0 reactions
every tech revolution used the last one's speed to fool us. this time we might not get 20 years to adapt
read something that made me uncomfortable. every major tech shift took longer than people thought to arrive, but once it did, we had time to build safety frameworks
steam engine to factory safety laws: 70 years
second industrial revolution to labor protections: 30 years
nuclear weapons to arms control treaties: 20 years
internet to basic regulations: 20 years
each time, society had a window to figure out guardrails
but each revolution also moved faster than the last. and we keep using the previous speed to estimate the next one
right now AI task completion time doubles every 7 months (according to some research group called Meter). early 2024 models could handle a few minutes of work. now they can do 5-10 hour tasks independently
if that curve continues, we're looking at models that can work for days or weeks without human intervention within a year or two
the uncomfortable part: we probably don't have 20 years to figure out safety frameworks this time. maybe not even 5 years
nuclear weapons gave us the cuban missile crisis. but before that, we had 20 years of smaller conflicts to learn boundaries. kennedy and khrushchev knew where the lines were because they'd spent two decades testing them
with AGI we might not get that learning period. the gap between "AI that needs supervision" and "AI that doesn't" could be really short
been thinking about this in my own work. using ai coding tools and the capability jump in just the last year is noticeable. stuff that needed constant hand-holding 6 months ago now runs mostly autonomous. tried cursor, verdent, couple others. all of them got way better at handling complex tasks without breaking things
not saying AGI is here. but the "we'll figure it out when we get there" approach feels riskier when "there" might arrive faster than the time it takes to build consensus on what "figured out" even means
the article mentioned something about trust being a slow variable. you can't speed up institutional trust or regulatory frameworks the way you can speed up model training
so what happens when the tech moves faster than our ability to build social/political structures around it
feels like we're in uncharted territory but maybe im wrong
#technology
source
read something that made me uncomfortable. every major tech shift took longer than people thought to arrive, but once it did, we had time to build safety frameworks
steam engine to factory safety laws: 70 years
second industrial revolution to labor protections: 30 years
nuclear weapons to arms control treaties: 20 years
internet to basic regulations: 20 years
each time, society had a window to figure out guardrails
but each revolution also moved faster than the last. and we keep using the previous speed to estimate the next one
right now AI task completion time doubles every 7 months (according to some research group called Meter). early 2024 models could handle a few minutes of work. now they can do 5-10 hour tasks independently
if that curve continues, we're looking at models that can work for days or weeks without human intervention within a year or two
the uncomfortable part: we probably don't have 20 years to figure out safety frameworks this time. maybe not even 5 years
nuclear weapons gave us the cuban missile crisis. but before that, we had 20 years of smaller conflicts to learn boundaries. kennedy and khrushchev knew where the lines were because they'd spent two decades testing them
with AGI we might not get that learning period. the gap between "AI that needs supervision" and "AI that doesn't" could be really short
been thinking about this in my own work. using ai coding tools and the capability jump in just the last year is noticeable. stuff that needed constant hand-holding 6 months ago now runs mostly autonomous. tried cursor, verdent, couple others. all of them got way better at handling complex tasks without breaking things
not saying AGI is here. but the "we'll figure it out when we get there" approach feels riskier when "there" might arrive faster than the time it takes to build consensus on what "figured out" even means
the article mentioned something about trust being a slow variable. you can't speed up institutional trust or regulatory frameworks the way you can speed up model training
so what happens when the tech moves faster than our ability to build social/political structures around it
feels like we're in uncharted territory but maybe im wrong
#technology
source
| 3,000 mlx | |
| 4 | |
| 0 |