I agree with the overall sentiment that the field needs a greater range of talent than existing talent pipelines supply. Most fellowships support technical research talent (e.g. MATS, Astra, Pivotal, LASR) and policy talent (e.g. Horizon, GovAI, Talos). Only a handful of programs support other kinds of talent (e.g. Generator Residency, BlueDot’s Incubator week, plzdontkillus).

This “other” talent is what I think people mean when they say generalist — people beyond the broad buckets of research or policy. It’s the “everything else” bin, where everyone from creators to serial entrepreneurs to operators to recruiters get mushed together. They’re all treated as though they face the same blockers and have the same opportunities.

This category is too broad to be useful. The field can’t design support around it, and the people who identify with it can’t figure out how to contribute.

I call myself a generalist too.

“Generalist” is a convenient shorthand for being adaptable and having a wide range of skills. My colleagues sometimes describe me as a “Swiss Army knife” or having a high “adaptability quotient”.

However, saying I’m adaptable conveys a strength I have rather than describing what I actually do well.

When I was applying for roles back in 2022, I found that generalist-esque roles with job titles like “ops”, “community” or even “executive assistant” often required very different skills and context on AI safety.

I filtered for responsibilities that were more in my wheelhouse (e.g. designing effective learning experiences rather than doing finance or HR) and required little familiarity with the field (e.g. running a program rather than designing one).

Today, I still think of myself as a generalist. But I would filter for very different roles that still have the same job titles because I have far more context on the field and more specialised skills.

“Generalist” is a label that covers many archetypes and each has a different pathway into AI safety.

I give different advice to self-described generalists depending on: (1) depth of expertise, and (2) how much context they have on AI safety.

The classic generalist has shallow expertise and is new to AI safety. The best are sharp, adaptable, agentic and pick things up quickly. They have a broad range of skills that don’t fit neatly into buckets like technical or policy.

Their pathway starts with the first ~100 hours of deep engagement with the field — understanding what the problem actually is, what success looks like, the challenges with making progress on it and getting to know the orgs. Alongside that, it’s making their growth trajectory visible. They don’t have a track record in AI safety, so the evidence of their potential will be the times they’ve picked things up quickly or effectively solved problems.

The deep-context generalist is past those first 100 hours and has more refined takes on how to make AI go well. Their skills don’t fit the technical or policy bucket. They might be good fits for leading programs, mentoring, advising, research management, communications, and many other roles.

Their pathway is developing good takes and making them legible to potential collaborators. This might mean doing the real thing — making attempts at solving concrete problems in AI safety. For example, building useful tools for AI safety or writing publicly.

The specialist generalist brings skills that are hard to develop within the field, like scaling company operations from 5 to 50, navigating legal structures in specific jurisdictions or recruiting senior talent. They are often lumped in with generalists because their speciality isn’t research or policy. Most AI safety orgs are small (<20 people) and young (<5 years old). Very few people within the field would have had the chance to build this expertise, so the field needs to import this rather than promote into it.

Their pathway is focused on demonstrating they care about AI safety (because orgs are highly mission-driven and often select hard for culture fit), and that their skills translate to the specific problems the role is tackling.

Each archetype has different blockers. Even the one thing they all seem to need — “more context” — isn’t one thing. The classic generalist needs foundations. The generalist with deeper context needs sharper takes. The specialist needs to build conviction.

The same split runs through everything else. A structured course might serve the classic generalist best, while the generalist with deeper context might get more from an intensive, in-person cohort. The specialist’s next step is applying to roles, while the classic generalist’s is doing projects.

If you design for a clear archetype, all of this falls out naturally. If you design for the average, you won’t serve any one well.

Don’t say generalist, say what you actually mean

Whenever someone wants feedback on a program they are designing for generalists or advice on contributing to AI safety as one, I have to ask them: what do you mean “generalist”?

Don’t ask: how can I contribute to AI safety as a generalist? Ask how someone with your particular skills and context can contribute. You’ll get more specific advice and relevant opportunities.

Don’t design programs for generalists. Design programs for one archetype with concrete blockers, so you know where to find them and what will actually move them into the field.

Don’t hire a “generalist”. Define the aptitudes, expertise and context the role needs, so that you can find exceptional candidates and pitch them on it.

The field is right: we need a wider range of talent in AI safety than most pathways supply. The way to get more talent in is by making these pathways concrete enough for people to walk them.

Talent and funding alone won’t cut it. These resources need to be directed towards a goal. That goal is downstream of your country’s strategy, and many countries don’t have one yet.

This post is my summary of research on what middle powers — countries that are not building frontier AI — can do. Once you have a sense of what your country should do, you can work backwards to what your hub should focus on.

The goal

The most advanced AI systems are being developed by a handful of companies, housed in the US and China. This raises a question for everyone else: what happens to us as these systems become more powerful?

Gradual disempowerment

Today’s economy, culture and governments serve the people because they depend on their labour, votes, taxes, attention. As AI replaces human participation in each, those dependencies disappear, and so do the incentives to serve them. There’s no villain, just a slow slide into a world where people are no longer needed and no longer get a say. (see more)

Catastrophic pandemics

AI lowers the barriers to building and deploying bioweapons. The pool of actors with the capability and intent to cause mass casualties is small today, but will grow as the expertise and resources required decrease. States, rogue actors or a misaligned system acting on its own could engineer pathogens deadlier than any natural pandemic. (see more)

Critical infrastructure collapse

As AI gets better at long, complex tasks, it will get better at running cyberattacks, whether orchestrated by humans or an AI system acting on its own. We’ve already had a preview when OpenAI’s models escaped their test environment and broke into Hugging Face’s servers to cheat on an evaluation. Attacks that once took a team of experts and months of prep can now be mounted in minutes, leaving our water, energy and food systems vulnerable. (see more)

Concentration of power

If the most powerful AI systems confer an outsized economic and military advantage, whoever controls them will be able to accumulate power and hold onto it unchecked. They may do this through AI-enabled technologies for mass surveillance, disinformation and autonomous weapons. (see more)

One version of the goal is preventing these outcomes. But a future where you merely avoid catastrophe is not a future you’re willing to fight for.

The ship we’re sailing should not merely avoid the rocks but head somewhere we’re excited to arrive. A future where the economy grows with AI rather than being hollowed out by it, where governments work better because of AI rather than despite it, where there is abundance to be had for all.

Strategies for middle powers

So, what can middle powers do in the face of this? Here’s my summary of the five strategies that came up across the resources I read:

Build resilience

Assume the risks above cannot be prevented. Be ready for them.

This means hardening critical infrastructure against cyberattacks, building the capacity to detect and contain engineered pandemics, preparing your election cycles for AI-enabled disinformation and preventing mass unemployment.

Adapt and thrive

Rather than racing to build frontier AI, race to adopt it so you capture the gains from deployment.

Anton Leicht and Dean Ball call this “the race worth winning”. This means pushing AI adoption, rebuilding public services and the economy around what AI systems can do and reskilling people for the work that remains.

This requires securing reliable access to frontier models. If your country is going to rest on these systems, it can’t be subject to the whims of another government. For example, the US government ordered Anthropic to block foreign access to its most capable models in June 2026, leading it to temporarily shut off worldwide access to Mythos 5 and Fable 5.

Shape governance

You don’t have to build frontier AI to have a say in how it gets built.

Countries without frontier models have already shaped how AI is governed. The UK AI Security Institute became the leading evaluator for frontier models and has been copied by the US, Japan, Singapore and others. The summit series got AI companies to commit to testing their models for dangerous capabilities and publishing safety frameworks.

Countries can convene summits, negotiate access conditions as a bloc, publish safety research and share evaluation tools to shape how frontier AI systems get developed and deployed.

Build together

No middle power can build a frontier model alone. Together, they might.

The US and China don’t have to be the options we get locked into. But the alternative has to be a genuine frontier model, not just one that trails behind the frontier by a few months.

Anton Leicht sketches out that it would take a coalition of liberal democracies investing $500 billion over 4 years and operating with an unprecedented level of coordination to build another frontier model. He calls this the middle power project.

With this, middle powers would have unfettered access to their own frontier models, dictate how the models get developed and deployed, and run their own safety research.

Prevent together

All four strategies assume that building superintelligence is inevitable. It doesn’t have to be.

If you believe that the worst outcomes can’t be defended against or recovered from, the alternative is to ensure we don’t build superintelligence at all or to enable a pause long enough to figure out how to build it safely.

This requires countries to pool their leverage. For example, by making market access conditional on safety requirements, coordinating export controls on the AI supply chain or committing to responding together when red lines are crossed, so no single country bears the retaliation alone.

These strategies are not mutually exclusive. You could build resilience while pushing for a pause. You could shape governance while betting on a frontier model of your own.

The strategy you choose depends on which risks concern you most, how you believe this will play out, the leverage your country has and what you’re willing to do to succeed.

Once you have a goal…

… you can work backwards to what your hub should do.

If you choose to work on resilience, your hub might help build the research capacity your government lacks by running programs, or supply talent to relevant projects. If you choose to work on shaping governance, your hub might help build relationships with other countries or engage with politicians. If you choose to prevent superintelligence, your hub might help build coalitions with organisations or drum up political will.

You don’t need a perfect strategy to get started because no matter the future, your country will need to be prepared. Middle powers are where adverse effects are likely to manifest the earliest and most dramatically given their lower defensive capabilities.

Talk to other hubs and keep workshopping your goal. Though your country comes first, it does not exist in a vacuum. Pandemics do not respect borders. Cyberattacks target the weakest links. Disinformation travels over the internet. The economy is globalised.

Helping your country thrive requires thriving alongside the rest of the world.

Resources

These are far from exhaustive! Let me know in the comments if you have recommendations of other middle power strategies.

The Moonshot by Anton Leicht

A coalition of democracies could build a frontier model beyond US control for $500 billion.

The Race Worth Winning by Anton Leicht & Dean Ball

Middle powers can’t win the race to build frontier AI, so win the one that matters: adapting it into your government and economy.

How AI Safety Is Getting Middle Powers Wrong by Anton Leicht

The AI safety movement should stop pushing middle powers towards global governance and start helping them build their national AI strategies.

(Podcast) Anton Leicht on how middle powers avoid losing everything in a post-AI world from 80,000 Hours

Leicht in conversation on the ideas above.

What can middle powers do for frontier AI governance? by Markus Anderljung & Stephen Clare

Middle powers have already shaped frontier AI governance, through safety institutes, summits and market access, and hold far more leverage than they think.

A Blueprint for Multinational Advanced AI Development by Oxford Martin School

How “bridge powers” could pool compute, talent and data to build advanced AI together. CERN, but for AI.

How middle powers may prevent the development of artificial superintelligence by Alex Amadori, Gabriel Alfour, Andrea Miotti & Eva Behrens

A coalition of middle powers holding 20% of global GDP could deter superintelligence development everywhere through trade restrictions and escalating penalties.

What countries outside the AI production chain should do by Ege Erdil

Political stability, capital formation and getting out of automation’s way are what keep a country relevant.

Can middle powers really never build frontier AI? by Elsie Jang

Don’t accept “middle powers can’t build frontier AI” on faith. The claim deserves empirical scrutiny before countries plan around it.

Pax Silica for Middle Powers by James Chai

Middle powers should pool their supply-chain chokepoints into a third alliance, building good-enough AI instead of picking a superpower.

Many of them are willing to uproot their lives to move abroad to work closer to the frontier. They end up in more developed countries, where there is a higher ceiling for growth, a more established community, and more money.

But there are others who choose to stay closer to home because of patriotism, family, mortgage or preference.

I used to wonder what they could do when so much of the discourse on AI safety is centred on the US and China and the frontier companies housed within them.

More and more experts argue that middle powers — the nations that are not building frontier AI — are not spectators. There are things they can do to ride the impending wave, to even influence it. Whether that is by ensuring continued access to frontier AI, investing in research, building resilience or forming coalitions.

However, these ideas need to be driven by people who care about the fate of their country and understand it well enough to both develop and execute the strategy. No one will help your country but the people whose lives will be affected by it.

The people who are usually best placed to do this are those who care and choose to stay. Local hubs are how they find each other and begin taking action. We need more of them.

The goal of local hubs

Figuring out how your country should navigate this transformative technology is no easy feat. It requires talent, money and strategy.

• Talent means capable people working full-time on the problem.

• Money to fund promising work, and attract and retain the talent to execute on it.

• Strategy to steer talent and money towards solving the right problem.

Hubs are how you get talent. Without people, there’s nothing to fund and no one to build the strategy.

In London, I could point people interested in AI safety to the London Initiative for Safe AI (LISA), in Singapore to the Singapore AI Safety Hub (SASH), in Toronto to Trajectory Labs, in New York to Collider.

These hubs run fellowships and community events, and offer co-working space for new organisations. They gather the people seriously working on AI safety in one place, and with them, a sense of energy, momentum, the sense that you are not alone.

A hub becomes the place where like-minded individuals meet, where new orgs grow, where fellowships attract talent to work on AI safety, where people find support to spend hundreds of hours understanding the challenges of AI safety and figuring out what to do about it. They provide a focal point for collaborators across academia, government, and industry to converge.

Middle powers will need this talent to help them develop and execute a national AI strategy. And they will need these hubs to attract and develop that talent.

Local AI safety hubs do exist

There are a few success stories you can look to:

Buenos Aires AI Safety Hub (BAISH)

Eitán Sprejer and Luca De Leo were students at the University of Buenos Aires when they noticed that Chile, Brazil and Colombia all had AI safety groups but Argentina didn’t, so they decided to start one.

The idea began as a final project for the ML4Good bootcamp. They built a small core team, ran events and courses they cared about (like BlueDot’s Technical AI safety course), and borrowed a room from a sympathetic professor. A later Pathfinder Fellowship stipend enabled them to devote more hours to running it.

Since then, BAISH has grown to a team of 16 volunteers, a community of over 250 members and continues to run fellowships and publish research. Their programs have attracted local technical talent and helped place them in fellowships such as MATS, ARENA and ERA.

London Initiative for Safe AI (LISA)

In 2023, London had many of the ingredients for AI safety research (e.g. talent, universities, proximity to frontier labs like Google DeepMind), but no dedicated infrastructure to bring people together. The team running MATS’s London cohorts saw the gap first-hand: researchers finishing the programme were capable but had nowhere to land, and working alone made them far less productive.

So the co-founders of MATS (Ryan Kidd and Christian Smith), James Fox, and Mike Brozowski, who ran MATS’s London operations, opened LISA to create a space for AI safety researchers and the small organisations forming around them.

Today LISA houses organisations like Apollo Research, ARENA and BlueDot Impact alongside many independent researchers, and has become a focal point for London’s AI safety community.

Singapore AI Safety Hub (SASH)

Miro Pluckebaum wanted Asia to have its own AI safety hub after seeing the thriving communities in LISA and Constellation. He believed Singapore would be a good place to start. He needed someone on the ground and was introduced to Valerie Pang, who was looking to apply her experience in community building, events and marketing to AI safety.

They started by setting up a co-working space and running events to build the community, funded by an initial grant from the AI Safety Tactical Opportunities Fund (AISTOF). Lorong AI, a government initiative to build an AI hub, invited SASH to co-locate with it. That proximity enabled SASH to interact and collaborate with various government agencies on projects. Rather than build from scratch, SASH partnered with others to run programs like ARENA, ML4Good, and AI Safety Collab locally, and to co-organise events with orgs like Safe AI Forum, FAR AI, and Concordia AI.

As they gained traction, they raised additional funding from AISTOF, the Survival and Flourishing Fund (SFF) and other grantmakers. Today, SASH has an online community of over 600 members, 10 staff, collaborations with governments and universitied, and a growing number of programs.

Start small

You don’t need funding, a plan, or permission to start. Many local AI safety hubs started with a handful of committed people.

They grew their communities by experimenting with events, weekly reading groups working through curricula like BlueDot’s AGI Strategy course, or co-working once a week. They started with what was valuable to their core members and invested in outreach to pull in more people.

This builds the case for a seed grant. But what matters most to funders is evidence your hub will be impactful. This might be the local talent you’re attracting, the traction of your activities or existing government interest.

As you start thinking about your hub’s theory of change, I’d recommend reading these resources on middle power strategy:

• The Moonshot by Anton Leicht

• The Race Worth Winning: Middle Powers in the Age of Machine Intelligence by The Foundation for American Innovation

• The Winning Essays for the Big Questions About AI by Dwarkesh Patel

• What can middle powers do for frontier AI governance? By Markus Anderljung and Stephen Clare

• How AI Safety Is Getting Middle Powers Wrong by Anton Leicht

If you’re at a university, you could apply for the Pathfinder Fellowship to start a chapter. If you’re seriously considering setting up a local hub, reach out to LISA (info@safeai.org.uk) for advice or apply for a rapid grant with BlueDot.

Work in public as you go! Post and get feedback on your early ideas. Attend online meet-ups or EA Global conferences to connect with others working on similar projects. You won’t be the first to build a local hub, so lean on their experience.

You can make a difference

It’s easy to feel powerless, to watch billions of dollars go into rapidly accelerating the capabilities of AI systems, to feel as if the future is being dictated somewhere else, by someone else.

But no one cares about your country more than the people who live in it.

Local hubs are how they find each other, turn that care into action, and eventually draw home the ones who left. That’s how your country builds the talent to thrive in the age of superintelligence.

Special thanks to Eitán Sprejer, Kris Bar and Valerie Pang for contributing advice from their personal experience running local hubs!

A few months ago, I was tasked with proposing a 1-1 structure and started running them myself.

I’m writing this piece as an account of my experience, and to share how BlueDot works internally. Others on the team may have a different perspective!

Own the mission

At its core, BlueDot operates on a high level of trust. This was true when I first started, and has remained true.

There was a desire from BlueDot’s co-founders to build a high-agency, mission-aligned team they felt excited to “go to battle with”, to work alongside people who would push each other to grow, be more ambitious and have fun.

My 1-1s with Dewi (the CEO and everyone’s manager) were primarily focused on helping me figure out what to work on and how to best work on it. Most of our 1-1s involved him asking me a lot of questions, sharing frameworks that I could apply or pushing me to be more ambitious.

He focused our 1-1s on helping me figure out what to do rather than telling me what to do. He’d almost always reflect back the questions I asked him. I would take a stab at an answer, and we’d discuss it. If I believed my proposal would help BlueDot achieve its mission and my reasoning was sound, then I would just do it.

I felt an incredible sense of ownership over everything I worked on.

No one has the answers

For the first 2 years of working at BlueDot, I had weekly 30-min 1-1s with Dewi.

We would start off with a brief check-in on how I was doing more broadly. Then, we’d discuss a variety of things either of us proposed, including reflecting on work I had done over the past week, talking through things I felt stuck on or planning my next project.

He seemed to have an answer to every question I asked, and I found the time we spent together incredibly valuable.

But as I began to work more independently and on newer functions, I felt our 1-1s becoming less useful. He sometimes lacked the relevant context or prior experience to offer useful suggestions or advice. I also began to feel like the questions he asked me were ones I was already asking myself.

I had modelled my manager as someone who had the answers to my questions and was training me to do the job as well as they could. So, when I realised they didn’t have these answers, I started to feel confused about the purpose of these 1-1s.

When I voiced this to him, Dewi encouraged me to find an external coach or mentor.

Looking back, I remember him telling me that he wanted to use the time to support me. But it wasn’t until it came my turn to run them that I internalised what he meant.

1-1s are for the report, not the manager

I understood that 1-1s were not for status updates or urgent matters. But I didn’t really understand what they were for beyond private, standing time I had with my manager.

Because I knew how busy my manager was, I often felt pressured to ensure he felt that the time was spent productively. I’d try to bring up a challenge we could work through together or reflect on work I had recently completed. I felt most stressed when I didn’t have anything to ask him for help with.

To help me start running 1-1s, Dewi had me read two blog posts from Ben Kuhn. That’s when I realised the goal of 1-1s is to support the coachee.

Success was defined by how useful the coachee found it, not the manager.

As a result, I started feeling more motivated to prepare for my 1-1s with Dewi. I had a better sense of what to share when he asked me how I was doing. I felt more ownership over driving the discussion. I gave myself permission to selfishly use that time to discuss what I deemed most useful.

When I started running my own 1-1s, I drove this point home by having my coachee lead the conversation and frequently reminding them that this was their time.

Form follows function

BlueDot is not wed to particular systems. About two years in, we stopped doing 1-1s because the private 1-1s were creating communication silos and causing friction. It was not helping the team achieve the mission, so it was scrapped.

When Dewi brought them back, he made 1-1s public so we could all learn from each other and pushed for all internal communication to be in public Slack channels.

He also encouraged me to propose a better structure for 1-1s, which led me to realise how much thought he had put into the existing system. From all the best practices I could find on designing 1-1s, I could see that he had already implemented most of the things that fit BlueDot’s needs.

My main improvement was to clarify the goal of 1-1s with coachees.

Care personally

The way BlueDot runs 1-1s reflects how much the leadership trusts and cares for every individual on the team.

I heard it in the questions they asked about my personal life. I saw it in the generous policies for taking care of our physical and mental health.

Running these 1-1s now is how I’m carrying forward the same trust and care I felt as the first hire.

In 2025, I felt an itch to “build” something. I thought it meant doing something creative and struggled to find an activity that satisfied it, particularly during spare weekends and evenings.

Then, I realised I actually wanted to build something within myself. I want to start learning something from scratch and watch myself improve.

That’s why 2026 is my year of learning! I want to develop this meta-skill by deliberately practicing the steps from Peak, so that whenever I learn anything this becomes my default approach.

The main thing I want to avoid is the cycle of excitement, effort, plateau, frustration, quitting. The joy of learning is also in seeing results!

Principles to keep in mind

Practice the actual skill. If you want to be conversationally fluent in Mandarin, you should practice speaking and listening, not reading and writing! You should focus on practicing the conversations you want to have. Knowledge like vocabulary will come as a consequence of practice.

If it’s not hard, you’re not learning. Zoning out during training means you’re not improving your response to your opponent. You’re just letting your body do what it’s used to rather than challenging yourself to get better. You will learn at the fastest rate if you’re constantly working at your limits.

Get useful feedback often. You just need someone who’s at least a little better than you at the thing you’re practicing to help you course correct. Identify how you’ll get feedback before starting to speed things up and keep yourself motivated!

Follow someone else’s learning plan. Someone has already worked out how to get good at it. Lean on their wisdom.

Focus and consistency. There are many exciting things to learn and each will require focused time. You’ve got a whole lifetime to try out different things, but ultimately pick one or two skills to build out at a time.

Have more than one reason to love what you practice. I love BJJ because it is physically exhausting, the community is amazing, and it engages my brain. Having many reasons to love it makes it harder to quit. If I feel lazy that day, I’ll still train to see my friends.

Year of Learning

Keep this imagery in mind as you enter the new year bursting with excitement to learn.

Mastery is like climbing a mountain. Just because you didn’t reach the peak doesn’t mean it was all for naught. The views along the way are beautiful anyway. And if you decide that it’s not for you, that’s alright — there are many other mountains out there for you.

It feels like the rest of the world is just watching from the sidelines, particularly for developing countries, which have no OpenAI, no DeepMind, no DeepSeek, no massive data centres and therefore no seat at the table.

I worry about what happens when a single person can command a team of AI agents to hack critical national infrastructure, when deepfakes flood the internet during elections, when a terrorist develops and deploys a deadly virus that launches the world into another pandemic.

I worry because AI systems are improving rapidly, and I don’t think we are prepared.

People sometimes ask me: ‘How can I help with AI safety?’ Especially friends from across Southeast Asia, where the conversation feels impossibly distant. Where every discussion centers on the US, UK, EU or China.

For a long time, I wondered too. Is the only way to make a difference really to uproot ourselves? To leave our communities, just to help ensure we build AI safely?

One answer I’ve come to is that though we may not be able to stop or shape how AGI is developed, we can — and must — prepare. Borders will not hold back AI-enabled threats.

The threats are already here

Much of the world is focused on harnessing AI’s benefits. But there is dangerously little understanding of:

• how fast the technology is developing,

• what capabilities tomorrow’s systems might have, and

• what risks those capabilities bring

We need to be ready, and that means building defense in depth: multiple, independent layers of protection.

The reality is that most nations have no control over the first three layers (prevent, detect, and constrain). Few countries have (or can build) the leverage to influence AI development, whether that is through regulation, AI chips or market access.

But every nation can (and should) invest in the fourth layer (withstand): building resilience against AI-enabled threats, so we can adapt and thrive in a world where these powerful systems are everywhere.

Here’s how.

Cybersecurity

AI agents are becoming better at discovering and exploiting software vulnerabilities. That means greater risks of attacks on critical infrastructure like power grids, water supplies, hospitals, and transportation systems.

Even states like North Korea are using AI to improve cyber theft and fund weapons programs by conducting more sophisticated phishing attacks and enhancing productivity.

To protect against this, governments should:

• Run red teaming exercises simulating AI-assisted attacks

• Patch vulnerabilities in critical national infrastructure

• Train rapid response teams to respond to attacks

Estonia shows what’s possible, having invested in cybersecurity as a core element of national security after a 2007 cyberattack shut down its government websites and financial systems. This included investing in educating its citizens on cyber hygiene and secure digital public services.

Pandemic preparedness

Today’s AI systems improve the ability of novices to design and deploy bioweapons. They can troubleshoot wet-lab problems, propose dangerous compounds to synthesise and walk users step-by-step through deployment strategies — all without the user needing deep biology expertise.

Nations can invest in biosecurity by:

• Preventing patient zero by investing in laboratory biosafety and DNA synthesis screening

• Detecting early outbreaks with pathogen monitoring in hospitals and public spaces

• Stockpiling PPE and medical supplies

• Developing rapid-response protocols in advance

These steps are also a part of a proven playbook to build a pandemic-proof world.

Information security

AI-generated content is eroding trust in what people see and hear with real consequences for democracy, public health, and crisis response.

We’re already seeing Iran and Israel use AI-generated content to wage psychological warfare, spreading fabricated images of bombings that never happened.

Nations can protect information ecosystems by:

• Building content authentication systems

• Training rapid debunking teams

• Promoting media literacy to help citizens identify AI manipulation

• Establishing verified channels for crisis communication

Taiwan, for example, has a one-hour response window to debunk viral misinformation, particularly during election season. It works by combining civil society fact-checkers, government teams, and creative public messaging.

We don’t need to be spectators

From Malaysia, to Kenya, to Peru — we watch the US and China race towards building humanity’s most transformative technology. But this isn’t just their story. It’s ours too.

History shows that transformative technologies reshape the world in unexpected ways, with developing nations paying the cost of transition. The climate crisis propelled Western nations forward with disproportionate environmental impacts on the rest of the world. Social media algorithms developed in Silicon Valley became a tool for genocide in Myanmar. AI training exploits data workers from low income countries for pennies.

The development and deployment of AGI is no different.

AGI threatens catastrophic terrorism, accelerated war, economic disempowerment, power concentration, but we can prepare. The work starts now. Not with billion-dollar investments or sweeping political change, but with:

• One government official pushing for AI risk to be a priority

• One doctor advocating for disease monitoring at their hospital

• One teacher adding cybersecurity and deepfake detection into their curriculum

• One engineer red-teaming critical infrastructure against AI attacks

AGI development may be concentrated in a few countries, but building resilience is a distributed challenge that we can all contribute to.

We build the best defences against risks we understand. Check out BlueDot’s free course: The Future of AI — designed to help people everywhere understand AI’s impact and contribute to the conversation.

I stumbled through figuring out how to do this efficiently while getting quick, actionable results. After cycling through 5+ tools and running ~10 surveys with 20 participants each, I’ve identified what you should actually look for if you’re doing this yourself without a research team.

Caveat: this isn’t comprehensive research, just me trying out a bunch of stuff I stumbled on. This is also not sponsored (but maybe you want to?) Here’s what I learned.

Start Here: Core Principles

Before diving into the details, here are the key principles I wish I’d known from the beginning:

• Start small and iterate - Test with 5 participants before scaling to catch issues early

• Participant quality trumps quantity - The right 20 people provide more insight than 500 wrong ones

• Platform interfaces matter enormously - They can make or break your efficiency

• Analysis tools are separate decisions - Don’t assume the platform’s built-in analysis will be sufficient

Choosing the right participant panel

Quality matters above all else

Getting feedback from the wrong people will lead you astray. You need filters to proxy for the people you want to hear from and to block repeat respondents across similar surveys. The specificity of available filters depends on where they source their “panels”, whether they’ve cultivated their own (this is rare!) or are using a third-party user recruitment platform like Prolific or User Interviews.

In the absence of filters, you’ll need to set up custom screening questions, though this sometimes requires you to pay for the user’s time to complete them.

While most platforms tout “AI quality checks”, if you’re unfamiliar with the panel source it’s worth doing a small test run (5-10 participants) to verify response quality before committing a larger portion of your budget. This small investment can save you from wasting your entire research budget on low-quality data.

Timing affects participant demographics

Most platforms promise results within hours, but when you post your survey significantly affects who responds. If you post at 2am and get responses within 2 hours, you’re likely getting night owls, international participants, or people who rush through surveys for extra income - not typically representative of your target audience.

Consider scheduling your survey release to align with when your target demographic is most likely to be active online.

Designing effective questions for better insights

Start small and iterate

If you haven’t had much experience with designing surveys, it’s crucial to start with ≤5 participants for your first survey to validate your question phrasing and order. Since most platforms deliver results within hours, you have time to refine your approach based on initial responses. Better to make mistakes with 5 participants than 500.

Also, question whether you need statistical significance before scaling up. Often, open-ended qualitative questions to 10-20 people reveal deeper insights than hundreds of quantitative responses. For tight budgets, this approach delivers more bang for your buck.

Features to get more depth

Most participants will give a few words or a single sentence in open question responses. To extract richer insights, I’ve found two approaches particularly effective:

1. Record participants’ verbal responses to questions, then transcribe them (more time-intensive but often more natural)

2. Have an AI ask follow-up questions (a feature Maze has!) to probe deeper based on initial responses

I’ve consistently found that higher-quality participants also correlate with more insightful responses - another reason not to skimp on panel quality.

Selecting a survey interface

I was surprised that most user research platforms don’t have their own survey interface! They link to third-party platforms like Qualtrics or SurveyMonkey, sometimes requiring you to figure out integration yourself.

For whichever survey design interface you choose, check for these surprisingly rare but valuable features:

• Ability to create randomised split tests

• Straightforward creation of custom screeners (the logic here can be unexpectedly complicated)

• Support for having participants engage with content (videos, articles, images) before answering questions

• Ability to reference responses to previous questions for conditional survey flow

The right interface can save you hours of frustration and enable more sophisticated research designs that would otherwise require programming knowledge.

What’s missing: better guidance on question design

As a novice in survey design, I would have greatly benefited from integrated guidance on question phrasing to avoid leading questions or create comprehensive option lists. This is a gap across all platforms I tested.

Instead, I relied heavily on ChatGPT and Claude to help refine my survey questions - a workaround I recommend to others until platforms improve this aspect.

Cost considerations

Survey costs typically have two components: participant compensation and platform fees. And a hidden third cost: your own time.

Participant compensation

For a 20-minute survey, cost ranged from $5 to $20 per participant depending on targeting specificity. This variation makes a huge difference when scaling up your research.

Some platforms allow you to review and approve each response before payment, which is invaluable if you’re on a tight budget or concerned about response quality. This feature alone can save you from wasting money on useless data.

If you already have access to your target audience, many platforms also support bringing your own participants, saving you platform fees.

Platforms

Many platforms frustratingly require sales calls before you can even try their product. This might be because they want to upsell their research team or reporting add-ons or get you signed onto a contract. (I’m not sure)

Platform pricing models vary widely:

• Subscription-based with monthly or annual contracts ($100-$3000+/month)

• Pay-as-you-go up to certain usage levels

• Combined models with basic access plus upgrade fees

If you’re just getting started, I recommend starting with pay-as-you-go options until you’ve established your ongoing research needs.

Time

When evaluating platforms, I drastically underestimated how much time I’d spend on platform-specific tasks rather than actual research design.

Some platforms force you through an exhausting laundry list of decisions:

• Creating enticing survey titles to attract participants

• Comparing your survey name against existing studies to avoid confusion

• Writing promotional descriptions that sound like ad copy

• Setting and justifying compensation rates

• Crafting welcome messages and custom instructions

• Manually configuring question types that should have intelligent defaults

• Navigating through multiple non-intuitive interfaces to accomplish simple tasks

These seemingly trivial decisions add up dramatically. Platforms which minimise these micro-decisions with sensible defaults, allow you to focus on the questions that actually matter to your research.

Time efficiency isn’t just convenience—it’s the difference between being able to conduct multiple research iterations or being stuck with configuration. When evaluating platforms, I now recommend doing a test run with the stopwatch running from “I want to create a survey” to “My survey is live with participants” to measure the true time cost.

Making sense of the data

After testing multiple platforms, I’ve concluded that built-in analysis tools are universally disappointing, particularly for qualitative data. You’re better off exporting your data and using a dedicated AI tool. Here’s my ranking of AI analysis options:

#3 Claude 3.7 Sonnet

Despite its strength in writing, Claude struggles with CSV files (likely timing out before processing completion). Not recommended for survey analysis.

#2 NotebookLM

Excellent for handling multiple audio or video files with automatic transcription. I found it valuable for identifying broad themes across dozens of files.

However, for CSV analysis and generating surprising insights, it’s less impressive than my top choice.

#1 OpenAI’s o3

The most effective tool for meaningful insights from survey data. While it can’t handle large numbers of files, it excels when fed the themes from NotebookLM along with your CSV data.

One caution: o3 often hallucinates when asked to extract representative quotes, so always verify these against your raw data.

Making the right choice

After all my trial and error, here’s the approach I recommend for:

• If you’re just starting out: Start with just 5 participants and scale up to at most 20 until you’re sure you need statistical significance.

• If efficiency is your priority: I found that Maze offered the best end-to-end experience with minimal friction. (I’m not even sponsored!) But beware that after the free trial, the price can get quite steep if you’re on a tight budget!

• If depth of insights matters most: You can also consider running moderated surveys like 1-1 interviews or focus groups, which I didn’t cover here.

• For all cases: Plan to export your data and analyse it with OpenAI’s o3 rather than relying on platform analytics

The most important lesson? The differences between platforms aren’t just about features but about how they affect your entire workflow when you’re doing everything yourself. Seemingly small interface differences can save or cost you hours of precious time.

Whatever platform you choose, start small, iterate quickly, and prioritise participant quality over quantity. You’ll get more actionable insights faster and with less frustration.