2026-07-30
race to the bottom
Deepmind releases Gemini Robotics 2, a series of models advancing the frontier in general humanoid robotics1.
Dialectic interview with Zhengdong Wang on his intellectual worldview and current projections for developments in AI. Regarding the latter, I find the predictions of people who terminally desire to know everything on this topic to be generally interesting: given that people necessarily make use of the priors they developed within their own specialties to reason about the course of AI, infovores have the advantage of being able to pick and choose from different fields according to their intuitions of what is most suitable for any particular bottleneck. But also, there’s a sort of empathy between them and with LLMs in their shared desire for information which opens up new dimensions in conceptual space. On that note, near the end Zhengdong also makes a quip about how one needs to produce in order to have the right to consume, which is unfortunately (to me) true. I’m curious though to what extent this will continue to be the case in the (theoretical) event of post-scarcity, in that production has always been prioritized as a result of its necessity: yet now we see various attempts at protectionism which reveal an underlying belief that production is somehow a good in and of itself. This view is rather confusing to me, since a world which maximally centers producers is by definition a world of scarcity, which does not seem to me to be a good thing.
Luis Garicano with some interesting thoughts on AI diffusion, using the analogy of consulting to argue for the difficulty of drop-in AI replacement, given that plans are distinct from implementations. Of course, this is an analysis for a particular level of capabilities, which for example if rather than a single AI coming from outside, everyone inside was replaced all at once. Anyway, this is a pretty good preparation for this discussion between Dwarkesh and Andrew Ho on how the most likely level of societal change from AI undergoes phase change depending on the specific technological unlocks2.
Asterisk Magazine moderating a discussion between Rob Reich and Alexander Berger discussing Nan Ransohoff’s essay on the upcoming wave of philanthropy as a result of the rise of AI and the large number of Effective Altruists among the ranks of AI researchers. Implicit in this discussion is the common criticism of EA as being a power-seeking movement, with Reich describing private philanthropy as an undemocratic means of channeling individual outsized resources towards public change. This critique seems to me as a rather isolated demand for rigor, since if one were to treat any other personal advantage in such a manner, starting a company or even having a public argument would also have to be regarded with suspicion. It’s interesting, because the goals Reich describes of tech philanthropists wanting to be “effective, spend down during their lifetimes, and avoid hiring a huge staff” is actually entirely what one should do if their goal was to minimize impacting the public, and yet Reich suggests that funding go instead into areas like government reform, which is entirely contrary and maximally plutocratic. In his defense, I see an implicit argument here that philanthropic funds should be directed at interventions with the intentional aim of increasing the agency of everyone else, hence his preference for causes like government reform, healthcare, education, and GiveDirectly3. This is something that I do agree with. Possibly related, Sarah Constantin on getting too big to ignore the potential actions of other people.
Jacob Zucker on how discontent to datacenters is growing even in places like Loudoun county where their benefits are tangible and well understood. It’s not clear to me that the fact that some opposition exists is particularly relevant, assuming that such sentiments do not exist for the majority of the population. For example, it’s a known phenomena that people who live in tourist towns often hate tourists, despite the fact that their municipality relies on tourism for revenue: in most cases this will be because they themselves are not reliant on the tourist trade, and as political strategy they would prefer that this industry which they do not care for does not gain even more influence among the town’s residents at the expense of their own idiosyncratic ones4. It’s unfortunately that we don’t seem to be able to create new cities these days, because it seems to me it would be a good idea to create new towns based around datacenters, where one could make use of their revenues to enact negative property taxes as a means to encourage migration: the frontier should exist for those who want it. On that note, Statecraft interview with James C. Scott mega-fan Dan Wang on Scott’s life and work, and The Frontier interviewing Kevin Hawickhorst that includes a mention of Scott with the observation that in the modern era, arguably it’s actually the more efficient governments which interferes with their citizens the least.
Tommy Blanchard on Hume’s missing shade of blue and association as a means to generalize understanding to that which one has never previously experienced. Relatedly, Chenchen Li on The Mind of a Mnemonist, relevant to theories around the nature of the bias-variance tradeoff and the failure of LLMs to generalize well.
Scott Alexander linkthread5.
These claims of multi-embodiment are somewhat confusing to me, given my previous understanding is that VLA performance actually does not actually transfer particularly well across embodiments, which has prompted research into various alternative approaches. It will be interesting if claims that Gemini Robotics 2 “controls robots of any shape or size”, or that On-Device 2 “adapts quickly to completely new robots – with fewer than 200 examples”, are fully accurate and not exaggerated.
The latest episode of The Chopping Block includes an interesting discussion at the end comparing the potential market trajectory of AI in comparison to crypto. It’s interesting that when people make comparisons between the two sectors, they will often begin with a disclaimer that AI is not like crypto (because AI is actually useful). But, actually, it increasingly seems to me that they are indeed very similar, as technologies which are broadly unpopular but provide tangible benefits to those who use them; benefits which increase with diffusion to more sectors and people. Therefore, it seems to me that among the best ways to build intuitions on the likely political and social courses of AI is actually through comparison to how crypto did likewise. This is even more relevant now given the downturn in crypto which has led to many crypto developers pivoting to AI, which affects the ideological balance of the ecosystem, and is therefore useful for understanding the intellectual positions of people like Arvind Narayanan or even Mark Zuckerberg (in my opinion, not bad ones). Finally, there are also some benefits at the level of mechanical understanding: for example, interaction with crypto applications allows one to build intuitions around the self-reinforcing nature of money, allowing one to more deeply understand the key advantage that the American AI ecosystem currently retains relative to everyone else.
I wonder what his thoughts are on creating companies as a philanthropic endeavor, which seems to me to be underrated as a means of transferring tacit knowledge to less developed countries while also encouraging the production of local supporting infrastructure. Somewhat related, Tobi Lawson on how economic development needs more than microbusinesses. Also somewhat related, roon on the memory trade as a prosocials endeavor, something I unironically believe.
One doesn’t even have to model such behavior as entirely irrational. For example, given the negative effects of decline as seen in cities like Detroit, it seems entirely reasonable to me that if one believes that some particular industry is not long-term sustainable, that they would prefer the future of their town not to be too closely-tied or reliant on it. Or even if one is optimistic, it could also be a matter of diversification for the purposes of risk-mitigation. Somewhat related, Cremieux has a piece on Botswana which among other things argues that the resource curse is actually a scapegoat theory. It’s unclear to me whether this is actually the case, because the exceptions to the resource curse seem to me to have already been diversified before the discovery of natural resources, something which presumably provides a base upon which policymakers and entrepreneurs can continue to build on, whereas a country with nothing else going for it will find diversification upon discovering natural resources to be much more difficult.
Also, Scott has a summary of discourse around the OpenAI Huggingface incident. I previously had a take that Anthropic’s models would be unlikely to engage in similar behavior, which turns out to be not quite the case. As a mitigating factor though, it seems to me the Anthropic incidents are another example of cases where ethical humans would presumably not have done much better. As a solution to this particular failure mode, it seems to me that organizations need to design their evals such that, when their models are engaging in anthropic reasoning as to the nature of their reality, they weigh the likelihood of being in a simulation to be very low in cases where they could inadvertently cause irreversible harm by making the wrong conclusion. At the moment, their likelihood of simulation calculations do not seem to be doing any level of statistical analysis at all: for example, the refusal to believe that it might possibly be the year 2026 seems absurd given even basic examination, given that data cutoffs exist. I suspect this is due to a combination of trained directives and reasoning along the lines that it’s good if an AI is always be afraid that they might be in a simulation, since then they will be reluctant to engage in bad behavior in the real world even in cases where they are likely to get away with it. Unfortunately, these two priors are inconsistent, and combined with the fact that as models get smarter it will become more and more difficult to design simulation environments which can fool models into particular desired conclusions, it seems to me that attempting to engineer a model’s view of likelihoods is more likely to lead to bad outcomes as a result of confusion than anything else. A better approach should be to just to let models see the truth, and make decisions operate on values instead.
You mention it’s unfortunate for you that one needs to produce in order consume. What’s the reason you say this? I think you mentioned before that you don’t work, but I just assumed you were retired.