Opens profile photo
Follow
Anders Sandberg
@anderssandberg
Academic jack-of-all-trades.
Joined September 2009

Anders Sandberg’s Tweets

I like this scheme, but I would add a seventh category of direct value change, eg. moral enhancement or modifying what gives us rewards.
Quote Tweet
What are the six ways that technology can change our moral beliefs and practices? @spillteori and I cover this in our recent paper. Here is a quick summary with some examples. As always, there is a lot more in the full paper (available here: link.springer.com/article/10.100)
Show this thread
3
8
This is actually a profound thing. As we progress, the space of possible action grows. That means new virtues and new responsibilities.
Quote Tweet
Stoicism teaches us to distinguish between what we can control and to accept things we cannot change. The problem is that mining stars is clearly not forbidden by the laws of physics. On a long enough time line, the set of things we cannot yet do will shrink to nothing.
Show this thread
1
37
This is an excellent overview of how nontrivial making current LLMs is. Especially the instruct approach is both very important for their practical utility, and makes the relationship between them and people convoluted.
Quote Tweet
The wisdom that "LLMs just predict text" is true, but misleading in its incompleteness. "As an AI language model trained by OpenAI..." is an astoundingly poor prediction of what a typical human would write. Let's resolve this contradiction — a thread:
Show this thread
19
First shocking adult realisation : "everybody's winging it all the time!" #2: "they are as dumb as me!" #3: "it still works, somehow?" #4: "the people who *think* they know what they are doing are really dangerous..."
Quote Tweet
One disappointing thing you discover about the Adult World is that the minimum competence level of professionals – doctors, lawyers – is much lower than you would hope Not a little bit lower, much lower
8
206
Researchers having to reformat papers to fit journals estimated to cost $230 million per year. Not sure I see that as a large problem, but it sure is annoying. It also feels like what AI might be very helpful for: can journal guidelines be adapted for AI?
2
19
Show this thread
This is very cool and useful. The real problem will be to get enough representation of different values in RLHF - both an issue of whose values get regarded as relevant and the effort to do RL - but that also suggests looking for techniques to make this cheaper may be good.
Quote Tweet
Human-aligned AI is a multi-objective problem. Yet, current RLHF prioritizes certain values when aligning LLMs, resulting in a lack of transparency and unfair representation of minorities. In our latest paper arxiv.org/abs/2306.04488), we embrace the diversity of human values.
Show this thread
1
6
Our lives form a joint 4D braid. Several billion threads weaving around each other, with dense tangles, many recurring patterns lasting weeks, years or decades. It also has an overall 4D shape set by human presence and absence.
Image
2
11
Show this thread
Our spatial presence is a kind of fractal, with long Levy flights to remote locations and dense clusters around home, work and friends. Here and there thin threads of hikes, exploration, or driving wrong.
Image
1
9
Show this thread
Generally, if we could see where we had ever been it would make a pattern where next to the many places we had been thousands of times there are places we have never been. Obviously a few meters up in the air, but also many closets, corners, and neighbouring flats.
2
15
Show this thread
Ok, seems to have been officially denied. So, assuming we trust this too (I bet there will be a string of confusing clarifications next), storm in water glass. Keep moving, nothing to see here. AI drones are completely safe and nothing to worry about.
8
17
Show this thread
Ok, seems to have been officially denied. So, assuming we trust this too (I bet there will be a string of confusing clarifications), storm in water glass. Keep moving, nothing to see here. AI drones are completely safe and nothing to worry about.
3
3
Show this thread
It seems this story is not true... or at least there is an official claim that it is not true/mistaken. Fine, my credulous bad! But I wonder why so many people keep on replying that it is not true: dont you read the other replies?
Quote Tweet
OK, this story is just unbelievably topical right now... and utterly an "I told you so" from the perspective AI safety community (especially the twist at the end). aerosociety.com/news/highlight
Show this thread
Image
8
33
Show this thread
Thinking about this kind of messiness is healthy and I hope makes one a bit of a better person. It also clashes with the often too glib Pride flagwaving we also want to do: we need heroes and uplifting stories of integrity. Reality is complex: that takes real courage to face.
Image
9
Show this thread