In today’s episode, host Luisa Rodriguez interviews Paul Niehaus — cofounder of GiveDirectly — on the case for giving unconditional cash to the world’s poorest households.
They cover:
The empirical evidence on whether giving cash directly can drive meaningful economic growth
How the impacts of GiveDirectly compare to USAID employment programmes
GiveDirectly vs GiveWell’s top-recommended charities
How long-term guaranteed income affects people’s risk-taking and investments
Whether recipients prefer getting lump sums or monthly instalments
How GiveDirectly tackles cases of fraud and theft
The case for universal basic income, and GiveDirectly’s UBI studies in Kenya, Malawi, and Liberia
The political viability of UBI
Plenty more
Producer and editor: Keiran Harris Audio Engineering Lead: Ben Cordell Technical editing: Dominic Armstrong and Milo McGuire Additional content editing: Katy Moore and Luisa Rodriguez Transcriptions: Katy Moore
In today’s episode, host Rob Wiblin speaks with repeat guest Ian Morris about what big-picture history says about the likely impact of machine intelligence.
They cover:
Some crazy anomalies in the historical record of civilisational progress
Whether we should think about today’s technology from an evolutionary perspective
Whether war will make a resurgence
Why we can’t end up living like The Jetsons
Whether stagnation or cyclical futures are realistic
What it means that over the very long term the rate of economic growth has increased
Whether violence between humans and powerful AI systems is likely
The most likely reasons for Rob and Ian to be really wrong about all of this
How professional historians react to this sort of talk
The future of Ian’s work
Plenty more
Producer and editor: Keiran Harris Audio Engineering Lead: Ben Cordell Technical editing: Milo McGuire Transcriptions: Katy Moore
In today’s episode, host Luisa Rodriguez interviews Seren Kell — Senior Science and Technology Manager at the Good Food Institute Europe — about making alternative proteins as tasty, cheap, and convenient as traditional meat, dairy, and egg products.
They cover:
The basic case for alternative proteins, and why they’re so hard to make
Why fermentation is a surprisingly promising technology for creating delicious alternative proteins
The main scientific challenges that need to be solved to make fermentation even more useful
The progress that’s been made on the cultivated meat front, and what it will take to make cultivated meat affordable
How GFI Europe is helping with some of these challenges
How people can use their careers to contribute to replacing factory farming with alternative proteins
The best part of Seren’s job
Plenty more
Producer and editor: Keiran Harris Audio Engineering Lead: Ben Cordell Technical editing: Dominic Armstrong and Milo McGuire Additional content editing: Luisa Rodriguez and Katy Moore Transcriptions: Katy Moore
In today’s episode, host Rob Wiblin gets the rare chance to interview someone with insider AI policy experience at the White House and DeepMind who’s willing to speak openly — Tantum Collins.
They cover:
How AI could strengthen government capacity, and how that’s a double-edged sword
How new technologies force us to confront tradeoffs in political philosophy that we were previously able to pretend weren’t there
To what extent policymakers take different threats from AI seriously
Whether the US and China are in an AI arms race or not
Whether it’s OK to transform the world without much of the world agreeing to it
The tyranny of small differences in AI policy
Disagreements between different schools of thought in AI policy, and proposals that could unite them
How the US AI Bill of Rights could be improved
Whether AI will transform the labour market, and whether it will become a partisan political issue
The tensions between the cultures of San Francisco and DC, and how to bridge the divide between them
What listeners might be able to do to help with this whole mess
Panpsychism
Plenty more
Producer and editor: Keiran Harris Audio engineering lead: Ben Cordell Technical editing: Simon Monsour and Milo McGuire Transcriptions: Katy Moore
In today’s episode, host Rob Wiblin speaks with repeat guest and audience favourite Anders Sandberg about the most impressive things that could be achieved in our universe given the laws of physics.
They cover:
The epic new book Anders is working on, and whether he’ll ever finish it
Whether there’s a best possible world or we can just keep improving forever
What wars might look like if the galaxy is mostly settled
The impediments to AI or humans making it to other stars
How the universe will end a million trillion years in the future
Whether it’s useful to wonder about whether we’re living in a simulation
The grabby aliens theory
Whether civilizations get more likely to fail the older they get
The best way to generate energy that could ever exist
Black hole bombs
Whether superintelligence is necessary to get a lot of value
The likelihood that life from elsewhere has already visited Earth
And plenty more.
Producer and editor: Keiran Harris Audio Engineering Lead: Ben Cordell Technical editing: Simon Monsour and Milo McGuire Transcriptions: Katy Moore
In today’s episode, host Luisa Rodriguez interviews Kevin Esvelt — a biologist at the MIT Media Lab and the inventor of CRISPR-based gene drive — about the threat posed by engineered bioweapons.
They cover:
Why it makes sense to focus on deliberately released pandemics
Case studies of people who actually wanted to kill billions of humans
How many people have the technical ability to produce dangerous viruses
The different threats of stealth and wildfire pandemics that could crash civilisation
The potential for AI models to increase access to dangerous pathogens
Why scientists try to identify new pandemic-capable pathogens, and the case against that research
Technological solutions, including UV lights and advanced PPE
Using CRISPR-based gene drive to fight diseases and reduce animal suffering
And plenty more.
Producer and editor: Keiran Harris Audio Engineering Lead: Ben Cordell Technical editing: Simon Monsour Additional content editing: Katy Moore and Luisa Rodriguez Transcriptions: Katy Moore
Effective altruism is associated with the slogan “do the most good.” On one level, this has to be unobjectionable: What could be bad about helping people more and more?
But in today’s interview, Toby Ord — moral philosopher at the University of Oxford and one of the founding figures of effective altruism — lays out three reasons to be cautious about the idea of maximising the good that you do. He suggests that rather than “doing the most good that we can,” perhaps we should be happy with a more modest and manageable goal: “doing most of the good that we can.”
Toby was inspired to revisit these ideas by the possibility that Sam Bankman-Fried, who stands accused of committing severe fraud as CEO of the cryptocurrency exchange FTX, was motivated to break the law by a desire to give away as much money as possible to worthy causes.
Toby’s top reason not to fully maximise is the following: if the goal you’re aiming at is subtly wrong or incomplete, then going all the way towards maximising it will usually cause you to start doing some very harmful things.
This result can be shown mathematically, but can also be made intuitive, and may explain why we feel instinctively wary of going “all-in” on any idea, or goal, or way of living — even something as benign as helping other people as much as possible.
Toby gives the example of someone pursuing a career as a professional swimmer. Initially, as our swimmer takes their training and performance more seriously, they adjust their diet, hire a better trainer, and pay more attention to their technique. While swimming is the main focus of their life, they feel fit and healthy and also enjoy other aspects of their life as well — family, friends, and personal projects.
But if they decide to increase their commitment further and really go all-in on their swimming career, holding back nothing back, then this picture can radically change. Their effort was already substantial, so how can they shave those final few seconds off their racing time? The only remaining options are those which were so costly they were loath to consider them before.
To eke out those final gains — and go from 80% effort to 100% — our swimmer must sacrifice other hobbies, deprioritise their relationships, neglect their career, ignore food preferences, accept a higher risk of injury, and maybe even consider using steroids.
Now, if maximising one’s speed at swimming really were the only goal they ought to be pursuing, there’d be no problem with this. But if it’s the wrong goal, or only one of many things they should be aiming for, then the outcome is disastrous. In going from 80% to 100% effort, their swimming speed was only increased by a tiny amount, while everything else they were accomplishing dropped off a cliff.
The bottom line is simple: a dash of moderation makes you much more robust to uncertainty and error.
As Toby notes, this is similar to the observation that a sufficiently capable superintelligent AI, given any one goal, would ruin the world if it maximised it to the exclusion of everything else. And it follows a similar pattern to performance falling off a cliff when a statistical model is ‘overfit’ to its data.
In the full interview, Toby also explains the “moral trade” argument against pursuing narrow goals at the expense of everything else, and how consequentialism changes if you judge not just outcomes or acts, but everything according to its impacts on the world.
Toby and Rob also discuss:
The rise and fall of FTX and some of its impacts
What Toby hoped effective altruism would and wouldn’t become when he helped to get it off the ground
What utilitarianism has going for it, and what’s wrong with it in Toby’s view
How to mathematically model the importance of personal integrity
Which AI labs Toby thinks have been acting more responsibly than others
How having a young child affects Toby’s feelings about AI risk
Whether infinities present a fundamental problem for any theory of ethics that aspire to be fully impartial
How Toby ended up being the source of the highest quality images of the Earth from space
Get this episode by subscribing to our podcast on the world’s most pressing problems and how to solve them: type ‘80,000 Hours’ into your podcasting app. Or read the transcript.
Producer and editor: Keiran Harris Audio Engineering Lead: Ben Cordell Technical editing: Simon Monsour Transcriptions: Katy Moore
Mustafa Suleyman was part of the trio that founded DeepMind, and his new AI project is building one of the world’s largest supercomputers to train a large language model on 10–100x the compute used to train ChatGPT.
But far from the stereotype of the incorrigibly optimistic tech founder, Mustafa is deeply worried about the future, for reasons he lays out in his new book The Coming Wave: Technology, Power, and the 21st Century’s Greatest Dilemma (coauthored with Michael Bhaskar). The future could be really good, but only if we grab the bull by the horns and solve the new problems technology is throwing at us.
On Mustafa’s telling, AI and biotechnology will soon be a huge aid to criminals and terrorists, empowering small groups to cause harm on previously unimaginable scales. Democratic countries have learned to walk a ‘narrow path’ between chaos on the one hand and authoritarianism on the other, avoiding the downsides that come from both extreme openness and extreme closure. AI could easily destabilise that present equilibrium, throwing us off dangerously in either direction. And ultimately, within our lifetimes humans may not need to work to live any more — or indeed, even have the option to do so.
And those are just three of the challenges confronting us. In Mustafa’s view, ‘misaligned’ AI that goes rogue and pursues its own agenda won’t be an issue for the next few years, and it isn’t a problem for the current style of large language models. But he thinks that at some point — in eight, ten, or twelve years — it will become an entirely legitimate concern, and says that we need to be planning ahead.
In The Coming Wave, Mustafa lays out a 10-part agenda for ‘containment’ — that is to say, for limiting the negative and unforeseen consequences of emerging technologies:
Developing an Apollo programme for technical AI safety
Instituting capability audits for AI models
Buying time by exploiting hardware choke points
Getting critics involved in directly engineering AI models
Getting AI labs to be guided by motives other than profit
Radically increasing governments’ understanding of AI and their capabilities to sensibly regulate it
Creating international treaties to prevent proliferation of the most dangerous AI capabilities
Building a self-critical culture in AI labs of openly accepting when the status quo isn’t working
Creating a mass public movement that understands AI and can demand the necessary controls
Not relying too much on delay, but instead seeking to move into a new somewhat-stable equilibria
As Mustafa put it, “AI is a technology with almost every use case imaginable” and that will demand that, in time, we rethink everything.
Rob and Mustafa discuss the above, as well as:
Whether we should be open sourcing AI models
Whether Mustafa’s policy views are consistent with his timelines for transformative AI
How people with very different views on these issues get along at AI labs
The failed efforts (so far) to get a wider range of people involved in these decisions
Whether it’s dangerous for Mustafa’s new company to be training far larger models than GPT-4
Whether we’ll be blown away by AI progress over the next year
What mandatory regulations government should be imposing on AI labs right now
Appropriate priorities for the UK’s upcoming AI safety summit
Get this episode by subscribing to our podcast on the world’s most pressing problems and how to solve them: type ‘80,000 Hours’ into your podcasting app. Or read the transcript.
Producer and editor: Keiran Harris Audio Engineering Lead: Ben Cordell Technical editing: Milo McGuire Transcriptions: Katy Moore
In today’s episode, host Luisa Rodriguez interviews economist Michael Webb of DeepMind, the British Government, and Stanford about how AI progress is going to affect people’s jobs and the labour market.
They cover:
The jobs most and least exposed to AI
Whether we’ll we see mass unemployment in the short term
How long it took other technologies like electricity and computers to have economy-wide effects
Whether AI will increase or decrease inequality
Whether AI will lead to explosive economic growth
What we can we learn from history, and reasons to think this time is different
Career advice for a world of LLMs
Why Michael is starting a new org to relieve talent bottlenecks through accelerated learning, and how you can get involved
Michael’s take as a musician on AI-generated music
And plenty more
If you’d like to work with Michael on his new org to radically accelerate how quickly people acquire expertise in critical cause areas, he’s now hiring! Check out Quantum Leap’s website.
Get this episode by subscribing to our podcast on the world’s most pressing problems and how to solve them: type ‘80,000 Hours’ into your podcasting app. Or read the transcript.
Producer and editor: Keiran Harris Audio Engineering Lead: Ben Cordell Technical editing: Simon Monsour and Milo McGuire Additional content editing: Katy Moore and Luisa Rodriguez Transcriptions: Katy Moore
In today’s episode, host Luisa Rodriguez interviews the head of research at Our World in Data — Hannah Ritchie — on the case for environmental optimism.
They cover:
Why agricultural productivity in sub-Saharan Africa could be so important, and how much better things could get
Her new book about how we could be the first generation to build a sustainable planet
Whether climate change is the most worrying environmental issue
How we reduced outdoor air pollution
Why Hannah is worried about the state of biodiversity
Solutions that address multiple environmental issues at once
How the world coordinated to address the hole in the ozone layer
Surprises from Our World in Data’s research
Psychological challenges that come up in Hannah’s work
And plenty more
Get this episode by subscribing to our podcast on the world’s most pressing problems and how to solve them: type ‘80,000 Hours’ into your podcasting app. Or read the transcript.
Producer and editor: Keiran Harris Audio Engineering Lead: Ben Cordell Technical editing: Milo McGuire and Dominic Armstrong Additional content editing: Katy Moore and Luisa Rodriguez Transcriptions: Katy Moore
In July, OpenAI announced a new team and project: Superalignment. The goal is to figure out how to make superintelligent AI systems aligned and safe to use within four years, and the lab is putting a massive 20% of its computational resources behind the effort.
Today’s guest, Jan Leike, is Head of Alignment at OpenAI and will be co-leading the project. As OpenAI puts it, “…the vast power of superintelligence could be very dangerous, and lead to the disempowerment of humanity or even human extinction. … Currently, we don’t have a solution for steering or controlling a potentially superintelligent AI, and preventing it from going rogue.”
Given that OpenAI is in the business of developing superintelligent AI, it sees that as a scary problem that urgently has to be fixed. So it’s not just throwing compute at the problem — it’s also hiring dozens of scientists and engineers to build out the Superalignment team.
Plenty of people are pessimistic that this can be done at all, let alone in four years. But Jan is guardedly optimistic. As he explains:
Honestly, it really feels like we have a real angle of attack on the problem that we can actually iterate on… and I think it’s pretty likely going to work, actually. And that’s really, really wild, and it’s really exciting. It’s like we have this hard problem that we’ve been talking about for years and years and years, and now we have a real shot at actually solving it. And that’d be so good if we did.
Jan thinks that this work is actually the most scientifically interesting part of machine learning. Rather than just throwing more chips and more data at a training run, this work requires actually understanding how these models work and how they think. The answers are likely to be breakthroughs on the level of solving the mysteries of the human brain.
The plan, in a nutshell, is to get AI to help us solve alignment. That might sound a bit crazy — as one person described it, “like using one fire to put out another fire.”
But Jan’s thinking is this: the core problem is that AI capabilities will keep getting better and the challenge of monitoring cutting-edge models will keep getting harder, while human intelligence stays more or less the same. To have any hope of ensuring safety, we need our ability to monitor, understand, and design ML models to advance at the same pace as the complexity of the models themselves.
And there’s an obvious way to do that: get AI to do most of the work, such that the sophistication of the AIs that need aligning, and the sophistication of the AIs doing the aligning, advance in lockstep.
Jan doesn’t want to produce machine learning models capable of doing ML research. But such models are coming, whether we like it or not. And at that point Jan wants to make sure we turn them towards useful alignment and safety work, as much or more than we use them to advance AI capabilities.
Jan thinks it’s so crazy it just might work. But some critics think it’s simply crazy. They ask a wide range of difficult questions, including:
If you don’t know how to solve alignment, how can you tell that your alignment assistant AIs are actually acting in your interest rather than working against you? Especially as they could just be pretending to care about what you care about.
How do you know that these technical problems can be solved at all, even in principle?
At the point that models are able to help with alignment, won’t they also be so good at improving capabilities that we’re in the middle of an explosion in what AI can do?
In today’s interview host Rob Wiblin puts these doubts to Jan to hear how he responds to each, and they also cover:
OpenAI’s current plans to achieve ‘superalignment’ and the reasoning behind them
Why alignment work is the most fundamental and scientifically interesting research in ML
The kinds of people he’s excited to hire to join his team and maybe save the world
What most readers misunderstood about the OpenAI announcement
The three ways Jan expects AI to help solve alignment: mechanistic interpretability, generalization, and scalable oversight
What the standard should be for confirming whether Jan’s team has succeeded
Whether OpenAI should (or will) commit to stop training more powerful general models if they don’t think the alignment problem has been solved
Whether Jan thinks OpenAI has deployed models too quickly or too slowly
The many other actors who also have to do their jobs really well if we’re going to have a good AI future
Plenty more
Get this episode by subscribing to our podcast on the world’s most pressing problems and how to solve them: type ‘80,000 Hours’ into your podcasting app. Or read the transcript below.
Producer and editor: Keiran Harris Audio Engineering Lead: Ben Cordell Technical editing: Simon Monsour and Milo McGuire Additional content editing: Katy Moore and Luisa Rodriguez Transcriptions: Katy Moore
Back in 2007, Holden Karnofsky cofounded GiveWell, where he sought out the charities that most cost-effectively helped save lives. He then cofounded Open Philanthropy, where he oversaw a team making billions of dollars’ worth of grants across a range of areas: pandemic control, criminal justice reform, farmed animal welfare, and making AI safe, among others. This year, having learned about AI for years and observed recent events, he’s narrowing his focus once again, this time on making the transition to advanced AI go well.
In today’s conversation, Holden returns to the show to share his overall understanding of the promise and the risks posed by machine intelligence, and what to do about it. That understanding has accumulated over around 14 years, during which he went from being sceptical that AI was important or risky, to making AI risks the focus of his work.
(As Holden reminds us, his wife is also the president of one of the world’s top AI labs, Anthropic, giving him both conflicts of interest and a front-row seat to recent events. For our part, Open Philanthropy is 80,000 Hours’ largest financial supporter.)
One point he makes is that people are too narrowly focused on AI becoming ‘superintelligent.’ While that could happen and would be important, it’s not necessary for AI to be transformative or perilous. Rather, machines with human levels of intelligence could end up being enormously influential simply if the amount of computer hardware globally were able to operate tens or hundreds of billions of them, in a sense making machine intelligences a majority of the global population, or at least a majority of global thought.
As Holden explains, he sees four key parts to the playbook humanity should use to guide the transition to very advanced AI in a positive direction: alignment research, standards and monitoring, creating a successful and careful AI lab, and finally, information security.
In today’s episode, host Rob Wiblin interviews return guest Holden Karnofsky about that playbook, as well as:
Why we can’t rely on just gradually solving those problems as they come up, the way we usually do with new technologies.
What multiple different groups can do to improve our chances of a good outcome — including listeners to this show, governments, computer security experts, and journalists.
Holden’s case against ‘hardcore utilitarianism’ and what actually motivates him to work hard for a better world.
What the ML and AI safety communities get wrong in Holden’s view.
Ways we might succeed with AI just by dumb luck.
The value of laying out imaginable success stories.
Why information security is so important and underrated.
Whether it’s good to work at an AI lab that you think is particularly careful.
The track record of futurists’ predictions.
And much more.
Get this episode by subscribing to our podcast on the world’s most pressing problems and how to solve them: type ‘80,000 Hours’ into your podcasting app. Or read the transcript below.
Producer: Keiran Harris Audio Engineering Lead: Ben Cordell Technical editing: Simon Monsour and Milo McGuire Transcriptions: Katy Moore
In Oppenheimer, scientists detonate a nuclear weapon despite thinking there’s some ‘near zero’ chance it would ignite the atmosphere, putting an end to life on Earth. Today, scientists working on AI think the chance their work puts an end to humanity is vastly higher than that.
In response, some have suggested we launch a Manhattan Project to make AI safe via enormous investment in relevant R&D. Others have suggested that we need international organisations modelled on those that slowed the proliferation of nuclear weapons. Others still seek a research slowdown by labs while an auditing and licencing scheme is created.
Today’s guest — journalist Ezra Klein of The New York Times — has watched policy discussions and legislative battles play out in DC for 20 years. Like many people he has also taken a big interest in AI this year, writing articles such as “This changes everything.” In his first interview on the show in 2021, he flagged AI as one topic that DC would regret not having paid more attention to.
So we invited him on to get his take on which regulatory proposals have promise, and which seem either unhelpful or politically unviable.
Out of the ideas on the table right now, Ezra favours a focus on direct government funding — both for AI safety research and to develop AI models designed to solve problems other than making money for their operators. He is sympathetic to legislation that would require AI models to be legible in a way that none currently are — and embraces the fact that that will slow down the release of models while businesses figure out how their products actually work.
By contrast, he’s pessimistic that it’s possible to coordinate countries around the world to agree to prevent or delay the deployment of dangerous AI models — at least not unless there’s some spectacular AI-related disaster to create such a consensus. And he fears attempts to require licences to train the most powerful ML models will struggle unless they can find a way to exclude and thereby appease people working on relatively safe consumer technologies rather than cutting-edge research.
From observing how DC works, Ezra expects that even a small community of experts in AI governance can have a large influence on how the the US government responds to AI advances. But in Ezra’s view, that requires those experts to move to DC and spend years building relationships with people in government, rather than clustering elsewhere in academia and AI labs.
In today’s brisk conversation, Ezra and host Rob Wiblin cover the above as well as:
Whether it’s desirable to slow down AI research
The value of engaging with current policy debates even if they don’t seem directly important
Which AI business models seem more or less dangerous
Tensions between people focused on existing vs emergent risks from AI
Two major challenges of being a new parent
Get this episode by subscribing to our podcast on the world’s most pressing problems and how to solve them: type ‘80,000 Hours’ into your podcasting app. Or read the transcript below.
Producer: Keiran Harris Audio Engineering Lead: Ben Cordell Technical editing: Milo McGuire Transcriptions: Katy Moore
In today’s episode, host Luisa Rodriguez interviews the Head of Policy at the Centre for the Governance of AI — Markus Anderljung — about all aspects of policy and governance of superhuman AI systems.
They cover:
The need for AI governance, including self-replicating models and ChaosGPT
Whether or not AI companies will willingly accept regulation
The key regulatory strategies including licencing, risk assessment, auditing, and post-deployment monitoring
Whether we can be confident that people won’t train models covertly and ignore the licencing system
The progress we’ve made so far in AI governance
The key weaknesses of these approaches
The need for external scrutiny of powerful models
The emergent capabilities problem
Why it really matters where regulation happens
Advice for people wanting to pursue a career in this field
And much more.
Get this episode by subscribing to our podcast on the world’s most pressing problems and how to solve them: type ‘80,000 Hours’ into your podcasting app. Or read the transcript below.
Producer: Keiran Harris Audio Engineering Lead: Ben Cordell Technical editing: Simon Monsour and Milo McGuire Transcriptions: Katy Moore
As AI advances ever more quickly, concerns about potential misuse of highly capable models are growing. From hostile foreign governments and terrorists to reckless entrepreneurs, the threat of AI falling into the wrong hands is top of mind for the national security community.
With growing concerns about the use of AI in military applications, the US has banned the export of certain types of chips to China.
But unlike the uranium required to make nuclear weapons, or the material inputs to a bioweapons programme, computer chips and machine learning models are absolutely everywhere. So is it actually possible to keep dangerous capabilities out of the wrong hands?
In today’s interview, Lennart Heim — who researches compute governance at the Centre for the Governance of AI — explains why limiting access to supercomputers may represent our best shot.
As Lennart explains, an AI research project requires many inputs, including the classic triad of compute, algorithms, and data.
If we want to limit access to the most advanced AI models, focusing on access to supercomputing resources — usually called ‘compute’ — might be the way to go. Both algorithms and data are hard to control because they live on hard drives and can be easily copied. By contrast, advanced chips are physical items that can’t be used by multiple people at once and come from a small number of sources.
According to Lennart, the hope would be to enforce AI safety regulations by controlling access to the most advanced chips specialised for AI applications. For instance, projects training ‘frontier’ AI models — the newest and most capable models — might only gain access to the supercomputers they need if they obtain a licence and follow industry best practices.
We have similar safety rules for companies that fly planes or manufacture volatile chemicals — so why not for people producing the most powerful and perhaps the most dangerous technology humanity has ever played with?
But Lennart is quick to note that the approach faces many practical challenges. Currently, AI chips are readily available and untracked. Changing that will require the collaboration of many actors, which might be difficult, especially given that some of them aren’t convinced of the seriousness of the problem.
Host Rob Wiblin is particularly concerned about a different challenge: the increasing efficiency of AI training algorithms. As these algorithms become more efficient, what once required a specialised AI supercomputer to train might soon be achievable with a home computer.
By that point, tracking every aggregation of compute that could prove to be very dangerous would be both impractical and invasive.
With only a decade or two left before that becomes a reality, the window during which compute governance is a viable solution may be a brief one. Top AI labs have already stopped publishing their latest algorithms, which might extend this ‘compute governance era’, but not for very long.
If compute governance is only a temporary phase between the era of difficult-to-train superhuman AI models and the time when such models are widely accessible, what can we do to prevent misuse of AI systems after that point?
Lennart and Rob both think the only enduring approach requires taking advantage of the AI capabilities that should be in the hands of police and governments — which will hopefully remain superior to those held by criminals, terrorists, or fools. But as they describe, this means maintaining a peaceful standoff between AI models with conflicting goals that can act and fight with one another on the microsecond timescale. Being far too slow to follow what’s happening — let alone participate — humans would have to be cut out of any defensive decision-making.
Both agree that while this may be our best option, such a vision of the future is more terrifying than reassuring.
Lennart and Rob discuss the above as well as:
How can we best categorise all the ways AI could go wrong?
Why did the US restrict the export of some chips to China and what impact has that had?
Is the US in an ‘arms race’ with China or is that more an illusion?
What is the deal with chips specialised for AI applications?
How is the ‘compute’ industry organised?
Downsides of using compute as a target for regulations
Could safety mechanisms be built into computer chips themselves?
Who would have the legal authority to govern compute if some disaster made it seem necessary?
The reasons Rob doubts that any of this stuff will work
Could AI be trained to operate as a far more severe computer worm than any we’ve seen before?
What does the world look like when sluggish human reaction times leave us completely outclassed?
And plenty more
Get this episode by subscribing to our podcast on the world’s most pressing problems and how to solve them: type ‘80,000 Hours’ into your podcasting app. Or read the transcript below.
Producer: Keiran Harris Audio mastering: Milo McGuire, Dominic Armstrong, and Ben Cordell Transcriptions: Katy Moore
Can there be a more exciting and strange place to work today than a leading AI lab? Your CEO has said they’re worried your research could cause human extinction. The government is setting up meetings to discuss how this outcome can be avoided. Some of your colleagues think this is all overblown; others are more anxious still.
Today’s guest — machine learning researcher Rohin Shah — goes into the Google DeepMind offices each day with that peculiar backdrop to his work.
He’s on the team dedicated to maintaining ‘technical AI safety’ as these models approach and exceed human capabilities: basically that the models help humanity accomplish its goals without flipping out in some dangerous way. This work has never seemed more important.
In the short-term it could be the key bottleneck to deploying ML models in high-stakes real-life situations. In the long-term, it could be the difference between humanity thriving and disappearing entirely.
For years Rohin has been on a mission to fairly hear out people across the full spectrum of opinion about risks from artificial intelligence — from doomers to doubters — and properly understand their point of view. That makes him unusually well placed to give an overview of what we do and don’t understand. He has landed somewhere in the middle — troubled by ways things could go wrong, but not convinced there are very strong reasons to expect a terrible outcome.
Today’s conversation is wide-ranging and Rohin lays out many of his personal opinions to host Rob Wiblin, including:
What he sees as the strongest case both for and against slowing down the rate of progress in AI research.
Why he disagrees with most other ML researchers that training a model on a sensible ‘reward function’ is enough to get a good outcome.
Why he disagrees with many on LessWrong that the bar for whether a safety technique is helpful is “could this contain a superintelligence.”
That he thinks nobody has very compelling arguments that AI created via machine learning will be dangerous by default, or that it will be safe by default. He believes we just don’t know.
That he understands that analogies and visualisations are necessary for public communication, but is sceptical that they really help us understand what’s going on with ML models, because they’re different in important ways from every other case we might compare them to.
Why he’s optimistic about DeepMind’s work on scalable oversight, mechanistic interpretability, and dangerous capabilities evaluations, and what each of those projects involves.
Why he isn’t inherently worried about a future where we’re surrounded by beings far more capable than us, so long as they share our goals to a reasonable degree.
Why it’s not enough for humanity to know how to align AI models — it’s essential that management at AI labs correctly pick which methods they’re going to use and have the practical know-how to apply them properly.
Three observations that make him a little more optimistic: humans are a bit muddle-headed and not super goal-orientated; planes don’t crash; and universities have specific majors in particular subjects.
Plenty more besides.
Get this episode by subscribing to our podcast on the world’s most pressing problems and how to solve them: type ‘80,000 Hours’ into your podcasting app. Or read the transcript below.
Producer: Keiran Harris Audio mastering: Milo McGuire, Dominic Armstrong, and Ben Cordell Transcriptions: Katy Moore
GiveWell is one of the world’s best-known charity evaluators, with the goal of “searching for the charities that save or improve lives the most per dollar.” It mostly recommends projects that help the world’s poorest people avoid easily prevented diseases, like intestinal worms or vitamin A deficiency.
But should GiveWell, as some critics argue, take a totally different approach to its search, focusing instead on directly increasing subjective wellbeing, or alternatively, raising economic growth?
Today’s guest — cofounder and CEO of GiveWell, Elie Hassenfeld — is proud of how much GiveWell has grown in the last five years. Its ‘money moved’ has quadrupled to around $600 million a year.
Its research team has also more than doubled, enabling them to investigate a far broader range of interventions that could plausibly help people an enormous amount for each dollar spent. That work has led GiveWell to support dozens of new organisations, such as Kangaroo Mother Care, MiracleFeet, and Dispensers for Safe Water.
But some other researchers focused on figuring out the best ways to help the world’s poorest people say GiveWell shouldn’t just do more of the same thing, but rather ought to look at the problem differently.
Currently, GiveWell uses a range of metrics to track the impact of the organisations it considers recommending — such as ‘lives saved,’ ‘household incomes doubled,’ and for health improvements, the ‘quality-adjusted life year.’ To compare across opportunities, it then needs some way of weighing these different types of benefits up against one another. This requires estimating so-called “moral weights,” which Elie agrees is far from the most mature part of the project.
The Happier Lives Institute (HLI) has argued that instead, GiveWell should try to cash out the impact of all interventions in terms of improvements in subjective wellbeing. According to HLI, it’s improvements in wellbeing and reductions in suffering that are the true ultimate goal of all projects, and if you quantify everyone on this same scale, using some measure like the wellbeing-adjusted life year (WELLBY), you have an easier time comparing them.
This philosophy has led HLI to be more sceptical of interventions that have been demonstrated to improve health, but whose impact on wellbeing has not been measured, and to give a high priority to improving lives relative to extending them.
An alternative high-level critique is that really all that matters in the long run is getting the economies of poor countries to grow. According to this line of argument, hundreds of millions fewer people live in poverty in China today than 50 years ago, but is that because of the delivery of basic health treatments? Maybe a little), but mostly not.
Rather, it’s because changes in economic policy and governance in China allowed it to experience a 10% rate of economic growth for several decades. That led to much higher individual incomes and meant the country could easily afford all the basic health treatments GiveWell might otherwise want to fund, and much more besides.
On this view, GiveWell should focus on figuring out what causes some countries to experience explosive economic growth while others fail to, or even go backwards. Even modest improvements in the chances of such a ‘growth miracle’ will likely offer a bigger bang-for-buck than funding the incremental delivery of deworming tablets or vitamin A supplements, or anything else.
Elie sees where both of these critiques are coming from, and notes that they’ve influenced GiveWell’s work in some ways. But as he explains, he thinks they underestimate the practical difficulty of successfully pulling off either approach and finding better opportunities than what GiveWell funds today.
In today’s in-depth conversation, Elie and host Rob Wiblin cover the above, as well as:
The research that caused GiveWell to flip from not recommending chlorine dispensers as an intervention for safe drinking water to spending tens of millions of dollars on them.
What transferable lessons GiveWell learned from investigating different kinds of interventions, like providing medical expertise to hospitals in very poor countries to help them improve their practices.
Why the best treatment for premature babies in low-resource settings may involve less rather than more medicine.
The high prevalence of severe malnourishment among children and what can be done about it.
How to deal with hidden and non-obvious costs of a programme, like taking up a hospital room that might otherwise have been used for something else.
Some cheap early treatments that can prevent kids from developing lifelong disabilities, which GiveWell funds.
The various roles GiveWell is currently hiring for, and what’s distinctive about their organisational culture.
Get this episode by subscribing to our podcast on the world’s most pressing problems and how to solve them: type ‘80,000 Hours’ into your podcasting app. Or read the transcript below.
Producer: Keiran Harris Audio mastering: Simon Monsour and Ben Cordell Transcriptions: Katy Moore
What is the nature of the universe? How do we make decisions correctly? What differentiates right actions from wrong ones?
Such fundamental questions have been the subject of philosophical and theological debates for millennia. But, as we all know, and surveys of expert opinion make clear, we are very far from agreement. So… with these most basic questions unresolved, what’s a species to do?
In today’s episode, philosopher Joe Carlsmith — Senior Research Analyst at Open Philanthropy — makes the case that many current debates in philosophy ought to leave us confused and humbled. These are themes he discusses in his PhD thesis, A stranger priority? Topics at the outer reaches of effective altruism.
To help transmit the disorientation he thinks is appropriate, Joe presents three disconcerting theories — originating from him and his peers — that challenge humanity’s self-assured understanding of the world.
The first idea is that we might be living in a computer simulation, because, in the classic formulation, if most civilisations go on to run many computer simulations of their past history, then most beings who perceive themselves as living in such a history must themselves be in computer simulations. Joe prefers a somewhat different way of making the point, but, having looked into it, he hasn’t identified any particular rebuttal to this ‘simulation argument.’
If true, it could revolutionise our comprehension of the universe and the way we ought to live.
The second is the idea that “you can ‘control’ events you have no causal interaction with, including events in the past.” The thought experiment that most persuades him of this is the following:
Perfect deterministic twin prisoner’s dilemma: You’re a deterministic AI system, who only wants money for yourself (you don’t care about copies of yourself). The authorities make a perfect copy of you, separate you and your copy by a large distance, and then expose you both, in simulation, to exactly identical inputs (let’s say, a room, a whiteboard, some markers, etc.). You both face the following choice: either (a) send a million dollars to the other (“cooperate”), or (b) take a thousand dollars for yourself (“defect”).
Joe thinks, in contrast with the dominant theory of correct decision-making, that it’s clear you should send a million dollars to your twin. But as he explains, this idea, when extrapolated outwards to other cases, implies that it could be sensible to take actions in the hope that they’ll improve parallel universes you can never causally interact with — or even to improve the past. That is nuts by anyone’s lights, including Joe’s.
The third disorienting idea is that, as far as we can tell, the universe could be infinitely large. And that fact, if true, would mean we probably have to make choices between actions and outcomes that involve infinities. Unfortunately, doing that breaks our existing ethical systems, which are only designed to accommodate finite cases.
In an infinite universe, our standard models end up unable to say much at all, or give the wrong answers entirely. While we might hope to patch them in straightforward ways, having looked into ways we might do that, Joe has concluded they all quickly get complicated and arbitrary, and still have to do enormous violence to our common sense. For people inclined to endorse some flavour of utilitarianism, Joe thinks ‘infinite ethics’ spell the end of the ‘utilitarian dream‘ of a moral philosophy that has the virtue of being very simple while still matching our intuitions in most cases.
These are just three particular instances of a much broader set of ideas that some have dubbed the “train to crazy town.” Basically, if you commit to always take philosophy and arguments seriously, and try to act on them, it can lead to what seem like some pretty crazy and impractical places. So what should we do with this buffet of plausible-sounding but bewildering arguments?
Joe and Rob discuss to what extent this should prompt us to pay less attention to philosophy, and how we as individuals can cope psychologically with feeling out of our depth just trying to make the most basic sense of the world.
In the face of all of this, Joe suggests that there is a promising and robust path for humanity to take: keep our options open and put our descendants in a better position to figure out the answers to questions that seem impossible for us to resolve today — a position he calls “wisdom longtermism.”
Joe fears that if people believe we understand the universe better than we really do, they’ll be more likely to try to commit humanity to a particular vision of the future, or be uncooperative to others, in ways that only make sense if you were certain you knew what was right and wrong.
In today’s challenging conversation, Joe and Rob discuss all of the above, as well as:
What Joe doesn’t like about the drowning child thought experiment
An alternative thought experiment about helping a stranger that might better highlight our intrinsic desire to help others
What Joe doesn’t like about the expression “the train to crazy town”
Whether Elon Musk should place a higher probability on living in a simulation than most other people
Whether the deterministic twin prisoner’s dilemma, if fully appreciated, gives us an extra reason to keep promises
To what extent learning to doubt our own judgement about difficult questions — so-called “epistemic learned helplessness” — is a good thing
How strong the case is that advanced AI will engage in generalised power-seeking behaviour
Get this episode by subscribing to our podcast on the world’s most pressing problems and how to solve them: type ‘80,000 Hours’ into your podcasting app. Or read the transcript below.
Producer: Keiran Harris Audio mastering: Milo McGuire and Ben Cordell Transcriptions: Katy Moore
Imagine you are an orphaned eight-year-old whose parents left you a $1 trillion company, and no trusted adult to serve as your guide to the world. You have to hire a smart adult to run that company, guide your life the way that a parent would, and administer your vast wealth. You have to hire that adult based on a work trial or interview you come up with. You don’t get to see any resumes or do reference checks. And because you’re so rich, tonnes of people apply for the job — for all sorts of reasons.
Today’s guest Ajeya Cotra — senior research analyst at Open Philanthropy — argues that this peculiar setup resembles the situation humanity finds itself in when training very general and very capable AI models using current deep learning methods.
As she explains, such an eight-year-old faces a challenging problem. In the candidate pool there are likely some truly nice people, who sincerely want to help and make decisions that are in your interest. But there are probably other characters too — like people who will pretend to care about you while you’re monitoring them, but intend to use the job to enrich themselves as soon as they think they can get away with it.
Like a child trying to judge adults, at some point humans will be required to judge the trustworthiness and reliability of machine learning models that are as goal-oriented as people, and greatly outclass them in knowledge, experience, breadth, and speed. Tricky!
Can’t we rely on how well models have performed at tasks during training to guide us? Ajeya worries that it won’t work. The trouble is that three different sorts of models will all produce the same output during training, but could behave very differently once deployed in a setting that allows their true colours to come through. She describes three such motivational archetypes:
Saints — models that care about doing what we really want
Sycophants — models that just want us to say they’ve done a good job, even if they get that praise by taking actions they know we wouldn’t want them to
Schemers — models that don’t care about us or our interests at all, who are just pleasing us so long as that serves their own agenda
In principle, a machine learning training process based on reinforcement learning could spit out any of these three attitudes, because all three would perform roughly equally well on the tests we give them, and ‘performs well on tests’ is how these models are selected.
But while that’s true in principle, maybe it’s not something that could plausibly happen in the real world. After all, if we train an agent based on positive reinforcement for accomplishing X, shouldn’t the training process spit out a model that plainly does X and doesn’t have complex thoughts and goals beyond that?
According to Ajeya, this is one thing we don’t know, and should be trying to test empirically as these models get more capable. For reasons she explains in the interview, the Sycophant or Schemer models may in fact be simpler and easier for the learning algorithm to creep towards than their Saint counterparts.
But there are also ways we could end up actively selecting for motivations that we don’t want.
For a toy example, let’s say you train an agent AI model to run a small business, and select it for behaviours that make money, measuring its success by whether it manages to get more money in its bank account. During training, a highly capable model may experiment with the strategy of tricking its raters into thinking it has made money legitimately when it hasn’t. Maybe instead it steals some money and covers that up. This isn’t exactly unlikely; during training, models often come up with creative — sometimes undesirable — approaches that their developers didn’t anticipate.
If such deception isn’t picked up, a model like this may be rated as particularly successful, and the training process will cause it to develop a progressively stronger tendency to engage in such deceptive behaviour. A model that has the option to engage in deception when it won’t be detected would, in effect, have a competitive advantage.
What if deception is picked up, but just some of the time? Would the model then learn that honesty is the best policy? Maybe. But alternatively, it might learn the ‘lesson’ that deception does pay, but you just have to do it selectively and carefully, so it can’t be discovered. Would that actually happen? We don’t yet know, but it’s possible.
In today’s interview, Ajeya and Rob discuss the above, as well as:
How to predict the motivations a neural network will develop through training
Whether AIs being trained will functionally understand that they’re AIs being trained, the same way we think we understand that we’re humans living on planet Earth
Stories of AI misalignment that Ajeya doesn’t buy into
Analogies for AI, from octopuses to aliens to can openers
Why it’s smarter to have separate planning AIs and doing AIs
The benefits of only following through on AI-generated plans that make sense to human beings
What approaches for fixing alignment problems Ajeya is most excited about, and which she thinks are overrated
How one might demo actually scary AI failure mechanisms
Get this episode by subscribing to our podcast on the world’s most pressing problems and how to solve them: type ‘80,000 Hours’ into your podcasting app. Or read the transcript below.
Producer: Keiran Harris Audio mastering: Ryan Kessler and Ben Cordell Transcriptions: Katy Moore
