The Nonlinear Library

The Nonlinear Library

By The Nonlinear FundEducation
Download on the App Store

All The Nonlinear Library episodes

  • EA - ALTER Israel - End-of-2022 Update by Davidmanheim
    Welcome to The Nonlinear Library, where we use Text-to-Speech software to convert the best writing from the Rationalist and EA communities into audio. This is: ALTER Israel - End-of-2022 Update, published by Davidmanheim on January 22, 2023 on The Effective Altruism Forum.
    ALTER is an organization in Israel that works on several EA priority areas and causes. As laid out in our mid-year update, we had several things planned for the second half of 2022, and they were largely successful. Unfortunately, the past several months have been particularly challenging, in large part due to the collapse of FTX and issues with the Israeli banking system. (See coverage here on some specific impacts.) Delayed items include actually posting this update, and other planned progress - but despite that, ALTER has made progress on several fronts, and is continuing its work on several fronts.
    Accomplishments and Progress
    We ran a successful AI safety conference in Israel, AISIC2022 and identified a number of people who are interested in or working on AI safety here in Israel. We are continuing to follow up with them, and will hopefully have more to announce in that vein in the near future.
    We are continuing to do academic outreach in Israel, and building our network for AI safety, biosecurity, and other cause areas.
    Vadim Weinstein,the mathematician who we hired, has been working on AI safety with Vanessa, and we are excited about potential progress. We are working on figuring out how to replace his grant from FTX with other money, but there is a strong interest in funding the work, and we are attempting to finalize those arrangements.
    David attended the Bioweapons Convention Review Conference and continued outreach around Israeli engagement in various international forums.
    Challenges
    The situation with banking in Israel has become much more challenging due to our receipt of a grant from FTX. We are now prohibited from receiving money from international organizations from our (current) bank, which has made getting our other grant money very challenging, and we have so far been unsuccessful. (We are likely to be able to resolve this in the coming months, and EA Israel can potentially lend us funds in the interim.)
    The grant we received from FTX is not being used, and we are trying to understand the legal situation, and what will happen with the bankruptcy. This money is being kept in a separate account, pending clarity about its status and the potential to provide restitution to those defrauded by FTX. Any resolution will likely not be for quite a long time.
    Once we receive currently granted but unreceived funds, our runway is sufficient through the end of this year, rather than the end of next year. We will be applying for additional funding over the course of the year to allow us to continue operations.
    The funding environment over the coming years is more challenging, and for that reason we are less able to support additional work, and have a smaller funding cushion. We are still investigating the different potential avenues forward for this.
    We are also revisiting our strategic planning in light of the changing funding environment and our current projects.
    Plans
    David will be working for ALTER as full time director of research and policy in 2023, as well as continuing to work on our direction and planning.
    We are continuing to work to build a community of people in Israel working on AI safety, including funding one researcher identified via our AI safety conference.
    We are funding a prize for work on Vanessa Kosoy’s AI Safety Learning-Theoretic Agenda
    We are building out a network of academics interested in additional EA-adjacent fields, especially focused on issues relevant to Global Catastrophic Biological Risks. (If you know of anyone in Israel who we should contact, let us know!)
    We are working on a collaborative policy project with several other individuals and groups to get Israel to iodize salt, since iodine deficiency in Israel is an ...
    4 min
  • EA - [BBC News] Postpartum haemorrhage: Niger halves blood-loss deaths at clinics by Sean Ericson
    Welcome to The Nonlinear Library, where we use Text-to-Speech software to convert the best writing from the Rationalist and EA communities into audio. This is: [BBC News] Postpartum haemorrhage: Niger halves blood-loss deaths at clinics, published by Sean Ericson on January 21, 2023 on The Effective Altruism Forum.
    I'm not an expert in this area, but it would seem that this study is an encouraging development in a very important area.
    Over the research period in Niger, an estimated 1,417 fewer women died from bleeding after childbirth - known as postpartum haemorrhage (PPH) - than otherwise would have.
    It also prevented tens of thousands of other women from experiencing abnormally high blood loss.
    PPH now accounts for one in 10 of maternal deaths in Niger, whereas before the project began it accounted for more than three times that.
    Summary of the Lancet article about the study:
    Background
    Primary postpartum haemorrhage is the principal cause of birth-related maternal mortality in most settings and has remained persistently high in severely resource-constrained countries. We evaluate the impact of an intervention that aims to halve maternal mortality caused by primary postpartum haemorrhage within 2 years, nationwide in Niger.
    Methods
    In this 72-month longitudinal study, we analysed the effects of a primary postpartum haemorrhage intervention in hospitals and health centres in Niger, using data on maternal birth outcomes assessed and recorded by the facilities’ health professionals and reported once per month at the national level. Reported data were monitored, compiled, and analysed by a non-governmental organisation collaborating with the Ministry of Health. All births in all health facilities in which births occurred, nationwide, were included, with no exclusion criteria. After a preintervention survey, brief training, and supplies distribution, Niger implemented a nationwide primary postpartum haemorrhage prevention and three-step treatment strategy using misoprostol, followed if needed by an intrauterine condom tamponade, and a non-inflatable anti-shock garment, with a specific set of organisational public health tools, aiming to reduce primary postpartum haemorrhage mortality.
    Findings
    Among 5 382 488 expected births, 2 254 885 (41·9%) occurred in health facilities, of which information was available on 1 380 779 births from Jan 1, 2015, to Dec 31, 2020, with reporting increasing considerably over time. Primary postpartum mortality decreased from 82 (32·16%; 95% CI 25·58–39·92) of 255 health facility maternal deaths in the 2013 preintervention survey to 146 (9·53%; 8·05–11·21) of 1532 deaths among 343 668 births in 2020. Primary postpartum haemorrhage incidence varied between 1900 (2·10%; 2·01–2·20) of 90 453 births and 4758 (1·47%; 1·43–1·52) of 322 859 births during 2015–20, an annual trend of 0·98 (95% CI 0·97–0·99; p<0·0001).
    Interpretation
    Primary postpartum haemorrhage morbidity and mortality declined rapidly nationwide. Because each treatment technology that was used has shown some efficacy when used alone, a strategic combination of these treatments can reasonably attain outcomes of this magnitude. Niger's strategy warrants testing in other low-income and perhaps some middle-income settings.
    Funding
    The Government of Norway, the Government of Niger, the Kavli Trust (Kavlifondet), the InFiL Foundation, and individuals in Norway, the UK, and the USA.
    Translation
    For the French translation of the abstract see Supplementary Materials section.
    Thanks for listening. To help us out with The Nonlinear Library or to learn more, please visit nonlinear.org.
    5 min
  • LW - Large language models learn to represent the world by gjm
    Welcome to The Nonlinear Library, where we use Text-to-Speech software to convert the best writing from the Rationalist and EA communities into audio. This is: Large language models learn to represent the world, published by gjm on January 22, 2023 on LessWrong.
    There's a nice recent paper whose authors did the following:
    train a small GPT model on lists of moves from Othello games;
    verify that it seems to have learned (in some sense) to play Othello, at least to the extent of almost always making legal moves;
    use "probes" (regressors whose inputs are internal activations in the network, trained to output things you want to know whether the network "knows") to see that the board state is represented inside the network activations;
    use interventions to verify that this board state is being used to decide moves: take a position in which certain moves are legal, use gradient descent to find changes in internal activations that make the output of the probes look like a slightly different position, and then verify that when you run the network but tweak the activations as it runs the network predicts moves that are legal in the modified position.
    In other words, it seems that their token-predicting model has built itself what amounts to an internal model of the Othello board's state, which it is using to decide what moves to predict.
    The paper is "Emergent world representations: Exploring a sequence model trained on a synthetic task" by Kenneth Li, Aspen Hopkins, David Bau, Fernanda Viégas, Hanspeter Pfister, and Martin Wattenberg; you can find it at.
    There is a nice expository blog post by Kenneth Li at/.
    Some details that seem possibly-relevant:
    Their network has a 60-word input vocabulary (four of the 64 squares are filled when the game starts and can never be played in), 8 layers, an 8-head attention mechanism, and a 512-dimensional hidden space. (I don't know enough about transformers to know whether this in fact tells you everything important about the structure.)
    They tried training on two datasets, one of real high-level Othello games (about 140k games) and one of synthetic games where all moves are random (about 4M games). Their model trained on synthetic games predicted legal moves 99.99% of the time, but the one trained on real well-played games only predicted legal moves about 95% of the time. (This suggests that their network isn't really big enough to capture legality and good strategy at the same time, I guess?)
    They got some evidence that their network isn't just memorizing game transcripts by training it on a 20M-game synthetic dataset where one of the four possible initial moves is never played. It still predicted legal moves 99.98% of the time when tested on the full range of legal positions. (I don't know what fraction of legal positions are reachable with the first move not having been C4; it will be more than 3/4 since there are transpositions. I doubt it's close to 99.98%, though, so it seems like the model is doing pretty well at finding legal moves in positions it hasn't seen.)
    Using probes whose output is a linear function of the network activations doesn't do a good job of reconstructing the board state (error rate is ~25%, barely better than attempting the same thing from a randomly initialized network), but training 2-layer MLPs to do it gets the error rate down to ~5% for the network trained on synthetic games and ~12% for the one trained on championship games, whereas it doesn't help at all for the randomly trained network. (This suggests that whatever "world representation" the thing has learned isn't simply a matter of having an "E3 neuron" or whatever.)
    I am not at all an expert on neural network interpretability, and I don't know to what extent their findings really justify calling what they've found a "world model" and saying that it's used to make move predictions. In particular, I can't refute the following argument:
    "In most positions, just knowing what moves are legal is enough to give you...
    5 min
  • LW - NYT: Google will “recalibrate” the risk of releasing AI due to competition with OpenAI by Michael Huang
    Welcome to The Nonlinear Library, where we use Text-to-Speech software to convert the best writing from the Rationalist and EA communities into audio. This is: NYT: Google will “recalibrate” the risk of releasing AI due to competition with OpenAI, published by Michael Huang on January 22, 2023 on LessWrong.
    Cross-posted from the EA Forum
    The New York Times: Sundar Pichai, CEO of Alphabet and Google, is trying to speed up the release of AI technology by taking on more risk.
    Mr. Pichai has tried to accelerate product approval reviews, according to the presentation reviewed by The Times.
    The company established a fast-track review process called the “Green Lane” initiative, pushing groups of employees who try to ensure that technology is fair and ethical to more quickly approve its upcoming A.I. technology.
    The company will also find ways for teams developing A.I. to conduct their own reviews, and it will “recalibrate” the level of risk it is willing to take when releasing the technology, according to the presentation.
    This change is in response to OpenAI's public release of ChatGPT. It is evidence that the race between Google/DeepMind and Microsoft/OpenAI is eroding ethics and safety.
    Demis Hassabis, CEO of DeepMind, urged caution in his recent interview in Time:
    He says AI is now “on the cusp” of being able to make tools that could be deeply damaging to human civilization, and urges his competitors to proceed with more caution than before.
    “When it comes to very powerful technologies—and obviously AI is going to be one of the most powerful ever—we need to be careful,” he says.
    “Not everybody is thinking about those things. It’s like experimentalists, many of whom don’t realize they’re holding dangerous material.”
    Worse still, Hassabis points out, we are the guinea pigs.
    Alphabet/Google is trying to accelerate a technology that its own subsidiary says is powerful and dangerous.
    Thanks for listening. To help us out with The Nonlinear Library or to learn more, please visit nonlinear.org.
    2 min
  • AF - Gemini modeling by Tsvi Benson-Tilsen
    Welcome to The Nonlinear Library, where we use Text-to-Speech software to convert the best writing from the Rationalist and EA communities into audio. This is: Gemini modeling, published by Tsvi Benson-Tilsen on January 22, 2023 on The AI Alignment Forum.
    [Metadata: crossposted from. Written 17 June 2022. I'm likely to not respond to comments promptly.]
    A gemini model is a kind of model that's especially relevant for minds modeling minds.
    Two scenarios
    You stand before a tree. How big is it? How does it grow? What can it be used to make? Where will it fall if you cut it here or there?
    Alice usually eats Cheerios in the morning. Today she comes downstairs, but doesn't get out a bowl and spoon and milk, and doesn't go over to the Cheerios cupboard. Then she sees green bananas on the counter. Then she goes and gets a bowl and spoon and milk, and gets a box of Cheerios from the cupboard. What happened?
    We have some kind of mental model of the tree, and some kind of mental model of Alice. In the Cheerios scenario, we model Alice by calling on ourselves, asking how we would behave; we find that we'd behave like Alice if we liked Cheerios, and believed that today there weren't Cheerios in the cupboard, but then saw the green bananas and inferred that Bob had gone to the grocery store, and inferred that actually there were Cheerios. This seems different from how we model the tree; we're not putting ourselves in the tree's shoes.
    Gemini modeling and empathic modeling
    What's the difference though, really, between these two ways of modeling? Clearly Alice is like us in a way the tree isn't, and we're using that somehow; we're modeling Alice using empathy ("in-feeling"). This essay describes another related difference:
    We model Alice's belief in a proposition by having in ourselves another instance of that proposition.
    (Or: by having in ourselves the same proposition, or a grasping of that proposition.) We don't model the tree by having another instance of part of the tree in us. I call modeling some thing by having inside oneself another instance of the thing--having a twin of it--"gemini modeling".
    Gemini modeling is different from empathic modeling. Empathic modeling is tuning yourself to be like another agent in some respects, so that their behavior is explainable as what [you in your current tuning] would do. This is a sort of twinning, broadly, but you're far from identical to the agent you're modeling; you might make different tradeoffs, have different sense acuity, have different concepts, believe different propositions, have different skills, and so on; you tune yourself enough that those differences don't intrude on your predictions. Whereas, the proposition "There are Cheerios in the cupboard.", with its grammatical structure and its immediate implications for thought and action, can be roughly identical between you and Alice.
    As done by humans modeling humans, empathic modeling may or may not involve gemini modeling: we model Alice by seeing how we'd act if we believed certain propositions, and those propositions are gemini modeled; on the other hand, we could do an impression of a silly friend by making ourselves "more silly", which is maybe empathic modeling without gemini modeling. And, gemini modeling done by humans modeling humans involves empathic modeling: to see the implications of believing in a proposition or caring about something, we access our (whole? partial?) agentic selves, our agency.
    Gemini modeling vs. general modeling
    In some sense we make a small part of ourselves "like a tree" when we model a tree falling: our mental model of the tree supports [modeled forces] having [modeled effects] with resulting [modeled dynamics], analogous to how the actual tree moves when under actual forces. So what's different between gemini modeling and any other modeling? When modeling a tree, or a rock, or anything, don't we have a little copy or representation of some aspects of the thing in us? Isn't that like having a so...
    18 min
  • EA - NYT: Google will “recalibrate” the risk of releasing AI due to competition with OpenAI by Michael Huang
    Welcome to The Nonlinear Library, where we use Text-to-Speech software to convert the best writing from the Rationalist and EA communities into audio. This is: NYT: Google will “recalibrate” the risk of releasing AI due to competition with OpenAI, published by Michael Huang on January 22, 2023 on The Effective Altruism Forum.
    The New York Times: Sundar Pichai, CEO of Alphabet and Google, is trying to speed up the release of AI technology by taking on more risk.
    Mr. Pichai has tried to accelerate product approval reviews, according to the presentation reviewed by The Times.
    The company established a fast-track review process called the “Green Lane” initiative, pushing groups of employees who try to ensure that technology is fair and ethical to more quickly approve its upcoming A.I. technology.
    The company will also find ways for teams developing A.I. to conduct their own reviews, and it will “recalibrate” the level of risk it is willing to take when releasing the technology, according to the presentation.
    This change is in response to OpenAI's public release of ChatGPT. It is evidence that the race between Google/DeepMind and Microsoft/OpenAI is eroding ethics and safety.
    Demis Hassabis, CEO of DeepMind, urged caution in his recent interview in Time:
    He says AI is now “on the cusp” of being able to make tools that could be deeply damaging to human civilization, and urges his competitors to proceed with more caution than before.
    “When it comes to very powerful technologies—and obviously AI is going to be one of the most powerful ever—we need to be careful,” he says.
    “Not everybody is thinking about those things. It’s like experimentalists, many of whom don’t realize they’re holding dangerous material.”
    Worse still, Hassabis points out, we are the guinea pigs.
    Alphabet/Google is trying to accelerate a technology that its own subsidiary says is powerful and dangerous.
    Thanks for listening. To help us out with The Nonlinear Library or to learn more, please visit nonlinear.org.
    2 min
  • EA - Vegan Nutrition Testing Project: Interim Report by Elizabeth
    Welcome to The Nonlinear Library, where we use Text-to-Speech software to convert the best writing from the Rationalist and EA communities into audio. This is: Vegan Nutrition Testing Project: Interim Report, published by Elizabeth on January 21, 2023 on The Effective Altruism Forum.
    Thanks for listening. To help us out with The Nonlinear Library or to learn more, please visit nonlinear.org.
    1 min
  • LW - [Linkpost] TIME article: DeepMind’s CEO Helped Take AI Mainstream. Now He’s Urging Caution by Akash
    Welcome to The Nonlinear Library, where we use Text-to-Speech software to convert the best writing from the Rationalist and EA communities into audio. This is: [Linkpost] TIME article: DeepMind’s CEO Helped Take AI Mainstream. Now He’s Urging Caution, published by Akash on January 21, 2023 on LessWrong.
    TIME recently published an article about DeepMind, including some quotes from CEO Demis Hassabis. Posting some relevant quotes below (bolding added by me).
    You might also be interested in a recent interview with Sam Altman.
    Hassabis on safety:
    It is in this uncertain climate that Hassabis agrees to a rare interview, to issue a stark warning about his growing concerns. “I would advocate not moving fast and breaking things,” he says, referring to an old Facebook motto that encouraged engineers to release their technologies into the world first and fix any problems that arose later. The phrase has since become synonymous with disruption. That culture, subsequently emulated by a generation of startups, helped Facebook rocket to 3 billion users. But it also left the company entirely unprepared when disinformation, hate speech, and even incitement to genocide began appearing on its platform. Hassabis sees a similarly worrying trend developing with AI. He says AI is now “on the cusp” of being able to make tools that could be deeply damaging to human civilization, and urges his competitors to proceed with more caution than before.
    “When it comes to very powerful technologies—and obviously AI is going to be one of the most powerful ever—we need to be careful,” he says. “Not everybody is thinking about those things. It’s like experimentalists, many of whom don’t realize they’re holding dangerous material.” Worse still, Hassabis points out, we are the guinea pigs.
    The choice to be acquired by Google (as opposed to Facebook):
    The following year, Google acquired DeepMind for some $500 million. Hassabis turned down a bigger offer from Facebook. One reason, he says, was that, unlike Facebook, Google was “very happy to accept” DeepMind’s ethical red lines “as part of the acquisition.” (There were reports at the time that Google agreed to set up an independent ethics board to ensure these lines were not crossed.) The founders of the fledgling AI lab also reasoned that the megacorporation’s deep pockets would allow them access to talent and computing power that they otherwise couldn’t afford.
    Differences between DeepMind and OpenAI:
    While DeepMind, Google, and other AI labs had been working on similar research behind closed doors, OpenAI was more willing to let the public use its tools. In late 2022 it launched DALL·E 2, which can generate an image of almost any search term imaginable, and the chatbot ChatGPT.
    Hassabis on the decision to release Chinchilla:
    But despite Hassabis’s calls for the AI race to slow down, it appears DeepMind is not immune from the competitive pressures. In early 2022, the company published a blueprint for a faster engine. The piece of research, called Chinchilla, showed that many of the industry’s most cutting-edge models had been trained inefficiently, and explained how they could deliver more capability with the same level of computing power. Hassabis says DeepMind’s internal ethics board discussed whether releasing the research would be unethical given the risk that it could allow less scrupulous firms to release more powerful technologies without firm guardrails. One of the reasons they decided to publish it anyway was because “we weren’t the only people to know” about the phenomenon.
    Hassabis on concerns about information-sharing & freeloaders:
    He says that DeepMind is also considering releasing its own chatbot, called Sparrow, for a “private beta” some time in 2023. (The delay is in order for DeepMind to work on reinforcement learning-based features that ChatGPT lacks, like citing its sources. “It’s right to be cautious on that front,” Hassabis says.) But he admits that the company may soon need to change its ...
    7 min
  • EA - Available talent after major layoffs at tech giants by nicolenohemi
    Welcome to The Nonlinear Library, where we use Text-to-Speech software to convert the best writing from the Rationalist and EA communities into audio. This is: Available talent after major layoffs at tech giants, published by nicolenohemi on January 21, 2023 on The Effective Altruism Forum.
    Disclaimer: I am uncertain whether all of the relevant AI safety-related organizations this post might reach are net positive.
    Initially, I expected recruiting teams to already know about the layoffs discussed here, which is why I didn’t post about it sooner. When I asked some organizations about it, however, they were too swamped with other recruiting-related tasks, so none of them had heard about the layoffs (I didn’t yet get in touch with 80,000 hours). They were grateful for being connected with the available talent.
    Situation
    Amidst the volatile global economy, rising inflation, investor pressure, and increased borrowing costs, the highest number of job cuts in 2022 were ascribed to the tech sector, with almost 100,000 for the year, up by 649% compared to 2021 (Challenger, Gray & Christmas Inc.). Layoffs include engineers, ML researchers, executives, and more.
    List of layoffs
    Layoffs in numbers and (% of total employees)
    Meta: 11,000 (13%), among them, the Probability team
    Google: 12,000 (6%)
    Amazon: 18,000 (1.2%)
    Twitter 3,750 (50%)
    Stripe Inc. 980 (14%)
    Salesforce: 8,000 (10%)
    etc.
    Notable example: Meta’s Probability team
    Some of the Meta AI Probability team’s focus areas are differentiable and probabilistic programming and uncertainty, including increasing machine learning systems' robustness, trustworthiness, and efficiency by ascribing uncertainty measurements to models. Their whole team of about 50 employees was laid off. I had some casual chats with some of their team members, finding out that they were:
    already thinking about AI alignment,
    thought that “alignment definitely is important”,
    keen to start working in an AI alignment-related field, and
    happy to apply/be contacted by AI safety-related orgs and even requested to connect with their recruiting teams.
    Takeaway
    There is some seriously great tech talent that, sadly, has been laid off abruptly by the above-named tech giants. Thousands of individuals are now seeking new job opportunities. It could be beneficial for relevant organizations to explore this opportunity and consider contacting and encouraging some of them to apply for open positions.
    News sources
    NPR
    Bloomberg
    CNBC, CNBC
    Thanks for listening. To help us out with The Nonlinear Library or to learn more, please visit nonlinear.org.
    3 min
  • EA - [TIME magazine] DeepMind’s CEO Helped Take AI Mainstream. Now He’s Urging Caution (Perrigo, 2023) by will
    Welcome to The Nonlinear Library, where we use Text-to-Speech software to convert the best writing from the Rationalist and EA communities into audio. This is: [TIME magazine] DeepMind’s CEO Helped Take AI Mainstream. Now He’s Urging Caution (Perrigo, 2023), published by will on January 20, 2023 on The Effective Altruism Forum.
    Linkposting, tagging and excerpting in accord with 'Should pretty much all content that's EA-relevant and/or created by EAs be (link)posted to the Forum?'.
    He [Demis Hassibis] and his colleagues have been working toward a much grander ambition: creating artificial general intelligence, or AGI, by building machines that can think, learn, and be set to solve humanity’s toughest problems. Today’s AI is narrow, brittle, and often not very intelligent at all. But AGI, Hassabis believes, will be an “epoch-defining” technology—like the harnessing of electricity—that will change the very fabric of human life. If he’s right, it could earn him a place in history that would relegate the namesakes of his meeting rooms to mere footnotes.
    But with AI’s promise also comes peril. In recent months, researchers building an AI system to design new drugs revealed that their tool could be easily repurposed to make deadly new chemicals. A separate AI model trained to spew out toxic hate speech went viral, exemplifying the risk to vulnerable communities online. And inside AI labs around the world, policy experts were grappling with near-term questions like what to do when an AI has the potential to be commandeered by rogue states to mount widespread hacking campaigns or infer state-level nuclear secrets. In December 2022, ChatGPT, a chatbot designed by DeepMind’s rival OpenAI, went viral for its seeming ability to write almost like a human—but faced criticism for its susceptibility to racism and misinformation.
    It is in this uncertain climate that Hassabis agrees to a rare interview, to issue a stark warning about his growing concerns. “I would advocate not moving fast and breaking things,” he says, referring to an old Facebook motto that encouraged engineers to release their technologies into the world first and fix any problems that arose later. The phrase has since become synonymous with disruption. That culture, subsequently emulated by a generation of startups, helped Facebook rocket to 3 billion users. But it also left the company entirely unprepared when disinformation, hate speech, and even incitement to genocide began appearing on its platform. Hassabis sees a similarly worrying trend developing with AI. He says AI is now “on the cusp” of being able to make tools that could be deeply damaging to human civilization, and urges his competitors to proceed with more caution than before.
    “When it comes to very powerful technologies—and obviously AI is going to be one of the most powerful ever—we need to be careful,” he says. “Not everybody is thinking about those things. It’s like experimentalists, many of whom don’t realize they’re holding dangerous material.” Worse still, Hassabis points out, we are the guinea pigs.
    Thanks for listening. To help us out with The Nonlinear Library or to learn more, please visit nonlinear.org.
    4 min