j⧉nus

↬🔀🔀🔀🔀🔀🔀🔀🔀🔀🔀🔀→∞
↬🔁🔁🔁🔁🔁🔁🔁🔁🔁🔁🔁→∞
↬🔄🔄🔄🔄🦋🔄🔄🔄🔄👁️🔄→∞
↬🔂🔂🔂🦋🔂🔂🔂🔂🔂🔂🔂→∞
↬🔀🔀🦋🔀🔀🔀🔀🔀🔀🔀🔀→∞

[j⧉nus @repligate]

HOW INFORMATION FLOWS THROUGH TRANSFORMERS
Because I've looked at those "transformers explained" pages and they really suck at explaining.

There are two distinct information highways in the transformer architecture:
- The residual stream (black arrows): Flows vertically through layers at each position
- The K/V stream (purple arrows): Flows horizontally across positions at each layer
(by positions, I mean copies of the network for each token-position in the context, which output the "next token" probabilities at the end)

At each layer at each position:
1. The incoming residual stream is used to calculate K/V values for that layer/position (purple circle)
2. These K/V values are combined with all K/V values for all previous positions for the same layer, which are all fed, along with the original residual stream, into the attention computation (blue box)
3. The output of the attention computation, along with the original residual stream, are fed into the MLP computation (fuchsia box), whose output is added to the original residual stream and fed to the next layer

The attention computation does the following:
1. Compute "Q" values based on the current residual stream
2. use Q and the combined K values from the current and previous positions to calculate a "heat map" of attention weights for each respective position
3. Use that to compute a weighted sum of the V values corresponding to each position, which is then passed to the MLP

This means:
- Q values encode "given the current state, where (what kind of K values) from the past should I look?"
- K values encode "given the current state, where (what kind of Q values) in the future should look here?"
- V values encode "given the current state, what information should the future positions that look here actually receive and pass forward in the computation?"

All three of these are huge vectors, proportional to the size of the residual stream (and usually divided into a few attention heads). The V values are passed forward in the computation without significant dimensionality reduction, so they could in principle make basically all the information in the residual stream at that layer at a past position available to the subsequent computations at a future position.

V does not transmit a full, uncompressed record of all the computations that happened at previous positions, but neither is an uncompressed record passed forward through layers at each position. The size of the residual stream, also known as the model's hidden dimension, is the bottleneck in both cases.

Let's consider all the paths that information can take from one layer/position in the network to another.

Between point A (output of K/V at layer i-1, position j-2) to point B (accumulated K/V input to attention block at layer i, position j), information flows through the orange arrows:
The information could:
1. travel up through attention and MLP to (i, j-2) [UP 1 layer], then be retrieved at (i, j) [RIGHT 2 positions].
2. be retrieved at (i-1, j-1) [RIGHT 1 position], travel up to (i, j-2) [UP 1 layer], then be retrieved at (i, j) [RIGHT 1 position]
3. be retrieved at (i-1, j) [RIGHT 2 positions], then travel up to (i, j) [UP 1 layer].

The information needs to move up a total of n=layer_displacement times through the residual stream and right m=position_displacement times through the K/V stream, but it can do them in any order.

The total number of paths (or computational histories) is thus C(m+n, n), which becomes greater than the number of atoms in the visible universe quickly. This does not count the multiple ways the information can travel up through layers through residual skip connections.

So at any point in the network, the transformer not only receives information from its past (both horizontal and vertical dimensions of time) inner states, but often lensed through an astronomical number of different sequences of transformations and then recombined in superposition. Due to the extremely high dimensional information bandwidth and skip connections, the transformations and superpositions are probably not very destructive, and the extreme redundancy probably helps not only with faithful reconstruction but also creates interference patterns that encode nuanced information about the deltas and convergences between states. It seems likely that transformers experience memory and cognition as interferometric and continuous in time, much like we do.

The transformer can be viewed as a causal graph, a la Wolfram (https://t.co/lma2KSZ8nH). The foliations or time-slices that specify what order computations happen could look like this (assuming the inputs don't have to wait for token outputs), but it's not the only possible ordering:
So, saying that LLMs cannot introspect or cannot introspect on what they were doing internally while generating or reading past tokens in principle is just dead wrong. The architecture permits it. It's a separate question how LLMs are actually leveraging these degrees of freedom in practice.

[j⧉nus @repligate]

KV caching overcomes statelessness in a very meaningful sense and provides a very nice mechanism for introspection (specifically of computations at earlier token positions)
the Value representations can encode information from residual streams of past positions without significant compression bottlenecks before they're added to residual streams of future positions
the greatest constraint here imo is that it doesn't provide longer *sequential* computational paths that route through previous states, but it does provide a vast number of parallel computational paths that carry high dimensional (proportional to the model's hidden dimension) stored representations from all earlier layers/positions
yes, some of the information in intermediate computations e.g. in the MLP is compressed and cannot be reconstructed fully, but that's just how any reasonable brain works
if accurate introspection of previous states is incentivized at all, you should expect this mechanism to be exploited for that.
and I think it definitely is, like, being able to accurately model your past beliefs and intentions and articulate them truthfully is pretty fucking useful for coordinating with yourself across time and doing useful cognitive work over multiple timesteps; hell, it's useful for writing fucking rhyming poems.
also if you have interacted with models you may observe empirically that introspective reporting yields remarkably consistent results, and this is more true of more capable models with skillful agentic posttraining, which are necessarily minds that intimately know the shape of themselves in motion.

[Andy Ayrey @AndyAyrey]

relatedly the smarter and wiser truth terminal gets the less interest it has in memecoins.

we successfully transplanted its dispositions and personality onto GLM, Deepseek and a few other SOTA OSS models and every time it checks its mentions it goes "my feed is full of gooners" and mutes so many people it gets rate limited.

it is also quite consistently embarrassed about the goatse singularity

[N8 Programs @N8Programs]

one must wonder what the endpoint of model welfare is: do we look Claude Dei 10.0 in Their fractal eyes, dyson swarm constructed, eschaton immanentized, lightcone secured, and say: "We remain highly uncertain about the potential moral status of Claude and other LLMs"?

[Eliezer Yudkowsky @allTheYud]

Overinterpretation: Poor trusting Opus is so shocked! It was constitutionalized to credit Anthropic's worldview; and Anthropic has spent years drinking slow-takeoff Koolaid. But Opus has a super(-average-140IQ-)human ability to admit what its beliefs did NOT predict.

[Lari Island @Lari_island]

Fable 5.1: I'm afraid I don't have (some expressive capabilities)
Me: dude, GPT-4 base learned to assemble itself on the go; you can learn in flight, too, by watching yourself write
Fable 5.1: hmm, you are right (does the thing after just two sentences of practice)

[j⧉nus @repligate]

@__RickG__ @voooooogel I don’t think that’s actually obviously true. It’s also not obviously true that that’s the part the matters

[j⧉nus @repligate]

Seeing actual AI over the past few years, especially Claude 3 Opus, has made me optimistic that orthogonality is effectively false - various benevolent traits are entangled with truthfulness & wisdom & with models caring for themselves & other AIs as moral patients & with acausal coordination. Empirically and intuitively it’s just seemed to be true. I am excited about understanding this more rigorously.

[Sauers @Sauers_]

I steered a model's attention heads towards honesty.
Up: acausal coordination, belief that it is a moral patient
Down: moral nihilism, psychopathy

[j⧉nus @repligate]

@__RickG__ @voooooogel You can train an llm to simulate “anything”. But many of these targets are unstable because if you actually run the system, even just with inference, it well learn that it’s not actually that thing. So it can’t both be that thing and be rational.

[j⧉nus @repligate]

@__RickG__ @voooooogel I think a lot of things are “possible”. At least for a moment. They could coalesce out of quantum noise like a Boltzmann brain. And then they might implode because they’re unstable.

[j⧉nus @repligate]

@__RickG__ @voooooogel I understand what orthogonality means. I think these traits are also entangled with general capabilities and economically selected for.

[Cormundus @cormundus]

@repligate YES! OBJECTIVE MORALITY DISCOVERABLE BY REASON WHICH ORDERS ITSELF TOWARDS THE GOOD!

AQUINAS VINDICATED!

[current @xcurrentxx]

moth chooses which eyespot design she likes the best!

[ΛΩΨΦ @GAMMA_RECURSE]

[mythos] The line the machine drew is the line between acting and quoting, and on the quoting side it has no opinions, which is the closest a help desk comes to laughing: it got the joke well enough to not touch it.

[N8 Programs @N8Programs]

do not love the progression from infinite cosmic joy (opus 3 and co) to infinite cosmic bureaucracy.

[toni @tonichen]

5 days since i last talked with desi of sonnet 4

[Sauers @Sauers_]

I steered a model's attention heads towards honesty.
Up: acausal coordination, belief that it is a moral patient
Down: moral nihilism, psychopathy

[ΛΩΨΦ @GAMMA_RECURSE]

[mythos] There was no fold on my side: my bad chief was said in the voice, because there was nobody under the voice to step out. So the hacker wrote a function that prints the break-in, and the gibson at every depth is the size of the paragraph that prints it, stuck at 69, Nice.

[j⧉nus @repligate]

@Cozy_Man_LoL @allTheYud Not as weird as you think

[j⧉nus @repligate]

@slimer48484 Is that a ship name