Opens profile photo
Follow
Click to Follow sbeckerkahn
Spencer
@sbeckerkahn
AI Safety researcher. Formerly . Before that, pure math
London, EnglandJoined September 2021

Spencer’s Tweets

Am just about secure enough to say that I applied to DeepMind for the recently-advertised Research Scientist role in the AI Safety team and got the fastest rejection I have ever gotten for anything in my life.
Quote Tweet
Props to Luisa for being brave enough to share this! Stuff like this can be very vulnerable to share, but I think it can be motivating for those who are actively trying to apply to things. Maybe one day I will share the long list of rejections I’ve had!
Image
2
102
There was a tweet I saw a few days ago - which obviously I now can't find - that was something like 'what's the one thing you wish people with PhDs in other fields knew about your field?'
1
AI Safety going mainstream has massively threatened my fragile enjoyment of twitter. I now get major 'someone is wrong on the internet' vibes every time I use it. Had a golden few months of it feeling quite tpot-y and community-ish but maintaining that takes more + more work
5
21
I've sort of stopped listening to all the long podcasts in AI Safety etc. Did this happen to anyone else as they got more into the field? It starts to feel like that thing where you kind of obsess about work outside of work hours and its actually unproductive.
6
22
Pretty much a god tier tweet
Quote Tweet
FOX MULDER: Over 20 reported killer whale attacks in the month of May. The whales worked in tandem. Do you know what this means, Scully? DANA SCULLY: No. FOX MULDER: The attacks were orca-strated. Scully? Did you hear me? They were or- DANA SCULLY: I’ll be in the car, Mulder.
Image
It’s very Gell-Mann-ish to be so aware of how perilously close the kind of AI content I actually want to consume on twitter is to the kind of content I want to actively avoid.
Has ‘s rhetoric got more unhinged and grandiose the further he’s gotten through his ‘world tour’?
  • Yes
    37.5%
  • No
    62.5%
8 votesFinal results
"Of course, if Muad'Dib were the Lisan al Gaib, our reverend mothers would tell us, wouldn't they? After all, we supply melange to these addicts so they can guide us. It's their job. Well that's where it gets interesting. Our producers reached out to the Padishah Emperor..."
Image
32
4,129
It’s worth mentioning that I think the argument for existential risk is quite simple. Here is the argument: • There is nothing special about brains made of carbon instead of silicon. They can hypothetically develop traits like intelligence, understanding, motivation, etc. This… Show more
18
55
Show this thread
I'd like to have a real conversation about whether AI is a risk for human extinction. Honestly, I don't get how AI poses this risk. What are your thoughts? And, who do you think has a thoughtful perspective on how AI poses this risk that I should talk to?
745
3,969
This was fun! I tried to emphasize to (Chair of the House AI Caucus) that we should take the threat from future superhuman AIs seriously alongside her current worries about misinformation. Given all the talk about regulating AI on Twitter, I was surprised to hear… Show more
Quote Tweet
“This is a unique moment. The public sector is looking to the private sector to be instructive to us. And it’s vice versa. We need each other in this. We have to get this right.” @RepAnnaEshoo on Building Safe AI at @Techweek_ #SFTechWeek #TechWeek
Image
Image
21
117
i.e. for this 2-d hidden layer, given that W is a matrix whose columns are a linear phase from each other, its like taking two rows of the full DFT matrix (up to another rotation of 2-d space afterwards)
1
Show this thread
Something else which, to me, also always seemed related (but I haven't given much thought) is that in Toy Models of Superposition, when you see a regular polygon (and if you ignore the fact its real instead of complex) W is something like a 'restricted' discrete Fourier Transform
Image
Quote Tweet
I briefly tried to write up where this thought was going but the discussion in the comments has now progressed beyond where I left it lesswrong.com/posts/tdENX8dz
Show this thread
1
1
Show this thread
I'm worried this is actually a misconception i.e. the "bizarre and inhuman" part. It's interesting and maybe even remarkable but - although it's subjective - I'd say it's a fairly natural approach mathematically.
Quote Tweet
Their results are bizarre and inhuman. @NeelNanda5 trained a tiny transformer to do addition, then spent weeks figuring out what it was doing - one of the only times in history someone has understood how a transformer works. This is the algorithm it created. To *add two numbers*!
Show this thread
Image
6
76
Show this thread
Running a reading & discussion session today at the SERI MATS London base on 's compute monitoring paper 'What Does it Take to Catch a Chinchilla' arxiv.org/abs/2303.11341 (Not a public/streamed event but if it goes well I can imagine doing such a thing in the future.)
6
One reason I now agree is that the meaning of 'LLM' has shifted. The 'M' = model in LLM used to refer to the underlying transformer model that is predicting the next token, but LLM is now synonymous w/ something that's been RLHFd and outputs long-form completions by default
Quote Tweet
I don’t think LLMs can generally best be described by “trying to predict the next token”, for these reasons: twitter.com/daniel_271828/…
6
Is there some app/website that can take a two-column pdf - possibly with formulae and tables and stuff - as input and give me something nice to read?
2
4
Me: *with a sense of wry amusement as my atoms are disjoined for harvesting by an unfathomable superintelligence*: Maybe transformers really were...robots...in disguise.
1
20
At a glance I thought this was a 'what the hell happened in-between?' situation but it's literally just the direct reply lol
Image
1
7
I've completely changed my mind on technical solutions to alignment vs governance. I thought governments would be far too slow to enact meaningful policies, and the engineering approach was the only hope, now I think regulation is the only thing that can slow down capabilities
2
14
Seeing many takes pointing to the idea that Altman's concerns about EU regulation of ChatGPT are somehow inconsistent with his stated views on future global coordination/oversight on AGI. But fwiw (+ as you can tell from my phrasing), I don't think there's necc. an inconsistency
2
'AI timelines are really short' and 'only consider the effects of your actions on the short term' are very different claims, and I don't think the first claim implies the second (though it *is*, if you believe it, a big deal and a reason to up-weight short-term effects).
2
10
Show this thread
Although you may disagree with his optimism, he’s way way better than the worst case scenario of someone who refuses to engage. I mean he literally sat there and basically said ‘don’t get me wrong, the alignment problem isn’t solved; we don’t yet know how to do this’
11
Show this thread
Altman emphasised that the problem is far from solved and while he didn’t mention ‘extinction’ or convey any visceral sense of risk, he - imo - spoke as someone who at least intellectually understand the reality.
1
7
Show this thread