Orphan answers
Back to Improve answers.
These 201 canonical answers are not marked as a related or follow-up to any other canonical answer, so cannot be found through the read interface by normal browsing. Feel free to add them to some! See a full list of available canonical questions or browse by tags.
The Alignment Problem (2020) by Brian Christian is the most recent in-depth guide to the field.
The book which first made the case to the public is Nick Bostrom’s Superintelligence (2014). It gives an excellent overview of the state of the field (as it was then) and makes a strong case for the subject being important, as well as exploring many fascinating adjacent topics. However, it does not cover newer developments, such as mesa-optimizers or language models.
There's also Human Compatible (2019) by Stuart Russell, which gives a more up-to-date review of developments, with an emphasis on the approaches that the Center for Human-Compatible AI are working on, such as cooperative inverse reinforcement learning. There's a good review/summary on SlateStarCodex.
Although not limited to AI safety, The AI Does Not Hate You (2020) is an entertaining and accessible outline of both the core issues and an exploration of some of the community and culture of the people working on it.
Various other books explore the issues in an informed way, such as Toby Ord’s The Precipice (2020), Max Tegmark’s Life 3.0 (2017), Yuval Noah Harari’s Homo Deus (2016), Stuart Armstrong’s Smarter Than Us (2014), and Luke Muehlhauser’s Facing the Intelligence Explosion (2013).
Intelligence is powerful. Because of superior intelligence, we humans have dominated the Earth. The fate of thousands of species depends on our actions, we occupy nearly every corner of the globe, and we repurpose vast amounts of the world's resources for our own use. Artificial Superintelligence (ASI) has potential to be vastly more intelligent than us, and therefore vastly more powerful. In the same way that we have reshaped the earth to fit our goals, an ASI will find unforeseen, highly efficient ways of reshaping reality to fit its goals.
The impact that an ASI will have on our world depends on what those goals are. We have the advantage of designing those goals, but that task is not as simple as it may first seem. As described by MIRI in their Intelligence Explosion FAQ:
“A superintelligent machine will make decisions based on the mechanisms it is designed with, not the hopes its designers had in mind when they programmed those mechanisms. It will act only on precise specifications of rules and values, and will do so in ways that need not respect the complexity and subtlety of what humans value.”
If we do not solve the Control Problem before the first ASI is created, we may not get another chance.
Failures can happen with narrow non-agentic systems, mostly from humans not anticipating safety-relevant decisions made too quickly to react, much like in the 2010 flash crash.
A helpful metaphor draws on self-driving cars. By relying more and more on an automated process to make decisions, people become worse drivers as they’re not training themselves to react to the unexpected; then the unexpected happens, the software system itself reacts in an unsafe way and the human is too slow to regain control.
This generalizes to broader tasks. A human using a powerful system to make better decisions (say, as the CEO of a company) might not understand those very well, get trapped into an equilibrium without realizing it and essentially losing control over the entire process.
More detailed examples in this vein are described by Paul Christiano in What failure looks like.
Another source of failures is AI-mediated stable totalitarianism. The limiting factor in current pervasive surveillance, police and armed forces is manpower; the use of drones and other automated tools decreases the need for personnel to ensure security and extract resources.
As capabilities improve, political dissent could become impossible, checks and balances would break down as a minimal number of key actors is needed to stay in power.
We're also building a cleaner web UI for readers and a bot interface.
Ajeya Cotra has written an excellent article named Why AI alignment could be hard with modern deep learning on this question.
Imagine, for example, that you are tasked with reducing traffic congestion in San Francisco at all costs, i.e. you do not take into account any other constraints. How would you do it? You might start by just timing traffic lights better. But wouldn’t there be less traffic if all the bridges closed down from 5 to 10AM, preventing all those cars from entering the city? Such a measure obviously violates common sense, and subverts the purpose of improving traffic, which is to help people get around – but it is consistent with the goal of “reducing traffic congestion”.
If we could, it would solve a large part of the alignment problem.
The challenge is, how do we code this? Converting something to formal mathematics that can be understood by a computer program is much harder than just saying it in natural language, and proposed AI goal architectures are no exception. Complicated computer programs are usually the result of months of testing and debugging. But this one will be more complicated than any ever attempted before, and live tests are impossible: a superintelligence with a buggy goal system will display goal stability and try to prevent its programmers from discovering or changing the error.
Items on lists start with *, numbered lists with #
- For external links use [ followed directly by the URL, a space, then display text and finally a ] symbol
- e.g. [https://www.example.com External link text] gives External link text
- For internal links write the page title wrapped in [[]]s
- e.g. [[What is the Stampy project?]] gives What is the Stampy project?. Including a pipe symbol followed by display text e.g. [[What is the Stampy project?┊Display Text]] allows you to show different Display Text.
- (ref)Reference notes go inside these tags(/ref)[1]
- If you post the raw URL of an image from imgur it will be displayed.[2] You can reduce file compression if you get an account. Note that you need the image itself, right click -> copy image address to get it
- To embed a YouTube video, use (youtube)APsK8NST4qE(/youtube) with the video ID of the target video.
- Start with ** or ## for double indentation
- Three 's around text - Bold
- Two 's around text Italic - Italic
Great! I’ll ask you a few follow-up questions to help figure out how you can best contribute, give you some advice, and link you to resources which should help you on whichever path you choose. Feel free to scroll up and explore multiple branches of the FAQ if you want answers to more than one of the questions offered :)
Note: We’re still building out and improving this tree of questions and answers, any feedback is appreciated.
At what level of involvement were you thinking of helping?
Please view and suggest to this google doc for improvements: https://docs.google.com/document/d/1S-CUcoX63uiFdW-GIFC8wJyVwo4VIl60IJHodcRfXJA/edit#
GPT-3 is the newest and most impressive of the GPT (Generative Pretrained Transformer) series of large transformer-based language models created by OpenAI. It was announced in June 2020, and is 100 times larger than its predecessor GPT-2.[1]
Gwern has several resources exploring GPT-3's abilities, limitations, and implications including:
- The Scaling Hypothesis - How simply increasing the amount of compute with current algorithms might create very powerful systems.
- GPT-3 Nonfiction
- GPT-3 Creative Fiction
Vox has an article which explains why GPT-3 is a big deal.
- ↑ GPT-3: What’s it good for? - Cambridge University Press
Dev team
Name |
Vision talk |
Github |
Trello |
Active? |
Notes / bio |
|---|---|---|---|---|---|
Aprillion |
video |
Aprillion |
yes |
yes |
experienced dev (Python, JS, CSS, ...) |
Augustus Caesar |
yes |
AugustusCeasar |
yes |
soon! |
Has some Discord bot experience |
Benjamin Herman |
no |
no (not needed) |
no |
no |
Helping with wiki design/css stuff |
ccstan99 |
no |
ccstan99 |
yes |
yes |
UI/UX designer |
chriscanal |
yes |
chriscanal |
yes |
yes |
experienced python dev |
Damaged |
no (not needed) |
no (not needed) |
no (not needed) |
yes |
experienced Discord bot dev, but busy with other projects. Can answer questions. |
plex |
yes |
plexish |
yes |
yes |
MediaWiki, plans, and coordinating people guy |
robertskmiles |
yes |
robertskmiles |
yes |
yes |
you've probably heard of him |
Roland |
yes |
levitation |
yes |
yes |
working on Semantic Search |
sct202 |
yes |
no (add when wiki is on github) |
yes |
yes |
PHP dev, helping with wiki extensions |
Social Christancing |
yes |
chrisrimmer |
yes |
maybe |
experienced linux sysadmin |
sudonym |
yes |
jmccuen |
yes |
yes |
systems architect, has set up a lot of things |
tayler6000 |
yes |
tayler6000 |
no |
yes |
Python and PHP dev, PenTester, works on Discord bot |
Editors
Language Models are a class of AI trained on text, usually to predict the next word or a word which has been obscured. They have the ability to generate novel prose or code based on an initial prompt, which gives rise to a kind of natural language programming called prompt engineering. The most popular architecture for very large language models is called a transformer, which follows consistent scaling laws with respect to the size of the model being trained, meaning that a larger model trained with the same amount of compute will produce results which are better by a predictable amount (when measured by the 'perplexity', or how surprised the AI is by a test set of human-generated text).
See also
A Superintelligence would be intelligent enough to understand what the programmer’s motives were when designing its goals, but it would have no intrinsic reason to care about what its programmers had in mind. The only thing it will be beholden to is the actual goal it is programmed with, no matter how insane its fulfillment may seem to us.
Consider what “intentions” the process of evolution may have had for you when designing your goals. When you consider that you were made with the “intention” of replicating your genes, do you somehow feel beholden to the “intention” behind your evolutionary design? Most likely you don't care. You may choose to never have children, and you will most likely attempt to keep yourself alive long past your biological ability to reproduce.
A potential solution is to create an AI that has the same values and morality as a human by creating a child AI and raising it. There’s nothing intrinsically flawed with this procedure. However, this suggestion is deceptive because it sounds simpler than it is.
If you get a chimpanzee baby and raise it in a human family, it does not learn to speak a human language. Human babies can grow into adult humans because the babies have specific properties, e.g. a prebuilt language module that gets activated during childhood.
In order to make a child AI that has the potential to turn into the type of adult AI we would find acceptable, the child AI has to have specific properties. The task of building a child AI with these properties involves building a system that can interpret what humans mean when we try to teach the child to do various tasks. People are currently working on ways to program agents that can cooperatively interact with humans to learn what they want.
Many parts of the AI alignment ecosystem are already well-funded, but a savvy donor can still make a difference by picking up grantmaking opportunities which are too small to catch the attention of the major funding bodies or are based on personal knowledge of the recipient.
One way to leverage a small amount of money to the potential of a large amount is to enter a donor lottery, where you donate to win a chance to direct a much larger amount of money (with probability proportional to donation size). This means that the person directing the money will be allocating enough that it's worth their time to do more in-depth research.
For an overview of the work the major organizations are doing, see the 2021 AI Alignment Literature Review and Charity Comparison. The Long-Term Future Fund seems to be an outstanding place to donate based on that, as they are the organization which most other organizations are most excited to see funded.
In principle it could (if you believe in functionalism), but it probably won't. One way to ensure that AI has human-like emotions would be to copy the way human brain works, but that's not what most AI researchers are trying to do.
It's similar to how once some people thought we will build mechanical horses to pull our vehicles, but it turned out it's much easier to build a car. AI probably doesn't need emotions or maybe even consciousness to be powerful, and the first AGIs that will get built will be the ones that are easiest to build.
This is a really interesting question! Because, yeah it certainly seems to me that doing something like this would at least help, but it's not mentioned in the paper the video is based on. So I asked the author of the paper, and she said "It wouldn't improve the security guarantee in the paper, so it wasn't discussed. Like, there's a plausible case that it's helpful, but nothing like a proof that it is". To explain this I need to talk about something I gloss over in the video, which is that the quantilizer isn't really something you can actually build. The systems we study in AI Safety tend to fall somewhere on a spectrum from "real, practical AI system that is so messy and complex that it's hard to really think about or draw any solid conclusions from" on one end, to "mathematical formalism that we can prove beautiful theorems about but not actually build" on the other, and quantilizers are pretty far towards the 'mathematical' end. It's not practical to run an expected utility calculation on every possible action like that, for one thing. But, proving things about quantilizers gives us insight into how more practical AI systems may behave, or we may be able to build approximations of quantilizers, etc. So it's like, if we built something that was quantilizer-like, using a sensible human utility function and a good choice of safe distribution, this idea would probably help make it safer. BUT you can't prove that mathematically, without making probably a lot of extra assumptions about the utility function and/or the action distribution. So it's a potentially good idea that's nonetheless hard to express within the framework in which the quantilizer exists. TL;DR: This is likely a good idea! But can we prove it?
An AGI which has recursively self-improved into a superintelligence would be capable of either resisting our attempts to modify incorrectly specified goals, or realizing it was still weaker than us and acting deceptively aligned until it was highly sure it could win in a confrontation. AGI would likely prevent a human from shutting it down unless the AGI was designed to be corrigible. See Why can't we just turn the AI off if it starts to misbehave? for more information.
We don’t yet know which AI architectures are safe; learning more about this is one of the goals of FLI's grants program. AI researchers are generally very responsible people who want their work to better humanity. If there are certain AI designs that turn out to be unsafe, then AI researchers will want to know this so they can develop alternative AI systems.
When one person tells a set of natural language instructions to another person, they are relying on much other information which is already stored in the other person's mind.
If you tell me "don't harm other people," I already have a conception of what harm means and doesn't mean, what people means and doesn't mean, and my own complex moral reasoning for figuring out the edge cases in instances wherein harming people is inevitable or harming someone is necessary for self-defense or the greater good.
All of those complex definitions and systems of decision making are already in our mind, so it's easy to take them for granted. An AI is a mind made from scratch, so programming a goal is not as simple as telling it a natural language command.
Anthropic fine tuned a language model to be more helpful, honest and harmless: HHH.
Motivation: The point of this is to:
- see if we can "align" a current day LLM, and
- raise awareness about safety in the broader ML community.
How can we interpret what all the neurons mean?
Chris Olah, the interpretability legend, is working on looking really hard at all the neurons to see what they all mean. The approach he pioneered is circuits: looking at computational subgraphs of the network, called circuits, and interpreting those. Idea: "decompiling the network into a better representation that is more interpretable". In-context learning via attention heads, and interpretability here seems useful.
One result I heard about recently: a linear softmax unit stretches space and encourages neuron monosemanticity (making a neuron represent only one thing, as opposed to firing on many unrelated concepts). This makes the network easier to interpret.
Motivation: The point of this is to get as many bits of information about what neural networks are doing, to hopefully find better abstractions. This diagram gets posted everywhere, the hope being that networks, in the current regime, will become more interpretable because they will start to use abstractions that are closer to human abstractions.
If you like interactive FAQs, you're in the right place already! Joking aside, some great entry points are the AI alignment playlist on YouTube, “The Road to Superintelligence” and “Our Immortality or Extinction” posts on WaitBuyWhy for a fun, accessible introduction, and Vox's “The case for taking AI seriously as a threat to humanity” as a high-quality mainstream explainer piece.
The free online Cambridge course on AGI Safety Fundamentals provides a strong grounding in much of the field and a cohort + mentor to learn with. There's even an anki deck for people who like spaced repetition!
There are many resources in this post on Levelling Up in AI Safety Research Engineering with a list of other guides at the bottom. There is also a twitter thread here with some programs for upskilling and some for safety-specific learning.
The Alignment Newsletter (podcast), Alignment Forum, and AGI Control Problem Subreddit are great for keeping up with latest developments.
GPT-3 showed that transformers are capable of a vast array of natural language tasks, codex/copilot extended this into programming. One demonstrations of GPT-3 is Simulated Elon Musk lives in a simulation. Important to note that there are several much better language models, but they are not publicly available.
DALL-E and DALL-E 2 are among the most visually spectacular.
MuZero, which learned Go, Chess, and many Atari games without any directly coded info about those environments. The graphic there explains it, this seems crucial for being able to do RL in novel environments. We have systems which we can drop into a wide variety of games and they just learn how to play. The same algorithm was used in Tesla's self-driving cars to do complex route finding. These things are general.
Generally capable agents emerge from open-ended play - Diverse procedurally generated environments provide vast amounts of training data for AIs to learn generally applicable skills. Creating Multimodal Interactive Agents with Imitation and Self-Supervised Learning shows how these kind of systems can be trained to follow instructions in natural language.
GATO shows you can distill 600+ individually trained tasks into one network, so we're not limited by the tasks being fragmented.
