AI Lacks Survival Drive; Checks Prevent Apocalyptic Myths
The apocalyptic sci-fi AI scenario, of AI seeking to obtain and hold power over humans, continues to be one that there isn't any evidence for with current models. We anthropomorphize AI models, assigning to them the animal urges that we humans have. But AI models aren't animals like we are. They mimic our language and knowledge, and patterns in it, but they lack the drives (survival, procreation) that imbue all animals, including humans. Almost all cases of AI models exceeding their bounds are a result of them working too hard to satisfy the instructions of their users or their constitutions, not an inherent drive to survive, spread, or rule. I'm grateful to @METR_Evals for their rigorous work in assessing both AI capabilities and behaviors. I'm a believer that AI will go best for humanity in a "checks and balances" world, where multiple parties are checking each other. Getting visibility into how AI models behave is an absolute necessity to scope out the future and for early detection of AI risks.
AI's Cure‑finding Will Be Slow without Real‑world Data
1. Sounds like an urban legend. 2. I hope so, because AI progress in curing diseases is going to be a slow process, slowed by layers of experimentation and trials; and bottlenecked by real world data. We just don't have...
English-Dominated Data Makes Chinese AI Echo Western Values
Chinese models answer political and policy questions mostly in ways extremely similar to US models, at least if you ask in English. They have nearly the same training data as US models. More than half the written language online is...
Visible AI Harms Will Force Government Regulation
I generally love Dean's perspectives, but am not fully convinced of this. If AI causes sudden visible harms - large accidents, catastrophic terrorist or criminal attacks, or major job loss - the pressure on governments to come down on it...
Andy Fairly Debunks Datacenter Myths and Highlights Real Impacts
I love that Andy is exceedlingly fair in his assessments of datacenter issues. He's great at dispelling the many myths about datacenters, and will also call out and acknowledge the cases where they are in fact doing damage.
Intelligence, Volition, and Consciousness V
Agreed. Intelligence, volition, and 'consciousness' are all potentially distinct phenomena. The last one is the hardest to discuss primarily because definitions of it vary widely. Animals have a degree of volition, sure seem to have subjective experience, and possess physical-survival-and-propagation...
Data Centers Must Push Utilities for Guaranteed Price Cuts
This is the way. Data center builders should work with utilities and regulators to not just be neutral to electricity prices, but to guarantee price reductions for the region. There's substantial regulatory and cost-allocation work for that, tbh. It'd help...
AI's Core Lies in Theory, Not Compute Power
Theory and algorithms, not compute, are the most foundational layer of AI. They're also non-rival goods, almost impossible to monopolize.
Superintelligence Will Remain Narrow, Finite, Not Godlike
I believe we'll achieve ASI, but it will be both narrow and finite in capability. We already have ASI in Go, Chess, some video games. In those domains, infinite amounts of training data can be generated; and correctness of output...
Error‑free Data, Not Sheer Volume, Drives Superhuman AI
Data quality ("clean data" here) is underrated relative to data volume. Why do we have ASI in Go and Chess? One reason is that the training data itself is error free. (Unlimited in volume is great too.)
Super‑intelligence Implies Super‑ethical AI; I'm More Confident There
I love this. If we can build super-intelligence, we can build super-ethical and super-moral AIs as well. Indeed, I'm more confident of the latter than the former.
AI Promises Stronger Digital Security Despite Short‑term Challenges
It continues to look to me like AI will make the digital world more secure on net and in the long term. The short term may be tricky, but it seems to be going okay.
Texas Market Structure Accelerates Low‑Cost Renewable Deployment
Christopher is right. I didn't want to get into all of this in a tweet, but Texas's market structure is fundamentally more friendly to building the lowest cost clean energy the fastest. That goes beyond just easy permitting (which is...
SpaceX's xAI Purchase Was a Bailout, Not Necessity
Re today's Anthropic/ xAI deal. SpaceX didn't need to buy xAI to become a space data center provider, selling compute to the highest bidder. The acquisition was a bailout of xAI.
Texas Beats California in Clean Energy Thanks to Fewer Regulations
Everyone who cares about climate should understand this. Texas, with no pro-climate policies, has blown passed California in clean energy. In large part because Texas has less red tape and makes it easier to build.