Mac Budkowski ᵏ pfp
Mac Budkowski ᵏ
@macbudkowski
I read this pmarca essay and apart from saying "people been scared of tech for ages", "it's hard to measure the problem" and the quote below he doesn't really address AI doomers arguments. I like his essays but got deeply disappointed with this one. https://pmarca.substack.com/p/why-ai-will-save-the-world
6 replies
0 recast
5 reactions
🤷🚂👲🧑‍🤝‍🧑 pfp
🤷🚂👲🧑‍🤝‍🧑
@m-j-r.eth
imho the main X-risk was sufficiently addressed, but the one thing that stuck out to me was the glossing over of "productivity != local compensation". don't think the essay confronted the imperative to maintain a domestic consumer base of physical goods/services over the global nonscarcity of productive software.
1 reply
0 recast
0 reaction
Mac Budkowski ᵏ pfp
Mac Budkowski ᵏ
@macbudkowski
Which argument do you think addressed the X-Risk properly? Maybe I missed something
1 reply
0 recast
0 reaction
🤷🚂👲🧑‍🤝‍🧑 pfp
🤷🚂👲🧑‍🤝‍🧑
@m-j-r.eth
all of AI Risk #1 & reference to criminal behavior at the end was sufficient. just because one force multiplier exists does not mean that the underlying risk is unlikely to occur. bio/misinfo/robo all rely on human commission of that "attack", all preexistant w/o multiplier. and PGP = munition is invalidated already.
2 replies
0 recast
0 reaction
Mac Budkowski ᵏ pfp
Mac Budkowski ᵏ
@macbudkowski
I think these are decent argument but for me the central question is: "How do you control something that's 100X smarter than you?". I haven't found an answer to that.
1 reply
0 recast
0 reaction
🤷🚂👲🧑‍🤝‍🧑 pfp
🤷🚂👲🧑‍🤝‍🧑
@m-j-r.eth
it hasn't been deployed as 100x INT w/o HITL design. autonomous LLM hasn't demonstrated 100x INT. and we should be focused on 100x efficiency for sum of several billion online user's INT. control is wicked easy when the interface & mass consensus are one and the same. who's throwing a datacenter at a monolith?
2 replies
0 recast
0 reaction
Mac Budkowski ᵏ pfp
Mac Budkowski ᵏ
@macbudkowski
Ofc it hasn't but that's the end-game. If you want to augment people's intelligence, you want to add as much as you can. If you don't, your competitor will do it and get you out of business. That's how incentives are designed.
1 reply
0 recast
0 reaction
🤷🚂👲🧑‍🤝‍🧑 pfp
🤷🚂👲🧑‍🤝‍🧑
@m-j-r.eth
I don't see the real end-game, most of the results are at the ground floor, in the hands of engineers stumped by JSON and memory bugs. 100X INT is soonest and likeliest from public market, so really the question of AI risk is "which misanthropic firm offers the most incentive for consumption?", not the inflection point
1 reply
0 recast
0 reaction
Mac Budkowski ᵏ pfp
Mac Budkowski ᵏ
@macbudkowski
We are at the early stage and there are def some huge challenges to solve on the way. They might take a few years to be fixed or a few decades. On the other hand, I spoke with one of the OpenAI Co-Founders few weeks ago, and he told me they expected to reach ASI in 10 years.
1 reply
0 recast
0 reaction
🤷🚂👲🧑‍🤝‍🧑 pfp
🤷🚂👲🧑‍🤝‍🧑
@m-j-r.eth
I think that if the public had a free market hub for multiparty, autonomous agent control, we'd get: - AGI in a couple of years - ASI before 2030 the issue has been the publication of supply chain (chipfab, robotics, etc) & the lack of incentive for crowdsourcing. 2020 gave us exascale Folding@Home, where's ML@Home?
0 reply
0 recast
0 reaction