Jacqueline Bellon (jacbellon.bsky.social) (@JacquesBellon) header
Jacqueline Bellon (jacbellon.bsky.social) (@JacquesBellon) profile picture
Joined January 2017
Photo by Jacqueline Bellon (jacbellon.bsky.social) (@JacquesBellon)Photo by Jacqueline Bellon (jacbellon.bsky.social) (@JacquesBellon)Photo by Jacqueline Bellon (jacbellon.bsky.social) (@JacquesBellon)Photo by Jacqueline Bellon (jacbellon.bsky.social) (@JacquesBellon)Photo by Jacqueline Bellon (jacbellon.bsky.social) (@JacquesBellon)Photo by Jacqueline Bellon (jacbellon.bsky.social) (@JacquesBellon)Photo by Jacqueline Bellon (jacbellon.bsky.social) (@JacquesBellon)Photo by Jacqueline Bellon (jacbellon.bsky.social) (@JacquesBellon)Photo by Jacqueline Bellon (jacbellon.bsky.social) (@JacquesBellon)Photo by Jacqueline Bellon (jacbellon.bsky.social) (@JacquesBellon)
Jacqueline Bellon (jacbellon.bsky.social) retweeted
why does this happen? the model believes there's a seahorse emoji, sure, but why does that make it output a *different* emoji? here's a clue from everyone's favorite underrated interpretability tool, logit lens! in logit lens, we use the model's lm_head in a weird way. typically, the lm_head is used to turn the residual (the internal state built up over the model layers) into a set of token probabilities after the final layer. but in logit lens, we use the lm_head after *every* layer - showing us what tokens the model would output if that layer were the final layer. for early layers, this results in hard-to-interpret states. but as we move through the layers, the model iteratively refines the residual first towards concepts useful for continuing the text, and then towards the final prediction. looking at the image again, at the final layer, we have the model's actual output - ĠðŁ, IJ, ł - aka, an emoji byte prefix followed by the rest of the fish emoji. (it looks like unicode nonsense because of a tokenization quirk - don't worry about it. if you're curious, ask claude about this line of code: `bytes([byte_decoder[c] for c in 'ĠðŁIJł']).decode('utf-8') == ' 🐠'`) but look what happens in the middle layers - we don't just get emoji bytes! we get those *concepts*, specifically the concept of a seahorse. for example, on layer 52, we get "sea horse horse". later, in the top-k, we get a mixture of "sea", "horse", and that emoji prefix, "ĠðŁ". so what is the model thinking about? seahorse + emoji! it's trying to construct a residual representation of a seahorse emoji. why would it do that? well, let's look at how the lm_head actually works. the lm_head is a huge matrix of residual-sized vectors associated with token ids. when a residual is passed into it, it's going to compare that residual with each token vector, and in coordination with the sampler, select the token id with a vector most similar to the residual. (more technically: it's a linear layer without a bias, so v @ w.T does dot products with each unembedding vector, then log_softmax and argmax/temperature sample.) so if the model wants to output the word "hello", it needs to construct a residual similar to the vector for the "hello" output token that the lm_head can turn into the hello token id. and if the model wants to output a seahorse emoji, it needs to construct a residual similar to the vector for the seahorse emoji output token(s) - which in theory could be any arbitrary value, but in practice is seahorse + emoji, word2vec style. the only problem is the seahorse emoji doesn't exist! so when this seahorse + emoji residual hits the lm_head, it does its dot product over all the vectors, and the sampler picks the closest token - a fish emoji. now, that discretization is valuable information! you can see in Armistice's example that when the token gets emplaced back into the context autoregressively, the model can tell it isn't a seahorse emoji. so it tries again, jiggles the residual around and gets a slightly different emoji, rinse and repeat until it realizes what's going on, gives up, or runs out of output tokens. but until the model gets the wrong output token from the lm_head, it just doesn't know that there isn't a seahorse emoji in the lm_head. it assumes that seahorse + emoji will produce the token(s) it wants. ------------------ to speculate (even more), i wonder if this a part of the benefit of RL - it gives the models information about their lm_head that's otherwise difficult to get at because it's at the end of the layer stack. (remember that base models are not trained on their own outputs / rollouts - that only happens in RL.)
44
150
50
1,281
180,843
Jacqueline Bellon (jacbellon.bsky.social) retweeted
Putting wheels on this cross to bear is the best possible metaphor for American Evangelicalism
1,216
18,069
867
186,685
5,426,026
Jacqueline Bellon (jacbellon.bsky.social) retweeted
🟠 En menos de un mes, los mayores especialistas en Simondon del planeta se reúnen en Lisboa. cful.letras.ulisboa.pt/praxi…
1
10
2
18
1,856
Jacqueline Bellon (jacbellon.bsky.social) retweeted
The #NobelPrizeinPhysics2024 for Hopfield & Hinton rewards plagiarism and incorrect attribution in computer science. It's mostly about Amari's "Hopfield network" and the "Boltzmann Machine." 1. The Lenz-Ising recurrent architecture with neuron-like elements was published in 1925 [L20][I24][I25]. In 1972, Shun-Ichi Amari made it adaptive such that it could learn to associate input patterns with output patterns by changing its connection weights [AMH1]. However, Amari is only briefly cited in the "Scientific Background to the Nobel Prize in Physics 2024." Unfortunately, Amari's net was later called the "Hopfield network." Hopfield republished it 10 years later [AMH2], without citing Amari, not even in later papers. 2. The related Boltzmann Machine paper by Ackley, Hinton, and Sejnowski (1985) [BM] was about learning internal representations in hidden units of neural networks (NNs) [S20]. It didn't cite the first working algorithm for deep learning of internal representations by Ivakhnenko & Lapa (Ukraine, 1965)[DEEP1-2][HIN]. It didn't cite Amari's separate work (1967-68)[GD1-2] on learning internal representations in deep NNs end-to-end through stochastic gradient descent (SGD). Not even the later surveys by the authors [S20][DL3][DLP] nor the "Scientific Background to the Nobel Prize in Physics 2024" mention these origins of deep learning. ([BM] also did not cite relevant prior work by Sherrington & Kirkpatrick [SK75] & Glauber [G63].) 3. The Nobel Committee also lauds Hinton et al.'s 2006 method for layer-wise pretraining of deep NNs (2006) [UN4]. However, this work neither cited the original layer-wise training of deep NNs by Ivakhnenko & Lapa (1965)[DEEP1-2] nor the original work on unsupervised pretraining of deep NNs (1991) [UN0-1][DLP]. 4. The "Popular information" says: “At the end of the 1960s, some discouraging theoretical results caused many researchers to suspect that these neural networks would never be of any real use." However, deep learning research was obviously alive and kicking in the 1960s-70s, especially outside of the Anglosphere [DEEP1-2][GD1-3][CNN1][DL1-2][DLP][DLH]. 5. Many additional cases of plagiarism and incorrect attribution can be found in the following reference [DLP], which also contains the other references above. One can start with Sec. 3: [DLP] J. Schmidhuber (2023). How 3 Turing awardees republished key methods and ideas whose creators they failed to credit. Technical Report IDSIA-23-23, Swiss AI Lab IDSIA, 14 Dec 2023. people.idsia.ch/~juergen/ai-… See also the following reference [DLH] for a history of the field: [DLH] J. Schmidhuber (2022). Annotated History of Modern AI and Deep Learning. Technical Report IDSIA-22-22, IDSIA, Lugano, Switzerland, 2022. Preprint arXiv:2212.11279. people.idsia.ch/~juergen/dee… (This extends the 2015 award-winning survey people.idsia.ch/~juergen/dee…)
208
1,161
490
5,318
1,177,650
Jacqueline Bellon (jacbellon.bsky.social) retweeted
Are we ready for a world where our data is exposed at a glance? @CaineArdayfio and I offer an answer to protect yourself here: tinyurl.com/meet-ixray
373
1,559
1,133
5,240
1,250,857
Jacqueline Bellon (jacbellon.bsky.social) retweeted
‼️🌍 Programa provisional del congreso de Lisboa por el centenario de Gilbert Simondon. Si estáis por la ciudad en esas fechas, no me lo perdería. Cartelazo. 1/3
1
7
1
18
1,899
Jacqueline Bellon (jacbellon.bsky.social) retweeted
Replying to @nixcraft
Claude is similarly spassy
5
7
2
203
24,162
Jacqueline Bellon (jacbellon.bsky.social) retweeted
😂
470
1,031
558
15,493
3,529,544
Jacqueline Bellon (jacbellon.bsky.social) retweeted
To all prospective PhD students of the world: If you generate your research statement using chatGPT, your grade will be 0, and you will not be invited to the interview. So, please save your and the committee's time.
1,038
306
784
6,630
3,184,156
Jacqueline Bellon (jacbellon.bsky.social) retweeted
Infinitely Recursive of Game of Life. Life and civilization emerges from self-organization at different levels of complexity.
42
348
64
1,752
268,104
Jacqueline Bellon (jacbellon.bsky.social) retweeted
Jacqueline Bellon spricht über gesellschaftliche Konsequenzen von #KI-Anwendungen und zur Frage: Wer überlistet wen? Mensch-Maschine-Verhältnisse, adversariale Attacken, Jailbreaking und Specification Gaming. #SympGS24 #Glücksspiel @UniHohenheim
1
3
8
306
Jacqueline Bellon (jacbellon.bsky.social) retweeted
Der 10 h Brüssel-Tripp hat sich gelohnt: Ich habe die eine, entscheidende Stimme für namentliche Abstimmungen im LIBE-Ausschuss geliefert - und dann haben wir bei Chatkontrolle/Massenüberwachung gegen Sozialdemokraten & Konservative & Rechtsradikale ALLES VERLOREN was man nur verlieren kann. Mit zwei oder drei Stimmen Unterschied. Immerhin können die freundlichen Kollegen von den Piraten Ihnen jetzt die Abstimmungsergebnisse namentlich dokumentieren... Und jetzt bitte die A6 wieder freimachen!
52
205
9
1,083
108,640
Jacqueline Bellon (jacbellon.bsky.social) retweeted
40 years of Boston Dynamics (1983 - 2023) [📹 World Data Center, HD]
136
1,314
147
6,811
976,689
Jacqueline Bellon (jacbellon.bsky.social) retweeted
I find convincing GPT-4 to do things it doesn’t realize it can do is a lot easier with the voice interface. Somehow speaking words of encouragement feels more natural.
37
80
36
722
151,726
Jacqueline Bellon (jacbellon.bsky.social) retweeted
Scientists captured the moments in which the brain makes memories by tracking beta-actin mRNA with fluorescent "tags" [📹 Albert Einstein College of Medicine]
70
497
62
3,212
310,822
Jacqueline Bellon (jacbellon.bsky.social) retweeted
This clip is absolutely terrifying to me xitter.gq/HumansNoContext/status…
5,704
17,173
6,919
173,999
38,291,120
Jacqueline Bellon (jacbellon.bsky.social) retweeted
Ok, this is just silly. The last one is probably the best. 😆 Try it 👇
28
23
2
180
22,958
Jacqueline Bellon (jacbellon.bsky.social) retweeted
Take a look at this news paper, and now think about what their announced partnership with OpenAI will look like. xitter.gq/OpenAI/status/17349404…
„Free Palestine ist das neue Heil Hitler“ to.welt.de/iq6xdzk
13
31
6,834
Jacqueline Bellon (jacbellon.bsky.social) retweeted
Why does the internet feel so dead? Well, it only feels dead if you're looking at feeds. garbageday.email/p/selling-y…
2
15
2
72
26,556
Jacqueline Bellon (jacbellon.bsky.social) retweeted
Google has scanned over 25M books, but you can’t read them. Fears over cannibalization of print sales & copyright concerns killed much of the project. But it turns out that digitization INCREASED print demand. The effects are big: an average of around $3,750 more sales per book.
27
244
32
1,321
425,519