Top all-time · #2137
Theia Vogel
Personal blog · vgel.me · on HN as vgelaiprogrammingscience
Theia Vogel's personal blog at vgel.me covers ai, programming and science. Since 2015, 37 different Hacker News readers have linked it 62 times, most often “Representation Engineering: Mistral-7B on Acid”. It is still writing: the last post was on 2026-08-25, and you can follow it by feed.
Blog about linguistics, programming, and my projects
Most-linked posts?
- Representation Engineering: Mistral-7B on Acid13 people · 15 mentions · first linked 2024-01-22
- I made a transformer by hand (2023)6 people · 6 mentions · first linked 2023-09-11
- GPT-3 will ignore tools when it disagrees with them4 people · 8 mentions · first linked 2023-03-06
- Small Models Can Introspect, Too (2025)3 people · 3 mentions · first linked 2025-12-21
- How to make LLMs go fast2 people · 2 mentions · first linked 2023-12-19
- Why do LLMs freak out over the seahorse emoji?2 people · 2 mentions · first linked 2025-10-06
- Getting a full PDF from a DRM-encumbered online textbook2 people · 2 mentions · first linked 2015-10-21
- Elven Rope, Ultra-High Molecular Weight Polyethylene, and LLMs2 people · 2 mentions · first linked 2025-11-26
- Writing a C compiler in 500 lines of Python (2023)2 people · 2 mentions · first linked 2023-09-04
What Hacker News says?
Working on a game called Themengi ( https://vgel.me/themengi/ ), where you learn an alien language to navigate a world via text commands and dialog. Currently taking a break from the game itself to rewrite the linguistic parser for the game in Rust, which you can follow in my blog series: https://vgel.me/posts/symbolic-linguistics-part1
Comment linking /posts/symbolic-linguistics-part1
on “Ask HN: What's your quarantine side project?”Submitted Why do LLMs freak out over the seahorse emoji? as a story · 734 points
HN discussionThis is one of the most interesting articles I've read ... maybe this year. The author also has another one that is super cool: https://vgel.me/posts/representation-engineering/ I wish I could take the time to learn more about this because it looks like a way to make LLMs more useful beyond additional prompting techniques.
Comment linking /posts/representation-engineering/
on “Why do LLMs freak out over the seahorse emoji?”Submitted Representation Engineering: Mistral-7B on Acid as a story · 360 points
HN discussionYou don't even need SAE for residual stream based steering, you can do so much more naively by just taking the average difference in activations of prompted vs. unprompted versions, then PCA to get a "control vector" that you can use to elicit the prompted behavior https://vgel.me/posts/representation-engineering/
Comment linking /posts/representation-engineering/
on “Qwenputer: A functioning, deterministic RISC-V CPU inside Qwen3.5”Submitted Writing a C compiler in 500 lines of Python (2023) as a story · 247 points
HN discussionReminds me of https://vgel.me/posts/representation-engineering/ . There they were adding a control vector, w' = cvec + w, here they are "ablating" it, w' = w - dot(w,cvec)*cvec. There is an interesting field of learning how to "brain chip" LLMs into doing what you want.
Comment linking /posts/representation-engineering/
on “Uncensor any LLM with abliteration”Submitted How to make LLMs go fast as a story · 215 points
HN discussionRelated: https://vgel.me/posts/representation-engineering/ https://github.com/vgel/repeng
Comment linking /posts/representation-engineering/
on “Persona vectors: Monitoring and controlling character traits in language models”Submitted Getting a full PDF from a DRM-encumbered online textbook as a story · 109 points
HN discussionGames typically need to redraw the entire screen anyways. Imagine a sidescroller where even a 1px movement requires the entire background and every object on top to be redrawn, or in 3D even a tiny camera move shifts the position of every triangle on the screen. Even very simple games without a lot of movement often have animations, where you then have to worry about transparency so you may have to redraw the region…
Comment linking /hoverator
on “Stage is a minimalistic 2D, cross-platform HTML5 game engine”Submitted Representation Engineering (2024) as a story · 39 points
HN discussionLLMs have emotions ;) https://vgel.me/posts/representation-engineering/
Comment linking /posts/representation-engineering/
on “How Chain-of-Thought Reasoning Helps Neural Networks Compute”Submitted Small Models Can Introspect, Too (2025) as a story · 5 points
HN discussionThanks for the blog post, I also enjoyed their previous post on making a transformer by hand: https://vgel.me/posts/handmade-transformer/
Comment linking /posts/handmade-transformer/
on “How to make LLMs go fast”Submitted Writing a C compiler in 500 lines of Python as a story · 5 points
HN discussionThank you for making it free! I've used it on my blog ( https://vgel.me ) forever, since back before I ever would've paid for analytics, and it's been rock-solid the whole time. Some of the most "set it and forget it" software I use. I definitely owe you a donation so I'm glad I saw this!
Comment linking /
on “Creator is hoping to raise at least €1k for basic living expense”Submitted I made a transformer by hand (2023) as a story · 4 points
HN discussionGreat article! Agree with everything mentioned. Good to share with all tech people who don't work in the AI space and "hallucinate" about working of LLMs. Having said that, although LLMs don't directly learn from what we say to them, they can for sure create a representation or persona using Representation Engineering. Quoting from an excellent article by Theia Vogel [1]: "In October 2023, a group of authors from th…
Comment linking /posts/representation-engineering/
on “Training is not the same as chatting: LLMs don’t remember everything you say”Submitted Representation Engineering Mistral-7B an Acid Trip as a story · 3 points
HN discussionNo, their revised understanding is more accurate. The model has internal representations of concepts; the seahorse emoji fails because it uses those representations and stumbles: https://vgel.me/posts/seahorse/
Comment linking /posts/seahorse/
on “What is intelligence? (2024)”Submitted Representation Engineering Mistral-7B an Acid Trip as a story · 3 points
HN discussionRelated: https://vgel.me/posts/representation-engineering/
Comment linking /posts/representation-engineering/
on “Scaling Monosemanticity: Extracting Interpretable Features from Claude 3 Sonnet”Submitted I made a transformer by hand (no training) as a story · 3 points
HN discussionIf you're interested, I wrote a blog post a few years ago about how to implement a syscall (getcwd) on Xv6: https://vgel.me/posts/pwd_command_xv6/
Comment linking /posts/pwd_command_xv6/
on “Xv6, a simple Unix-like teaching operating system”Submitted Small Models Can Introspect, Too (2025) as a story · 2 points
HN discussionI wonder if we will eventually get a chat agent that just has a slider for "wokeness". Possibly using something like this "Representation Engineering": https://vgel.me/posts/representation-engineering/ Just as a thought experiment, project that to the hypothetical existential risk scenario. Maybe there is a slider for aggression (hopefully not) or literalness of interpretation of instructions.
Comment linking /posts/representation-engineering/
on “Yishan Wong: "Google's Gemini issue is not about woke/DEI"”Submitted Gyre as a story · 2 points
HN discussionThis one isn't new at least. https://vgel.me/posts/tools-not-needed/
Comment linking /posts/tools-not-needed/
on “OpenAI Function calling and API updates”Submitted Small Models Can Introspect, Too as a story · 2 points
HN discussion
