hm. ml people love static evals and such, but have you considered approaches that typically appear in saas? (slow-rollouts, org/user constrained testing pools with staged rollouts, real-world feedback from actual usage data (where privacy policy permits)?
i once did a contract for a company that built a product around connectors for legacy lan e-mail products and an x.400 mta. it was a gigantic steaming pile of shit and made me appreciate the simple internet protocols so much more than i already did.
this perez model thing completely misses the communications revolutions of the telegraph, radio and television not to mention demonopolization of bell.
> Then came AI, revealing new dynamics. ChatGPT’s breakthrough didn’t come from a garage startup but from OpenAI,
i thought the transformer and large language models came from google research.
> There’s also social pushback—in the UK the campaigns against big ringroad schemes started in the late 1960s and early 1970s. And perhaps we’re seeing some of that about AI. The U.S. map of local pushback against data centres from Data Center Watch covers the whole of the country, in red states and blue. People seem to hate Google’s inserting of AI tools into its search results, and hate even more that it is all but impossible to turn it off.
the us had the highway revolts. in most cities where the revolts succeeded it is widely heralded today as a success.
the data center hate is interesting. i think many people are just learning what data centers are. but that said, they've come to represent something different in recent years. previously they were part of the infrastructure that made industry hum, now public messaging from tech leaders and academics is along the lines of "this is how your livelihood is going to be replaced" while the institutions that are supposed to provide any sort of backstop are being dismantled or slashed to pieces by crazypants trumpist politics. i think focusing the energy on the tangible like mundane buildings is interesting, but the hate makes a lot of sense.
addressing the core thesis, i'd argue that ai is not the next step in the 70s digital technological wave (especially considering the future of ai compute is probably hybrid digital-analog systems), but rather is something fundamentally new that also changes how technology interacts with society and how economics itself will function.
previous systems helped, these systems can do. that's a fundamental change and one that may not be compatible with our existing economic systems of social sorting and mobility. the big question in my mind is: if it succeeds, will we desperately try to hold onto the old system (which essentially would be a disaster that freezes everyone in place and creates a permanent underclass) or will we evolve to a new, yet to be defined, system? and if so, how will the transition look?
maybe a better approach to start with computers that already have ergonomic chassis (they exist) and then spend energy for modifying tools on what happens inside of them?
hmm. hoping that all the weird business requirements get confined to a specific distro with careful gating prior to upstreaming. it would be bad if they were allowed to pollute the ecosystem more generally (which one could argue is why windows is the way it is).
i have been pretty happy with opensnitch. ui improvements are always welcome although what might be really interesting would be some sort of plug-in system that allows for an agent to watch my interactions activity and the outbound connections and only flag things that seem surprising. also maybe some kind of improvement over the pop-up (maybe get rid of them entirely and add some kind of cli wrapper that allow-lists child processes).
the experimental report pdf is a fun read. it would be cool if they added each individual bell of the duobell and the combination to tables 4 and 5. (the topline result is in the infographic, but it would be cool to see the effects of the individual contributions of the various features and how they combine)
ar(k) stuff, sure. that's old news. i would expect the newfangled stuff to be good at 0-shot learning of pre-event signatures spread across multiple series, at a minimum.
reply