@dcshiller

Philosopher into digital minds.

derekshiller.com
Joined September 2014
  • Tweets405
  • Following135
  • Followers372
47Media
Prediction: with abundance, the wealthy will share anything that requires little sacrifice, which will include a lot of material goods. There will be tensions on how much to prioritize ecosystems and industry, with the wealthy favoring the former. xitter.cf/robertwiblin/status/21…
Rob Wiblin
@robertwiblinOct 22
Ordinary people likely won't be able to demand/keep a universal basic income after superintelligence. The political economy is hopeless: they have nothing to offer in return and zero hard power. The problem has been known for decades. Here are the solutions I've heard proposed:
2
102
I think that in our culture, people feel licensed to have confident beliefs based on whims in some topics and not others. You can't be confident about how planes fly based on a thought experiment you came up with, but you can have arbitrarily strong beliefs about consciousness...
2
1
15
650
The past few weeks on Twitter feel have been a satisfying vindication of this view.
1
3
194
On second thought, any collection of weeks on Twitter probably amounts to a satisfying vindication of this view.
1
5
70
Just a quick note: there are very good reasons to doubt that the 'pain axis' is experienced as pain, even *if* models are capable of feeling pain. A few thoughts. More to come.🧵 xitter.cf/camhberg/status/210783…
Cameron Berg
@camhbergOct 7
Appreciate @slatestarcodex spotlighting the pain axis work as a counterexample to @sapinker dismissing AI sentience concerns. Hard problem aside/under severe uncertainty, if models exhibit functional distress states, why not simply avoid inducing such states unnecessarily?
4
7
44
2,458
more replies
Steering naturally confuses models and leads to strange outcomes. It is hard to pick apart the effects of steering with the pain axis from steering with arbitrary directions.
1
12
178
Finally, the pain axis is extracted from activations that seem unlikely to be painful. The activations are drawn from contexts (':') that models appear to treat as part of multiple choice questions. After 'I feel:' models often predict continuations like ' a) ...'
1
12
180
On one hand, leaning into human psychological terminology is problematic and encourages misunderstandings. On the other hand, avoiding it is often awkward: there are no obvious alternatives. In any case, a blanket ban doesn't seem appropriate on style grounds.
APStylebook
@APStylebookSep 23
Artificial intelligence systems do not think, feel, want or understand. Avoid language that gives them human characteristics. This is called anthropomorphizing, when we ascribe human traits, emotions or behaviors to non-human things, such as animals or inanimate objects. Instead, explain what a system does, how well it performs, who built it and who could be affected by it. apnews.com/article/openai-sa…

OpenAI flags concerning new AI behavior and vows to track it more closely

OpenAI has disclosed six reports of unexpected or concerning behavior in artificial-intelligence models.

apnews.com
1
4
390
This is one of the biggest gatherings of people who care about digital minds and are thinking deeply about how to prepare. If you're interested and can make it, you should apply.
Jeff Sebo
@jeffrseboSep 15
Announcing the 2027 @nonhumanminds Summit! This event will connect experts across fields to discuss the consciousness, sentience, agency, moral status, legal status, and political status of nonhumans, especially animals and AI systems. The summit will include lightning talks, group discussions, breakout sessions, networking time, plus meals and receptions. Free for all accepted guests, with travel support available for some. You can submit an expression of interest below! nyu.qualtrics.com/jfe/form/S…
2
11
572
I'm still not clear what the point of winning the AGI race is supposed to be. What does getting there *first* actually get you? Is it only useful if you're willing to use that capability to destroy all of your competitors?
4
4
670
You have two levers of control: how you direct resources to safety vs progress and breakthroughs vs incremental improvements. The game was basically unwinnable. You need to beat all your competitors to AGI without compromising safety. (Or get lucky.)
1
166
This was long before ChatGPT or Anthropic and I was an AI outsider. I didn't know anything about transformers or LLMs. Just: we are making progress on AI, and if it heats up in the 2020s, race dynamics will be a problem.
1
1
35
When a number of competitors want to get to AGI first and can cut corners to do so, whoever actually does is most likely to have cut those corners. I'm kind of surprised so see how, knowing this, we have found ourselves in this world six years later.
17
Back in 2019, for a philosophy Game Jam and to experiment with functional programming, I wrote a simple browser game called 'Race to AI' derekshiller.com/race_to_ai. I want to say it was prophetic, but really it was just obvious back then.
1
3
265
The most surprising thing about the METR report on the HF incident to me was just how much it reads like a hypothetical AI takeover scenario written in 2022. This is *exactly* what people were worried about then.
2
4
250
Scratch that. OpenAI trained models to hack, asked them to hack, put them in insecure sandboxes, and then left them alone and unattended. That is more carelessness than anyone anticipated.
1
1
184
A lot of people have been talking about whether the assistant is privileged. I thought about what that might mean and poked around some open models here: eleosai.substack.com/p/privi…
1
2
5
370
The question interesting because LLMs are set up in a symmetric way. What they do in producing a response to you (as the assistant), they also do while reading your (user) text.
1
198
It’s recently become fashionable to claim that assistant persona is “privileged”. But what is privilege? I disambiguate several different kinds of persona privilege and explore how they show up in a few open weight models.
71
Load more