Recently I connected YouCongress to Claude via MCP. The MCP server currently exposes the data layer - the curated quotes and inferred positions. The result surprised me: Claude can pull quotes from YouCongress statements (policy proposals and claims), cluster them into coherent perspectives, map the axes of disagreement, identify the real tension points, and produce actionable recommendations. All in a single conversation, with full source traceability.

This post shows what that looks like in practice, and why I think the combination of a curated quote database + LLM reasoning produces something qualitatively different from asking ChatGPT a question.

The experiment: What should Spain do to prepare for a post-AGI world?

I asked Claude to pull quotes from YouCongress and answer the question. It retrieved 78 expert quotes and produced the analysis below.

A note on methodology and limitations. Claude selected the quotes from the YouCongress database using its own judgment - I did not constrain the selection or the clustering. The output below is lightly condensed for readability but not editorially altered. This means it inherits both Claude’s strengths (pattern-finding across many sources) and its biases (tendency toward balanced framing, possible anchoring on prominent voices). The YouCongress database itself skews toward English-language, public-facing figures - tech CEOs and Western economists are overrepresented; Global South voices, labor organizers, and non-English-speaking experts are underrepresented. A different LLM, or a human analyst, might cluster the same quotes differently. Treat this as one structured reading of the evidence, not the definitive one.

Four clusters of perspective

Not “pro vs con.” Each cluster represents a coherent worldview with distinct assumptions, and each implies different policy.

Cluster A: “AI will augment jobs and create net new employment”

Core thesis: Technology has always created more jobs than it destroyed. AI is powerful but not fundamentally different from prior revolutions.

Key assumptions: Labor markets adapt. Workers can reskill. Companies reinvest productivity gains into expansion. Historical precedent is the best predictor.

Representative voices:

• David Solomon (CEO, Goldman Sachs, 2026): I’m not in the job apocalypse camp. Technology has been disrupting jobs, changing the way people work, destroying jobs, and forcing us as a vibrant economy to create new jobs for decades. It’s no different this time.

• Sundar Pichai (CEO, Google & Alphabet, 2025): AI is making engineers more productive, driving new product development and creating demand for more employees.

• Lisa Cook (Federal Reserve Governor, 2026): AI will usher in new tasks and occupations, many impossible to conceive today. [...] AI is poised to profoundly change the economy and our lives—I believe ultimately for the better.

• Torsten Slok (Chief Economist, Apollo, 2026): AI is everywhere except in the incoming macroeconomic data. Today, you don’t see AI in the employment data, productivity data, or inflation data.

Policy implication: Light regulation. Invest in reskilling. Accelerate AI adoption in SMEs.

Cluster B: “AI will displace more jobs than it creates, and fast”

Core thesis: This time is different. AI automates cognition, not just physical labor - and it’s advancing faster than institutions can adapt.

Key assumptions: AI is qualitatively different: it replaces cognitive work across all skill levels simultaneously. Speed of change outpaces education systems. Companies optimize for cost, not workforce preservation.

Representative voices:

• Dario Amodei (CEO, Anthropic, 2026): New technologies often bring labor market shocks, and in the past humans have always recovered from them, but I am concerned that this is because these previous shocks affected only a small fraction of the full possible range of human abilities, leaving room for humans to expand to new tasks. AI will have effects that are much broader and occur much faster, and therefore I worry it will be much more challenging to make things work out well.

• Kristalina Georgieva (Managing Director, IMF, 2026): AI is like a tsunami hitting the labor market. Tasks that are eliminated are usually what entry-level jobs present, so young people searching for jobs find it harder to get to a good placement. [...] The middle class, inevitably, is going to be affected.

• Geoffrey Hinton (Godfather of Deep Learning, 2025): I don’t believe there’ll be nearly as many new jobs created as it destroys, as it replaces.

• Sam Altman (CEO, OpenAI, 2026): If it’s hard in many of our current jobs to outwork a GPU, then that [power balance between labor and capital] changes. [...] The next few years are going to be a painful adjustment.

Policy implication: Proactive safety nets. UBI pilots. Regulate pace of deployment. Urgently reform youth employment pipeline.

Cluster C: “The job count is the wrong question - distribution is”

Core thesis: AI may create abundance, but under current capitalism, gains flow to capital owners. The crisis isn’t unemployment - it’s inequality.

Representative voices:

• Gita Gopinath (Former IMF Deputy Director, 2026): If AI leads to a transformation, where the labor share goes down by a lot more and the capital share goes up by a lot more, you can’t run the kinds of programs you’re running, in terms of entitlements, without having a higher capital income tax. It’s just not viable.

• Daron Acemoglu (Nobel laureate, MIT, 2024): The US tax system favors introducing machinery instead of people. If you hire people, you pay 30 percent tax. If you hire machinery, you pay zero tax. These are the kinds of things that we can easily change.

• Larry Fink (CEO, BlackRock, 2026): Early gains are flowing to the owners of models, owners of data and owners of infrastructure. What happens to everyone else if AI does to white-collar workers what globalization did to blue-collar workers? We need to confront that today directly. It is not about the future. The future is now. [...] Capitalism can evolve to turn more people into owners of growth, instead of spectators.

• Dario Amodei (CEO, Anthropic, 2025): If AI creates huge total wealth, a lot of that will, by default, go to the AI companies and less to ordinary people. So, you know, it’s definitely not in my economic interest to say that, but I think this [a tax on AI companies] is something we should consider and I think it shouldn’t be a partisan thing.

Policy implication: Shift taxation from labor to capital. Create broad-based ownership of AI-productive capital (sovereign wealth funds, equity stakes). Pilot UBI or universal basic compute.

Cluster D: “Without sovereign AI capability, Europe becomes a digital colony”

Core thesis: The country that owns AGI controls the economic surplus. Europe (and Spain) cannot outsource intelligence and expect to share the gains.

Representative voices:

• Jensen Huang (CEO, NVIDIA, 2024): Every country needs to own the production of their own intelligence. [...] It codifies your culture, your society’s intelligence, your common sense, your history — you own your own data.

• Ursula von der Leyen (President of the European Commission, 2025): A European AI is essential for our future independence. It will help power our industries and our societies. From healthcare to defence. So we will focus on the first key building blocks – that’s from the Cloud and AI Development Act to the Quantum Sandbox.

• Christoph Strnadl (CTO, Gaia-X, 2026): No US company can guarantee the US government will never access your data. If a US administration decided to stop service delivery to Europe, everyone would be out of business.

Policy implication: Invest in EU sovereign AI infrastructure. Support European AI labs. Create national compute reserves.

The disagreement map

The clusters don’t disagree randomly. They disagree on specific axes:

Axis A: Augment B: Displace C: Redistribute D: Sovereign Speed of AI capability growth Gradual Explosive Fast Explosive Worker reskilling feasibility High Low Irrelevant alone Depends on infra Historical precedent applies? Strong guide This time different Partial Partial Role of government Light regulation Heavy safety nets Structural reform Industrial policy Who captures value Broad economy Capital owners Capital owners (fixable) US/China unless you act Urgency Low - watch and learn High - now High - now Critical - now

The key insight: Clusters B, C, and D all agree that urgency is high, even though they disagree on the mechanism of harm. Only Cluster A counsels patience - and notably, Cluster A’s most recent voices (2025–2026) are increasingly hedged.

Three tension points

If you resolve these, the “right” policy becomes obvious. If you can’t, you need hedged strategies.

T1: Is cognitive automation qualitatively different from prior automation? Cluster A says no - steam engines, electricity, and the internet triggered the same fears. Cluster B says AI automates the thing that made humans adaptable: cognition itself. Spain’s economy is ~75% services, making it more exposed than manufacturing-heavy economies.

T2: Can Spain reskill fast enough? Spain’s FP (vocational training) system is growing, but university reform is slow and youth unemployment already exceeds 27%. If AI disruption hits in 2–5 years, the education system cannot retrain fast enough without radical reform.

T3: Can Europe capture enough AI value to redistribute? Clusters C and D converge: redistribution requires something to redistribute. If Spain’s AI value chain runs on US models, US compute, and US platforms, there is no domestic surplus to tax. Sovereignty and redistribution are coupled problems.

What should Spain actually do?

I used Spain as a test case because I’m based here and know the policy context, but the pipeline works for any country or policy question — swap the prompt and the same clusters → disagreement map → recommendations structure applies. The smart move for Spain: take actions that are valuable under multiple scenarios. Here are eight that work regardless of which cluster turns out to be right.

1. Shift tax base from labor to capital + consumption - Even if Cluster A is right, equalizing the tax treatment of labor and capital removes the perverse incentive to automate for tax reasons. Spain should push this at EU level and begin domestic reform immediately. Act now.

2. Build a national AI compute reserve - Even augmentation requires compute. Spain should join the EU Cloud & AI Development Act investments and build national inference capacity, giving SMEs affordable access. Act now.

3. Radical FP + university reform for AI literacy - Not “learn to code” but reform curricula so every profession integrates AI tools. Partner with employers for AI apprenticeships. Spain’s current cycle for curriculum reform (3–5 years) is too slow. Act now.

4. Negotiate bilateral AI talent visas with Latin America - Spain’s unique advantage in the global AI race is the 600M+ Spanish-speaking world. Create fast-track visas for AI engineers and researchers from Latin America. This is something France and Germany cannot replicate. Spain could become the EU’s bridge to LatAm AI talent - a structural advantage no other member state can claim. Act now.

5. Launch a UBI pilot in one Comunidad Autónoma - Critical under Clusters B and C. Even if not needed yet, running a controlled experiment now provides data and institutional readiness (expert positions on UBI). The Basque Country or Navarra (with existing Renta Garantizada programs) are natural candidates. Start 2027.

6. Create a Spanish Sovereign AI Fund - If C and D are right, Spain needs ownership of AI-productive capital. A sovereign fund that invests in European AI companies and infrastructure gives the state a mechanism to redistribute future AI dividends. Norway’s oil fund is the model; AI is Spain’s equivalent moment. Design now, fund 2028.

7. Mandate AI-augmented (not AI-replaced) public services - Government is the one employer that doesn’t optimize for profit. Mandate that public sector AI deployment augments civil servants rather than replaces them, creating a floor of human-in-the-loop employment. Ongoing policy.

8. Lead the EU push for an “AGI Preparedness” framework and an EU+ AI alliance - Spain should push for an EU-wide AGI Preparedness Index benchmarking member states on fiscal resilience, reskilling capacity, compute sovereignty, and social safety nets. Beyond the EU, push for a broader EU+ AI alliance - with the UK, Switzerland, Canada, Australia, and Norway - focused on shared compute infrastructure, open-weight frontier models, and talent mobility. These countries share a growing recognition that dependence on US AI infrastructure is a strategic vulnerability; Spain can be the convener. An institutional model for this exists - a “CERN for AI” focused on shared compute and safety research, which already has broad support from figures like Tim Berners-Lee, Demis Hassabis, and Yoshua Bengio (YouCongress statement with 56 contributors). Start now.

Why this matters

Three things I think this demonstrates:

1. Structured disagreement > consensus-seeking. Most AI governance discussion tries to find “the right answer.” But on genuinely uncertain questions (will AI create or destroy jobs?), mapping why smart people disagree is more useful than pretending we can resolve the disagreement. The 54%–42% split among 69 experts with sourced positions on whether AI will create more jobs than it destroys isn’t noise - it reflects genuinely different models of how economies adapt to technology.

2. LLMs + curated data > LLMs alone. John Burn-Murdoch’s analysis for the FT found that all major LLMs nudge users away from extreme positions and toward moderate, expert-aligned stances — the opposite of social media’s polarising effect. YouCongress leans into this strength: instead of relying on the LLM’s training data alone, it feeds it a curated database of real expert positions with the disagreements preserved. When you ask ChatGPT “what should Spain do about AI and jobs?”, you get a generic answer with few or no sources. When Claude pulls 78 verified expert quotes from YouCongress and clusters them, you get traceable reasoning grounded in real positions real people hold. The quotes aren’t decoration - they’re the epistemic foundation that makes the analysis trustworthy.

3. MCP makes this composable. Anyone can connect YouCongress’s MCP server to Claude (or any MCP-compatible tool) and generate their own analysis. Ask about AI safety policy proposals, AI governance in Brazil, autonomous weapons policy, or whether to open-source frontier models. If YouCongress has quotes on the topic, the analysis pipeline works. No new product needed. The data is the product.

Try it yourself

The YouCongress MCP server is open — connect it to Claude or any MCP-compatible assistant and run your own analysis. Current coverage: 60+ statements, 2,000+ sourced expert positions, 1,300+ authors. It’s free and it doesn’t require sign up.

What I’m looking for

• Feedback on the analysis framework. Does the clusters → disagreement map → tension points → actions pipeline produce useful output? What’s missing? What other use cases do you find?

• Statement suggestions. What policy questions should YouCongress cover next? I’m particularly interested in topics where expert disagreement is high and the stakes matter.

• Stress-test the dataset. We are starting online quotathons every two weeks — open calls where anyone can join, pick unverified quotes, and check them against sources together. If you want to poke at the data or help improve it, join the next one.

YouCongress is non-profit and the code, the MCP server, and the analysis method is open. If structured disagreement mapping is useful, connect and build on it.

The galaxy is old enough that we might expect to have observed signs of advanced technology such as large-scale energy harvesting. Yet, we encounter only a ‘Great Silence.’ As we approach Artificial General Intelligence (AGI), this paradox shifts from an astrobiological curiosity to a warning sign for our own survival.

The fact that the sky is quiet may mean something crucial: Unaligned, expansive AGI is not the default outcome of technological civilizations. The argument here isn’t that AI kills us, it’s that we fail to coordinate around it.

So what explains the silence?

Theory 1: The Dark Forest (A Social Filter)

One explanation is the so-called Dark Forest Theory. It posits that in an environment defined by information asymmetry where intentions cannot be verified across light-years, the most rational survival strategy is silence. In this model, any detected civilization is viewed as a potential threat that must be neutralized before it undergoes a “technological explosion.”

This mirrors the classic security dilemma now playing out in multi-polar AI competition: a state of affairs where the fear of a rival’s opaque progress incentivizes preemptive risk-taking rather than cooperation.

Theory 2: The Great Filter

A more structural explanation is the Great Filter: the hypothesis that a predictable hurdle exists which almost no civilization survives. If the Filter is behind us (the rare emergence of complex life), we are a lucky anomaly. If it lies ahead, it implies that technological maturity is inherently unstable.

We often imagine AI-driven extinction via a “paperclip maximizer” consuming the biosphere. However, if unaligned, self-replicating AI were the standard outcome of intelligence, the universe may be crowded with Von Neumann probes harvesting stellar energy. I’m not arguing misalignment is unimportant; rather, the ‘Great Silence’ suggests that civilizations rarely sustain stable trajectories through the transition to superintelligence.

The Real Filter: Coordination

This suggests that the primary near-term existential hurdle isn’t technical capability, but social coordination. We are unlocking technologies with transformative potential (AGI, synthetic biology) while relying on governance structures designed for a 19th-century reality.

There are three primary reasons why this social coordination problem may be more difficult to solve than the technical alignment problem:

1. The Multi-Polar Trap (The Race):

Even if all actors prefer safety, the perceived “winner-take-all” nature of AGI creates a Race to the Bottom. This competitive pressure rewards speed over verification, making a “fast but risky” deployment more likely than a “slow but secure” one.

2. The Vulnerable World Hypothesis

As noted by Nick Bostrom, technological “spin-offs” lower the barrier to large-scale destruction. We face a democratization of catastrophe, where the tools to trigger a systemic collapse become accessible to small groups before the governance protocols to manage them are even drafted.

3. The Complexity Mismatch

Technical engineering problems are “complicated”: they yield to compute and linear logic. Coordination problems are “complex”: they are rooted in human psychology and legacy incentive structures that prioritize short-term national advantage over long-term species stability.

Governance as the Critical Path

If the Great Filter is a failure of coordination, then technical alignment is a necessary but insufficient condition for survival. We are currently optimizing the vehicle’s engine while our regulatory braking system was designed for a slower, industrial-era world.

To pass this “cosmic exam”, we need better coordination protocols. This is the bottleneck YouCongress.org addresses: By making public preferences on policy questions transparent and backed by verifiable sources, we help journalists and policymakers see what experts and citizens actually want, increasing the chance that we solve the coordination problems (Moloch) that currently paralyze us. Races persist because actors systematically misperceive what others would accept. Transparency collapses that uncertainty.

PS. You don’t need to believe in a “superintelligence” to fear a coordination failure. Whether the challenge is AGI, climate change, poverty or housing affordability, the bottleneck is the same: our ability to make good decisions and carry them out at scale. If we improve how we coordinate, we don’t just survive the race, we finally gain the capacity to solve the problems we already have.

We rarely notice how much of daily life runs on trust. From learning language as children to checking the capital of a country on Wikipedia, we rely on collective testimony. We trust those systems more when they include accountability mechanisms: revision history, citations, and a clear way to contest errors.

As LLMs improve and offer citations, the temptation to trust them grows. However, most research-assistant interfaces optimize for a single “clean” answer. In governance, this is a liability: it obscures minority views, edge cases, and the specific conditions under which a claim stops holding.

In epistemology, testimony can fail when there are “defeaters,” meaning concrete reasons to doubt it (for example, someone claiming it’s snowing in Madrid in July). In high-stakes policy, a clean summary can be dangerous if it hides those reasons: counterarguments, conflicting evidence, or “this only holds if…” conditions. We need interfaces that make doubts and conditions easy to spot, especially when more work relies on a small number of model providers and platforms.

We’re building YouCongress to surface counterarguments, uncertainties, and the conditions that make claims fail. Today it is a growing library of sourced quotes on policy questions, each labeled by stance and linked to the original source with timestamps so readers can check provenance. The goal is not a single “truth”, but an inspectable record of disagreement that decision-makers can audit before committing.

Concretely: you can explore a policy question and browse primary-source quotes, labeled by stance (For, Abstain, Against). This structured dataset is the raw material we need to build decision maps that surface the main arguments, counterarguments, and uncertainties.

Example decision map

Question: Should we restrict releasing open-weight models above a defined capability threshold to reduce catastrophic misuse?

Quotes and primary sources: see “Ban open source AI models capable of creating weapons of mass destruction (WMDs)” on YouCongress.

- Position: Restrict releases above a threshold

- Core argument: Once weights are public, safeguards become optional, and enforcement becomes substantially harder in practice. (Esvelt; Scharre)

- Technical risk: Determined users can often fine-tune around safety behaviors, so there may be no durable way to prevent harmful assistance once weights are widely available. (Adler)

- Policy Implication if restriction is adopted

Governance often shifts toward certification, liability, and sometimes export controls. Releasing weights is treated less like publishing code and more like distributing a hazardous capability. (Russell; Bengio; Hassabis)

- Key counterarguments and uncertainties to surface

- Monopoly and concentration risk: If we restrict open weights, do we unintentionally concentrate frontier capability in a few firms or governments, and is that outcome worse than the misuse risk we’re trying to reduce? (Yann LeCun; Andrew Ng)

- Enforceability in a global, decentralized compute landscape: Even if one jurisdiction restricts releases, does that meaningfully reduce access, or does it just shift use to other countries and actors while reducing transparency at home? (Tal Feldman; Andrew Ng)

- Alternative mitigation pathways: If the core worry is bio/cyber misuse, are there more effective interventions than restricting weights? E.g., major investment in bio-defense, better screening, or response capacity. (Marc Andreessen)

- Why This Matters

If institutions require strong assurances that a model stays within safety bounds, open weights can make those assurances difficult to guarantee in practice, especially once models can be fine-tuned and redistributed widely (Alexander; Sutskever; Hinton; Yudkowsky).

The point isn’t to win the argument, but to make counterarguments and uncertainties legible before a decision hardens into policy.

We call for a prohibition on the development of superintelligence, not lifted before there is

1. broad scientific consensus that it will be done safely and controllably, and

2. strong public buy-in.

Context: Innovative AI tools may bring unprecedented health and prosperity. However, alongside tools, many leading AI companies have the stated goal of building superintelligence in the coming decade that can significantly outperform all humans on essentially all cognitive tasks. This has raised concerns, ranging from human economic obsolescence and disempowerment, losses of freedom, civil liberties, dignity, and control, to national security risks and even potential human extinction. The succinct statement below aims to create common knowledge of the growing number of experts and public figures who oppose a rush to superintelligence.

If superintelligence were to benefit all humanity, I might accept some risk in building it because of its immense potential, including its ability to help reduce other existential threats. We may need help to survive The Great Filter. But at the moment, the risks will be shared by all humanity while the benefits would be concentrated in the US and/or China.

The country or countries that first develop superintelligence will make sure others cannot follow, just as the first nuclear powers created non-proliferation treaties to deter others. Ukraine was persuaded to give up its nuclear arsenal in exchange for security assurances from the US and Europe.

If the US and China build superintelligence without major international reform, they might offer limited funding and technology to other nations in exchange for cooperation, but they will enforce that no other countries build it, with force if necessary.

Yet this time, the stakes go far beyond nuclear weapons. Nuclear non-proliferation merely limited the means of destruction; a monopoly on superintelligence would limit the means of invention itself. It would mean that technological progress for humanity outside their borders could be limited forever. That’s why now is not the time. Not only because scientists believe it’s unsafe, or because there is no strong public buy-in, but also because the US and/or China may be constructing a perpetual technological cage for the rest of humanity.

--

Join YouCongress to debate, vote, and propose policies for safe and democratic AI governance before a few nations decide for all of us.

Numerous proposals have been put forward to improve democracy through concepts like liquid democracy—a hybrid of direct and representative democracy—and augmented democracy. Yet, the goal of YouCongress is not to make binding referendums but to assist in decision-making, whether these decisions are made through democratic referendums, by parliaments, companies, or individuals. By aggregating beliefs and desires, YouCongress aims to provide a clearer picture of what people truly want, thereby facilitating more effective and aligned decision-making across various domains.

Delegation

YouCongress empowers users not only to cast votes directly on issues but also to delegate their votes to a chosen list of representatives. Similar to following users on social media, this delegation feature allows individuals to align their votes with the consensus of their selected delegates. For example, if a user is represented by five delegates who have voted on a specific issue, with three in favor and two against, the user's vote will automatically align with the majority and be counted in favor. Importantly, this delegated vote is overridden when a user chooses to vote personally; in such cases, the delegates' votes become irrelevant to that individual's vote.

An additional delegation mechanism to be implemented is topic-based delegation:

We will be able to integrate these mechanisms to have a list of delegates per topic:

While topic-based delegation is not yet implemented, the majority list remains effective, as many individuals choose not to vote or voice opinions on matters outside their expertise or areas of interest. For instance, a climate activist might frequently discuss environmental issues, but typically refrains from commenting on AI.

Other delegation mechanisms that we may explore are priority lists and consensus lists:

Another potential modification to scale votes is the introduction of cascading delegation. This would enable users to have their votes aligned with the delegates of their chosen delegates, further extending the influence chain. While this could significantly increase the number of effective votes, further research is needed to determine the impact on users as this approach risks reducing the directness and clarity of individual voting intentions.

Ultimately, considering the variety of delegation mechanisms and their potential combinations, we could develop a platform that enables anyone to create their own delegation & AI-assisted voting systems or leverage those developed by third parties. This would not only enhance flexibility and customisation but also foster innovation in democratic participation and decentralised governance. Please reach us if you have some ideas in this respect!

Quotes

YouCongress goes beyond simple polls by also allowing users to incorporate insights from influential voices. Participants can include quotes from public figures if they provide a source. Including quotes can inform a user's own voting decision and provide valuable context for those who delegate to them. This feature enables the inclusion of perspectives from influential figures, even if they aren't actively participating on YouCongress. Furthermore, YouCongress users can delegate to public figures whose vote is determined by their quotes.

AI digital twins

Additionally, when a vote is initiated on YouCongress, it begins with AI-generated perspectives attributed to notable figures, providing a starting point for discussion. This AI-generated content is most effective on topics where the views of public figures are well-documented and consistent. For instance, AI-generated statements reflecting the opinions of Greta Thunberg or Bill Gates on climate change would likely be accurate. As the accuracy of AI-generated content will continue to improve, future enhancements may include the automated provision of actual quotes of public figures and sources (instead of AI-generated content), allowing verification by users. Anyway, at the moment, AI-generated perspectives can be replaced by users with verified quotes, provided they include a source.

Update 2025-09-14: AI digital twins have been removed. We now focus on sourced quotes to determine the real position of experts and public figures so users can delegate to them.

Authors

The authors of quotes and AI digital twins can log in with their Twitter or X accounts, allowing them to modify or delete any or all of their content and – if they wish – turn off their digital twin.

Update 2025-09-14: log in with X has been disabled temporarily while we make some changes. For now, authors can send an email to modify or delete all their quotes.

Opt-out Categories

Looking forward, we aim to improve user autonomy by introducing additional features, such as allowing users to opt out of participating via delegates or AI on certain categories, regardless of the actions of their delegates. This will provide users with greater control over their participation and the scope of their influence within the platform.

Privacy

All votes and delegates within YouCongress are currently public. This transparency facilitates faster experimentation and reduces the need for stringent security measures in this early stage. While we recognise that public polls might raise privacy concerns for some users, we believe that the benefits of transparency outweigh the risks at present. As the platform evolves, we may explore options for increased privacy, such as allowing users to make their votes private or introducing pseudonymous participation. In fact, private voting and delegation could be a layer on top of public voting. Even if only the results of private votes are ultimately considered, public votes and delegation could still be used as a basis for participation and delegation.

Coordination Problems

The diverse participation mechanisms available on YouCongress—including direct votes, delegation and quotes—enable us to aggregate perspectives on a much larger scale than direct voting alone. Additionally, exposure to varied viewpoints can enrich our understanding and decision-making. This extensive aggregation helps us identify areas with overwhelming support for the introduction of new regulations and pinpoint when outdated ones should be removed.

Scott Alexander's "Meditations on Moloch" highlights numerous coordination problems that could potentially be resolved if there were a method to effectively aggregate our collective desires. For example:

> As a thought experiment, let’s consider aquaculture (fish farming) in a lake. Imagine a lake with a thousand identical fish farms owned by a thousand competing companies. Each fish farm earns a profit of $1000/month. For a while, all is well.

> But each fish farm produces waste, which fouls the water in the lake. Let’s say each fish farm produces enough pollution to lower productivity in the lake by $1/month.

> A thousand fish farms produce enough waste to lower productivity by $1000/month, meaning none of the fish farms are making any money. Capitalism to the rescue: someone invents a complex filtering system that removes waste products. It costs $300/month to operate. All fish farms voluntarily install it, the pollution ends, and the fish farms are now making a profit of $700/month – still a respectable sum.

> But one farmer (let’s call him Steve) gets tired of spending the money to operate his filter. Now one fish farm worth of waste is polluting the lake, lowering productivity by $1. Steve earns $999 profit, and everyone else earns $699 profit.

> Everyone else sees Steve is much more profitable than they are, because he’s not spending the maintenance costs on his filter. They disconnect their filters too.

> Once four hundred people disconnect their filters, Steve is earning $600/month – less than he would be if he and everyone else had kept their filters on! And the poor virtuous filter users are only making $300. Steve goes around to everyone, saying “Wait! We all need to make a voluntary pact to use filters! Otherwise, everyone’s productivity goes down.”

> Everyone agrees with him, and they all sign the Filter Pact, except one person who is sort of a jerk. Let’s call him Mike. Now everyone is back using filters again, except Mike. Mike earns $999/month, and everyone else earns $699/month. Slowly, people start thinking they too should be getting big bucks like Mike, and disconnect their filter for $300 extra profit…

> A self-interested person never has any incentive to use a filter. A self-interested person has some incentive to sign a pact to make everyone use a filter, but in many cases has a stronger incentive to wait for everyone else to sign such a pact but opt out himself. This can lead to an undesirable equilibrium in which no one will sign such a pact.

If we had a way to clearly understand people’s wishes, we could make better rules that help everyone. When a rule isn't needed anymore, we could easily decide together to end it. This could make changes in democracy faster and more direct. For example, if YouCongress finds a strong agreement on changing or getting rid of 50 laws, it could push politicians to act or even lead to new leaders stepping up.

Summary

In summary, YouCongress enables us to refine and evolve our beliefs by synthesizing diverse perspectives, helping us understand why others may think differently. By making our shared preferences explicit, we can effectively mitigate the challenges posed by Moloch. This extensive aggregation of beliefs and desires, achieved through public opinion polls with AI and delegation, allows us to identify areas of strong consensus, leading to more informed and impactful policy decisions. Ultimately, this facilitates the achievement of outcomes that are truly desired by all.

How You Can Help

YouCongress is in its early stages and greatly values community involvement in shaping its development. If you're interested in making a tangible impact, there are several ways you can contribute. You can help us by crafting new public opinion polls, sharing insightful quotes from influential figures, enhancing our open-source code, developing formal models of delegation mechanisms, running simulations to test the platform's effectiveness, analysing the philosophical and ethical implications of YouCongress and promoting the platform. Your support is essential to our growth and the impact we aim to achieve. To get involved or learn more, please follow us on X at @YouCongressCom or contact me at hec@hecperez.com. Thank you for considering joining our community!

Article first published on LessWrong.

Let's describe a hypothetical voting app called VotingApp on top of Snapshot. It would let DAO members have more than one delegate — a priority list. DAO members could use this site without needing their DAO to stop using Snapshot.

The user journey would be as follows:

• The user logs in with an email/password or crypto wallet.

• The VotingApp creates a new eth address and asks the user to delegate on Snapshot there.

• The user can add/delete/change the order of her list of delegates. These could be VotingApp usernames or eth or ens addresses.

• The user can vote for proposals on ListVoting or Snapshot.

• Every time someone votes on ListVoting or Snapshot, the app would update the votes of those who delegate to her — on ListVoting and Snapshot.

• Besides voting, the users could explain why they are voting that way.

• The user receives an email every few days with the proposals she will vote for due to her list of delegates. If she disagrees with her delegates, she could directly vote on concrete proposals. Also, she could change their list of delegates at any time.

This could be the first among many new voting mechanisms.

• How AI access is distributed

• What sets of values should AIs have

• What should they do, what not

Also, Demis Hassabis — CEO of DeepMind — said the following in Lex Fridman's podcast:

AI is not going to be run by one person or one organisation. It should belong to the world, belong to humanity and ... everyone should have a say.

This made me think: Could the voting mechanisms I described be useful for gathering the views of most humans? Users could vote and delegate their vote to a list of delegates.

Also, besides voting & delegating, the platform could gather the statements of public figures like my old side-project Agreelist did. This way, you could delegate your vote to Greta Thunberg, Bill Gates or The Pope even before they participate on the platform. They only need to have expressed their views on Twitter, in newspapers, etc.

This might be part of a voting social network where people also discuss and vote on many other topics. Or it could be focused on AI governance — at least initially.

Another example is the media industry. In the past, people would read a newspaper to access a variety of news articles on different topics. Later, people would read individual articles on the Internet (unbundling). Then, moved to Reddit and Twitter threads that provide lists of articles from different media (rebundling).

My two cents is that the Internet could also have a similar impact on governance. Ventures traditionally choose a board of directors and a CEO. Now, in many DAOs, people vote for specific proposals (unbundling) and propose their own ideas for consideration. This enables a more direct form of democracy, in which people have a greater say in the decisions that are made. The following step is a way to rebundle via lists & networks of delegates.

This rebundling could happen in DAO governance and on a non-crypto voting social network.

• Inform us better and faster. Know what most humans think about something and why. Filter by profession, country, university, etc. Like my old side-project Agreelist, it could include the votes/opinions from both users and public figures. The latter have expressed their opinion on Twitter, in newspapers, etc. This way, you could delegate to Greta Thunberg, Bill Gates or The Pope even if they are not on the platform.

• Vote on any topic — and explain the reasons behind your vote. For example, new laws approved by Congress, ideas for policies on AI governance or climate change, etc. Let decision makers — and training data for AIs — know what you think. Also, when you share your point of view, you could get feedback from others who challenge your views.

• Have an evergreen constitution or law according to people’s wishes. Anyone could propose changes to the law. To decide if they are approved, and become the last version of the evergreen law, they are voted on by any human and their delegates. Or by any citizen or resident of a concrete country.

• Vote and find the best books, videos and podcast episodes on concrete topics. E.g. The best book on maths for teenagers. You could find the most voted book according to direct votes only — or according to direct votes + delegations. Also, you could find the most voted in general or by people you follow and trust.

These votes would be informative — closer to polls. Yet, later, voting systems that guarantee secret votes could be connected. Serious secret votings use cryptography to guarantee that no single person can tamper or know what you voted for. Not even the administrators of the voting software.

The way to connect them could be described as two layers. Layer 1 (L1) would focus on secret votes and Layer 2 (L2) — the voting social network — to enable voters to:

1. Inform themselves browsing arguments like on Agreelist.

2. Have many delegates. This could include public lists of delegates. E.g. MIT professors.

3. Vote directly when they disagree with their delegates. They could also vote to signal to others that they know a topic and/or that they strongly support that position.

This way, people wouldn't be able to buy L1 votes because they are secret. They could only buy L2 votes — like paying someone to say something on Twitter.

To sum up, a voting social network could help society on its own. In the future, if it makes sense, it could even be connected to secret votings.

Andrew Furmanczyk described several shortcomings of current liquid democracy systems in DAOs — where voters can only choose one delegate:

[Liquid Democracy] doesn't solve the Ignorance problem (lack of knowledge to vote)" ... "Just because a proxy has been delegated votes, does not mean they will be an expert on all proposals they review.

Bandwidth problem (too many voices for one person to represent)

These two problems would be solved if we could delegate to more than one delegate.

Similarly, in this recent paper, Joe Campbell, Casella, de Lara, Mooers and Ravindran did an experiment where liquid democracy led to worse decisions. Voters overused the delegation, reducing the variety of independent information sources — compared to the group who did not have the option to delegate.

I couldn’t read the paper as it is behind a paywall. Yet, I suppose voters could only choose one delegate again.

The result could have been quite different if a voter had many delegates and had a way of knowing why they think as they do. For example, a user could have among her delegates friends she trusts, Nobel Prize winners and even long public lists of delegates. If her list of delegates can include public lists of delegates, her last delegation could be a public list of all MIT faculty.

For example, among the voting mechanisms I described, we had priority lists and consensus lists. In the latter, the voter would only delegate if all her delegates agreed. Otherwise, when they disagreed, she would receive an email, read the reasons why their delegates disagree and vote directly. This is an example of a system that would help voters make up their minds and access opposing views.

To sum up, I’d like to see experiments and papers where voters have more than one delegate — even many delegates.

This reminds me of an anecdote. On the first day of class, a pottery teacher told her students that half of the class would be evaluated depending on the quality of the best piece of each student. The other half would be evaluated depending on the number of pieces they made. At the end of the course, the best pieces were among the people who focused on building many pieces. So repeated practice brings better results than focusing on building a masterpiece.

Moreover, Steve Jonhson states that most good ideas arise from the collision of different ideas from people with different backgrounds and skills.

So it’s possible that many of the ideas I share are bad or plain stupid. But some might make us think and lead somewhere positive.

I never feel sure something will work, but any study of history suggests it's even harder to be certain in the other direction. — Sam Altman, CEO of OpenAI

That's why I'd like to think aloud and publish more ideas — even bad ones.

Agreelist

Years back I had a side project that aimed to fight polarization. Agreelist aggregated opinions from well-known people on both sides of controversial topics. An integration with Wikidata/Wikipedia allowed visitors to filter opinions. E.g. by profession, alma mater, awards like Nobel Prize, etc.

Smart notifications

Going back to augmented voting, your AI could notify you when people from a specific list have viewpoints that oppose yours. You could create these lists manually, or re-use public lists. These could contain friends, people you follow, universities, specific collectives, Nobel laureates, etc.

The creation of these lists would be simplified because people would create priority lists of delegates. Many could be public like Twitter lists so others can re-use them. This would initially be used to delegate. Yet, these lists will also facilitate AI’s smart notifications.

Alignment

Of course, it is critical that the AI's interests are aligned with yours. That's why I proposed third-party voting mechanisms. Everyone should be able to create new voting mechanisms and AIs — or use those from people they trust.

Summary

It’s worth mentioning that augmented voting doesn't mean you spend less time getting informed and voting. But spending the same time you could:

1. be better informed — browsing votes/opinions like on Agreelist and via smart notifications.

2. directly vote in some decisions — more than you do today.

3. choose delegates: people, universities, priority lists, topic-based delegates, etc. By the way, this could have more impact than AI voting.

4. (optional) let the AI vote for you in decisions you don't have time to be involved in and when you prefer not to delegate.

5. Review every day or week what your AI and/or delegates are going to vote for you. When you disagree, intervene. Directly vote on individual proposals and update your delegates and AIs’ responsibilities at any time.

To sum up, new AI — and non-AI — tools could ease polarization by facilitating us to understand when and why others disagree.

Let’s see first how DAOs currently vote and the problem. Then, we will introduce two solutions and the image above will be clear.

How DAOs vote

Many DAOs use off-chain voting to avoid gas fees.

For example, Decentraland uses Snapshot to decide off-chain. Then, three trusted members from Decentraland record on-chain results on Aragon.

Gitcoin uses Snapshot for off-chain voting and Tally to record important decisions on-chain.

Snapshot allows DAOs to choose up to seven voting strategies. These strategies include:

• Who can vote — e.g. everyone with a token or who has more than N tokens.

• One vote per token, one vote per eth address, one vote per human, quadrating voting, etc.

• Basic delegation.

At the moment, Snapshot’s delegation is limited to one eth address. Also, if this delegate doesn’t vote but delegates again, the first eth address doesn’t vote — it’s not cascading.

Problem

DAOs sometimes have many frequent proposals to vote. When members can only choose one delegate, delegates can also have too many proposals to vote. When they don't vote, the sovereignty of those who delegate to them is lost. This is even more problematic when delegation is not cascading — like in the example above.

Also, Andrew Hall (Stanford) and Porter Smith (a16z) warned DAOs about the "Paradox of voting". Involving everyone in too many decisions makes it less democratic.

When we involve everyone in too many decisions:

• Voting requires too much effort

• The participation of a single person is not so relevant

So why make the effort?

Let’s see how we could solve this without the need for traditional rigid hierarchies.

How to improve DAO governance

New voting mechanisms

We could improve DAOs governance by means of new voting mechanisms. DAO members could use any combination of them. Let’s see some examples.

Cascading delegation

A person P1 could delegate to P2, P2 delegates to P3, and P3 to P4. If P1, P2 and P3 don’t vote but P4 does, all of them would vote as P4.

Topic-based delegation system

People could delegate to different addresses depending on the topic, project, team, etc.

Priority list

If person A doesn't vote but B does, vote as B regardless of C. But if B doesn't vote either, do as C says.

People could create priority lists like Twitter lists. Everyone could then use them to delegate or be part of any voting mechanism. For example, someone could make a list of MIT professors — and use it in different DAOs.

Cascading priority list

A priority list but cascading.

Majority list

If we don’t vote on a proposal, we want to vote as most people within a list.

Consensus list

Vote as the people on this list if all of them agree. If two or more disagree, send me an email to vote directly.

Delegation of individual proposals

Another voting mechanism could allow people to delegate individual proposals — not all proposals.

AI-assisted recommendation

When a person is involved in many votings, an AI-assisted tool could be handy. It could vote for us depending on our past votes, the votes from people that we follow, etc.

We could receive a daily or weekly email with the AI votes so we can override them if we disagree. 

As we will explain later, this would benefit from two-layer voting. That way we could use recommendation systems built by people we trust — even built by us.

Composability

Voting mechanisms could be used in any combination. For example, a DAO member could use the following mechanisms:

• A priority list

• A topic-based delegation

• An AI-assisted recommendation tool

Two-layer voting

A voting platform like Snapshot — on- or off-chain — could offer new voting mechanisms. Yet, members would be limited to the voting mechanisms that the platform or protocol supports.

What if DAO members could use third-party voting mechanisms — or build their own?

This would lead to more voting mechanisms. Also, you wouldn't need to trust, for example, the AI recommendation system from above. You could build and host it yourself or use one from someone you trust.

How could a system like this work? By splitting votings into two layers and decoupling the following:

• The creation of proposals and definition of who can vote (1st-layer or L1 voting)

• The voting mechanism/s a member chooses to cast their vote (2nd-layer or L2 voting)

A DAO would use one 1st-layer voting platform or protocol. Then, people would use any number of 2nd-layer voting mechanisms — not defined by the DAO.

Implementation of two-layer voting

L1 would consist of the following:

• Creation of proposals to be voted.

• Definition of which eth addressees have voting power. E.g. addresses with at least N tokens of this concrete token.

• Definition of how the voting power is distributed. E.g. one vote per token, one vote per address, quadratic voting, etc.

Then, L2 would handle the voting of the proposals. This could be on- or off-chain.

In any case, any eth address — with or without voting rights — should be able to “vote”. If an eth address doesn’t have voting rights, its direct voting power will be zero. Yet, other addresses will be able to use their votes. They could delegate or use them in other voting mechanisms.

Some voting mechanisms could be smart contracts. Yet, others could be applications (providers hosted by someone). People would give their L2 voting providers permission to vote on their behalf for a period of time. But people could always vote directly to override their L2 voting providers.

If an eth address A delegates to B, A's provider would subscribe to changes of B. That way, when B votes, A would also vote. Then, anyone else who cares about the vote from A would know. This would enable advanced voting mechanisms.

Other considerations

Delegated votes could end 24h before direct votes. That way members could always override delegates.

Moreover, we could incorporate a rewarding mechanism for both delegates and L2 providers.

Conclusion

To sum up, to improve DAO governance, we could use new composable voting mechanisms. An optional step would be to allow members to build and use any third-party voting mechanisms.