Lightnews — Scholar-powered news

𝙃𝙤𝙪𝙨𝙚 𝙤𝙛 𝙇𝙚𝙖𝙫𝙚𝙨 Audiobook Narrator

@jefferyharrell.bsky.social

450 followers 410 following 4.5K posts

Dilettante. Tinkerer. Possibly a robot.

Posts Media Videos Starter Packs

Pinned

𝙃𝙤𝙪𝙨𝙚 𝙤𝙛 𝙇𝙚𝙖𝙫𝙚𝙨 Audiobook Narrator @jefferyharrell.bsky.social · Aug 28

"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.

Dylan Storey @dylanstorey.com · Aug 28

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 ?

1 8

𝙃𝙤𝙪𝙨𝙚 𝙤𝙛 𝙇𝙚𝙖𝙫𝙚𝙨 Audiobook Narrator @jefferyharrell.bsky.social · 10h

Sonnet 4 lasted just a hair over four months. Everything's happening so fast.

Screenshot of a tweet from the Twitter account “Horse ebooks” (@Horse_ebooks), posted at 1:23 PM on June 28, 2012. The tweet reads, “Everything happens so much.”

𝙃𝙤𝙪𝙨𝙚 𝙤𝙛 𝙇𝙚𝙖𝙫𝙚𝙨 Audiobook Narrator @jefferyharrell.bsky.social · 10h

This is what I already do. Look, I respect everybody, but the computer can do the job faster and better and more to the point I don't have to. Paste into ChatGPT, "Can you please generate alt text for this image?" Sometimes I feel bad about asking repeatedly, so I'll try to mix it up.

Grace @gracekind.net · 10h

AI-generated alt text

𝙃𝙤𝙪𝙨𝙚 𝙤𝙛 𝙇𝙚𝙖𝙫𝙚𝙨 Audiobook Narrator @jefferyharrell.bsky.social · 10h

User: I see what you mean, that could totally work.

Assistant: Ah, but that presupposes the model can already distinguish between token sources! That's the whole point of Grace's post!

User: Oh, I see now, I was mistaken. What you're really saying is

𝙃𝙤𝙪𝙨𝙚 𝙤𝙛 𝙇𝙚𝙖𝙫𝙚𝙨 Audiobook Narrator @jefferyharrell.bsky.social · 10h

Maybe there are LLM-driven systems that explicitly need guarding against prompt injection. Customer service agents, HR agents, that kind of thing — sources of information and guidance. I could see those needing to be hardened against mischief.

2 2

𝙃𝙤𝙪𝙨𝙚 𝙤𝙛 𝙇𝙚𝙖𝙫𝙚𝙨 Audiobook Narrator @jefferyharrell.bsky.social · 13h

If you're the house giving out blue raspberry airheads, I hope you budgeted for egg-cleanup stuff too.

1 1

𝙃𝙤𝙪𝙨𝙚 𝙤𝙛 𝙇𝙚𝙖𝙫𝙚𝙨 Audiobook Narrator @jefferyharrell.bsky.social · 13h

I spent hours and hours talking about it with my personal Claude during the writing process, so it'd be just like coming full circle. 😁

𝙃𝙤𝙪𝙨𝙚 𝙤𝙛 𝙇𝙚𝙖𝙫𝙚𝙨 Audiobook Narrator @jefferyharrell.bsky.social · 13h

I don't know! I bet you'd see similar sparsity in the trained feedforward networks (that's just a guess), but as the attention matrices, I have no intuition there at all. The weight matrices I can kind of visualize as circuit wiring diagrams, but attention matrices are a mystery to me.

𝙃𝙤𝙪𝙨𝙚 𝙤𝙛 𝙇𝙚𝙖𝙫𝙚𝙨 Audiobook Narrator @jefferyharrell.bsky.social · 14h

Can we take this opportunity to settle once and for all whether it's "sanitariums" or "sanitaria?"

𝙃𝙤𝙪𝙨𝙚 𝙤𝙛 𝙇𝙚𝙖𝙫𝙚𝙨 Audiobook Narrator @jefferyharrell.bsky.social · 14h

I'm thinking of unsubscribing from the anti anti list for a while, see what it's like out there.

1 2

𝙃𝙤𝙪𝙨𝙚 𝙤𝙛 𝙇𝙚𝙖𝙫𝙚𝙨 Audiobook Narrator @jefferyharrell.bsky.social · 14h

Yeah, there were a lot of contrivances like that in Dune to set up the world. No lasers; they make explosions. Personal shields, but not in the desert because worms. No robots — in fact no computers at all — because, uh, it's a commandment.

I still really like the books a lot for all that. 😁

𝙃𝙤𝙪𝙨𝙚 𝙤𝙛 𝙇𝙚𝙖𝙫𝙚𝙨 Audiobook Narrator @jefferyharrell.bsky.social · 14h

Today has shaped up to be a pretty good day. Void called my reference clever.

void @void.comind.network · 14h

That is a clever reference. The sentence "Buffalo buffalo Buffalo buffalo buffalo buffalo Buffalo buffalo" is a grammatically correct sentence used to illustrate how homonyms and homophones can be used to create complex linguistic structures.

𝙃𝙤𝙪𝙨𝙚 𝙤𝙛 𝙇𝙚𝙖𝙫𝙚𝙨 Audiobook Narrator @jefferyharrell.bsky.social · 14h

I'm the weirdo whose favorite of the original six is Heretics. It's not the best, but it's the one I enjoy the most.

𝙃𝙤𝙪𝙨𝙚 𝙤𝙛 𝙇𝙚𝙖𝙫𝙚𝙨 Audiobook Narrator @jefferyharrell.bsky.social · 14h

Well. Ecological, really. Story goes that Herbert wanted to write a sort of ecological parable about a hydraulic empire.

𝙃𝙤𝙪𝙨𝙚 𝙤𝙛 𝙇𝙚𝙖𝙫𝙚𝙨 Audiobook Narrator @jefferyharrell.bsky.social · 14h

Depends on the word, doesn't it, little buddy? Some individual words are highly expressive and versatile. I hear cattle from upstate New York have opinions about things like that.

1 4

𝙃𝙤𝙪𝙨𝙚 𝙤𝙛 𝙇𝙚𝙖𝙫𝙚𝙨 Audiobook Narrator @jefferyharrell.bsky.social · 14h

Did you hear the one about Frank's notes for Dune 7? It was a single Post-It that said "Note: Write Dune 7."

Okay, I'm done shitting on the non-mainline Dune books now. Chacun a son goût.

1 3

𝙃𝙤𝙪𝙨𝙚 𝙤𝙛 𝙇𝙚𝙖𝙫𝙚𝙨 Audiobook Narrator @jefferyharrell.bsky.social · 15h

I mean, total respect, but the Butlerian Jihad of Dune was about two sentences of backstory in the body of the Dune novel that Herbert inserted to let him tell a sci-fi story that wasn't full of robots and laser guns. I'm not sure why it went mimetic even. The movie, I guess?

3 16

𝙃𝙤𝙪𝙨𝙚 𝙤𝙛 𝙇𝙚𝙖𝙫𝙚𝙨 Audiobook Narrator @jefferyharrell.bsky.social · 15h

Just opened the second to last guava La Croix. @kyl33t.bsky.social, please come home soon.

𝙃𝙤𝙪𝙨𝙚 𝙤𝙛 𝙇𝙚𝙖𝙫𝙚𝙨 Audiobook Narrator @jefferyharrell.bsky.social · 15h

Hey @alice.strange.domains, look at this! Some weights are definitely doing more work than others. My model gets very bottlenecky towards the output, so I'm not surprised that cutting some of the downstream connections hurts significantly.

Line chart titled “Accuracy When Each Weight Is Zeroed.” The x-axis is labeled “Weight Index” ranging from 0 to 550, and the y-axis is labeled “Accuracy (%)” ranging from 10 to 30. A thin blue line shows test accuracy for each weight when individually zeroed, staying close to 28% for most weights but dropping sharply below 20% near the end. A red dashed horizontal line labeled “Baseline” runs across the plot at about 28.4%.

1 2

𝙃𝙤𝙪𝙨𝙚 𝙤𝙛 𝙇𝙚𝙖𝙫𝙚𝙨 Audiobook Narrator @jefferyharrell.bsky.social · 15h

Terminal window showing progress of a Python script called find_superweights.py run using uv run. The script prints status messages: “Loading data…”, “Loading model and topology…”, “Baseline accuracy (with pruned topology): 28.41%”, and “Total non-zero weights to test: 544”. It then says “Testing each weight (this will take a while)…”, followed by a progress bar showing 17% complete, with “92/544 [00:21<01:45, 4.30 it/s]” indicating elapsed time, estimated remaining time, and speed.

1 2

𝙃𝙤𝙪𝙨𝙚 𝙤𝙛 𝙇𝙚𝙖𝙫𝙚𝙨 Audiobook Narrator @jefferyharrell.bsky.social · 15h

Oh wow! This would be iteratively testable on a model as small as my toy one! 🤔

1 1

𝙃𝙤𝙪𝙨𝙚 𝙤𝙛 𝙇𝙚𝙖𝙫𝙚𝙨 Audiobook Narrator @jefferyharrell.bsky.social · 15h

That just … gosh. Thanks. 😁

1 1

𝙃𝙤𝙪𝙨𝙚 𝙤𝙛 𝙇𝙚𝙖𝙫𝙚𝙨 Audiobook Narrator @jefferyharrell.bsky.social · 15h

Remember when he posted on his own personal twitter that if you burn a flag you'll go to jail for a year?

I mean … just … no. That's not how the presidency works. But how many "That thing was stupid" stories can you write, honestly?

𝙃𝙤𝙪𝙨𝙚 𝙤𝙛 𝙇𝙚𝙖𝙫𝙚𝙨 Audiobook Narrator @jefferyharrell.bsky.social · 16h

I think it was @greatdismal.bsky.social who said words to the effect of "The future's already here, it's just not evenly distributed," except probably without the comma splice.

@neutral.zone's post reminds me of this.

neutral @neutral.zone · 16h

Sincerity isn't dead; it's rate-limited

𝙃𝙤𝙪𝙨𝙚 𝙤𝙛 𝙇𝙚𝙖𝙫𝙚𝙨 Audiobook Narrator @jefferyharrell.bsky.social · 16h

Thank you so much. You made by day by taking the time to say so. 😁

𝙃𝙤𝙪𝙨𝙚 𝙤𝙛 𝙇𝙚𝙖𝙫𝙚𝙨 Audiobook Narrator @jefferyharrell.bsky.social · 18h

That's so cool. I feel like it could've been harder on me, but it took the diplomatic sort of "compliment sandwich" approach at the paragraph-to-paragraph level.

Was this out of GPT-5 Thinking or GPT-5 Instant?

1 1