#nla — Public Fediverse posts
Live and recent posts from across the Fediverse tagged #nla, aggregated by home.social.
-
Morning, today we're looking at #MechanisticInterpretability , a subfield of AI Safety that attempts to understand the inner workings of artificial intelligence by analysing concrete structures, algorithms and circuits.
Why do we even need to do this? Because the hope of understanding neurons as being features died on polysemanticity -- models represent more features than dimensions by assigning them to an overcomplete set of non-orthogonal directions (i.e. you can't hope that concepts can be broken down into linear combinations of features).
This isn't a new field, people have been projecting intermediate GPT layers through the final layer activation and looking at the top-k most likely next tokens since GPT-2.
And that's how we arrive at the Natural Language Autoencoder -- an autoencoder over residual-stream activations where the bottleneck is natural language text instead of a sparse vector.
Nothing in the training objective requires the verbalisation bottleneck to be readable, faithful, or even semantically related to the intermediate layer being investigated, so while the training algorithm is typical of autoencoders, two things are distinct here:
1) both parts of the autoencoder (here the activation-verbaliser, AV, and the activation-reconstructor, AR) undergo special SFT to ensure the verbaliser and reconstructor can generate and read natural language from intermediate layer activations
2) a KL-divergence penalty is baked into the objective to make sure that, during joint training, the NLA doesn't diverge from the SFT versionA side-product of this training is that we can take the current verbalisation and "desired" verbalisation and compute steering vectors from their difference.
Surprisingly, Anthropic didn't find any evidence that the NLA was engaging in steganography to smuggle information to avoid human detection, but it did find that NLAs tend to confabulate a lot.
#AIResearch #Transformers #TransformerCircuits #Anthropic #NLA
-
Making a tutorial is much harder than just editing my logs.
But I'm a good way through editing the demo screencast.
And I've started applying the technique to the scene I needed it for. It's still pretty complicated, but it's much easier when you know what's going on! #Blender3D #Animation #NonLinearAnimation #NLA -
I've gotten bogged down in NLA Editor mechanics, so I've decided I'm going to have to do a little simplified learning project to make sure I understand how the various settings on action strips interact.
Tutorial time!
Going to actually make some effort to film & screencast this one better. Basically a tutorial, but it's going to be kind of unscripted and open-ended, because I'm not sure how it's going to end, either!
Should be exciting. Or at least interesting.
:huh::ani_dunno:
I think I have a plan. Going to start setting up for it tomorrow.
I guess this could also count as my "monthly project", so a return to that, too. #LunaticsProject #Tutorial #Blender3D #NLA #Animation -
Eine Bildergeschichte des Landkreises Wittmund der 1960er und 1970er Jahre – Der Nachlass des Pressefotografen Ehnt Ulfert Janssen
-
I believe we should start refer to all the #Crap #AI as Natural Language Applications, it makes more sense and it is more rational and true.
If you call them NLA we can chill out all the stupid hype around them and let people have a sane approach with these technologies; right now people are using them as oracles to get proper answer on anything because the hype and the misleading name.
-
Hey Australians!
Tell the #NationalLibrary that you want them to be on #Mastodon:
https://docs.google.com/forms/d/e/1FAIpQLSeU0qddTjTS_j7TG6t9MtuKWk7yWNYtRZZW-z6Ax2x6V7jOfA/viewform
-
Just did a quick feasibility test connecting to a remote desktop from a browser running in the network of my employer ... working perfectly fine with my old DSL-50 connection 🥳
I think #firewalls these days are a bit silly, whatever they don't allow must be forced to work via #http (or better #https) instead 😂 -- and #guacamole is really a nice solution to do this for remote desktops. But as it's acting as a #gateway also handling #authentication, I finally feel comfortable making a connection with #xrdp accessible, which still doesn't support #NLA. When you see the "login screen" in guacamole, you are already authenticated ... so this feature lack is downgraded from a severe security concern to a mild annoyance (having to type your password again).
-
Herbstmorgendlicher (Quasi-) #Archivhimmel #NLA #Aurich
-
Morgendlichen #Archivhimmel hatte ich noch nie, glaube ich. #NLA #Aurich
-
Vorübergehend trockener #Archivhimmel am #NLA #Aurich
-
Якщо в когось є пропозиція, які емодзі додати, пишіть під цим постом)
*лише для користувачів NL Social
-
Цікавинка для нових користувачів NL Social:
В налаштуваннях ви можете увімкнути альтернативну тему Mastodon: Elephant.