Not really - its vector matrix multiplication, not matrix matrix mult.

This is the answer, but

Essentially, the brain is massively underclocked because of design-space restrictions imposed by biology and evolution

The main restriction is power efficiency: the brain provides a great deal of intelligence for a budget of only ~20 watts. Spreading out that power budget over a very wide memory operating at very slow speed just turns out to be the most power efficient design (vs a very small memory running at very high speed), because memory > time.

Would have made much more sense (visually and otherwise) to show graphs in log space. Example: https://www.openphilanthropy.org/research/modeling-the-human-trajectory/

The effectiveness of weight sharing (and parameter compression in general) diminishes as you move the domain from physics (simple rules/patterns tiled over all of space/time) up to language/knowledge (downstream facts/knowledge that are far too costly to rederive from simulation).

BNNs cant really take advantage of weight sharing so much, so ANNs that are closer to physics should be much smaller parameter wise, for the same compute and capability. Which is what we observer for lower level sensor/motor modalities.

The single factor prime causative factor driving the explosive growth in AI demand/revenue is and always has been the exponential reduction in $/flop via moore's law, which simply is jevon's paradox manifested. With more compute everything is increasingly easy and obvious; even idiots can create AGI with enough compute.

Abilities/intelligence come almost entirely from pretraining, so all the situation awareness and scheming capability that current (and future similar) frontier models possess is thus also mostly present in the base model.

Yes, but for scheming, we care about whether the AI can self-locate itself as an AI using its knowledge. The fact that (at a minimum) sampling from the system is required for it to self-locate as an AI might make a big difference here.

So if your 'yes' above is agreeing that capabilities - including scheming - come mostly from pretraining, then I don't see how relevant it is whether or not that ability is actually used/executed much in pretraining, as the models we care about will go through post-training and I doubt you are arguing post-training will reliably remove scheming.

I also think it seems probably very hard to train a system capable of obsoleting top human experts which doesn't understand that it is an AI even if you're willing to take a big competitiveness hit.

Indeed but that is entirely the point - by construction!

Conceptually we have a recipe R (arch, algorithms, compute, etc), and a training dataset which we can parameterize by time cutoff T. Our objective (for safety research) is not to train a final agent, but instead to find a safe/good R with minimal capability penalty. All important results we care about vary with R independently of T, but competitiveness/dangerousness does vary strongly with T.

Take the same R but vary the time cutoff T of the training dataset: the dangerousness of the AI will depend heavily on T, but not the relative effectiveness of various configurations of R. That is simply a restatement of the ideal requirements for a safe experimental regime. Models/algos that work well for T of 1950 will also work for T of 2020 etc.

Training processes with varying (apparent) situational awareness

• 1:2.5 The AI seemingly isn't aware it is an AI except for a small fraction of training which isn't where much of the capabilities are coming from. For instance, the system is pretrained on next token prediction, our evidence strongly indicates that the system doesn't know it is an AI when doing next token prediction (which likely requires being confident that it isn't internally doing a substantial amount of general-purpose thinking about what to think about), and there is only a small RL process which isn't where much of the capabilities are coming from.

Abilities/intelligence come almost entirely from pretraining, so all the situation awareness and scheming capability that current (and future similar) frontier models possess is thus also mostly present in the base model.  The fact that you need to prompt them to summon out a situationally aware scheming agent doesn't seem like much of a barrier, and indeed strong frontier base models are so obviously misaligned/jail-breakable/dangerous that releasing them to the public is PR-harmful enough to motivate RLHF post training purely for selfish profit-motives.

> This implies that restricting when AIs become (saliently) aware that they are an AI could be a promising intervention, to the extent this is possible without greatly reducing competitiveness.

Who cares if it greatly reduces competitiveness in experimental training runs?

We need to figure out how to align superhuman models - models trained with > 1e25 efficient flops on the current internet/knowledge, which requires experimental iteration.  We probably won't get multiple iteration attempts for aligning SI 'in prod', so we need to iterate in simulation (what you now call 'model organisms').

We need to find alignment training methods that work even when the agent has superhuman intelligence/inference.  But 'superhuman' hear is relative - measured against our capabilities.  The straightforward easy way to accomplish this is training agents in simulations with much earlier knowledge cutoff dates, which isn't theoretically hard - just requires constructing augmented historical training datasets.  So you could train on a 10T+ token dataset of human writings/thoughts with cutoff 2010, or 1950, or 1700, etc.  These base models wouldn't be capable of simulating/summoning realistic situationally aware agents, their RL derived agents wouldn't be situationally sim-aware either, etc.

 

Input vs output tokens are both unique per agent history (prompt + output), so that differentiation doesn't matter for my core argument about the RAM constraint. If you have a model which needs 1TB of KV cache, and you aren't magically sharing that significantly between instances, then you'll need at least 1000 * 1TB of RAM to run 1000 inferences in parallel.

The 3x - 10x cost ratio model providers charge is an economic observation that tells us something about the current cost vs utility tradeoffs, but it's much complicated by oversimpliciation of the current pricing models (they are not currently charging their true costs, probably because that would be too complicated, but also perhaps reveal too much information - their true cost would be more like charging rent on RAM for every timestep). It just tells you that very roughly, that on average, the mean (averaged over many customer requests) flop utilization of the generation phase (parallel over instances) is perhaps 3x to 10x lower than the prefill phase (parallel over time) - but it doesn't directly tell you why.

This is all downstream dependent on model design and economics. There are many useful requests that LLMs can fulfill without using barely any KV cache - essentially all google/oracle type use cases where you are just asking the distilled wisdom of the internet a question. If those were all of the request volume, then the KV cache RAM per instance would be inconsequential, inference batch sizes would be > 1000, inference flop utilization would be the same for prefill vs generation, and providers would charge the same price for input vs output tokens.

On the other extreme, if all requests used up the full training context window, then the flop utilization of inference would be constrained by approximately (max_KV_cache_RAM + weight_RAM / max_KV_cache_RAM ) / alu_ratio. For example if the KV cache is 10% of RAM, and alu_ratio is 1000:1, generation would have max efficiency of 1%. If infill efficiency was 30%, then output tokens would presumably be priced 30x more than input tokens.

So the observed input:output token pricing is dependent on the combination of KV_cache RAM fraction (largely a model design decision), current efficiency of implementations of infill vs generation, and most importantly - the distribution of request prompt lengths, which itself is dependent on the current economic utility of shorter vs longer prompts for current models.

In practice most current models have a much smaller KV cache to weight RAM fraction than my simple 1:1 example, but the basic point holds: training is more flop & interconnect limited, inference is more RAM and ram bw limited. But these constraints already shape the design space of models and how they are deployed.

LLMs currently excel at anything a human knowledge worker can do without any specific training (minimal input prompt length), but largely aren't yet competitive with human experts at most real world economic tasks that require significant unique per-job training. Coding is a good example - human thoughtspeed is roughly 9 token/s, or 32K/hour, or 256K per 8 hour work day, or roughly 1M tokens per week.

Current GPT4-turbo (one of the current leaders for coding), for example, has a max context length of 128K (roughly 4 hours). But if you actually use all of that for each request for typical coding requests that generate say 1K of useful output (equivalent to a few minutes of human thought), that will cost you about $1.25 for the input tokens, but only about $0.03 for the output tokens. That costs about as much as a human worker, per minute of output thought tokens. The cost of any LLM agent today (per minute of output thought) increases linearily with input prompt length - ie the agent's unique differentiating short term memory. Absent more sophisticated algorithms, the cost of running a react-like LLM agent thus grows quadratically with time, vs linear for humans (because each small observe-act time step has cost proportional to input context length, which grows per time step).

Human programmers aren't being replaced en masse (yet) in part because current models aren't especially smarter than humans at equivalent levels of job-specific knowledge/training.

Normalized for similar ability, LLMs currently are cheaper than humans at most any knowledge work that requires very little job-specific knowledge/training, and much more expensive than humans for tasks that require extensive job-specific knowledge/training - and this has everything to do with how transformers currently consume and utilize VRAM.

Not for transformers, for which training and inference are fundamentally different.

Transformer training parallelizes over time, but that isn't feasible for inference. So transformer inference backends have to parallelize over batch/space, just like RNNs, which is enormously less efficient in RAM and RAM bandwidth use.

So if you had a large attention model that uses say 1TB of KV cache (fast weights) and 1TB of slow weights, transformer training can often run near full efficiency, flop limited, parallelizing over time.

But similar full efficient transformer inference would require running about K instances/agents in parallel, where K is the flop/mem_bw ratio (currently up to 1000 on H100). So that would be 1000 * 1TB of RAM for the KV cache (fast weights) as its unique per agent instance.

This, in a nutshell, is part of why we don't already have AGI. Transformers are super efficient at absorbing book knowledge, but just as inefficient as RNNs at inference (generating new experiences, which is a key bottleneck on learning from experience).

Thus there is of course much research in more efficient long kv cache, tree/graph inference that can share some of the KV cache over similar branching agents, etc

Due to practical reasons, the compute requirements for training LLMs is several orders of magnitude larger than what is required for running a single inference instance.  In particular, a single NVIDIA H100 GPU can run inference at a throughput of about 2000 tokens/s, while Meta trained Llama3 70B on a GPU cluster^[1] of about 24,000 GPUs.  Assuming we require a performance of 40 tokens/s, the training cluster can run $\frac{2000}{40}\times 24000=1,200,000$ concurrent instances of the resulting 70B model.

I agree direction-ally with your headline, but your analysis here assumes flops is the primary constraint on inference scaling.  Actually it looks like VRAM is already the more important constraint, and would likely become even more dominant if AGI requires more brain-like models.

LLMs need VRAM for both 'static' and 'dynamic' weights.  The static weights are the output of the long training process, and shared over all instances of the same model or fine tune (LORAs share most).  However the dynamic 'weights' - in the attention KV cache - are essentially unique to each individual instance of the model, specific to its current working memory context and chain of thought.

So the key parameters here are total model size and dynamic vs static ratio (which depends heavily on context length and many other factors).  But for example if dynamic is 50% of the RAM usage then 1M concurrent instances would require almost as many GPUs.

If AGI requires scaling up to very large brain-size models ~100T params (which seems likely), and the dynamic ratio is even just 1%, then 1M concurrent instances would require on order 10M GPUs.