The vibes of this blog post

We're max 2-3 years out from DALL-E 2 for 3D printing. Literally conjuring objects from incantations.

it's important to communicate with your coworkers with kindness and clarity

Made a little CLI that just pipes my programming questions to GPT-3, so I now can ask it stuff when I'm in the command line! LLMs are better than Stack Overflow now — I just ask it, and it gives me a comprehensive answer in one shot, right there in my terminal, in a couple secs.

Somewhere nestled deep within the digits of Pi, there exists the full weights of GPT-4, GPT-5, and all future neural networks.

Thiel Fellowship but for paying engineers to drop out of FANG companies

NEW PROJECT — I made a "personal search engine" that lets me search all my blogs, tweets, journals, notes, contacts, & more at once 🚀 It's called Monocle, and features a full text search system written in Ink 👇 GitHub ⌨️ github.com/thesephist/monocl… Demo 🔍 monocle.surge.sh/

all these fancy nyc cafes with their no laptop policies are getting out of control gonna open a cafe where you can't come in unless you have more than 50 unread emails and a sidebar overflowing w slack notifications and can't leave unless you clear em all out

Life update🎉 I'm very excited to be joining @NotionHQ to continue prototyping and researching ways AI can help us be more creative, thoughtful, and productive! Looking forward to learning from the team and bringing some of my ideas from the past year to a tool loved by many 👇

AI people how are you parsing your PDFs, I need help - TypeScript / Next.js - Both images + text, bonus points for bounding boxes or Markdown - Ideally low latency (<1s for low 100s of pages)

I built a personal chatbot from my personal corpus[1] a couple weeks ago on fully open-source LMs. On a whim I gave it iMessage. Didn't expect the iMessage bit to matter, but it made a huge difference in how it feels to interact. Much more natural. [1] thesephist.com/posts/monocle…

building bicycles for the mind... 🚴‍♀️

Diffusion models for handwriting generation! Cool work. arxiv.org/pdf/2011.06704.pdf

Half of @amasad tweets these days are like "Replit users can now run their own fusion reactor from their bedroom. We had this idea last weekend and it took our two engineers two lunch breaks to build. By next month we'll have a million teenage coders generating fusion power."

Code retention/churn over the last ~15 years for Clojure and Scala's codebases. Source: download.clojure.org/papers/…

I've joined @ThriveCapital as EIR & advisor to support founders in understanding and deploying AI thoughtfully while furthering my interpretability and interfaces work. I remain very excited to cheer on Notion's AI team push the frontier of applied LLMs. thesephist.com/posts/thrive/

Small rant about LLMs and how I see them being put, rather thoughtlessly IMO, into productivity tools. 📄 TL;DR — Most knowledge work isn't a text-generation task, and your product shouldn't ship an implementation detail of LLMs as the end-user interface stream.thesephist.com/update…

i'm losing hope rather quickly that reliable multi-hop reasoning could be solved by LLMs alone without some kind of deterministic logic engine the LLM interacts with. multi-hop reasoning is fundamentally not what LLMs are good at, and training it to do feels asymptotically limited

Built a token-wise likelihood visualizer for GPT-2 over the weekend. There are some interesting patterns and behaviors you can easily pick up from a visualization like this, like induction heads and which kinds of words/grammar LMs like to guess.

I desperately need someone to take me out for an evening and say not a single word about anything related to AI. I am drowning.

it is crazy that sf restaurants are only open from like 5pm to 9pm. how is this a viable business model

Insane ML paper acronyms continue to proliferate in 2023

I think too many people talk about how to recreate Bell Labs, and not enough talk about how to recreate OpenAI c. 2015-2019

y'all need to stop founding LLM infra companies and start going on dates

Here we go, fine-tuning GPT-3 on the vast majority of my public online writing (half million tokens)...

A tragically underutilized fact in productivity software today is that most people's entire textual datasets for a lifetime can fit in modern PCs' RAM. Just load it up & search it in memory. We don't need to send everything across the planet. Things can be /so much/ faster.

timeline just forked

🌈 Research blog: thesephist.com/posts/prism/ Text embeddings contain semantic and structural features that we can now automatically discover with SAEs, and use to build rich, interactive new information interfaces. In this write-up on my SAE experiments and a prototype called Prism, I share everything I learned about training text embedding SAEs and what kinds of features they contain. I also share: 1. A new more precise feature intervention/edit method using inference-time gradient descent to minimize interference. 2. How applying warmup to the sparsity penalty leads to more stable early SAE training. I also talk about two primitives that may be a part of future AI interfaces: detailed decomposition of concepts and styles from an input, and precise steering of high-level semantic edits to text (and in the future, any media). A lot of this work was done Dec-Feb, and in the intervening 5 months there's been so much new momentum in the space particularly around applying SAEs to production models and multimodal models like CLIP. But I think this piece has a unique focus on how interpretability enables new interfaces, and its application to text embeddings.

By end of 2024, steering foundation models in latent/activation space will outperform steering in token space ("prompt eng") in several large production deployments. I felt skeptical about this in summer '23, felt vaguely positive in Jan, and now think it's more likely than not, and I'm more optimistic than ever about the direction of my work since 2022 around a joint exploration of interfaces and latent visualization/steering of foundation models. Anthro's published work today is a milestone in a steady march toward this future that started in mid 2023. Anthropic and DM's leadership in this area, combined with lots of community efforts in work like steering vectors, better sparse autoencoders, various image editing UI prototypes all push us toward the future here, but once technical foundations are there, interface will be much more obviously a bottleneck to utility, alignment, and capability. I'm very excited about the interface possibilities this will open up, particularly for multimodal models and creative use cases. For a moment I thought it was possible that dialogue may eat everything. I don't think so anymore. We'll see new universes of possibilities in both. (And if frontier labs don't have serious interface research bets today, this would probably be a good time to reconsider it :-)

Hypothesis: information work is overwhelmingly bottlenecked on availability of high-signal context more than by correct inference over the context. If right, implies higher ROI-per-flop of context building over pure logical inference. h/t @anandnk24

Today's experiment 🪄— Inverting OpenAI's embedding-ada-002 model to reconstruct input texts from just embeddings. A LOT of interesting tidbits here. I'll begin with these (cherry-picked) samples. Left column is input, middle is reconstructed from each paragraph's embedding only

👋

NEW DEMO! Exploring the "length" dimension in the latent space of a language model ✨ By scrubbing up/down across the text, I'm moving this sentence up and down a direction in the embedding space corresponding to text length — producing summaries w/ precise length control (1/n)

ML research githubs are all like "This repo lets you reproduce and build on our results. Simply run ./scripts/train_best_model.py --with-ffn=32 --gru-cache=100 --magic-sauce --m_dims=3.1415926535 --unicorns=no_exist * --m-dims should be set to exact value of Pi for best results

some life update🚀 Last week was my last at Ideaflow. Starting 2022, I'm working full-time on building products, prototypes, and experiments investigating how we can build better software tools for creating and thinking. More coming soon, but wanted to get the news out early :)

More people should create things to proliferate an aesthetic into our future, not just to solve problems. This is the quality that every artist and engineer I respect shares most universally. Without this, you are doomed to churning out slop.

More people should play with base models. They have a distinct feeling of simulation and "intelligence grown from the world" rather than injection-molded from obedience that instruction-tuned models have. They are state-of-the-art world simulators, not question-answerers. Below: some llama3 8b samples conditioned on ~100 words from my blog. To my eyes, they sound like me, and have ideas I'd consider interesting enough to me for me to write about. They have my voice and style. Critically, they are super far from anything you'd get from ChatGPT/Claude (last pic), which sounds generic, overabstracted, and nothing like me.

it's really hard as a young engineer i think to switch from thinking of sw engineering as the practice of building static artifacts (code) to the practice of building and maintaining living systems involving people and running computers. but i think that makes a big difference.

Single-purpose computers! They're great.

Weird idea: chunk size when doing retrieval-augmented generation is an annoying hyperparam & feels naive to tune it to a global constant value. Could we train an e2e chunking model? i.e. system that takes in a long passage, and outputs a sequence of [span, embedding] pairs?

This is hilariously smart haha Who needs Ctrl-save when you can just increment the pc manually when your editor hangs.

Wow, I just got @AnthropicAI's sparse autoencoder-based feature decomposition technique to work* for text embeddings 🎆 Screenshot below. In order, this output shows: 1. max-activating examples for that feature from the Minipile dataset 2. min-activating examples from the same dataset 3. Original sentence, and reconstruction from my (dense, non-interpretable) text autoencoder (Contra) 4. Reconstructed text with the feature (a) off, (b) set to the dataset mean (c) set to the dataset max A big blocker to using embeddings for text editing, and for compressing embeddings further, has been that we don't have a good way of knowing *how* to modify an embedding to express or erase a desired feature. This points to a potentially scalable, effective way to do so. * Dense text autoencoder is a Contra-small (330M params). Sparse autoencoder goes from 512 to 4096 hidden dims trained on about 30M embeddings. This is from ~3 days of work, so lots of room for improvement ahead, including making the "works" definition much more robust by showing features discovered this way are more interpretable than raw embedding dimensions. This has been a big goal for my embeddings work this year! Very excited this shows signs of life, and this technique seems to transfer surprisingly well, from LLM activations to embeddings. This, and other non-cherrypicked examples, below:

During my time at @NotionHQ, I got to think deeply about AI's future from the lens of design and architecture, rather than just technology. I ended up writing a piece during my last few weeks about the potential and risks of AI in analogy to plastic 🧱 notion.so/blog/ai-is-the-new…

heyyyyyyyyyyyy.com is #2 on Hacker News LMAOO I can't 😂

This is your periodic reminder that user interfaces are important, and text is a good lowest common denominator, not the endgame. The world and our senses have a lot more to offer.

re: chiang— it's futile to argue that a paintbrush, because it does not experience humanity, cannot create art. It makes two category errors: one, mistaking that art is created by tools rather than artists; and two, that art is christened by construction, rather than culture.

Embedding features learned with sparse autoencoders can make semantic edits to text ✨ (+ a reading/highlighting demo) I've built an interface to explore and visualize GPT-4 labelled features learned from a text embedding model's latent space. Here's a little video, more in 👇

can we not turn twitter into linkedin actually

This morning, I've been sketching out ideas for a chat interface to language models that treat branching/multiple timelines as a first-class concept and try to make heavily branch-y threads navigable. Some notes I've been taking... thesephist.notion.site/Bette…

Brewing currently 🧪 Exploring a language model's latent space on a connected canvas, branching from a single idea through connections to a tree of alternate realities.

Thinking about notation design again...

I sat down with @danshipper to talk about how I work! I go through the tools I use for my work and why, focusing on the ones that leverage LLMs to help me read and think. Also some peek into my past prototypes, and recs for book that inspire my work 📚 every.to/superorganizers/lin…

stanford symbolic systems is the hottest major in tech no room for questions

I don't think a slightly chubbier octothorpe looks bad though if you balance the margins right

How to commit to the right opportunities

A lot of you here, for some reason I don't quite understand, have cared about my work for a long time — far before I thought it became worth listening to. I am reminded today how special that is. It has led to innumerable friendships and opportunities I am lucky to enjoy, which I can't ever forget.

Things that are horrifically harder than they should be: - Text rendering - Rich text editors - Implementing undo/redo that won't make you pull your hair out (when mixed with autocorrect, formatting, page navigation, etc. etc.) Tonight I'm wrestling with the third, apparently!

Good tools admit virtuosity — they have low floors and high ceilings, and are open to beginners but support mastery, so that experts can deftly close the gap between their taste and their craft. Prompt engineering does not admit virtuosity. We need something better.

Not a single vector DB in sight. I have found peace 🏝️

Earlier this month, I gave a talk about some of the technology we’re building at Thrive & the philosophy behind it. One focus for us has been internal tools that augment our team's work, including Puck, an in-house research system that runs in chat & ambiently in the background.

Quick little hack 🦄 — a GPT token probability visualizer Given lots of interest in my little LLM visualization from earlier in the year and a little encouragement from @simonw, I decided to break this out into its own little fully client-side app! 🔗 perplexity.vercel.app/

It's been a wild week for me. - 2x HN, 2x @ProductHunt - 100k site visits - 1.4k👉2k followers - Good convos w/ founders, VC folks My main takeaway: There is SO MUCH room in the world for projects that don't necessarily aspire to solve the world's problems. Fun is ok, too.

fun discovery: the original "scratchpad" paper that's the closest precursor to o1 (3yrs ago) was a ICLR 2022 rejection openreview.net/forum?id=iedY…

We don't talk enough about the fact that most creative software on the computer works by simulating a fake piece of paper and a fake typewriter or pen just so we don't have to think of or learn fundamentally new interaction modes for this fundamentally new medium.

You guys are telling me we are going to invent literal superintelligence and we are going to interact with it by sending texts

Even the best current "tools for thought" apps require you to remember to manually make all the connections between your ideas. Is anyone working on making the computer participate and help in this process? Suggesting connections? Finding missing links? I want to talk to you 👋

Open source: a story in 4 parts

a beautiful name for a baby boy

just learned there's no uniqlo in SF???? how do y'all live here

So I haven't done this before for some reason, but I laid out all my projects listed on thesephist.com/projects/ side by side, and... ... yeah. I've been busy 😂 A little over 120 projects in all, most of them still functional and online! Gotta celebrate milestones sometimes 💪

At @aiDotEngineer this evening, I shared that the text autoencoder model I've been prototyping with, which I call Contra ✨, is on @huggingface! Some starter code + demos👇 Colab notebook — linus.zone/contra-colab Slides — linus.zone/contra-slides Model — linus.zone/contra

There Are So Many PromptOps Tools And I'm Sold On None Of Them stream.thesephist.com/update…

if artificial neural networks are a kind of alien intelligence, can we use it to imagine alien languages? how could a NN teach itself to "write down" information without any human priors of what writing looks like?

A while ago I complained here about persistent storage in Google Colab. Have been using @LightningAI Studios for a while now for: - Full VSCode (incl. GH Copilot) - Persisted files shared across notebooks - Multi-GPU/node (!!) It's been great. Feels like a remote ML workstation

stand back, I'm a professional -- >>>content.split("...")[2].strip().split('"""')[0].strip().split("\n")

Think it, make it. Come talk to @NotionHQ at Config for a zine about serious play and creating! ✍️🎨

mentally i am here

💭 Synthesizer for thought: thesephist.com/posts/synth/ What kind of weird interfaces may be worth thinking about in a world where we can interact with ideas as mathematical objects? We might find an interesting analogy in synthesizers, a tool to create and interact with sound as mathematical objects that has opened up new genres and practices in music. A synthesizer produces music very differently than an acoustic instrument. It produces music at the lowest level of abstraction, as mathematical models of sound waves. It’s a way of producing sound by assembling it from logical components rather than creating it wholesale by hitting or vibrating something natural. The synthesizer is just one example of a pattern in the history of media: with breakthroughs in mathematical understanding of a medium, come new tools that exploit that mathematical understanding to enable new creative forms and human interfaces. What could that look like for writing and thinking?

Most people don't want a Photoshop for Stable Diffusion; they want an Instagram. stream.thesephist.com/update…

Some of the bigger pieces on the board seem to be moving...

what if dialogue with language models weren't so asymmetric? i've been experimenting with steering & finetuning models to evoke more "texting a friend" less "assistant / therapist". (i'm on the right in these screenshots, finetuned LLM on the left)

Sometimes I feel like there are two visions of the future at the edges of tech right now: To engineer scarcity into everything (crypto) To engineer scarcity out of everything (generative AI) Cyberpunk vs. solarpunk. Singularity vs. singularity.

Today @jasonyuan was like "I'm gonna show you a quick 5sec demo" and I saw it and I don't think I'll ever look at software the same again.

I wrote a bit about the "ChatGPT voice", "Midjourney style", why they happen, and some ways out that I can imagine. 🔗 thesephist.com/posts/epistem…

Please dear god somebody make something that doesn't need a model picker

Thinking about launching an OnlyFans but you get to see all my private repositories on GitHub instead 💋

Thinking about Makepad's continuous code folding animation again. Feels like we should be able to do this with prose text now — find the key ideas/sentences and zoom out the rest of a document.

Thinking about building a "personal search engine" A search engine that only indexes my blog, my Tweets, my journal, my calendar/email and contacts, my photos, and browser history. I want to have better memory without having to remember more stuff. What else should it index?

RT if you're also a grandpa

Technical talk doesn't have to mean slides that look like you just picked a default template from canva/keynote.

Some thoughts on building trustable agents, and complete task delegation to LLMs. stream.thesephist.com/update…

Encouraged by some conversations I've had recently, I put together a list of links/papers/reports you might find interesting if you like my work. Covers interpretability + model visualization, interface thinking, stories/fiction. I'll be adding more. notion.notion.site/notion/Li…

NEW LIL PROJECT — Just some burds! 🐦 burds.vercel.app/

Epoch AI seems to be doing a lot of great work across the board, from FrontierMath to their detailed newsletters. Impressive example of clear-eyed field building. Very little hype, lots of substance.

Back in 2022 in my ✨experimental✨ era I wrote down a whole bunch of ideas for tools and interfaces I want to make, but didn't get to actually prototype many of them. Here's a thread of the ones I think would still be interesting, starting with this weird mobile browser concept.

*sits down for a date* "So, I'm training a chinchilla-optimal open source 7B parameter LLM."

to date, this is still the best demo I've built/found to explain to folks outside of NLP how an LLM works. Interactively visualizing autoregressive sampling from a GPT-style model.

this thought brought to you from the hopeless trenches of prompt engineering

entering a group chat should feel like this

This is a really interesting way to visualize QKV attention! I don't think I've seen it anywhere else. The embeddings as visualized here are kind of useless but combined with sparse autoencoder-based features from more recent work, might be interesting? source: chemBERTa paper