@ValerioCapraro the question is how any type of deterministic result can be encouraged when the context window [changed when the user "converses" with the llm] is inherently going to shift the region of the latent space being sampled.
@policytensor (2/2) why is the distinction is important? because anthropomorphic reasoning obscures the mechanism by which LLM alignment fails. “Rogue AI societies” better thought of as how Prions behave in biological systems- driven by a stochastic feedback process
@policytensor (1/2) Models do not think. Stringing LLMs together is not a society just because emergent phenomena comes out the other end.
Anthropomorphizing LLMs is not needed to understand that their use can, via stochastic dynamics, produce discoveries or dangerous outcomes
Non working women who can afford and want GLP1s are such a different population than the ones who don’t that this is entirely meaningless confounded data
@matthewstoller@DKThomp also the threat of a government investigation itself motivates behavioral change, doesnt matter if the court case is won. this would be way more effective than AI lobbying for a law they want
@matthewstoller been saying this for a while- make it clear these firms will be prosecuted for breaking laws that already exist. add new laws if needed. governments already let them illegaly scrape all of human knowledge w/out penalty
@zephyr_z9 not surprising at all if you understand experimental science. dont need 10 trillion parameters to model a relatively low dimensional process like material science.
@philippilk petrodollar isn't as important as it used to be, the dollar is kept afloat by US current account defecit - we pay for imports in dollars and the globe recycles those dollars into dollar denominated financial assets
@Rory_Johnston as a chem engineer, these steady state processes are often not specced to actually run at capacity for extended periods of time. not saying that this is why the plant went down, but this will increasingly happen as utilization is pushed to the max
1/ The fundamental flaw in AI-doomer scenarios is that they treat civilization as a frictionless API rather than what it is: a thicket of incompatible systems, rival actors and physical bottlenecks. Intelligence may exploit this structure. It cannot abolish its friction.
@matthewstoller The other issue is that climate change is a much larger risk, and is being completely ignored. For the climate, we have a century of empirical data to suggest increasingly instability in our food/water systems, whereas for AI, we have worried nerds.
The Trump administration appears to believe there is no lasting price for isolating Canada. That is profoundly shortsighted.
Canada and the U.S. traded roughly US$872 billion in goods and services in 2025. Canada supplies about 4.1 million barrels of crude oil per day, along with natural gas, electricity, uranium, aluminum, steel, potash and critical minerals. Our countries also share NORAD, Arctic surveillance, defence production and intelligence through Five Eyes.
The world is watching Washington threaten the sovereignty and economic security of its closest ally while disregarding an agreement Trump himself signed.
The message to every other ally is unmistakable:
American commitments last only until the next demand.
Canada will pay a price in the short term. But the long-term price for America could be higher: more expensive energy and supply chains, weakened continental defence, lost investment, reduced diplomatic influence and allies deliberately building a future that depends less on the United States.
America can pressure Canada. It cannot force Canada, or the rest of the world, to trust it again. 🇨🇦🇺🇸