creative @joinFAI
District of Columbia, USA
- Tweets 59,298
- Following 1,693
- Followers 5,432
- Likes 180,974
This post has been widely discussed, but I think worth highlighting the important implications it has for Christian alignment and computational theology research more generally.
First, the delicious implication that the puzzle of polysemanticity was a function of relatively primitive models means that the looming normative debate over alignment is closer than we might have imagined. Enormity helps. If more modern models naturally evolve towards interpretability, then one might suddenly end up with the unexpected luxury of asking not just "how" but "who" models are aligned towards. Lacking the excuse of inability, whether such powerful systems should answer to a company, a nationality, an ideology, or a theology comes into focus as the critical governance question.
Second, if the recursive generalization motif is the building block of neuronal meaning, then it may turn out that religious representations are a particularly interesting target from a mech interp point of view. Christian iconography and theology are by virtue of their long history highly compact computational objects: the Virgin Mary is the product of innumerable generalizations sprouting away from monosemantics. The complex "control flow" in Nick's vision governed by such resulting religious neurons represents the new challenging frontier as clean understandings of floppy eared dogs and automobile shapes come into focus.
interpretability might be the most important scientific problem in the world right now, as I don't think we can solve alignment if we don't know what we're building. I think I have good news about it. I co-authored early work in mechanistic interpretability at OpenAI and came away pessimistic, because neural networks appeared to fold into arbitrary tangled knots. years later, I started openai's first large-scale automated interpretability project, finding about 2% of neurons legible with the primitive tools we had
with modern agentic tools and a much closer look, the picture looks totally different. the vast majority in models up to trillions of parameters are legible, and their circuits beautiful. many famous examples of tangled neurons turned out to be misunderstood. The roadblocks I expected mostly aren't there. I now think ambitious interpretability will be solved by the time we need it most, and possibly quite soon, as automated scientists keep improving. it may be a bitter lesson for interpretability: the simple thing, at absurd scale, works
already-clean.goodfire.com/
3
2
51
4,808
Chick-fil-A® Sauce
Spicy Chick-fil-A® Sauce
Polynesian Sauce
Garden Herb Ranch Sauce
Honey Mustard Sauce
Barbeque Sauce
Honey Roasted BBQ Sauce
Sweet & Spicy Sriracha Sauce
Zesty Buffalo Sauce
What do these products have in common? One company could soon control them all.
Chick-fil-A is adding some heat to its signature sauce.
The chain has announced Spicy Chick-fil-A Sauce, its first new permanent sauce in 10 years. It mixes the familiar flavor of its classic sauce with spicy peppers and a touch of sweetness. Its last permanent sauce addition was Sweet & Spicy Sriracha in 2016.
(via @USATODAY )
2
2
28
1,429
Chris retweeted
Counter argument: Not only is taste subjective, it's a moving target. As soon as [tasteful thing] becomes easily replicable, it stops being tasteful.
Taste is a status game with a strong defensibility mechanism. That's why the term is so fuzzy and ill-defined. Vagueness is the whole point. It's meant to be hard to copy.
I no longer see any compelling reason to believe that taste is some permanent boundary LLMs won’t cross. It’s already happening.
104
160
1,569
74,469
commentators have successfully exposed 37 of the most recent 2 AI firm conspiracies
Are AI companies hyping up their technology's catastrophic tail-risks to distract media attention from its actual harms?
Probably not. And if so, the strategy is failing.
vox.com/politics/505834/ai-d…
3
413
Fascinating reading about how we used to write about malware and viruses back in 1999. In a decade I suspect we'll look back at how we talk about AI safety today and it will feel similar. schneier.com/crypto-gram/arc…
11
43
241
14,569
Chris retweeted
An admirable stance (one I wish more publications would hold the line on).
FAI's policy: we encourage full fluency with the latest AI tools (and deployment to help with research, ideation, typo-checks, microsites, etc.) but no writing for publication may be AI-generated.
The nib of it is that you cannot hold a computer responsible for what it writes, and journalism is about responsibility.
5
14
91
8,926
