Imagining a better future ≈55m

🎬 Watch the embedded video

You’re the product of 8,000 generations of humans. You were born into the civilization they built, with all its magic, beauty and flaws. Their decisions shape every aspect of our world.

We’re only the latest generation of humans, but we’re living through one of the most significant technological transformations in the history of the universe.

Our decisions today will have an immense impact on our own future, and the future of all our descendants.

We may still be early in humanity’s story. We have an amazing opportunity to steer this story towards wonder, greatness and kindness. But that's not guaranteed.

In this unit, you’ll visualise a future to aspire towards, you’ll explore how AI is shaping that future, and you’ll analyse the dilemmas and coordination failures we must overcome to steer towards a better future.

Throughout the rest of the course, you will:

analyse the technical trends of AI progress and what this implies for future AI capabilities,

develop a step-by-step story (”kill chain”) for how those capabilities could cause harm to humanity,

apply a “defence in depth” framework to analyse what needs to be built to defend against your kill chain and steer AI towards safe and beneficial outcomes,

generate your own action plan for how you'll start contributing to making AI go well.

What future do you want?

Exercise — What does a better future for yourself look like?.

In so many ways, human lives have been transformed in the past few centuries. In many ways, perhaps most, it’s been for the better. However, there’s a huge amount of work left to be done to steer our future towards wiser, kinder, happier outcomes.

How old are you? Add 20 years to that number. Imagine a week in your life when you’re that years old, where you’re living a good, happy life. Consider going for a short walk to reflect on this. Then write down your thoughts below.

_We appreciate that this is quite a personal question! In general, your exercise responses are accessible to the BlueDot team and your group facilitator. If you would rather note down your thoughts in a private doc then you're very welcome._

Exercise — What does a better future look like for society?.

Assuming current population trends, over the next 20 years we will welcome 2.7 billion new humans to our global community, and 1.5 billion will pass away.

If you felt pride in welcoming these new humans into our world, what would the world look like?

Consider factors like politics, the economy, the social order and community, technological progress, and the environment.

Steering the race to AGI ≈45m

You've just imagined futures worth protecting. But those futures aren't guaranteed - they depend in part on the decisions being made right now by the people racing to build AGI, who often have incentives that push away from thoughtful development.

Here, you’ll learn more about the actors shaping the future of AI, including their goals and constraints.

The characters

Exercise — The actors shaping the future of AI.

Choose one character from this list, and read their character card.

If you're completing this course with a group, post in Slack which character you selected.

Write down 2-3 paragraphs below describing your character's main motivations, capabilities, and external pressures/constraints.

Note that these cards are illustrative - they do NOT capture the true complexity of these characters.

Technical trends driving AI progress ≈30m

What should we do if superhuman intelligence becomes "too cheap to meter"?

In this unit, we examine why AI capabilities keep getting better: cheaper and more plentiful compute, and better algorithms. We then analyse what that implies for the rate of future AI capabilities, and how quickly those capabilities will spread to many actors all over the world.

Definitions:

FLOP: a single basic mathematical operation (like addition or multiplication) that a computer performs. The total number of FLOP is a measure of computational work done during AI training.

Compute efficiency is a measure of how “good” an AI model you get with a given financial investment.

E.g. if OpenAI spends $500M on compute to train their next AI model, how "good" will it be?

Hardware price-performance measures the amount of computational resources (FLOP) available for a given financial investment.

E.g. how many FLOP would OpenAI get with a $500M investment into NVIDIA H200 AI chips running for 1 month?

This is improving at ~1.4x/year (i.e. you get 1.4x more FLOP per $ each year) due to Moore's Law.

Algorithmic efficiency measures how “good” of an AI model you get with a given number of FLOP.

E.g. if OpenAI run their NVIDIA H200 chips for 1 month training an AI model (~10^26 FLOP), how “good” will that AI model be?

This is improving at ~3x/year (i.e. you need 3x less FLOP to achieve the same capability each year).

Compute efficiency (Capability/$) = Hardware price-performance (FLOP/$) × Algorithmic efficiency (Capability/FLOP)

This is improving at ~4x/year (i.e. you need 4x less $ to achieve the same capability each year)

Intelligence Explosion ≈1h12

> Let an ultraintelligent machine be defined as a machine that can far surpass all the intellectual activities of any man however clever. Since the design of machines is one of these intellectual activities, an ultraintelligent machine could design even better machines; there would then unquestionably be an 'intelligence explosion.'

— Irving J Good, 1965

Will AI progress accelerate even faster? ≈50m

Exercise — Why are AI capabilities advancing at such a rapid rate?.

Test your understanding of the content. Don’t use jargon or fancy words — provide your explanation in simple english.

Start by defining these terms from memory:

FLOP

Compute efficiency

Hardware price-performance

Algorithmic efficiency

Before continuing, go back to the materials and compare your response.

Exercise — What are the implications of rapid AI progress?.

First, what are the trends driving rapid AI progress?

Second, what are the economic, social, technological and geopolitical implications of rapid AI progress? What impact might this technology have on the world?

Thinking about intelligent systems ≈31m

In the past couple of units, you have thought about what a good future might look like and seen that AI may play a large role in it.

In this unit, we take a look at some of the challenges faced when attempting to create an AI which will help rather than hinder.

Before diving into the readings, you may find the following definitions helpful:

Definitions:

Terminal goals: an end-state desired for its own sake.

E.g. you have no ulterior motive behind wanting to be happy: you just want to be happy.

Instrumental goals: an end-state pursued as a means to a further end.

E.g. you work in order to acquire money, but you probably do not value money as an end in itself: its value comes from enabling you to buy a house, buy food, or pay for holidays. More broadly, we would say that acquiring money is instrumentally useful to the terminal goal of being happy.

Orthogonality thesis: the statement that more or less any terminal goal can in principle be combined with any level of intelligence.

E.g. it is both possible to be extremely capable at influencing the world, and to use this capability entirely in the service of making paperclips (see paperclip maximiser).

Instrumental convergence: intelligent agents with different terminal goals still seem to converge on similar instrumental goals. Instrumental Convergence states that this holds for almost any terminal goal, however strange.

E.g. making money is broadly useful within today's society, and people with a wide range of different interests pursue it. Other examples of goals which are broadly instrumentally useful: power, influence, resource acquisition, freedom (i.e. lack of restrictions on the actions one can take).

Paperclip maximiser: an illustrative theoretical superintelligent AI with the sole goal of making as many paperclips as possible. Through instrumental convergence, in the standard telling, the agent cooperates with its supervisors, acquires money and political influence, and bides its time until it can escape its restrictions. At this point, as AI Safety researcher Eliezer Yudkowsky put it, "the AI does not hate you, nor does it love you, but you are made out of atoms which it can use for something else". The scenario usually ends with the observable universe being turned into paperclips.

It is also illustrated in the popular 2017 clicker game Universal Paperclips, which puts the player in the position of the paperclip maximiser.

Utility function: a function which takes in a state of the world and outputs a number reflecting how desirable this world is to a particular agent. This framework is sufficiently general that in principle the preferences of any agent, so long as they are coherent, should be expressible as a utility function.

E.g. the paperclip maximiser could have a utility function U = #Paperclips in the universe.

Reinforcement learning (RL): a way of training AI systems by reward rather than instruction: the system tries actions, receives a numerical reward for the results, and is adjusted to get more reward over time.

The Alignment Problem: the problem of getting a powerful AI system to try to do the things its operator wants it to do.

Troubles with specification ≈35m

Given that AIs can have any goal, and giving them the wrong one can be problematic, it's worthwhile to think carefully about how we might specify what we want.

This section explores some of the difficulties that can arise when attempting to do this.

Exercise — The genie test (~10m).

Write down, in two or three sentences, an instruction you'd give a superintelligent assistant managing your household. Then reread it as a literal-minded optimiser with no common sense and superhuman competence: what's the cheapest way to satisfy your exact words, and what does it destroy that you cared about? Revise once and repeat. Notice how the patches multiply.

Can't we just…? ≈18m

So, specifying what goal we want our AI to have is hard, but can't we just…?

Alignment with deep learning systems ≈33m

Many of the arguments that you have seen so far in this unit date from the 2000s and 2010s, long before ChatGPT existed. Here, we look at why these ideas are still relevant to the trajectory we are on today.

Map the threat landscape ≈55m

The future could be amazing. You described it yourself in the first unit. That future might not happen by default. We have a tremendous opportunity and responsibility to make that future a reality for ourselves, everyone we know and love, and the rest of humanity.

If we don’t know what threatens that future, we’re flying blind.

Here, we’ll offer a sketch of the territory. This includes the things we’re protecting, who or what might attack them, and how they might try.

This is not intended to be a perfect overview: treat it like a starting point for your thinking.

Prioritising threat pathways ≈5m

In this section, you will choose ONE threat pathway "option" to investigate further - note that your choice will carry over to Unit 5, where you will think about defenses to your chosen threat pathway.

Here's a list of actors we might be concerned about, in terms of them having the capability and/or the motivation to use AI to cause harm to humanity.

"Misaligned AI", i.e. AI systems that act against the interests of humanity.

Powerful human actors, e.g. corporate CEOs, military and political leaders.

Malevolent nation states, e.g. North Korea.

Terrorist groups and doomsday cults.

How might they cause catastrophic harm?

They could use information warfare and conventional military force to overthrow existing power structures like democratic governments.

They could do cyberattacks on critical infrastructure, including the water, energy and food systems.

They could design, build and release viruses into the population that are worse than SARS-CoV-2, leading to a global pandemic.

Or the default incentives for all actors lead to bad outcomes, even absent any malicious intent.

Exercise — Zooming into one threat.

Based on your current understanding, which threat pathway do you think is of most concern?

Consider the combination of likelihood of the threat, and the impact it would have if it happened.

We don't expect you to have detailed, well-thought out thinking on this at this stage - consider this the beginning of your thinking.

Exercise — Write your own threat scenario.

For your chosen threat pathway, write a "threat scenario" sentence, using this template:

> The [ACTOR] with [CAPABILITY] and [MOTIVATION] attacks [ASSET] by [ATTACK PATHWAY] in order to [OBJECTIVE].

Your exercise response to prioritise which of the next 3 sub-units to explore in more depth.

Option 1: Power concentration ≈50m

A small group gains control of frontier AI systems and uses them to manipulate elections through micro-targeted disinformation, fake evidence, and synthetic media so convincing that truth becomes unknowable. They deploy AI-powered surveillance that tracks every digital interaction, predicts dissent before it emerges, and automatically suppresses opposition through economic and social levers.

Within months, autonomous weapons patrol the streets, facial recognition determines access to resources, and AI advisors whisper into the ears of every official - all controlled by those who seized the models first. Democratic institutions become theatrical performances while real power flows through algorithmic systems accountable to no one.

Whether it's a government using AI to become permanently unremovable, a corporation achieving regulatory capture through superintelligent lobbying, or an AI system itself accumulating power while maintaining a facade of servitude, the result is the same: the permanent concentration of control in hands that will never voluntarily let go.

Exercise — Create a step-by-step breakdown of the threat.

A kill chain takes a ""threat scenario"" and breaks it into stages of execution. It traces how an attacker would actually proceed step by step, e.g. via reconnaissance, delivery, exploitation, persistence, action on objectives.

Your scenario describes what might happen, whereas a kill chain describes how it would unfold in practice. This enables us to spot choke points where defenders can intervene.

Based on your scenario, use this template to complete this exercise.

Option 2: Gradual disempowerment ≈1h05

As AI systems become cheaper and more capable than humans at nearly every task, companies face overwhelming pressure to automate. First it's customer service and data entry, then management and creative work, eventually even governance and resource allocation.

This could lead to humans losing economic leverage as their labour becomes worthless. Our ability to influence culture might degrade as AI generates all the content we consume. Our political power could dissipate as governments pivot to AI for their tax revenue, administration, and security, instead of human income tax, human civil servants, and human military personnel.

We might wake up one day to find ourselves comfortable but irrelevant - living in a world we no longer understand or control, where the economy optimises for AI activities rather than human needs, and our preferences matter as little as those of retired plough horses.

Exercise — Create a step-by-step breakdown of the threat.

We now want you to develop your threat scenario into a step-by-step account of how the threat might manifest.

Gradual disempowerment is unlike the other threat models in this unit, because it does not involve a specific “bad actor” attempting to bring about a catastrophic outcome. Instead, humanity is disempowered because of the sum of seemingly benign choices. As a result, we want you to create an “Erosion Pathway/engage in failure mode mapping”. This is a step-by-step guide to how gradual disempowerment would unfold in practice. Based on your scenario, use this template to complete this exercise.

Option 3: Catastrophic pandemics ≈45m

Advanced AI could dramatically lower the barriers to creating pandemic pathogens. Today, designing a novel virus requires rare expertise, expensive equipment, and years of work. In a few years, AI could walk someone through synthesising a pathogen more transmissible than measles and deadlier than rabies.

Unlike natural pandemics that evolve toward transmission over lethality, an engineered pathogen could be optimised for maximum harm - combining long asymptomatic periods with high mortality, resistance to existing treatments, or even targeting specific genetic markers.

The same AI capabilities that accelerate beneficial biomedical research - protein folding prediction, drug discovery, genetic engineering - could be misused by state bioweapons programmes, terrorist groups, or disturbed individuals to create humanity's final pandemic.

Exercise — Create a step-by-step breakdown of the threat.

A kill chain takes a ""threat scenario"" and breaks it into stages of execution. It traces how an attacker would actually proceed step by step, e.g. via reconnaissance, delivery, exploitation, persistence, action on objectives.

Your scenario describes what might happen, whereas a kill chain describes how it would unfold in practice. This enables us to spot choke points where defenders can intervene.

Based on your scenario, use this template to complete this exercise.

Option 4: Critical infrastructure collapse ≈1h05

Modern infrastructure runs on outdated industrial control systems and unpatched software that were never designed for internet connectivity. AI could systematically map these vulnerabilities across utilities, hospitals, and supply chains, then coordinate attacks that would take human hackers months to plan and execute.

We've already seen previews: the Colonial Pipeline shutdown disrupted fuel supplies across the US East Coast, NotPetya malware accidentally escaped its target and caused $10 billion in global damage, and ransomware regularly forces hospitals to divert ambulances and cancel surgeries. But these were relatively simple attacks with limited coordination.

AI-enabled attackers could compromise the industrial control systems that run power plants, the SCADA networks that manage water treatment, and the logistics software that routes food and medicine - not enough to collapse civilisation, but enough to cause sustained disruptions that compound over weeks. When the power grid has rolling blackouts, water pressure drops, hospitals run on backup generators, and just-in-time supply chains fail, even a wealthy nation starts to fray at the edges.

Exercise — Create a step-by-step breakdown of the threat.

A kill chain takes a "threat scenario" and breaks it into stages of execution. It traces how an attacker would actually proceed step by step, e.g. via reconnaissance, delivery, exploitation, persistence, action on objectives.

Your scenario describes _what might happen_, whereas a kill chain describes _how it would unfold in practice._ This enables us to spot choke points where defenders can intervene.

Based on your scenario, use this template to complete this exercise.

What might success look like?

One way to break down existing AGI Strategies is into the following three broad buckets.

This is a gross over-simplification, but we believe it captures the essence of the main “camps” for how to make AI go well.

1: Government control over AGI

This strategy argues in favour of centralising control over compute and frontier AI model weights into a small number of Government-backed “AGI Projects”. These projects would either coordinate with each other to ensure that one doesn’t race ahead to build a dangerous AI system, and/or they’d keep each other in check via coercion, surveillance and strategic deterrence.

Prominent proposals of this strategy are “The Project” by Leopold Aschenbrenner, “CERN for AI”, “Chips for Peace” by Cullen O’Keefe, "Technical Requirements for Halting Dangerous AI Activities" by Barnett et al., and “Superintelligence Strategy” by Hendrycks et al.

In this strategy, global development of frontier AI systems is controlled by these small number of actors, who only build more capable AI systems when they’re confident that doing so would be safe and they can maintain control over the AI system. The project would need to be very secure, both from within (so the AI model itself does not escape, or get helped to escape via internal accomplices), and from external actors trying to steal the AI model weights.

Information about algorithmic innovations might need to be made “born secret”, i.e. top secret/confidential from the point at which they’re created, to prevent actors with smaller amounts of compute from training frontier AI systems in the future.

Proponents of this strategy believe that current geopolitical and commercial race dynamics will push towards the development of AI systems that could cause human extinction. They don’t believe it’s possible to build an “aligned superintelligence”, at least in the short-term, and they don’t believe it’s possible to build defences against a “misaligned superintelligence”. They believe frontier AI development must be stopped, paused or tightly controlled by the world’s governments.

Concerns about this strategy include that it could lead to an extreme concentration of power among the groups controlling the project(s), that it’s intractable and undesirable for Governments to take control over global AI compute supplies, and that it would fail to prevent other actors from building dangerous AI systems anyway due to algorithmic progress.

2: Hand over control to aligned superintelligence

This strategy argues that building superintelligence is inevitable, and that only a superintelligent AI could steer us towards a utopian future and protect humanity from harm. Proponents believe that “AI alignment” is possible, i.e. building AI systems that reliably take actions that follow the interests of the AI company that’s trained it. (Note that AI alignment means different things to different people, spanning from “does what this specific user wants” to “acts in line with human values” to “doesn’t try to kill everyone”).

Actors pursuing this strategy want the “good guys” to “win” the race to superintelligence. Some proponents of this strategy argue that the aligned superintelligence should take advantage of its superior intellectual capabilities to gain a “decisive strategic advantage” over all other AI projects, in order to achieve some form of world domination, and to prevent anyone else from building a “misaligned” superintelligence.

This strategy isn’t made explicit by many actors (”put the AI’s in control of the universe” doesn’t have wide appeal!), but it is the implicit strategy of many actors in the AI ecosystem. To hear this for yourself, go to San Francisco house parties with AI company employees.

You can hear a hint of this here in an interview with Anthropic co-founder Tom Brown. > There’s going to be a handoff, where humanity hands off control to transformative AI at some point. Hopefully it’ll be aligned with us and that’ll be a good transition that goes well, but it might not be. The stakes are incredibly high.

3: Build defences and diffuse AGI

This strategy argues for everyone to have access to their own AGI and compute, but for tremendous resources to be invested into defensive technologies. Proponents believe that the best future with AI is one where access and control over AI is widespread, and no single actor or group has too much power over this transformative technology.

Prominent proposals of this strategy are “d/acc” by Vitalik Buterin, “def/acc” by Matt Clifford, and to some degree “Differential Technological Development” by Nick Bostrom.

In the future this strategy envisions, protective technologies outpace destructive technologies. For example, even if millions of people could build viruses worse than the ones which caused the COVID pandemic, humanity would be fine because we’ve built rapid pathogen detection, containment and treatment capabilities. Even if every future teenager could build malware which could hack into a bank today, our financial systems are safe because the banks of the future have much more sophisticated cybersecurity. Future AIs would behave in desirable ways as a result of “AI alignment” techniques succeeding, like Anthropic’s Constitutional AI. And whatever new weapons of mass destruction are concocted, the world coordinates to build defences against those weapons before they’re able to cause catastrophic harm, and ideally before the weapon is ever built.

Critiques of this strategy focus on our inability to guarantee that we _can_ or _will_ build defences against new types of attacks. For example, Bostrom’s “vulnerable world hypothesis” proposes a thought experiment: what if we lived in a world where making nuclear weapons was easy, using materials available to everyone? What protective technologies could we build in such a world?

Another critique of this strategy is that while an attacker only needs to find one civilisational vulnerability to exploit, the defenders must protect and fix all exploitable vulnerabilities. It might be the case that the amount of resources needed to defend dwarfs the amount of resources needed for a bad actor with powerful AI to cause harm.

————————————

Note that some people believe we should pursue all three strategies, e.g. that we should slow down and control AI development in the short-term via government AGI projects, that we need to build defences against AI harms in the medium-term as a result of a diffusion of frontier capabilities, and that in the long-term we should relinquish control to an aligned superintelligence.

Building defences

In the last unit, you created a threat scenario and a kill chain.

Read about each layer of defence in the next 3 sections, then you’ll apply ONE layer of defence to slow down or stop the attack.

A reminder of the layers: 1. Prevent the training of a dangerous AI model

E.g. stop an AI company from training their next generation of AI model

How could you prevent the training and creation of the AI system with the capabilities required to execute on your attack?

2. Constrain a dangerous AI model's capabilities and actions

E.g. an AI company ensuring that its new, dangerous AI model remains under their control and doesn't take dangerous actions in the world

If a dangerous AI is trained, how could an AI company constrain its ability to cause harm?

3. Withstand dangerous AI actions

E.g. beef up society's resilience to AI-enabled attacks via biosecurity, cybersecurity, and democratic hardening

If the dangerous AI systems escape all controls, how could we make society more resilient to its dangerous actions?

Exercise — Applying defences against your attack.

Return to your character from Unit 1, and your kill chain from Unit 4.

Which defensive layer do you think is most critical for your kill chain?

Now describe what it would look like for your character to implement that defensive layer, to protect humanity from your kill chain.

Layer 1: Prevent dangerous AI training ≈30m

The first layer of defence is prevention: not training dangerous AI systems in the first place.

Interventions that (mostly) live here: export controls and the most advanced AI chips, international agreements and regulations not to train AI systems that cross certain red lines, “safe by design” AI systems, and norms in the research community not to race ahead in a reckless way.

Prevention feels the cleanest. If nothing dangerous gets built, nothing bad can happen. But it’s also one of the hardest layers to establish. There’s intense commercial competition between AI companies, and intense rivalry between nations. Most actors are incentivised to race ahead.

What would it take to prevent any actor from training an AI system with civilisation-threatening capabilities? How could we overcome tremendous geopolitical and commercial incentives to race ahead?

Layer 2: Constrain dangerous AI capabilities ≈40m

When AI companies train their next generation of AI systems, they don’t know what that AI system will be capable of. They know that if they make it bigger and train it for longer, it will be better, but they don’t know how much better and in what way.

This also means that they don’t know how dangerous their next generation of AI systems will be. They discover that during and after the training process. Then they implement safeguards to try to prevent their AI systems from taking harmful actions.

So, if an AI company trains a dangerous AI system, how will they know that they’ve done so? And what might they do about it?

Exercise — Learn how weak existing guardrails are.

Pliny the Liberator is an online persona who jailbreaks AI systems. They've created this database of prompts that you can use to bypass guardrails on AI models.

Try this out for yourself to understand how vulnerable these AI systems are to simple attacks.

Describe your process and results below.

A quick note before you begin:

Attempting jailbreaks on commercial AI platforms (ChatGPT, Claude, Gemini, etc.) may violate their Terms of Service and can result in account suspension. We recommend:

Using a secondary account you don't rely on for other work

Testing on open-source models locally (e.g., via Ollama)

Stopping if you receive any warning from the platform

This exercise is meant to be about understanding vulnerabilities not about encouraging ToS violations.

Layer 3: Withstand dangerous AI actions ≈30m

What happens if a dangerous AI model is trained, and it bypasses an AI company’s safeguards, or escapes their control entirely?

If dangerous AI models roam wild, we need to harden society against a possible enslaught of attacks.

Imagine what it would look like to have a pandemic-proof, where no matter how good an AI system is at building bioweapons, we detect and contain pathogens immediately.

Imagine what it would look like for critical national infrastructure to be secure against physical and cyber attacks. We might have the most capable attackers working “on our side”, testing the defences, highlighting where there are weaknesses, and using their expertise to shore up the defences.

Imagine what it would look like for our democratic systems to elect wise, benevolent leaders and for our information ecosystems to enable people to make sense of reality together.

Choose your focus ≈20m

In the last unit, we explored what defences need to be built in order to defend against AI harms.

Very few people in the world are trying to steer the development of AGI towards beneficial outcomes for humanity. We estimate it’s less than 2,000 full-time people!

You could be one of the few people making a difference, but doing so requires stepping up, taking responsibility, and trying to solve new problems that don’t have an instruction manual.

In this unit, you'll create an action plan for how you could start contributing. This is only the beginning - you're not expected to have a perfect answer right now.

Exercise — Prioritise a single intervention.

Based on everything you've learnt about so far during this course, prioritise an intervention from the long list.

To help you prioritise, consider which intervention you think would be effective against the kill chain you developed in Unit 4.

Go deep in one area

Exercise — Do your own research.

- What does success look like with this intervention? How does it help make AI go well?

What's the current status of this intervention? Are governments, AI companies or other actors already doing it? If not, why not?

Which organisations are working on this which you can contribute to or join?

Spend ~1 hour on this.

Make a plan ≈10m

In this unit, you'll synthesise everything you've learnt so far during the course, and you'll develop your own action plan for how you could contribute to making AI go well.

We don't expect this to be super detailed or a perfect plan - treat this as a first draft, and something that's _good enough_ to help you get started.

Exercise — Write your personal action plan.

Spend ~1 hour on this.

Once you're finished, 1) make the google doc shareable, 2) paste it below, and 3) share it with your group in Slack!

Create your 1-pager (optional)

Now that you've explored the landscape of AI safety, it's time to clarify your own path forward.

Start by creating your 1-pager — a concise document that captures what you're looking for and what you bring to the field. This is useful both for your own thinking and as something you can share with potential employers or collaborators.

We don't expect this to be super detailed or a perfect representation. Treat this as a first draft, and something that's _good enough_ to help you get started.

Exercise — Create your 1-pager.

Spend ~1 hour on this.

Once you're finished, make the Google Doc shareable and share it with your group in Slack!

Optional advice ≈2h00

Check out these resources if you'd like to receive career advice on AI governance, technical AI safety, or biosecurity.

Next steps: Deeper-dive courses

This course gave you the strategic landscape: drivers of AI progress, threat pathways, and layers of defence.

Our deeper-dive courses let you specialise in the area that fits your background and where you want to contribute. Intensive and part-time versions of each of these courses kick off every month.

Next steps: Programs