Thanks Steven for clearly making this point. I understand and agree with the point that weight update is important for true incremental learning. As you imply, weight updates give the opportunity for the model to represent information in more multidimensional way than simple context allows. It may be that something beyond transformers plus scaffolding is needed to get to 'real' continual learning, but I'm interested in comments about transformer-based possibilities.
Models could learn by retraining curated samples from prior models - like the agent rollouts...
I think that, even if LLM's don't smoothly evolve into AGI then ASI, an alternative 'brain-like' AGI will have a similar progress ramp that allows for alignment learning-by-doing in a very meaningful way. To explain this, let's discuss the LLM path a bit. OpenAI's deliberative alignment and Anthropic's more sober discussion of the ongoing alignment challenge both highlight the effort that companies today put in to understanding and improving LLM alignment. Alignment work is progressing through improved training, RLHF, RLAIF, Constitutional Classifiers, etc...
Interesting - I too suspect that good world models will help with data efficiency. Even using the existing training paradigm where a lot of data is needed to get the generalization to work well, if an AI has a good internal world model it could generate usable synthetic examples for incremental training. For example, when a child sees a photo of some strange new animal from the side, the child likely surmises that the animal looks the same from the other side; if the photo only shows one eye, the child can imagine that looking head on into the animal's fac...
Thanks for the post, and especially the causal graph framework you use to describe and analyze the categories of motivations. It feels similar to Richard Dawkins 'The Selfish Gene' work in so far as it studies the fitness of motivations in their selection environment.
One area I think it could extend to is around the concepts of curiosity, exploration, and even corrigibility. This relates to 'reflection' as mentioned by habryka. I expect that as AI moves towards AGI it will improve its ability to recognize holes in its knowledge and take actions to close th...
Thanks for the interesting peak into your brain. I have a couple thoughts to share on how my own approaches relate.
The first is related to watching plenty of sci-fi apocalyptic future movies. While it's exciting to see the hero's adventures, I'd like to think that I'd be one of the scrappy people trying to hold some semblance of civilization together. Or the survivor trying to barter and trade with folks instead of fighting over stuff. In general, even in the face of doom, just trying to help minimize suffering unto the end. So the 'death with dignity' eth...
That idea of catching bad mesa-objectives during training sounds key and I presume fits under the 'generalization science' and 'robust character training' from Evan's original post. In the US, NIST is working to develop test, evaluation, verification and validation standards for AI and it would be good to include this concept into that effort.
The data relating exponential capabilities to Elo is seen over decades in the computer chess history too. From the 1960s into the 2020's, while computer hardware advanced exponentially at 100-1000x per decade in performance (and SW for computer chess advanced too), Elo scores grew linearly at about 400x per decade, taking multiple decades to go from 'novice' to 'superhuman'. Elo scores have a tinge of exponential to them - a 400 point Elo advantage is about a 10:1 chance for the higher scored competitor to win, and an 800 point Elo is about 200:1, etc. It ...
Thanks Evan for an excellent overview of the alignment problem as seen from within Anthropic. Chris Olah's graph showing perspectives on alignment difficulty is indeed a useful visual for this discussion. Another image I've shared lately relates to the challenge of building inner alignment illustrated by figure 2 from Contemplative Artificial Intelligence:
In the images, the blue arrows indicate our efforts to maintain alignment on AI as capabilities advance through AGI to ASI. In the left image, we see the case of models that generalize in misaligned ways ...
I concur with that sentiment. GPUs hit a sweet spot between compute efficiency and algorithmic flexibility. CPUs are more flexible for arbitrary control logic, and custom ASICs can improve compute efficiency for a stable algorithm, but GPUs are great for exploring new algorithms where SIMD-style control flows exist (SIMD=single instruction, multiple data).
I would include "constructivist learning" in your list, but I agree that LLMs seem capable of this. By "constructivist learning" I mean a scientific process where the learning conceives of an experiment on the world, tests the idea by acting on the world, and then learns from the result. A VLA model with incremental learning seems close to this. RL could be used for the model update, but I think for ASI we need learnings from real-world experiments.
This post provides a good overview of some topics I think need attention by the 'AI policy' people at national levels. AI policy (such as the US and UK AISI groups) has been focused on generative AI and recently agentic AI to understand near-term risks. Whether we're talking LLM training and scaffolding advances, or a new AI paradigm, there is new risk when AI begins to learn from experiments in the world or reasoning about its own world model. In child development, imitation learning focuses on learning from examples, while constructivist learning focuses...
From what I gather reading the ACAS X paper, it formally proved a subset of the whole problem and many issues uncovered by using the formal method were further analyzed using simulations of aircraft behaviors (see the end of section 3.3). One of the assumptions in the model is that the planes react correctly to control decisions and don't have mechanical issues. The problem space and possible actions were well-defined and well-constrained in the realistic but simplified model they analyzed. I can imagine complex systems making use of provably correct...
Steve, thanks for your explanations and discussion. I just posted a base reply about formal verification limitations within the field of computer hardware design. In that field, ignoring for now the very real issue of electrical and thermal noise, there is immense value in verifying that the symbolic 1's and 0's of the digital logic will successfully execute the similarly symbolic software instructions correctly. So the problem space is inherently simplified from the real world, and the silicon designers have incentive to build designs that are easy to tes...
Thanks for the study Andrew. In the field of computer hardware design, formal verification is often used on smaller parts of the design, but randomized dynamic verification (running the model and checking results) is still necessary to test corner cases in the larger design. Indeed, the idea that a complex problem can be engineered so as to be easier to formally verify is discussed in this recent paper formally verifying IEEE floating point arithmetic. In that paper, published in 2023, they report using their divide-and-conquer approach on the problem resu...
An interesting grouping of fields. While my recent work is in with AI and machine learning, I used to work in the field of computer hardware engineering and have thought there are a lot of key parallels:
1) Critical usage test occurs after design and validation
In the case of silicon design, the fabrication of the design is quite expensive and so a lot of design techniques are included to facilitate debug after manufacturing and also a lot of effort is put in to pre-silicon validation. In the case of transformative AI, using...
The Tom's Hardware article is interesting, thanks. It makes the point that the price quoted may not include the full 'cost of revenue' for the product in that it might be the bare die price and not the tested and packaged part (yields from fabs aren't 100% so extensive functional testing of every part adds cost). The article also notes that R&D costs aren't included in that figure; the R&D for NVIDIA (and TSMC, Intel, AMD, etc) are what keep that exponential perf-per-dollar moving along.
For my own curiosity, I looked into current and past inc...
This is very interesting work, showing the fractal graph is a good way to visualize the predictive model being learned. I've had many conversations with folks who struggle with the idea 'the model is just predicting the next token, how can it be doing anything interesting'?. My standard response had been that conceptually the transformer model matches up tokens at the first layer (using the key and query vectors), then matches up sentences a few layers up, and then paragraphs a few layers above that; hence the model, when presented with an input, was not j...
Thanks for the interesting and thoughtful article. As a current AI researcher and former silicon chip designer, I'd suspect that our perf-per-doller is trending a bit slower than exponential now and not a hyperexponential. My first datapoint in support of this is the data from https://en.wikipedia.org/wiki/FLOPS which shows over 100X perf/dollar improvement from 1997 to 2003 (6 years), but the 100X improvement from 2003 is in 2012 (9 years), and our most recent 100X improvement (to the AMD RX 7600 the author cites) took 11 years. This aligns with TOP500 co...
Overall I like this plan and think it raises good points and shows possible paths. Thanks for the great work.
1) The pause time shown is from 2035 to 2040, which feels OK. One concern I have with longer pauses is the usual innovation of HW and SW, which tend generally to range from 100 to 1000X per decade each. So a 5 year pause means all non-controlled AI get 100-1000X more compute per dollar - systems that cost $1 billion to train and run well drop to 10 or even 1 million, which adds to the problem of enforcing a ban. You may want to note this cost reduct... (read more)