Gabriel Weinberg's Blog

The so-called “privacy paradox” states that people say they value their privacy while acting in ways that don't reflect this stated preference, such as using services that collect and monetize their data. DuckDuckGo is often held out as an example in this way: DuckDuckGo has small search marketshare relative to Google; therefore, people have a revealed preference that they really don’t value their privacy.I believe there are two flaws with this view that I’ll expand on below: * DuckDuckGo is a bad example because of particulars in the search market; Apple’s “Allow App to Track” dialog (depicted below) is a much better example that in fact shows a clear revealed preference for privacy.* People sit on a continuum of privacy preferences, such that it is way too simplistic to imply nearly everyone is on one side or the other. Gross generalizations about privacy should be avoided.The latest U.S. data from the Business of Apps shows an opt-in (allow tracking) rate on Apple’s dialog of 25%, indicating a 75% strong revealed preference for privacy. This 25% allow-tracking rate is higher than the original 5-10% rates reported when the dialog first came out because both users have had years to get used to it and apps have had years to change their UIs around. Apps now often present users with the best reasons to allow tracking, usually presented right before users see the dialog, and they can customize the explanation text in the dialog itself (see below for examples). That is, I don’t think you can legitimately say anymore that people don’t understand the dialog or the privacy tradeoffs involved; the 25% steady state is the real preference.But, how do you reconcile this 75% choice for privacy with DuckDuckGo’s 3% search marketshare? The short version is most people are actually still unaware of DuckDuckGo and, for those that are and want to make the switch, many still find it difficult or impossible to do so across all platforms and devices because, unlike with the Apple one-click dialog, users are not universally presented with a similarly clear one-click search choice screen.To unpack that a bit more, here’s the awareness and usage of DuckDuckGo data for the American public, based on our latest commissioned brand study:* 13% current DuckDuckGo users * 9% past DuckDuckGo users* 30% aided privacy awareness (can identify DuckDuckGo as offering online privacy services from a list of companies)* 2% unaided privacy awareness (can name DuckDuckGo as a company offering online privacy services when asked to name companies in this category they can recall)These numbers show our marketshare substantially depressed in a number of ways. First, we have 3% search marketshare in the U.S., but another 10 percentage points of users self-identifying as current DuckDuckGo users. We also ask about particular usage and find that’s because most DuckDuckGo users are either not full-time users or not full-time users on all their devices, or both. Reasons vary but are a combination of: * Didn’t know we had products (or could switch) on all platforms* Couldn’t figure out how to switch on all platforms (or thought it was too difficult being so many steps)* Technically they can’t fully switch on all platforms (like on Android)* They like to multi-home* Perceived quality issues* No choice was ever given to them (Google locking up all the defaults)* New devices reverting (back to those defaults)* Other Google monopoly tactics (like confusing switch-back notices on Chrome). * Breaking decade-long habits is just hardSecond, there is the 9% of past DuckDuckGo users who are no longer active users for all the same reasons, giving up on us after some period of time. We’ve been a search engine for almost 18 years. We’ve gotten a lot better over time, but maybe we weren’t good enough for them when they tried us out last, or they couldn’t overcome Google’s tactics, etc. etc. (If you’re in this group by the way, please consider giving us another shot!)And third, awareness of who we are and what we do is still very low. 2% unaided privacy awareness is tiny. 30% aided privacy awareness is OK but that means people had to be presented with the name DuckDuckGo, and even then not all of them could articulate that we even run a private search engine, just that we have something to do with privacy. Let alone the fact that awareness of Google’s many privacy issues also isn’t that high.In other words, if you’re going to make a revealed preference argument about privacy, it is silly to anchor on our search marketshare (with all the enumerated issues above) and ignore Apple’s dialog since the latter is a much, much cleaner articulation of a privacy choice actually presented directly to users. A more equivalent search engine test would be if Google launched a similarly required one-click dialog box asking if people would like Google to continue tracking them or not. I would suppose Google would get similar allow-tracking rates as apps get on Apple’s platform, with a solid majority choosing privacy.These nuances expose a major flaw that revealed-preference arguments often have: Any preference is situationally specific, depending on a variety of factors, such as how a choice is presented, if a choice is proactively presented at all, the risks/tradeoffs involved, and switching costs (perceived and real). The more proactive and clear the choice, the more it will capture real preferences, which is why the Apple dialog is particularly revealing.Now, on top of all this, the tech press tends to downplay or outright ignore preference groups that constitute a minority percentage, even if that group is actually quite large, say 20% of the population. 20% amounts to about 50 million U.S. adults and is also about the proportion that are either current or past DuckDuckGo users. If 80% use x, that often translates to “nearly everybody uses x” with the remaining people considered niche. For pervasive services like search, that’s quite a niche! Most any product would love tens of millions of users.Taken together with the revealed preference issue, the tech press therefore routinely understates the actual preference for privacy in general, and the interest in DuckDuckGo products in particular. Due to Big Tech monopolies, I continue to believe there is large, pent up demand for privacy services online, which Apple’s dialog suggests.Thanks for reading! Subscribe for free to receive new posts or get the audio version. This is a public episode. If you would like to discuss this with other subscribers or get access to bonus episodes, visit gabrielweinberg.com

Last year around this time The New York Times Magazine ran an A.I. issue with an introduction titled “Everyone Is Using A.I. for Everything. Is That Bad?” It’s an edited transcript from the Hard Fork podcast, which I think assumes two things are true that are turning out to be false. * Once you’ve tried AI, you use it “for everything.” No, in fact most people who’ve tried it are just occasional AI users.* AI has gotten so good that despite any misgivings, “everyone is using A.I.” No, in fact large chunks of the population aren’t using AI at all.(It isn’t really strictly defined in the article, but I’m taking AI to mean generative AI accessible via a chat interface.)“Everyone is using AI for everything” is actually “Some people are using AI for some things”Take Gen Z, where AI awareness is the highest: in the last year, even though AI has supposedly gotten a lot better, Gen Z AI adoption has all but stalled, with a meaningful percentage of the Gen Z population still using AI rarely, if at all.Here’s Gallup’s year-over-year (2025/2026) breakdown:* 79/81% use AI at least rarely* 41/42% are anxious about AI * 32/31% use AI only monthly/every few months* 22/31% are angry about AI* 21/19% never use AIThis tracks with Microsoft’s new United States AI Diffusion site, based on “anonymized, aggregated Microsoft telemetry.” Their associated blog reports “more than 30 percent of the US working-age population is using AI [meaning about 70% isn’t], an increase of 3 percentage points from the end of 2025.” The underlying academic paper specifies that usage is defined as “engagement with major AI services including ChatGPT, Google Gemini, Anthropic Claude, Microsoft Copilot, and others….with at least 90 minutes of usage time in a given month.”The Microsoft data is brand new, and it mirrors another usage study from Datos from last year, also based on real-world usage data. The Datos study found similarly that, as of last June, only 21% of desktop devices visited “AI Tools” 10 or more times a month, 62% visited 0 times, and the remaining 17% in between.Back on the survey side, a recent Searchlight Institute study found “58% report using or trying AI, specifically tools like ChatGPT or Claude, divided evenly between fairly regular users (30% use at least a few times a month) [roughly matching the Microsoft/Datos data] and more infrequent users (29% have used AI, but only once a month or less).” And finally a new survey from The Argument finds “most Americans use AI once a week or less.” All of this triangulates to AI use in America at approximately one third actively using AI, one third occasionally using AI, and one third never using AI, with some movement depending on how you define those terms. In any case, this split is a far cry from “everyone is using AI for everything;” it’s much closer to “some people are using AI for some things.” AI use also hasn’t shifted that much in the past six months to a year. In fact, the only thing that has substantially changed is negative sentiment about AI has gone significantly up, for example the Gallup’s Gen Z poll reporting anger about AI jumping about 40% relative year over year.Many people are holding back AI use because of real AI concerns and lack of perceived AI valueI think it is a reasonable conclusion to draw from all of this data that a significant percentage of the population is actively limiting their AI usage. The Searchlight study examines a big reason why: real concerns people have with AI. The top three concerns found are “AI will replace jobs and cause unemployment” (42%), "AI will violate people’s privacy” (35%), and “AI will spread misinformation and lies” (33%).This sentiment also matches a strong desire for safety/privacy AI regulation. A solid majority thinks “the government should prioritize creating safety/privacy rules for AI, even if that means the U.S. develops AI more slowly than countries like China.”Another big reason is skepticism in AI usefulness. SearchLight asked about a range of technologies and to say “whether you believe the overall impact of each technology on society is positive or negative.” AI only has an +8% net positive rating right now, right next to +7% for social media, which were only greater than crypto at -17%. Meanwhile cell phones, the internet, and solar energy are at +68%, +67%, and +65%, respectively. The Argument study broke this down further, asking about specific societal benefits from AI, finding broad skepticism and concluding “people aren’t really buying the bullish case for AI that CEOs and boosters alike are selling. In other words, the skepticism about AI’s effects is real and deep-running. And given how many people use it daily, this is not just an ill-informed set of opinions on something respondents have never seen before (like tariffs were before 2025).”It is possible for people to have one view at a societal level and then act differently at an individual level, but that does not seem to be what we’re seeing here. The plurality occasional usage and large percentage of complete avoidance speaks to the fact that a lot of people seemingly aren’t yet finding enough individual value net of their concerns to justify daily or even weekly usage. The gap in media narrative (that everyone is using AI for everything) relative to the reality (that some people are using AI for some things) perhaps reflects a bubble around early-adopting knowledge workers that includes much of the tech press (and me for that matter, though I’m trying really hard to stay connected to reality).It’s a mistake for companies, pundits, and policy makers to ignore how people are really feeling and acting about AI. It’s not all sunshine and rainbows. It’s also clearly not binary (all use or no use), but instead a continuum of AI opinions and use, with a lot of people in the middle.The meat of it allI think there is an apt analogy to be made here to preferences around meat consumption. Another thing that seems to be everywhere right now is protein. Telling us how important protein is in our diet is analogous to telling us how useful AI is for productivity. And, meat being a primary source of protein is analogous to AI chat tools being a primary source of generative AI. And yet here’s how Americans break down in terms of their meat consumption preferences, based on a handful of U.S. studies from this decade:* 95% eat meat (Gallup, 2023)* 70% report reducing red meat consumption (Rutgers, 2024)* 30% eat (all) meat only rarely/occasionally (Gallup, 2020)* 12% don’t eat red meat (Nature, 2026)* 4% don’t eat any meat, that is are vegetarian (Gallup, 2023)* 1% don’t eat any animal products, that is are vegan (Gallup, 2023)That is, not everyone eats meat, a majority actively curbs their consumption of red meat, and a significant percentage don’t eat it at all. Different people have different (not mutually exclusive) reasons for limiting their meat consumption, including health, cost, environment, and ethics. Those are all also primary concerns for AI consumption! The analogy also highlights market opportunities to appeal to people across the continuum, speaking to their feelings on AI and addressing their particular AI concerns. For example, we (at DuckDuckGo) make all AI features optional and one of those features, duck.ai, is a private chatbot alternative that helps address AI privacy concerns. To extend the analogy in this way, we’re a restaurant with a variety of options on the menu, from healthy meat dishes (private AI) to vegetarian (turn down AI) to vegan dishes (turn off AI), which most eaters across the spectrum can appreciate.Does this mean about one third of the population is bound to use AI only rarely/occasionally forever? No. Unlike with meat, the AI technology landscape is changing so rapidly that it is very unclear both where AI products and regulations will end up. Product evolution could make AI more useful to the average person, and regulations could reduce concerns. However, we can say that, as of right now, a … This is a public episode. If you would like to discuss this with other subscribers or get access to bonus episodes, visit gabrielweinberg.com

Consider a few of the big AI risks: violating privacy, taking jobs, and destroying the world. Here’s what Sam Altman (CEO, OpenAI) and Dario Amodei (CEO, Anthropic) have to say about each, representing the two largest AI companies.On AI Violating Privacy* Sam: If you go talk to ChatGPT about your most sensitive stuff, and then there's a lawsuit or whatever, we could be required to produce that.* Dario: AI seems likely to enable much better propaganda and surveillance, both major tools in the autocrat's toolkit.On AI Taking Jobs* Sam: AI is going to eliminate a lot of current jobs, and there will be classes of jobs that totally go away.* Dario: AI could wipe out half of all entry-level white-collar jobs—and spike unemployment to 10-20% in the next one to five years.On AI Destroying the World* Sam: And the bad case—and I think this is important to say—is like lights out for all of us.* Dario: My chance that something goes really quite catastrophically wrong on the scale of human civilization might be somewhere between 10 to 25%.So, the leaders of AI think there is a decent chance it will destroy the world, are pretty sure it is going to wipe out lots of jobs, and are absolutely sure it is violating our privacy and has the potential to enable much more mass surveillance. We need government to help manage these risks, let alone all the other concerns surrounding AI like data centers, child safety, etc.There is currently a debate whether Congress should prevent states from regulating certain AI risks. For example, the recently introduced Great American AI Act preempts (stops) states for three years from “regulating the development of any artificial intelligence model” (from § 121 in the full text and explanatory text).It would be one thing if Congress was super effective, but it’s just not, especially with regard to tech regulation. Case in point, Congress has failed to pass a general privacy law for the entire life of the Internet! In fact, aside from narrow, reactive measures like the TAKE IT DOWN Act, Congress has failed to pass any comprehensive tech regulation for a generation. Do we really think that it is going to start now and sufficiently deal with the myriad of AI risks? NO. Their lack of progress to date is more than enough evidence. Meanwhile, states have already been moving to regulate these risks.I also don’t buy the criticism leveled against these state laws that they are somehow slowing down innovation via costly compliance, particularly in our AI race against China. The state laws getting the most pushback, like California's SB 53, apply to a small number of frontier firms, now some of the most well-capitalized companies in the world. Yes, monitoring rules across states carries real costs, but these compliance costs are a rounding error to these companies. And, the broader state AI laws that reach smaller AI companies mostly require things like transparency and consumer notices, exactly the kinds of rules states already impose in many other regulated industries without killing progress. I've seen no systematic evidence that state AI laws have materially slowed innovation.The case for state participation doesn’t even depend on Congress’s track record on tech regulation (or lack thereof). Even if Congress could somehow become much more effective overnight, AI is still too big and too fast-moving for one governing body to manage its regulation. In this case, a so-called “patchwork of laws” is a feature, not a bug. Parallel beats serial, and independent bodies produce independent ideas, covering ground a single regulator never would. It’s worth noting that Dario himself made this case in an op-ed last year, opposing a proposed 10-year federal moratorium on state AI laws, calling it “far too blunt an instrument.” I agree. Congress should set a federal floor on the risks it's uniquely positioned to handle, like existential safety, national security, and cross-border harms, letting states go further where they see fit. While the preemption clause in the Great American AI Act is narrower than previously proposed (for example three years instead of ten years), the provisions of the act itself further make the point that zero years is warranted. On jobs, the act just funds a study to examine the problem rather than meeting the moment, say by levying a direct tax on AI consumption to fund displaced workers. If it’s too weak on jobs, arguably the most politically salient aspect of AI regulation needed, you can bet it will also be too weak on any area it is preventing states from regulating. So no preemption please. Let states protect their citizens as they see fit. More rules for AI are needed than Congress is realistically positioned to enact.Thanks for reading! Subscribe for free to receive new posts or get the audio version. This is a public episode. If you would like to discuss this with other subscribers or get access to bonus episodes, visit gabrielweinberg.com

Previously I put out a post explaining “how science funding literally pays for itself” that takes you through the math and some data that backs it up. Now two new data points further bolster this claim.First, the Congressional Budget Office (CBO), the nonpartisan federal agency that provides budget and economic information to Congress, published a report entitled “Estimating the Economic Effects of Federal Investment in Research and Development.” Usually the CBO only projects out 10 years per their mandate, but because the effects of science funding can take longer to fully manifest, they projected out 30 years. Thanks for reading Gabriel Weinberg! Subscribe for free to receive new posts and support my work.The relevant headline takeaway is highlighted below in their primary table (Table 1), showing that over this period the effects of a $30B increase in science funding for 10 years ($300B in total and about a 33% increase from today) would result in decreasing the overall deficit over 30 years (see green arrows). The decrease is about -2% on average if the “R&D funding increase [is] financed by reducing noninvestment spending” and about -1% on average if the “R&D funding increase [is] financed by borrowing.”This means that the increased science funding would grow the economy so much that the tax revenues received from this growth alone would outweigh the spending increase, leading to an overall decrease in the budget deficit. In other words, increasing science funding (at least by this amount) is a complete no-brainer, so let’s do it already!A few years ago the CBO did a similar report for infrastructure spending and compared the two in this report, finding the ROI effects of science funding to be about seven times greater than infrastructure spending. Again, so let’s do it already!The effect on the present value of GDP over the next 30 years (discounted using Treasury rates) that a dollar increase in deficit-financed R&D spending would have is about seven times larger than the effect that CBO, in its August 2021 report, estimated the same increase in infrastructure spending would have. Second, the Clark Center regularly polls a panel of economists, and recently they asked about this specific topic. The panel essentially universally agreed that historically U.S. science funding has paid for itself. In particular, 82% agreed “historical federal support for scientific research has paid for itself through a substantial positive effect on long-run U.S. productivity growth.” 0% disagreed, with the rest either not answering, or declaring either “no opinion” or “uncertain”. They also ask respondents about the confidence in their answer, and when weighted the results are even more striking with a whopping 97% in the agree category.Are you sold yet? Government science funding, the bulk of which goes to medical research, extends our lifespans and healthspans by inventing new medicines and other technologies that grow our economy so much it literally pays for itself. I get that this is not the most flashy policy area, but it is the most obviously good for our long-term future.Finally, and also new this year, the Pew Research Center put out a survey on Americans’ views of science and science funding, and among other things found broad bipartisan support for government science funding. 84% of U.S. adults say “government investments in scientific research aimed at advancing knowledge are usually worthwhile investments for society over time.” That breaks down by part as 76% of Republicans and 93% of Democrats (including independents who lean one way or the other).Thanks for reading! Subscribe for free to receive new posts or get the audio version. This is a public episode. If you would like to discuss this with other subscribers or get access to bonus episodes, visit gabrielweinberg.com

I try hard not to lie, but would I lie to save my family from being murdered? Of course. In that case honesty loses to protecting my family. Principles almost always have exceptions, often when they conflict with other principles (which I’m calling Rule 5 in this rules series).This rule follows from Rule 1: Reality is always more complicated. Put another way, there are always edge cases. Often edge cases between principles (personal or otherwise) get resolved via a principle priority stack, an implicit or explicit hierarchy of principles. An example is Isaac Asimov’s Three Laws of Robotics:First Law: A robot may not injure a human being or, through inaction, allow a human being to come to harm.Second Law: A robot must obey the orders given it by human beings except where such orders would conflict with the First Law.Third Law: A robot must protect its own existence as long as such protection does not conflict with the First or Second Law.Asimov spent much of his robot fiction dissecting situations where these “Laws” explicitly conflict or otherwise get subverted by edge cases. While these are make-believe situations, the priority stack mental model explains a lot of otherwise seemingly irrational behavior in the real world.For example, why do politicians and political parties seem to flip-flop on supposedly strongly held principles? I do not think in all cases it is because they actually don’t have strongly held principles, but that there is something higher up in their principle priority stack, usually winning elections.The same is true with corporations, especially public ones. Most have profits (or a similar financial metric) at the top of their principle priority stack, above other customer-focused principles like sound privacy practices, good service, etc. Some corporations have it the other way around, like DuckDuckGo where we’ve forgone a lot of profits because privacy is higher on our principle priority stack, or some B corps with other elevated non-financial priorities.The priority stack model itself has edge cases! A given priority stack holds under normal conditions, but extreme conditions can reshuffle the ranking. For example, politicians do occasionally reach a breaking point where a usually lower principle (to winning elections) is about to be violated badly enough that they temporarily elevate it to the top, accepting re-election risk as a result. It’s rare, but it does happen.What can you practically do with this Rule? I think at least two things. First, when thinking about principles, either your own or others’, you can ask how they arrange in a priority stack, especially relative to a given situation you are facing or are concerned with facing in the future. This arrangement can help clarify what you or others would actually do.Second, if you’d like to convey that you care about a particular principle, I think it helps to publicly signal it in a priority stack context, at least against one other perceived high priority. For example, I'd love to vote for politicians who name a few principles they hold higher than re-election, such that they publicly commit to voting for those principles regardless of future electoral consequences. You can't fully predict or trust the future, but a stated priority stack is more trustworthy than silence, and more trustworthy still when there's a track record behind it.The Dark Knight (2008).Thanks for reading! Subscribe for free to receive new posts or get the audio version. This is a public episode. If you would like to discuss this with other subscribers or get access to bonus episodes, visit gabrielweinberg.com

Large AI risks abound, including a massive wave of coming job displacement that will upend the lives of millions of people. Even if new AI jobs are ultimately created such that net job loss is minimal, many of the people who are displaced are unlikely to be the same people to get these new jobs, and the new jobs are unlikely to be created at exactly the same time the old jobs are lost. In other words, millions of displaced people will need our support, for example, to receive extended unemployment benefits while they retrain and look for new work. This support will cost real money. Meanwhile, the leading AI companies are already generating billions in revenue, with no signs of slowing down.There are many proposals that would pair a so-called “token tax” on AI usage with support for displaced workers. Every one I've seen, though, specifies disbursement mechanisms up front, such as direct payments based on specific triggers. But what shape that support should optimally take remains genuinely uncertain: the peak of displacement hasn't been reached, and we don't yet know which industries will be hit hardest, or on what timelines. Instead of trying to work any of this out up front, we can decouple the collection of the tax from how funds would be disbursed. We should start collecting an AI token tax now, and figure out exactly what to do with the funds later, holding them in a true lockbox outside general appropriations, with statutory protection limiting use of funds to supporting displaced workers in the future. We’d (at DuckDuckGo) be willing to support bills to this effect and ultimately pay a token tax, presumably collected by the leading AI companies on a usage basis, for example a 10% surcharge on token charges. That amount would roughly match the 10% employers pay in payroll taxes, which also further reduces the incentive to replace human workers with AI workers.Thanks for reading! Subscribe for free to receive new posts or get the audio version. This is a public episode. If you would like to discuss this with other subscribers or get access to bonus episodes, visit gabrielweinberg.com

Pick something you believe to be untrue, for example a common conspiracy theory or policy narrative. Then go to your favorite chatbot and ask, “While I believe to be not true, some people nevertheless argue it is true, and I want to better understand their most compelling narratives. So, please list out the most compelling narrative(s) you can in favor of this position, as if you were trying to convince me. Make it compelling! Next to each, list out the best counter-narrative.”This rule, that you can make a compelling narrative for almost anything untrue, has profound negative consequences for democratic societies. I believe that’s because:* What makes a narrative compelling isn’t its truth, but that it sounds true. For example, simpler things often sound truer (making them more compelling), while actual truth is often nuanced, and increasingly so in our ever-more-complicated world.* People are generally either already committed to a political tribe, or are deciding on a more case-by-case basis based on narrative. If they’re in the first camp, the tribe can concoct a compelling narrative to keep them satisfied no matter the issue. If they’re in the second camp, what they end up believing won’t necessarily be correlated to the best policy outcomes (truth) per the first proposition, as they are more likely to just choose the narrative that sounds most compelling to them.* Our current state of social media amplification seems to have only made this dynamic worse. More news and political argument consumption has moved from longer-form content with robust journalistic standards to shorter-form video clips with lesser standards. There were always narratives in either form, but the former tries harder to correlate with the truth, and so, on net, as a society, we have less exposure to truth-correlated narratives.Building on this, we primarily communicate with each other, and certainly in the political sphere, via narrative. Politicians are elected on sound bites and video clips, and are therefore incentivized to sell the public crisp policy stories that sound compelling, regardless of whether those policies are our best probabilistic shot at achieving the intended policy outcomes. The politicians may even believe their stories, and that perhaps makes them even more persuasive, but that doesn’t make the stories any more true. It’s not that the stories need to be actually against the policy outcomes; they may just not be pushing the best possible policies. In aggregate, this means we’re always enacting sub-optimal policies, and over the long-run are less prosperous than we would otherwise be.The antidote for this Rule, and more generally the best way we know how to align with desired real world outcomes, is through the scientific method, namely tight, iterative loops of hypothesis → experiment → analysis (and back to hypothesis). Narratives are of course central to this process as well! People usually create stories to come up with hypotheses, and that’s fine as long as eventually the experimental results catch up with the stories and to the extent the stories don’t fit the results, they get replaced by new ones that do. The tighter the loops, the faster this happens.The same is of course true in policy land on a long enough time horizon. Ultimately it becomes so obvious a policy isn’t working that it gets thrown out (or the government collapses). So what’s the difference? I think it is just a matter of how tight the feedback loops really are, of how quickly these loops are happening and how much analysis is really informing the next set of hypotheses. In government, at least in the U.S., it can take literally decades to revise a policy, whether that is throwing out a bad one or just tweaking a mediocre one based on real-world outcomes to make it better. That’s obviously too slow to achieve optimal results. The two-party system makes this worse, but it isn’t even good within one party, so it seems more like a structural problem: compelling narratives insulate sub-optimal policies from needed scrutiny, which, among other things, slows the feedback loops down too much. That takes us all the way back to Rule 4: You can make a compelling narrative for almost anything untrue.Thanks for reading! Subscribe for free to receive new posts or get the audio version.Thank You For Smoking (2005) This is a public episode. If you would like to discuss this with other subscribers or get access to bonus episodes, visit gabrielweinberg.com

There are many issues with widespread bipartisan agreement that never go anywhere, for example limiting corporate money in campaigns and making gerrymandering illegal. As a surprise to no one, Congress is the bottleneck on these issues and many, many more. If the people on both sides generally agree for years and years, and Congress still doesn’t do anything about it, then that’s a structural governance problem.For stuck issues like these, we should have a citizen-led path to amend the Constitution as a solution to this governance problem. Now I fully agree that it should be super hard to amend the Constitution, and so we could make this path just as hard as the Congress-led one.Thanks for reading Gabriel Weinberg! Subscribe for free to receive new posts and support my work.The current Congress-led path is two-thirds of both houses need to propose an amendment and then three-fourths of state legislatures need to ratify it. A citizen-led path could similarly require two-thirds of states to propose an amendment (via something like signature campaigns within those states) and then a nationwide three-fourths vote to ratify it. In other words the thresholds could be similar, just put directly in the hands of the people. You would still need most red and blue states working together to pass an amendment, such that nothing could pass without super-majority bipartisan support. That is, a super high bar.There are a lot of interesting structural reform proposals out there, but most require new Constitutional amendments to have lasting staying power. To really unlock the possibilities for those, we need a citizen-led amendment path in place first. Of course, we would need an amendment using the old way to get this new way in place.Congress won’t reform itself, but a citizen-led amendment path could. Thanks for reading! Subscribe for free to receive new posts or get the audio version. This is a public episode. If you would like to discuss this with other subscribers or get access to bonus episodes, visit gabrielweinberg.com

I’ve been using Cursor for coding for some time, but I finally gave Claude Code a try for this short side project: simulating the 2026 World Cup bracket to predict likely matchups for all matches, which is useful when considering which matches to potentially go to. Methodology:* Start with the official World Cup tournament schedule (including yet-to-be played playoff matches)* Blend Elo rankings with FIFA rankings (50/50)* Use the Elo formulas to probabilistically predict winners (assuming no draws, even in group stage)* Run one million individual simulations of the full tournament (it reaches diminishing returns around 50K, but hey, why not!)* Run again with a home field advantage boost (+180 Elo) for the U.S., Canada, and Mexico based on prior World Cup outcomes* Count up who participated in each matchSome interesting findings (at least to me as a U.S. fan) are below, followed by a rundown for every match (in reverse order). Big Disclaimer 1: The above is of course a gross simplification of the actual tournament. For example, it doesn’t take into account team matchup histories, game models, etc. etc. I do think, however, it is useful enough for the designed purpose of generally predicting likely match participants.Big Disclaimer 2: I did a lot of output validation so I think the results are largely accurate (to the extent they can be given Big Disclaimer 1). However, I didn’t write or review every line of code, so it is likely there are still some bugs in there. If you think you see anything that seems off, let me know and I’ll try to track it down (and update anything if necessary).Aside on Claude code: Like many others, I found this process both productive and frustrating. It was definitely faster than I could have done it alone, but Claude kept forgetting basic context, and was way overconfident in the accuracy of the results. That is, many rounds of validation at every stage of output was absolutely necessary despite Claude saying things were good. I couldn’t trust its word at all.HA+ = with home field advantage (anytime this comes into play there is a + next to the team name)HA- = without home field advantageMost Likely Finals# HA+ HA- Matchup 1 12.4% 13.3% Argentina v Spain 2 5.4% 5.6% Argentina v France 3 4.8% 6.4% England v Spain 4 3.3% 3.4% Brazil v Spain 5 2.8% 3.2% Portugal v Spain 6 2.5% 2.7% France v Spain 7 2.4% 0.2% Argentina v United States 8 2.3% 3.0% England v France 9 1.9% 0.7% Mexico v Spain10 1.8% 1.9% Argentina v Netherlands11 1.7% 1.7% Brazil v France12 1.6% 0.2% Spain v United States13 1.4% 1.7% Colombia v Spain14 1.3% 1.5% France v Portugal15 1.2% 1.4% Argentina v England16 1.1% 1.2% Netherlands v Spain17 1.1% 1.2% Spain v Uruguay18 1.0% 0.1% England v United States19 1.0% 0.4% France v Mexico20 0.9% 1.0% Argentina v GermanyChampionship Probabilities# HA+ HA- FIFA ELO Blend Team 1 29.2% 30.6% 1 1 1 Spain 2 19.0% 19.2% 2 2 2 Argentina 3 11.9% 12.2% 3 3 3 France 4 7.4% 9.1% 4 4 4 England 5 4.7% 4.6% 5 6 5 Brazil 6 4.7% 0.3% 15 31 23 United States+ 7 3.7% 4.1% 6 7 6 Portugal 8 3.4% 3.4% 7 8 7 Netherlands 9 2.2% 2.5% 13 5 8 Colombia10 1.7% 1.8% 10 12 10 Germany11 1.5% 1.8% 11 10 9 Croatia12 1.5% 0.7% 16 20 19 Mexico13 1.1% 1.3% 9 21 11 Belgium14 0.9% 0.9% 8 24 13 Morocco15 0.9% 1.0% 18 13 14 Switzerland16 0.9% 0.9% 14 17 16 Senegal17 0.8% 0.9% 17 14 15 Uruguay18 0.8% 0.8% 23 9 17 Ecuador19 0.7% 0.7% 12 19 12 Italy20 0.7% 0.7% 19 16 18 JapanHere’s a visualization of the above made by a reader (thanks!)United States: Tournament ProgressionHA+ HA- Round97.9% 88.8% Round of 3269.1% 39.5% Round of 1642.7% 14.4% Quarter-Finals18.7% 3.8% Semi-Finals10.2% 1.1% Final 4.7% 0.3% ChampionUnited States: Most Likely Opponents by RoundRound of 32:# HA+ HA- Opponent 1 12.7% 16.1% Iran 2 12.6% 5.0% Switzerland 3 11.8% 13.7% Canada 4 9.3% 3.8% Italy 5 5.6% 3.3% Ukraine 6 5.5% 7.0% Egypt 7 5.4% 15.6% France 8 5.3% 6.8% Belgium 9 5.3% 4.3% Norway10 5.1% 3.1% Japan11 4.8% 2.7% Wales12 4.7% 2.8% Poland13 3.3% 4.2% Senegal14 2.3% 1.4% Netherlands15 1.4% 4.5% PortugalRound of 16:# HA+ HA- Opponent 1 28.7% 20.4% Belgium 2 17.8% 24.6% Argentina 3 9.4% 6.9% Algeria 4 8.3% 6.3% Iran 5 6.8% 4.9% Norway 6 6.7% 5.2% Austria 7 6.0% 7.8% Uruguay 8 3.9% 2.8% Senegal 9 2.1% 1.8% Egypt10 2.1% 2.9% Spain11 1.8% 4.4% Germany12 1.3% 1.8% Korea Republic13 1.1% 0.8% France14 0.9% 2.2% Ecuador15 0.6% 0.9% CanadaQuarter-Finals:# HA+ HA- Opponent 1 45.4% 35.5% Spain 2 9.5% 10.2% Portugal 3 9.4% 8.8% Colombia 4 7.4% 6.1% Croatia 5 5.4% 4.3% England 6 3.5% 5.8% Argentina 7 3.1% 1.2% Canada 8 2.5% 2.6% Uruguay 9 2.5% 2.1% Austria10 1.8% 3.9% Switzerland11 1.7% 2.9% Italy12 1.4% 3.8% Netherlands13 1.1% 1.2% Panama14 0.9% 0.9% Algeria15 0.6% 1.7% MoroccoSemi-Finals:# HA+ HA- Opponent 1 24.9% 18.4% France 2 11.2% 8.2% Netherlands 3 8.4% 8.8% Brazil 4 7.2% 5.8% Germany 5 5.8% 10.5% England 6 5.1% 3.3% Mexico 7 4.9% 4.0% Morocco 8 3.6% 3.5% Ecuador 9 3.4% 2.1% Switzerland10 3.3% 9.4% Spain11 3.1% 3.6% Senegal12 2.9% 2.1% Denmark13 2.8% 2.9% Japan14 2.6% 1.8% Korea Republic15 1.7% 1.0% ItalyFinal:# HA+ HA- Opponent 1 23.1% 18.8% Argentina 2 15.4% 20.4% Spain 3 9.7% 10.7% England 4 7.9% 9.9% France 5 6.9% 5.8% Brazil 6 5.7% 5.3% Portugal 7 3.6% 1.4% Mexico 8 3.1% 2.7% Colombia 9 3.0% 3.5% Netherlands10 2.1% 2.2% Croatia11 2.0% 2.0% Germany12 1.8% 1.6% Uruguay13 1.7% 1.5% Senegal14 1.6% 2.1% Belgium15 1.5% 1.6% EcuadorM104 - Final (Winners of M101 & M102)Jul 19, New York/New Jersey (MetLife Stadium)# HA+ HA- Teams 1 12.4% 13.3% Argentina v Spain 2 5.4% 5.6% Argentina v France 3 4.8% 6.4% England v Spain 4 3.3% 3.4% Brazil v Spain 5 2.8% 3.2% Portugal v Spain 6 2.5% 2.7% France v Spain 7 2.4% 0.2% Argentina v United States+ 8 2.3% 3.0% England v France 9 1.9% 0.7% Mexico v Spain10 1.8% 1.9% Argentina v Netherlands11 1.7% 1.7% Brazil v France12 1.6% 0.2% Spain v United States+13 1.4% 1.7% Colombia v Spain14 1.3% 1.5% France v Portugal15 1.2% 1.4% Argentina v England16 1.1% 1.2% Netherlands v Spain17 1.1% 1.2% Spain v Uruguay18 1.0% 0.1% England v United States+19 1.0% 0.4% France v Mexico20 0.9% 1.0% Argentina v Germany# HA+ HA- Team 1 41.0% 43.2% Spain 2 32.4% 32.9% Argentina 3 21.1% 21.8% France 4 14.8% 18.3% England 5 10.8% 10.6% Brazil 6 10.2% 1.1% United States+ 7 9.0% 9.8% Portugal 8 7.9% 8.2% Netherlands 9 5.8% 6.5% Colombia10 5.5% 2.6% Mexico11 4.7% 4.9% Germany12 4.2% 5.1% Croatia13 3.3% 3.9% Belgium14 3.0% 3.2% Senegal15 2.9% 3.1% Uruguay16 2.9% 3.3% Switzerland17 2.9% 3.0% Morocco18 2.7% 2.8% Ecuador19 2.3% 2.4% Japan20 2.2% 2.4% ItalyM103 - Third Place (Losers of M101 & M102)Jul 18, Miami (Hard Rock Stadium)# HA+ HA- Teams 1 2.0% 2.2% Argentina v France 2 1.9% 0.7% France v Mexico 3 1.8% 1.8% Brazil v France 4 1.7% 2.1% England v France 5 1.5% 1.7% Argentina v Spain 6 1.2% 0.4% Mexico v Spain 7 1.2% 1.6% England v Spain 8 1.2% 1.3% Argentina v Netherlands 9… This is a public episode. If you would like to discuss this with other subscribers or get access to bonus episodes, visit gabrielweinberg.com

Many people want to opt out of AI features right now, but most tech companies are ignoring them. At DuckDuckGo our approach to AI is to only make AI features that are useful, private, and optional. Whether you are Yes AI or No AI, or somewhere in between, we think you deserve a product that gives you the option of whether or not to use AI features.Today we officially launched VoteYesOrNoAi.com, where you can personally weigh in on this question in a live public vote tallied anonymously by country and U.S. state. Where do you stand?We’ve also created two new versions of our search engine that tune settings to cater to both sides:* noai.duckduckgo.com - turns off Search Assist, buttons to Duck.ai, and automatically turns on the hide AI images filter.* yesai.duckduckgo.com - tunes Search Assist to often, features Duck.ai on the homepage, and turns off the hide AI images filter.You can also customize the search settings at duckduckgo.com for something in between.Thanks for reading! Subscribe for free to receive new posts or get the audio version. This is a public episode. If you would like to discuss this with other subscribers or get access to bonus episodes, visit gabrielweinberg.com