d/acc: one year later

2025 Jan 05 See all posts


d/acc: one year later

Special thanks to Liraz Siri, Janine Leger and Balvi volunteers for feedback and review

About a year ago, I wrote an article on techno-optimism, describing my general enthusiasm for technology and the massive benefits that it can bring, as well as my caution around a few specific concerns, largely centered around superintelligent AI, and the risk that it may bring about either doom, or irreversible human disempowerment, if the technology is built in the wrong ways. One of the core ideas in my post was the philosophy of : decentralized and democratic, differential defensive acceleration. Accelerate technology, but differentially focus on technologies improve our ability to defend, rather than our ability to cause harm, and in technologies that distribute power rather than concentrating it in the hands of a singular elite that decides what is true, false, good or evil on behalf of everyone. Defense like in democratic Switzerland and historically quasi-anarchist Zomia, not like the lords and castles of medieval feudalism.

In the year since then, the philosophy and ideas have matured significantly. I talked about the ideas on 80,000 Hours, and have seen many responses, largely positive and some critical. The work itself is continuing and bearing fruit: we're seeing progress in verifiable open-source vaccines, growing recognition of the value of healthy indoor air, Community Notes continuing to shine, a breakout year for prediction markets as an info tool, ZK-SNARKs in government ID and social media (and securing Ethereum wallets through account abstraction), open-source imaging tools with applications in medicine and BCI, and more. In the fall, we had the first significant d/acc event: "d/acc Discovery Day" (d/aDDy) at Devcon, which featured a full day of speakers from all pillars of d/acc (bio, physical, cyber, info defense, plus neurotech). People who have been working on these technologies for years are increasingly aware of each other's work, and people outside are increasingly aware of the larger story: the same kinds of values that motivated Ethereum and crypto can be applied to the wider world.



Table of contents

What d/acc is and is not

It's the year 2042. You're seeing reports in the media about a new pandemic potentially in your city. You're used to these: people get over-excited about every animal disease mutation, and most of them come to nothing. The previous two actual potential pandemics were detected very early through wastewater monitoring and open-source analysis of social media, and stopped completely in their tracks. But this time, prediction markets are showing a 60% chance of at least 10,000 cases, so you're more worried.

The sequence for the virus was identified yesterday. Software updates for your pocket air tester to allow it to detect the new virus (from a single breath, or from 15 minutes of exposure to indoor air in a room) are available already. Open-source instructions and code for generating a vaccine using equipment that can be found in any modern medical facility worldwide should be available within weeks. Most people are not yet taking any action at all, relying mostly on widespread adoption of air filtering and ventilation to protect them. You have an immune condition so you're more cautious: your open-source locally-running personal assistant AI, which handles among other tasks navigation and restaurant and event recommendation, is also taking into account real-time air tester and CO2 data to only recommend the safest venues. The data is provided by many thousands of participants and devices using ZK-SNARKs and differential privacy to minimize the risk that the data can be leaked or abused for any other purpose (if you want to contribute data to these datasets, there's other personal assistant AIs that verify formal proofs that these cryptographic gadgets actually work).

Two months later, the pandemic disappeared: it seems like 60% of people following the basic protocol of putting on a mask if the air tester beeps and shows the virus present, and staying home if they test positive personally, was enough to push the transmission rate, already heavily reduced due to passive heavy air filtering, to below 1. A disease that simulations show might have been five times worse than Covid twenty years ago turns out to be a non-issue today.

Devcon d/acc day


One of the most positive takeaways from the d/acc event at Devcon was the extent to which the d/acc umbrella successfully brought people together from very different fields, and got them to actually be interested in each other's work.

Creating events with "diversity" is easy, but making different people with different backgrounds and interests actually relate to each other is hard. I still have memories of being forced to watch long operas in middle school and high school, and personally finding them boring. I knew that I was "supposed to" appreciate them, because if I did not then I would be an uncultured computer science slob, but I did not connect with the content on a more genuine level. d/acc day did not feel like that at all: it felt like people actually enjoyed learning about very different kinds of work in different fields.

If we want to create a brighter alternative to domination, deceleration and doom, we need this kind of broad coalition building. d/acc seemed to be actually succeeding at it, and that alone shows the value of the idea.

The core idea of d/acc is simple: decentralized and democratic differential defensive acceleration. Build technologies that shift the offense/defense balance toward defense, and do so in a way that does not rely on handing over more power to centralized authorities. There is an inherent tie between these two sides: any kind of decentralized, democratic or liberal political structure thrives best when defense is easy, and suffers the most challenge when defense is hard - in those cases, the far more likely outcome is some period of war of all against all, and eventually an equilibrium of rule by the strongest.

The core principle of d/acc extends across many domains:


Chart from My Techno-Optimism, last year


One way to understand the importance of trying to be decentralized, defensive and acceleration-minded at the same time, is to contrast it with the philosophy that you get when you give up each of the three.

With d/acc, we want to:

Another way to think about d/acc is to go back to a frame from the Pirate Party movements in Europe in the late 00s: empowerment.



The goal is to build a world where we preserve human agency, achieving both the negative freedom of avoiding active interference (whether from other people acting as private citizens, or from governments, or from superintelligent bots) with our ability to shape our own destinies, and the positive freedom of ensuring that we have the knowledge and resources to. This echoes a centuries-long classical liberal tradition, which also includes Stewart Brand's focus on "access to tools" and John Stuart Mill's emphasis on education alongside liberty as key components of human progress - and perhaps, one might add, Buckminster Fuller's desire to see the process of global solving be participatory and widely distributed. We can see d/acc as a way of achieving these same goals given the technological landscape of the 21ˢᵗ century.

The third dimension: survive and thrive

In my post last year, d/acc specifically focused on the defensive technologies: physical defense, bio defense, cyber defense and info defense. However, decentralized defense is not enough to make the world great: you also need a forward-thinking positive vision for what humanity can use its newfound decentralization and safety to accomplish.

Last year's post did contain a positive vision, in two places:

  1. Focusing on the challenges of superintelligence, I proposed a path (far from original to me) of how we can have superintelligence without disempowerment:
    1. Today, build AI-as-tools rather than AI-as-highly-autonomous-agents
    2. Tomorrow use tools like virtual reality, myoelectrics and brain-computer interfaces to create tighter and tighter feedback between AI and humans
    3. Over time proceed toward an eventual endgame where the superintelligence is a tightly coupled combination of machines and us.
  2. When talking about info-defense, I also tangentially mentioned that in addition to _defensiv_e social technology that tries to help communities maintain cohesion and have high-quality discourse in the face of attackers, there is also progressive social technology that can help communities more readily make high-quality judgements: pol.is is one example, and prediction markets are another.

But these two points felt disconnected from the d/acc argument: "here are some ideas for creating a more democratic and defense-favoring world at the base layer, and by the way here are some unrelated ideas for how we might do superintelligence".

However, I think in reality there are some very important connections between what labelled above as "defensive" and "progressive" d/acc technology. Let's expand the d/acc chart from last year's post, by adding this axis (also, let's relabel it "survive vs thrive") to the chart and seeing what comes out:



There is a consistent pattern, across all domains, that the science, ideas and tools that can help us "survive" in one domain are closely related to the science, ideas and tools that can help us "thrive". Some examples:

In addition to this, there are also important cross-dependencies between subject areas:

The hard question: AI safety, short timelines and regulation


Different people have very different AI timelines. Chart from Zuzalu in Montenegro, 2023.


The argument against my post last year that I found most compelling was a critique from the AI safety community. The argument goes: "sure, if we have half a century until we get strong AI, we can concentrate our energies and build all of these good things. But actually it's looking likely we have three year timelines until AGI, and another three years until superintelligence. And so if we don't want the world to be destroyed or otherwise fall into an irreversible trap, we can't just accelerate the good, we also have to slow down the bad, and this means passing powerful regulations that may make powerful people upset". In my post last year, I did indeed not call for any specific strategy to "slow down the bad", beyond vague appeals to not build risky forms of superintelligence. And so here, it's worth addressing the question directly: if we are living in the least convenient world, where AI risk is high and timelines are potentially five years away, what regulation would I support?

First, the case for caution around new regulations

Last year, the main proposed AI regulation was the SB-1047 bill in California. SB-1047 required developers of the most powerful models (those that take over $100M to train, or over $10M in the case of fine-tunes) to take some safety-testing measures before releasing. In addition, it imposed liability on developers of AI models if they take insufficient care. Many detractors argued that the bill was "a threat to open source"; I disagreed, because the cost thresholds meant that it affected only the most powerful models: even LLama3 was probably under the threshold. Looking back, however, I think there was a larger issue with the bill: like most regulation, it was overfitted to the present-day situation. The focus on training cost is proving fragile in the face of new technology already: the recent state-of-the-art quality Deepseek v3 model was trained at a cost of only $6 million, and in new models like o1 costs are shifting from training to inference more generally.

Second, the most likely actors who would actually be responsible for an AI superintelligence doom scenario are realistically militaries. As we have seen in the last half-century of biosecurity (and beyond), militaries are willing to do scary things, and they can easily make mistakes. AI military use is advancing rapidly today (see Ukraine, Gaza). And any safety regulation that a government passes, by default would exempt their own military, and corporations that cooperate closely with the military.

That said, these arguments are not reasons to throw up our hands and do nothing. Rather, we can use them as a guide, and try to come up with rules that would trigger these concerns the least.

Strategy 1: Liability

If someone acts in some way that causes legally actionable damage, they could be sued. This does not solve the problem of risks from militaries and other "above-the-law" actors, but it is a very general-purpose approach that avoids overfit, and is often supported by libertarian-leaning economists for this exact reason.

The primary targets for liability that have been considered so far are:

Putting liability on users feels most incentive-compatible. While the link between how a model is developed and how it ends up being used is often unclear, the user decides exactly how the AI is used. Liability on users creates a strong pressure to do AI in what I consider the right way: focus on building mecha suits for the human mind, not on creating new forms of self-sustaining intelligent life. The former responds regularly to user intent, and so would not cause catastrophic actions unless the user wanted them to. The latter would have the greatest risk of going off and creating a classic "AI going rogue" scenario. Another benefit of putting liability as close to end usage as possible is that it minimizes the risk that liability will lead to people taking actions that are harmful in other ways (eg. closed source, KYC and surveillance, state/business collusion to clandestinely restrict users as with eg. debanking, locking out large regions of the world).

There is a classic argument against putting liability solely on users: users may be regular individuals without too much money, or even anonymous, leaving no one that could actually pay for a catastrophic harm. This argument can be overstated: even if some users are too small to be held liable, the average customer of an AI developer is not, and so AI developers would still be incentivized to build products that can give their users assurance that they won't face high liability risk. That said, it is still a valid argument, and needs to be addressed. You need to incentivize someone in the pipeline who has the resources to take the appropriate level of care to do so, and deployers and developers are both easily available targets who still have a lot of influence over how safe or unsafe a model is.

Deployer liability seems reasonable. A commonly cited concern is that it would not work for open-source models, but this seems manageable, especially since there is a high chance that the most powerful models will be closed source anyway (and if they turn out to be open, then while deployer liability does not end up very useful, it also does not cause much harm). Developer liability has the same concern (though with open source models there is some speed-bump of needing to fine-tune a model to cause it to do some originally disallowed thing), but the same counterargument applies. As a general principle, putting a "tax" on control, and essentially saying "you can build things you don't control, or you can build things you do control, but if you build things you do control, then 20% of the control has to be used for our purposes", seems like a reasonable position for legal systems to have.

One idea that seems under-explored is putting liability on other actors in the pipeline, who are more guaranteed to be well-resourced. One idea that is very d/acc friendly is to put liability on owners or operators of any equipment that an AI takes over (eg. by hacking) in the process of executing some catastrophically harmful action. This would create a very broad incentive to do the hard work to make the world's (especially computing and bio) infrastructure as secure as possible.

Strategy 2: Global "soft pause" button on industrial-scale hardware

If I was convinced that we need something more "muscular" than liability rules, this is what I would go for. The goal would be to have the capability to reduce worldwide available compute by ~90-99% for 1-2 years at a critical period, to buy more time for humanity to prepare. The value of 1-2 years should not be overstated: a year of "wartime mode" can easily be worth a hundred years of work under conditions of complacency. Ways to implement a "pause" have been explored, including concrete proposals like requiring registration and verifying location of hardware.

A more advanced approach is to use clever cryptographic trickery: for example, industrial-scale (but not consumer) AI hardware that gets produced could be equipped with a trusted hardware chip that only allows it to continue running if it gets 3/3 signatures once a week from major international bodies, including at least one non-military-affiliated. The signatures would be device-independent (if desired, we could even require a zero-knowledge proof that they were published on a blockchain), so it would be all-or-nothing: there would be no practical way to authorize one device to keep running without authorizing all other devices.

This feels like it "checks the boxes" in terms of maximizing benefits and minimizing risks:

Hardware regulation is being strongly considered already, though generally through the frame of export controls, which inherently have a "we trust our side, but not the other side" philosophy. Leopold Aschenbrenner has famously advocated that the US should race to get a decisive advantage and then essentially force China to sign a protocol limiting how many boxes they are allowed to run. To me, this approach seems risky, and could combine the flaws of multipolar races and centralization. If we have to limit people, it seems better to limit everyone on an equal footing, and do the hard work of actually trying to cooperate to organize that instead of one party seeking to dominate everyone else.

d/acc technologies in AI risk

Both of these strategies (liability and the hardware pause button) have holes in them, and it's clear that they are only temporary stopgaps: if something becomes possible to do on a supercomputer at time T, it will likely be possible on a laptop at time T + 5 years. And so we need something more stable to buy time for. Many d/acc technologies are relevant here. We can look at the role of d/acc tech as follows: if AI takes over the world, how would it do so?

As briefly mentioned above, liability rules are a naturally d/acc-friendly style of regulation, because they can very efficiently motivate all parts of the world to adopt these defenses and take them seriously. Taiwan has been experimenting with liability for false advertising recently, which can be viewed as one example of using liability to encourage info defense. We should not be too enthusiastic about putting liability everywhere, and remember the benefits of plain old freedom in enabling the little guy to participate in innovation without fear of lawsuits, but where we do want a stronger push to be secure, liability can be quite flexible and effective.

The role of crypto in d/acc

Much of d/acc goes far beyond typical blockchain topics: biosecurity, BCI and collaborative discourse tools seem far away from things that a crypto person normally talks about. However, I think there are some important ties between crypto and d/acc, particularly:


Cursive, an app that uses fully homomorphic encryption (FHE) to allow users to identify areas of common interest with other users, while preserving privacy. This was used at Edge City, one of the many offshoots of Zuzalu, in Chiang Mai.


In addition to these direct intersections, there is also another crucial shared point of interest: funding mechanisms.

d/acc and public goods funding

One of my ongoing interests is coming up with better mechanisms to fund public goods: projects that are valuable to very large groups of people, but that do not have a naturally accessible business model. My past work on this includes my contributions to quadratic funding and its use in Gitcoin Grants, retro PGF, and more recently deep funding.

Many people are skeptical of public goods as a concept. The skepticism generally comes from two sources:

  1. The fact that public goods have historically been used as a justification for heavy-handed central planning and government intervention in society and economy.
  2. A general perception that public goods funding lacks rigor and is run on social desirability bias - what sounds good, rather than what is good - and favors insiders who can play the social game.

These are important critiques, and good critiques. However, I would argue that strong decentralized public goods funding is essential to a d/acc vision, because a key d/acc goal (minimizing central points of control) inherently frustrates many traditional business models. It is possible to build successful businesses on open source - several Balvi grantees are doing so - but in some situations it is hard enough that important projects needs extra ongoing support. Hence, we have to do the hard thing, and figure out how to do public goods funding in a way that addresses both of the above critiques.

The solution to the first problem is basically credible neutrality and decentralization. Central planning is problematic because it gives control to elites who might turn abusive, and because it often overfits to the present-day situation and becomes less and less effective over time. Quadratic funding and similar mechanisms were precisely about funding public goods in a way that is as credibly neutral and (architecturally and politically) decentralized as possible.

The second problem is more challenging. With quadratic funding, a common critique has been that it quickly becomes a popularity contest, requiring project funders to spend a lot of effort publicly campaigning. Furthermore, projects that are "in front of people's eyeballs" (eg. end-user applications) get funded, but projects that are more in the background (the archetypal "dependency maintained by a guy in Nebraska") don't get any funding at all. Optimism retro funding relies on a smaller number of expert badge holders; here, popularity contest effects are diminished, but social effects of having close personal ties with the badge holders are magnified.

Deep funding is my own latest effort to solve this problem. Deep funding has two primary innovations:

  1. The dependency graph. Instead of asking each juror a global question ("what is the value of project A to humanity?"), we ask a local question ("is project A or project B more valuable to outcome C? And by how much?"). Humans are notoriously bad at global questions: in a famous study, when asked how much money they would pay to save N birds, responders answered roughly $80 for N=2,000, N=20,000 and N=200,000. Local questions are much more tractable. We then combine local answers into a global answer by maintaining a "dependency graph": for each project, what other projects contributed to its success, and how much?
  2. AI as distilled human judgement. Jurors are only each assigned a small random sample of all questions. There is an open competition through which anyone can submit AI models that try to efficiently fill in all the edges in the graph. The final answer is the weighted sum of models that is most compatible with the jury answers. See here for a code example. This approach allows the mechanism to scale to a very large size, while requiring the jury to submit only a small number of "bits" of information. This reduces opportunity for corruption, and ensures that each bit is high-quality: jurors can afford to think for a long time on each question, instead of quickly clicking through hundreds. By using an open competition of AIs, we reduce the bias from any one single AI training and administration process. Open market of AIs as the engine, humans as the steering wheel.



But deep funding is only the latest example; there have been other public goods funding mechanism ideas before, and there will be many more in the future. allo.expert does a good job of cataloguing them. The underlying goal is to create a societal gadget that can fund public goods with a level of accuracy, fairness and open entry that at least approximates the way that markets fund private goods. It does not have to be perfect; after all, markets are far from perfect themselves. But it should be effective enough that developers working on top-quality open-source projects that benefit everyone can afford to keep doing so without feeling the need to make unacceptable compromises.

Today, the leading projects in most d/acc subject areas: vaccines, BCI, "borderline BCI" like wrist myoelectrics and eye tracking, anti-aging medicines, hardware, etc, are proprietary. This has big downsides in terms of securing public trust, as we have seen in many of the above areas already. It also shifts attention toward competitive dynamics ("OUR TEAM must win this critical industry!"), and away from the larger competition of making sure these technologies come fast enough to be there to protect us in a world of superintelligent AI. For these reasons, robust public goods funding can be a strong booster of openness and freedom. This is another way in which the crypto community can help d/acc: by putting serious effort into exploring these funding mechanisms and making them work well within its own context, preparing them for much wider adoption for open-source science and technology more generally.

The future



The next few decades bring important challenges. There are two challenges that have recently been on my mind:

However, each of these challenges has a silver lining. First, we now have very powerful tools to do our remaining work more quickly:

Second, now that many principles that we hold dear are no longer occupied by a few particular segments of the old guard, they can be reclaimed by a broad coalition that anyone in the world is welcome to join. This is probably the biggest upside of recent political "realignments" around the world, and one that is worth taking advantage of. Crypto has already done an excellent job taking advantage of this and finding global appeal; d/acc can do the same.



Access to tools means that we are able to adapt and improve our biologies and our environments, and the "defense" part of d/acc means that we are able to do this without infringing on others' freedom to do the same. Liberal pluralist principles mean we can have a lot of diversity on how this is done, and our commitment to common humanity goals means it should get done.

We, humans, continue to be the brightest star. The task ahead of us, of building an even brighter 21ˢᵗ century that preserves human survival and freedom and agency as we head toward the stars, is a challenging one. But I am confident that we are up to it.