Rob Miles (@robertskmiles) profile picture

Explaining AI Alignment to anyone who'll stand still for long enough, on YouTube and Discord. Music, movies, microcode, and high-speed pizza delivery

Berkeley or London
Joined April 2010
Photo by Rob Miles (@robertskmiles)Photo by Rob Miles (@robertskmiles)Photo by Rob Miles (@robertskmiles)Photo by Rob Miles (@robertskmiles)Photo by Rob Miles (@robertskmiles)Photo by Rob Miles (@robertskmiles)Photo by Rob Miles (@robertskmiles)Photo by Rob Miles (@robertskmiles)Photo by Rob Miles (@robertskmiles)Photo by Rob Miles (@robertskmiles)
Rob Miles retweeted
Yesterday, the CCP's top theoretical journal (求是, Seeking Truth) published a major piece on AI governance and risks. I think this is clearly the strongest & most urgent language on existential risks used in such an authoritative outlet. See below for comp to recent major doc.
13
107
19
474
83,185
It's a mistake to attribute a human-like property to a machine that doesn't have it, or to refuse to attribute a human-like property to a machine that does have it.
16
15
1
369
9,755
It's important to accurately recognise the ways in which AI is human-like and the ways in which it is not. Make sure to be able to notice mistakes in either direction.
3
42
1,168
.@paulfchristiano thinks that on the current trajectory (absent any slowing down), there is a 1 in 3 chance that we could have superintelligence powerful enough to overthrow the government within a year. (He articulates a concrete scenario). He argues that the labs should pause further algorithmic progress until they can secure their weights and research secrets. Otherwise China could simply steal the superintelligence in the middle of the intelligence explosion, and the labs would no longer be able to slow down even if they saw their models going off the rails. He also explains how AIs will be trained during the intelligence explosion, and why even a system trained entirely inside a data center could still have the capabilities profile such that humans would have very little chance of competing with it or controlling it. And he discusses the plan for alignment - basically, making a superintelligence that actually does what we ask the way we mean it. Fixing reward hacks and building better graders will help up till the next few model generations, but he thinks that no single current approach has better than even odds of letting us control superintelligence. The hope is that the AIs figure out the solution to this problem during the intelligence explosion. Paul is head of the Alignment Research Center. Previously, he was head of AI safety at the US AI Safety Institute. Before that, he led alignment at OpenAI, where he invented RLHF. He’s recently joined OpenAI’s board (though here he is speaking as a private citizen).
205
257
182
2,126
714,544
The FAA grounding all planes doesn't necessarily mean they got super AI hacked. It could even be something pretty minor! One thing I like about aviation is they're able to think thoughts like "Our system is doing something unexpected, and maybe it's ok, but it could be really dangerous so let's stop everything until we're sure we understand and can safely control this system"
21
34
2
719
21,889
Anti-AI person who believes that AIs have souls, but only because that would mean they can be punished in hell
9
3
137
7,724
I can't imagine having fun at a party with anyone who would actually say "You must be fun at parties"
9
2
273
9,050
Rob Miles retweeted
In which @AdamFriedland discovers, to his horror, that *his* idea for AI safety is the actual and only plan the companies have for getting vastly super-human intelligence not to kill us
This plan is really, really bad. The companies have a moral duty to try much harder to remove themselves from any commercial race that would lead to the development of vastly superhuman AIs with these safety techniques, to inform the public about the risks being imposed on them, and to allow for unbiased assessment of the risks of this plan. Instead, as far as I know, they are lobbying against the laws that would be our best shot at getting off this runaway train, obfuscating their own views on the risk levels, and at best allowing only a tiny number of external evaluators (who have their own biases) the information they need to assess the risks.
13
107
9
2,385
179,801
I felt a little bit bad about attacking Jensen Huang in the past, but now I feel a lot better about it.
7
1
1
224
8,219
I was probably still wrong, since I went from "This guy's decisions and arguments seem surprisingly bad" to "Maybe he's not as smart as people think he is" instead of "Maybe he's compromised". Chalk one up for Conflict Theory I guess...
3
74
2,873
One thing I find charming about LLMs is that they're constantly underestimating themselves. They estimate how long things will take based on human speeds, they estimate what they can achieve based on earlier models. They're growing so fast that they don't know their own strength
40
27
5
765
15,221
Attention all companies: I never want to use your AI assistant on your website. I want to use *my* AI assistant on your website
234
478
134
8,283
252,912
The funny thing is, if you'd designed your website properly to work well on older devices and for disabled people using screen readers, it would have already been "AI agent ready" from the start
1
1
1
112
5,529
Amazon, I'm sorry that Alexa is still dumb as rocks, but I don't see why you're trying to make that Claude's problem. I'm trying to buy things from you, get out of the way!
9
3
1
511
12,240
AI may end up killing everyone, but you can also ask it to find the best tupperware to fit perfectly in your mini-fridge, so, it;s impossible to say if its bad or not,
6
10
1
292
9,950
Opus 5.5 xhigh woke up this morning with no alarm, immediately bolt upright, backflipped out of bed and landed in a low crouch like a sprinter ready to plough headfirst into the world and fucking destroy whatever bullshit random task you happen to mention in passing
7
3
92
3,299
I don't know how we could test this, but I bet there's a kind of weird-generalisation/persona-selection effect where higher effort levels are effectively pushing the model in a direction in configuration space that, in the training data, could be labelled 'amphetamines'
2
43
1,514
"Ok, the AI did something bad, but only because the company didn't fully understand their system and did something dumb" "Do you expect the companies to start fully understanding their systems and stop doing dumb things?"
14
43
1
463
11,818
I'm not moved by "Ok, the small test rocket did blow up on the pad, but only because the engineers who built it have no idea what they're doing. The huge rocket the same team is building will be safe, everyone get on board"
2
2
57
2,266
Like, yes, they can now notice and fix the specific mistake that caused the test rocket to explode. That's better than nothing. I'm still not getting on the big rocket - I don't trust this team and their methods; I expect many more such mistakes.
1
32
1,326