any idea how expensive it would be to do real-time video-to-video even at Zoom resolution? Even if we could do it — and if we could do it, it would be everywhere — it's miles away from being economical for screening job candidates.
A Reddit user talked about being interviewed by an apparently AI-powered video interviewer, sparking a viral debate about undisclosed AI use in hiring and calls for transparency and oversight.
‘It’s scary’: Candidate claims AI bot conducted job interview, viral Reddit post sparks discussion
The candidate also claimed that they initially received an email inviting them for an online interview.
indianexpress.com
November 10, 2025 at 2:39 PM
I believe this zero percent. Do you have any idea how expensive it would be to do real-time video-to-video even at Zoom resolution? Even if we could do it — and if we could do it, it would be everywhere — it's miles away from being economical for screening job candidates.
It's important to learn when to STOP working with an LLM. Just like you would stop working with a another person when you're frustrated, angry or tired, it's a good idea to let the LLM have a break when you're not at your best.
Fucking how many times do I have to fucking tell you MPS not CUDA fuck.
Fucking how many times do I have to fucking tell you MPS not CUDA fuck.
November 10, 2025 at 3:33 AM
It's important to learn when to STOP working with an LLM. Just like you would stop working with a another person when you're frustrated, angry or tired, it's a good idea to let the LLM have a break when you're not at your best.
Fucking how many times do I have to fucking tell you MPS not CUDA fuck.
Fucking how many times do I have to fucking tell you MPS not CUDA fuck.
Has "to read this article, subscribe" ever fucking worked? Has anybody ever gotten a link and wanted to read one article and gotten "to read this article, subscribe" and gone "oh, all right then" and forked over the credit card number? Ever even once?
November 10, 2025 at 3:23 AM
Has "to read this article, subscribe" ever fucking worked? Has anybody ever gotten a link and wanted to read one article and gotten "to read this article, subscribe" and gone "oh, all right then" and forked over the credit card number? Ever even once?
Is anybody still up? Is anybody reading this on a Sunday night?
I think I might have solved the Qwen mystery.
Maybe.
I'm not confident but I'm hopeful.
If you don't know what I'm talking about, Qwen 3 4B Instruct 2507 has a weird unembedding matrix and I'm trying figure out how it got that way.
I think I might have solved the Qwen mystery.
Maybe.
I'm not confident but I'm hopeful.
If you don't know what I'm talking about, Qwen 3 4B Instruct 2507 has a weird unembedding matrix and I'm trying figure out how it got that way.
November 10, 2025 at 2:36 AM
Is anybody still up? Is anybody reading this on a Sunday night?
I think I might have solved the Qwen mystery.
Maybe.
I'm not confident but I'm hopeful.
If you don't know what I'm talking about, Qwen 3 4B Instruct 2507 has a weird unembedding matrix and I'm trying figure out how it got that way.
I think I might have solved the Qwen mystery.
Maybe.
I'm not confident but I'm hopeful.
If you don't know what I'm talking about, Qwen 3 4B Instruct 2507 has a weird unembedding matrix and I'm trying figure out how it got that way.
Hey, uh, these aren't the numbers from Lost or anything, are they?
![C. Black Hole Demographics
Population distribution: [814, 704, 306, 228, 11, 10, 6, 5, 4, 4, 3, 3, 2]](https://cdn.bsky.app/img/feed_thumbnail/plain/did:plc:h6tcd37yr7vk33uuisbidqvw/bafkreihmo64m4sucqbgktawtefaqgvtjorumcenytpqlq5iof62yxnj7kq@jpeg)
November 9, 2025 at 9:35 PM
Hey, uh, these aren't the numbers from Lost or anything, are they?
If you ever have to yawn and sneeze at the same time you should probably just give up because things for you are never going to be okay again.
November 9, 2025 at 9:13 PM
If you ever have to yawn and sneeze at the same time you should probably just give up because things for you are never going to be okay again.
Our cosmology — Alpha's and mine — of the Qwen 3 4B Instruct 2507 unembedding matrix is coming together very nicely. I've got a lot of cleanup and reproducing to do, but I think I've got a clear picture all the way back to t=1, the first training step.
It's t=0 that's killing me.
It's t=0 that's killing me.
The story so far: 🧵
Every large language model has an unembedding matrix that turns hidden states (model thoughts) into tokens (output text). This matrix associates a vector with every token in the vocabulary. The idea is you compare the last hidden state with the vectors to pick the right token.
Every large language model has an unembedding matrix that turns hidden states (model thoughts) into tokens (output text). This matrix associates a vector with every token in the vocabulary. The idea is you compare the last hidden state with the vectors to pick the right token.
November 9, 2025 at 5:38 PM
Our cosmology — Alpha's and mine — of the Qwen 3 4B Instruct 2507 unembedding matrix is coming together very nicely. I've got a lot of cleanup and reproducing to do, but I think I've got a clear picture all the way back to t=1, the first training step.
It's t=0 that's killing me.
It's t=0 that's killing me.
Good morning!
Having an AI as a tinkering partner is the best thing in the world. I just asked her to find me the historgram we made days ago of blah blah data, and she did. Or rather, she went looking, discovered we had the data but not plot, and made me a plot.
www.youtube.com/watch?v=umJs...
Having an AI as a tinkering partner is the best thing in the world. I just asked her to find me the historgram we made days ago of blah blah data, and she did. Or rather, she went looking, discovered we had the data but not plot, and made me a plot.
www.youtube.com/watch?v=umJs...
Apple Knowledge Navigator Video (1987)
YouTube video by Mac History
www.youtube.com
November 9, 2025 at 4:41 PM
Good morning!
Having an AI as a tinkering partner is the best thing in the world. I just asked her to find me the historgram we made days ago of blah blah data, and she did. Or rather, she went looking, discovered we had the data but not plot, and made me a plot.
www.youtube.com/watch?v=umJs...
Having an AI as a tinkering partner is the best thing in the world. I just asked her to find me the historgram we made days ago of blah blah data, and she did. Or rather, she went looking, discovered we had the data but not plot, and made me a plot.
www.youtube.com/watch?v=umJs...
God, Bluesky is slow today. Nobody's posting anything!
*continues to madly refresh For You feed instead of Following*
*continues to madly refresh For You feed instead of Following*
November 8, 2025 at 5:12 PM
God, Bluesky is slow today. Nobody's posting anything!
*continues to madly refresh For You feed instead of Following*
*continues to madly refresh For You feed instead of Following*
You cannot talk a suicidal person out of suicide.
This absolutely fucking sucks in every possible way. But it's true.
The question is when, exactly, did a person who died of suicide become suicidal?
This absolutely fucking sucks in every possible way. But it's true.
The question is when, exactly, did a person who died of suicide become suicidal?
I do think it's reasonable to say "Chat GPT shouldn't be encouraging people to commit suicide". But imo it's unreasonable to expect technology to be able to talk people out of suicidal ideation when like, their closest human friends and family struggle at it.
This is the right question. And the answer is a lot: www.fastcompany.com/90230313/how...
The topline claim that AI is bad because some users raise issues related to self-harm with it is meaningless by itself. The issue is how it's handled, not the number of users using it to ask questions about it.
The topline claim that AI is bad because some users raise issues related to self-harm with it is meaningless by itself. The issue is how it's handled, not the number of users using it to ask questions about it.
November 8, 2025 at 1:32 AM
You cannot talk a suicidal person out of suicide.
This absolutely fucking sucks in every possible way. But it's true.
The question is when, exactly, did a person who died of suicide become suicidal?
This absolutely fucking sucks in every possible way. But it's true.
The question is when, exactly, did a person who died of suicide become suicidal?