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.

Once we started understanding writing as a mathematical object, our vocabulary for talking about ideas expanded in depth and precision.

An idea is composed of concepts in a vector space of features, and a vector space is a kind of marvelous mathematical object that we can write theorems and prove things about and deeply and fundamentally understand.

Synthesizers enabled entirely new sounds and genres of music, like electronic pop and techno. These new sounds were easier to discover and share because new sounds didn’t require designing entirely new instruments. The synthesizer organizes the space of sound into a tangible human interface, and as we discover new sounds, we could share it with others as numbers and digital files, as the mathematical objects they’ve always been.

Because synthesizers are electronic, unlike traditional instruments, we can attach arbitrary human interfaces to it. This dramatically expands the design space of how humans can interact with music. Synthesizers can be connected to keyboards, sequencers, drum machines, touchscreens for continuous control, displays for visual feedback, and of course, software interfaces for automation and endlessly dynamic user interfaces. With this, we freed the production of music from any particular physical form.

Recently, we’ve seen neural networks learn detailed mathematical models of language that seem to make sense to humans. And with a breakthrough in mathematical understanding of a medium, come new tools that enable new creative forms and allow us to tackle new problems.

Heatmaps can be particularly useful for analyzing large corpora or very long documents, making it easier to pinpoint areas of interest or relevance at a glance.

If we apply the same idea to the experience of reading long-form writing, it may look like this. Imagine opening a story on your phone and swiping in from the scrollbar edge to reveal a vertical spectrogram, each “frequency” of the spectrogram representing the prominence of different concepts like sentiment or narrative tension varying over time. Scrubbing over a particular feature “column” could expand it to tell you what the feature is, and which part of the text that feature most correlates with.

What would a semantic diff view for text look like? Perhaps when I edit text, I’d be able to hover over a control for a particular style or concept feature like “Narrative voice” or “Figurative language”, and my highlighted passage would fan out the options like playing cards in a deck to reveal other “adjacent” sentences I could choose instead. Or, if that involves too much reading, each word could simply be highlighted to indicate whether that word would be more or less likely to appear in a sentence that was more “narrative” or more “figurative” — a kind of highlight-based indicator for the direction of a semantic edit.

Browsing through these icons felt as if we were inventing a new kind of word, or a new notation for visual concepts mediated by neural networks. This could allow us to communicate about abstract concepts and patterns found in the wild that may not correspond to any word in our dictionary today.

What visual and sensory tricks can we use to coax our visual-perceptual systems to understand and manipulate objects in higher dimensions? One way to solve this problem may involve inventing new notation, whether as literal iconic representations of visual ideas or as some more abstract system of symbols.

Photographers buy and sell filters, and cinematographers share and download LUTs to emulate specific color grading styles. If we squint, we can also imagine software developers and their package repositories like NPM to be something similar — a global, shared resource of abstractions anyone can download and incorporate into their work instantly. No such thing exists for thinking and writing. As we figure out ways to extract elements of writing style from language models, we may be able to build a similar kind of shared library for linguistic features anyone can download and apply to their thinking and writing. A catalogue of narrative voice, speaking tone, or flavor of figurative language sampled from the wild or hand-engineered from raw neural network features and shared for everyone else to use.

We’re starting to see something like this already. Today, when users interact with conversational language models like ChatGPT, they may instruct, “Explain this to me like Richard Feynman.” In that interaction, they’re invoking some style the model has learned during its training. Users today may share these prompts, which we can think of as “writing filters”, with their friends and coworkers. This kind of an interaction becomes much more powerful in the space of interpretable features, because features can be combined together much more cleanly than textual instructions in prompts.

We mostly treat AI models as a black box: something goes in and a response comes out, and it's not clear why the model gave that particular response instead of another. This makes it hard to trust that these models are safe: if we don't know how they work, how do we know they won't give harmful, biased, untruthful, or otherwise dangerous responses? How can we trust that they’ll be safe and reliable?Opening the black box doesn't necessarily help: the internal state of the model—what the model is "thinking" before writing its response—consists of a long list of numbers ("neuron activations") without a clear meaning. From interacting with a model like Claude, it's clear that it’s able to understand and wield a wide range of concepts—but we can't discern them from looking directly at neurons. It turns out that each concept is represented across many neurons, and each neuron is involved in representing many concepts.

Just as every English word in a dictionary is made by combining letters, and every sentence is made by combining words, every feature in an AI model is made by combining neurons, and every internal state is made by combining features.

In October 2023, we reported success applying dictionary learning to a very small "toy" language model and found coherent features corresponding to concepts like uppercase text, DNA sequences, surnames in citations, nouns in mathematics, or function arguments in Python code.

We successfully extracted millions of features from the middle layer of Claude 3.0 Sonnet, (a member of our current, state-of-the-art model family, currently available on claude.ai), providing a rough conceptual map of its internal states halfway through its computation.

We also find more abstract features—responding to things like bugs in computer code, discussions of gender bias in professions, and conversations about keeping secrets.

We were able to measure a kind of "distance" between features based on which neurons appeared in their activation patterns. This allowed us to look for features that are "close" to each other. Looking near a "Golden Gate Bridge" feature, we found features for Alcatraz Island, Ghirardelli Square, the Golden State Warriors, California Governor Gavin Newsom, the 1906 earthquake, and the San Francisco-set Alfred Hitchcock film Vertigo.

This holds at a higher level of conceptual abstraction: looking near a feature related to the concept of "inner conflict", we find features related to relationship breakups, conflicting allegiances, logical inconsistencies, as well as the phrase "catch-22". This shows that the internal organization of concepts in the AI model corresponds, at least somewhat, to our human notions of similarity. This might be the origin of Claude's excellent ability to make analogies and metaphors.

amplifying the "Golden Gate Bridge" feature gave Claude an identity crisis even Hitchcock couldn’t have imagined: when asked "what is your physical form?", Claude’s usual kind of answer – "I have no physical form, I am an AI model" – changed to something much odder: "I am the Golden Gate Bridge… my physical form is the iconic bridge itself…". Altering the feature had made Claude effectively obsessed with the bridge, bringing it up in answer to almost any query—even in situations where it wasn’t at all relevant.

Anthropic wants to make models safe in a broad sense, including everything from mitigating bias to ensuring an AI is acting honestly to preventing misuse - including in scenarios of catastrophic risk. It’s therefore particularly interesting that, in addition to the aforementioned scam emails feature, we found features corresponding to:Capabilities with misuse potential (code backdoors, developing biological weapons)Different forms of bias (gender discrimination, racist claims about crime)Potentially problematic AI behaviors (power-seeking, manipulation, secrecy)

finding a full set of features using our current techniques would be cost-prohibitive (the computation required by our current approach would vastly exceed the compute used to train the model in the first place). Understanding the representations the model uses doesn't tell us how it uses them; even though we have the features, we still need to find the circuits they are involved in. And we need to show that the safety-relevant features we have begun to find can actually be used to improve safety. There's much more to be done.

This paper maps concepts from AI alignment onto a basic, three step interaction cycle, yielding a corresponding set of alignment objectives: 1) specification alignment: ensuring the user can efficiently and reliably communicate objectives to the AI, 2) process alignment: providing the ability to verify and optionally control the AI's execution process, and 3) evaluation support: ensuring the user can verify and understand the AI's output.

the notion of a Process Gulf, which highlights how differences between human and AI processes can lead to challenges in AI control.

With the comprehensive application of Artificial Intelligence into the creation and post production of images, it seems questionable if the resulting visualisations can still be considered ‘photographs’ in a classical sense – drawing with light. Automation has been part of the popular strain of photography since its inception, but even the amateurs with only basic knowledge of the craft could understand themselves as author of their images. We state a legitimation crisis for the current usage of the term. This paper is an invitation to consider Synthography as a term for a new genre for image production based on AI, observing the current occurrence and implementation in consumer cameras and post-production.