Dario Amodei, Chris Olah, Jacob Steinhardt
Paul Christiano, John Schulman, Dan Mané
2016 · arXiv
Rapid progress in machine learning and artificial intelligence (AI) has brought increasing attention to the potential impacts of AI technologies on society. In this paper we discuss one such potential impact: the problem of accidents in machine learning systems, defined as unintended and harmful behavior that may emerge from poor design of real-world AI systems. We present a list of five practical research problems related to accident risk, categorized according to whether the problem originates from having the wrong objective function ("avoiding side effects" and "avoiding reward hacking"), an objective function that is too expensive to evaluate frequently ("scalable supervision"), or undesirable behavior during the learning process ("safe exploration" and "distributional shift"). We review previous work in these areas as well as suggesting research directions with a focus on relevance to cutting-edge AI systems. Finally, we consider the high-level question of how to think most productively about the safety of forward-looking applications of AI.
Metadata: arXiv. Abstracts retain their original authors’ rights.
Original HN comments, with labels from simple text rules. A label describes the whole comment and may miss context or irony. Read the discussion to make up your own mind. About the labels
13 selected excerpts · 13 total mentions
I would add, as a compromise with the superintelligence true believers, "Concrete Problems in AI Safety" is quite good, and worth reading and thinking about. (It's especially relevant to RL systems that interact with the world.) https://arxiv.org/pdf/1606.06565v2.pdf
Ok but Dario has been thinking about AI Safety since 2016 [1], before even GPT-1. I think the simplest explanation is that the Anthropic folks genuinely believe what they say, it just happens to also help their business a lot. [1]: https://arxiv.org/abs/1606.06565
Can you be more specific on what you mean - how would you prove what anyone believes about something short of reading minds and what would proof look like to you? Here is Dario writing about AI safety in 2016 [1]. Dario and others in the Anthropic circle have long been associated [2] with the effective altruism movement, which whatever you think about them, they are very concerned about AI existential risks. Ronan Farrow & the New Yorker did a mega deep dive on Sam Altmans history of being dishonest and manipulative, and it credibly reports that Anthropic cofounders left due to not trusting leadership there with safety. On the other hand we have decades long evidence of…
Altman and co. being dishonest and employing people like Chris Lehane [3]. I have no affiliation with Anthropic beyond being a user of their products to be clear but it feels like HN is going to end up on the wrong side of this one. [1]. https://arxiv.org/abs/1606.06565 [2]. https://forum.effectivealtruism.org/posts/53Gc35vDLK2u5nBxP/... [3]. https://en.wikipedia.org/wiki/Chris_Lehane
The dangers of AGI I don't think are very related with consciousness or sentience (maybe the saddest outcome if it killed us all and weren't at all). I'm still catching up on my reading, but the existential risk of AGI as I understand it is chiefly that we don't know how to align AI (and keep it aligned) to be beneficial/benevolent for humanity, but the risk is contingent also on a few other things things - that we develop AI that agentic and able to develop and pursue it's own goals and sub-goals, and that it has power-seeking or game-theory strategic awareness/planning. The extra dangerous situation is if it is better than humans at making better AGIs and allowed to…
improve itself (or build its successors). Regardless of when we reach AGI, I think the alignment issue is a key one for both AI ethics and safety. As for the rest, while they aren't necessarily things that will happen, since it seems to the first thing people most want to build - the other day, the top three projects on https://github.com/trending were all projects to try to make ChatGPT autonomous/agentic. FWIW, I don't think the 6-month ban is a good idea, but I also think there's a lot of xrisk w/ AGI and it's better to be thinking about it seriously sooner, rather than later. What I've been reading: Ngo, Richard, Lawrence Chan, and Sören Mindermann. “The Alignment Problem from a Deep Learning Perspective.” arXiv, February 22, 2023. http://arxiv.org/abs/2209.00626. Carlsmith, Joseph. “Is Power-Seeking AI an Existential Risk?” arXiv, June 16, 2022. http://arxiv.org/abs/2206.13353. Yudkowsky, Eliezer. “AGI Ruin: A List of Lethalities.” Accessed April 8, 2023. https://www.alignmentforum.org/posts/uMQ3cqWDPHhjtiesc/agi-r.... June 6, 2022. paulfchristiano. “What Failure Looks Like.” Accessed April 6, 2023. https://www.alignmentforum.org/posts/HBxe6wdjxK239zajf/what-.... March 18, 2019. Soares, Nate. “The Value Learning Problem.” In Artificial Intelligence Safety and Security, edited by Roman V. Yampolskiy, 1st ed., 89–97. First edition. | Boca Raton, FL : CRC Press/Taylor & Francis Group, 2018.: Chapman and Hall/CRC, 2018. https://doi.org/10.1201/9781351251389-7. Amodei, Dario, Chris Olah, Jacob Steinhardt, Paul Christiano, John Schulman, and Dan Mané. “Concrete Problems in AI Safety.” arXiv, July 25, 2016. http://arxiv.org/abs/1606.06565.
Some overlap here with the 2017 paper from Google Brain et al: Concrete Problems in AI Safety: https://arxiv.org/pdf/1606.06565.pdf A nice video from the wonderful AI safety communicator Robert Miles outlines those problems: https://www.youtube.com/watch?v=AjyM-f8rDpg
Their paper "Concrete problems in AI safety"[1] is interesting. Could be more concrete. They're run into the "common sense" problem, which I sometimes define, for robots, as "getting through the next 30 seconds without screwing up". They're trying to address it by playing with the weighting in goal functions for machine learning. They write "Yet intuitively it seems like it should often be possible to predict which actions are dangerous and explore in a way that avoids them, even when we don’t have that much information about the environment." For humans, yes. None of the tweaks on machine learning they suggest do that, though. If your constraints are in the objective…
function, the objective function needs to contain the model of "don't do that". Which means you've just moved the common sense problem to the objective function. Important problem to work on, even though nobody has made much progress on it in decades. [1] https://arxiv.org/pdf/1606.06565.pdf
Yes, this is one specific safety problem -- there are many other RL safety problems that deserve high quality benchmarks too. See eg https://arxiv.org/pdf/1606.06565.pdf or https://medium.com/@deepmindsafetyresearch/building-safe-art... for discussions of the problem space.
I think the answer is that "just building in costs" is actually rather hard to get right. Check out how Concrete Problems in AI Safety (Section 6 in particular is about safe exploration) https://arxiv.org/pdf/1606.06565.pdf Quote: In practice, real world RL projects can often avoid these issues by simply hard-coding an avoidance of catastrophic behaviors. For instance, an RL-based robot helicopter might be programmed to override its policy with a hard-coded collision avoidance sequence (such as spinning its propellers to gain altitude) whenever it’s too close to the ground. This approach works well when there are only a few things that could go wrong, and the designers…
know all of them ahead of time. But as agents become more autonomous and act in more complex domains, it may become harder and harder to anticipate every possible catastrophic failure. The space of failure modes for an agent running a power grid or a search-and-rescue operation could be quite large. Hard-coding against every possible failure is unlikely to be feasible in these cases, so a more principled approach to preventing harmful exploration seems essential. Even in simple cases like the robot helicopter, a principled approach would simplify system design and reduce the need for domain-specific engineering
No, I wasn't aware of that. This is a great link, thanks! To anyone else here is their excellent 2016 paper linked from the above: "Concrete Problems in AI Safety" https://arxiv.org/abs/1606.06565 and direct link to 29-page PDF: https://arxiv.org/pdf/1606.06565.pdf [PDF!] However, that PDF does not mention "AGI" even a single time, except in reference 17 "The AGI containment problem". That is actually also avaialable online here - https://arxiv.org/abs/1604.00545 but doesn't seem to be what they have in mind. So it seems the use of the term "safe" is actually much more narrow in AI literature and probably in the mind of the person who wrote that sentence, than I as a lay reader thought of reading it applied to AGI. It's an interesting idea.
This is Dario Amodei, head of the OpenAI safety team. We are devoting a substantial fraction of the organization’s bandwidth to safety, both on the side of technical research (where we have several people currently working on it full-time, and are very actively hiring for more: https://openai.com/jobs/), and in terms of what we think the right policy actions are (several of us, including me, have been doing a lot of talking to people in government about what we can expect from AI and how the government can help). Beyond that it’s just a very central part of our strategy — it’s important for our organization to be on the cutting edge of AI, but as time goes on, safety will…
become more and more central, particularly as we have more concrete and powerful systems that need to be made safe. Past examples of our papers include this: https://blog.openai.com/deep-reinforcement-learning-from-hum... (with the associated system released here: https://blog.openai.com/gathering_human_feedback/) and this https://arxiv.org/abs/1606.06565 (done while I was still at Google, but in collaboration with OpenAI people right before I joined OpenAI). Based on the current projects going on in my group, I am hopeful we’ll have several more papers out soon.
"but little or no scientific work to be done." Quite a lot has been written about what scientific work needs to be done. These papers try to summarize possible research directions: https://arxiv.org/pdf/1606.06565.pdf https://intelligence.org/files/TechnicalAgenda.pdf
According to this paper https://arxiv.org/pdf/1606.06565.pdf, OpenAI (John Schulman and Paul Christiano) explicitly indicate interest in researching transparency and strongly support work on fairness.
Paper: https://arxiv.org/pdf/1606.06565v1.pdf Google Post: https://research.googleblog.com/2016/06/bringing-precision-t... It was a pleasure for us to work on this with OpenAI and others. John/Paul/Jacob are good friends, and wonderful colleagues! :)
More from the same topics.
Sébastien Bubeck, Varun Chandrasekaran, Ronen Eldan +11
Artificial intelligence (AI) researchers have been developing and refining large language models (LLMs) that exhibit remarkable capabilities across a variety of domains and tasks, challenging our understanding of learning and cognition.
Tom B. Brown, Benjamin Mann, Nick Ryder +28
Recent work has demonstrated substantial gains on many NLP tasks and benchmarks by pre-training on a large corpus of text followed by fine-tuning on a specific task.
Chiyuan Zhang, Samy Bengio, Moritz Hardt +2
Despite their massive size, successful deep artificial neural networks can exhibit a remarkably small difference between training and test performance.
A. Pluchino. A. E. Biondo, A. Rapisarda
The largely dominant meritocratic paradigm of highly competitive Western cultures is rooted on the belief that success is due mainly, if not exclusively, to personal qualities such as talent, intelligence, skills, efforts or risk taking.