The best things I read about Responsible AI in 2022.
Judging from my conversations with friends and family over the holiday season, it seems like suddenly, there’s a massive interest in learning more about AI and its potential risks — thank you, ChatGPT, for increasing public awareness of a.) the incredible potential of advanced AI systems and b.) the terrifying reality that as a society, we aren’t fully prepared to grapple with the potential risks and harms of such powerful systems.
So if you, like many of the folks I know, want to learn more about Responsible AI in 2023 — or perhaps even make moves to start working in this field — I have one last holiday gift for you: my list of the best things I read this year about AI. From drafting effective AI policy to dissecting how large language models work to managing burnout as someone working on RAI, these are my favorite articles, papers, and books.
👀 We Need to Talk About How Good AI Is Getting by Kevin Roose (NYT): when this article was published back in August, I sent it around to folks with the line, “Wow — this is actually a really great NYT article on the current state of AI. Please read it.” The reality is that most mainstream reporting on AI gets a lot wrong, between overhyping nascent technology and playing into exaggerated fears of Terminator-esque AI. But this article is truly a fantastic overview of where we are today — or rather, where we were in August 2022, because even since then, a tremendous amount of progress has been made.
📐 Aligning Language Models to Follow Instructions by Ryan Lowe and Jan Leike (OpenAI): whenever folks ask me, “How does ChatGPT actually work?” I point them to this great post from the OpenAI blog, which was published well before ChatGPT was released but is a fantastic, easy-to-understand overview of some of the incredible recent advancements in large language models. In particular, this post focuses on reinforcement learning through human feedback (RLHF), which is a powerful technique that allows us to fine-tune language models to be more aligned with our goals and needs.
🌸 BLOOM is the Most Important Model of the Decade by Alberto Romero (Medium): everyone who is even remotely online these days has heard of GPT-3, but have you heard of BLOOM? BLOOM is a large language model developed by the BigScience Research Workshop, and it’s an incredible effort to make advanced AI more open, inclusive, and responsible — in many ways, the polar opposite of models developed by private companies funded by private capital with little incentive to ensure that their technology benefits all of humanity. This explainer post is a great overview of why BLOOM is so exciting, particularly for those of us in the Responsible AI field.
📊 ChatGPT & the Professional’s Guide to Using AI by Allie Miller (LinkedIn Pulse): this is one of the most phenomenal overviews of how advanced AI systems — and even more specifically, generative AI systems like OpenAI’s ChatGPT — can make the lives of professionals in every field easier. This should be required reading for anyone and everyone, because we’re all going to be impacted by AI in the workplace, whether we work in tech or not.
⚖️ AI Governance in the Time of Generative AI by Ian Eisenberg (Credo AI): if the last few articles got you excited about generative AI (and maybe also a little alarmed), then this blog post from my colleague and our Head of Data Science at Credo AI Ian Eisenberg should be your next read. In this post, Ian does a great job of explaining why we need to focus on governing these powerful systems and starts to lay out how we might go about doing so — with follow up posts coming in early January of next year, this is a blog post series that you shouldn’t miss.
😧 Predictability & Surprise in Large Language Models by the Anthropic AI team (arXiv): while academic papers aren’t for everyone, I highly recommend this paper if you’re interested in going deeper into the risks of large language models like ChatGPT; in particular, the fact that these large language models have highly unpredictable “emergent capabilities” makes it very difficult for us to predict what larger models will be capable of once developed. If that doesn’t freak you out, I don’t know what will.
📜 Explainer: Impact Assessments for Artificial Intelligence by Sean Long, Jeremy Pesner, and Tom Romanoff (Bipartisan Policy Center): this article summarizes current efforts to define and implement “AI Impact Assessments” or “Algorithmic Impact Assessments” as a kind of standardized way of reporting on the potential harms and risks of AI systems; much like nutrition labels have helped to standardize the way we evaluate the potential harms of foods as consumers, the hope is that impact assessments (if standardized and implemented at scale) will help consumers of AI systems evaluate the potential harms — and will help regulators manage what kinds of systems should or should not be allowed to be put on the market.
🔥 Responsible AI Has a Burnout Problem by Melissa Heikkila (MIT Technology Review): The end of 2022 wasn’t a great time for folks working on Responsible AI teams in large tech companies, as many organizations cut their RAI teams from the budget. But even before the recent layoffs, many folks working on RAI efforts in the tech sector have faced a lot of backlash from their colleagues and their companies. This article from the MIT Technology Review (which I think publishes some of the best articles on Responsible AI!) shines light on the very real struggles of working on RAI from within. As interest in and awareness of RAI grows, I hope that we can give more support to the folks who are often invisibly working to solve the most dire and pressing issues with this technology at the source.
🦜 Bonus! On the Dangers of Stochastic Parrots: Can Large Language Models Be Too Big? By Emily Bender, Timnit Gebru, Angelina McMillan-Major, and Margaret Mitchell (arXiv): remember when Google engineer Blake Lemoine made headlines earlier this year for arguing that LaMDA, Google’s large language model chatbot, had achieved sentience? Regardless of where you stand on his argument, it’s important to know that this question of sentience was one of the precise risks called out by some of my all-time RAI heroes in the seminal paper that ultimately got Timnit and Margaret Mitchell fired from Google. While this is a paper from 2021, it absolutely deserves a spot on this list, and anyone who wants to go deeper into the world of Responsible AI in the age of LLMs absolutely must read it.
