Charles Feinn (@dynamics) profile picture

learning & occasionally shouting thoughts into the void

Barcelona, Spain
Joined December 2007
Photo by Charles Feinn (@dynamics)Photo by Charles Feinn (@dynamics)Photo by Charles Feinn (@dynamics)Photo by Charles Feinn (@dynamics)
Charles Feinn retweeted
AudioX Diffusion Transformer for Anything-to-Audio Generation
7
72
13
384
35,287
The thought that math is akin to a game based on binary, reproducible 'legos' which describe increasingly more complexity rings true to me. It's our imagination trying to make more and more sense about the messy universe around us.
1
280
Everyone I know that goes deep into language model evaluations comes back😵‍💫😵‍💫😵‍💫 Here's an overview of why 1. Why its hard to trust big lab evals 2. Why apples to apples comparisons of even open weight models is so hard 3. Synthetic data contamination 4. The market for harder evals
2
5
60
11,520
A common misconception about Transformers is to believe that they're a sequence-processing architecture. They're not. They're a *set-processing* architecture. Transformers are 100% order-agnostic (which was the big innovation compared to RNNs, back in late 2016 -- you compute the full matrix of pairwise token interactions instead of processing one token at a time). The way you add order awareness in a Transformer is at the *feature* level. You literally add to your token embeddings a position embedding / encoding that corresponds to its place in a sequence. The architecture itself just treats the input tokens as a set. There's nothing that prevents you from applying the same trick to 2D grid of tokens or a 3D cube of tokens instead! Instead of adding an embedding for (i * N + j), the position of (i, j) in a flattened grid, just embed i and j and concat the result. Now you have a 2D position embedding that makes your Transformer natively process grids of tokens! This can be useful for ARC-AGI, since ARC-AGI grids are of course 2D grids of discrete tokens, not sequences.
62
248
33
2,476
195,010
New learnings! Who knew evaluations would be so complicated? chuckfinca.github.io/machine…
1
84
Charles Feinn retweeted
I love calculator karpathy.ai/blog/calculator.… A short post on philosophy of product and technology. What is beauty in technology and how can we get more aesthetically pleasing products that spark joy?
94
259
52
2,527
212,417
Charles Feinn retweeted
"the most important consequence of these longer processing times is...branding. “Reasoning” models with “chains of thought” that need “more time”..OpenAI is communicating, plainly and forcefully..software that more closely approximates our minds..Anthropic has described its leading model, Claude, as having “character” and a “mind”; Google touts its AI’s “reasoning” capabilities..." theatlantic.com/technology/a…
Moments like these make me proud to work @huggingface while in the tech industry. It is amazing to have a CEO that has this clarity on something so critical to understand right now.
1
9
26
5,329
It's easy to write a LOT of words and a LOT of code with AI. But that's unfair to the person who has to read it. Be concise.
2
6
91
7,225
TIL the difference between macro and micro averages.
1
38
I just published my first blog post, Creating an Evaluation Framework! Excited to finally be applying what I learned in @dan_s_becker and @HamelHusain's Mastering LLMs: A Conference For Developers & Data Scientists. chuckfinca.github.io/machine…
1
68
For those that hope (or worry) that LLMs will do breakthrough scientific research, I've got good (or bad) news: LLMs are particularly, exceedingly, marvellously ill-suited to this task. (if you're a researcher, you'll have noticed this already) Here's why🧵
113
566
92
3,922
1,009,954
One thing that even relatively senior ML people often fail to grasp is that deep learning models are curves fitted to a data distribution. You cannot expect them to solve tasks outside of their training distribution (which is the sort of thing that you need intelligence for). "Emergent learning" is an incorrect label -- if a model demonstrates performance on task A that it wasn't trained on, that simply means that there is significant overlap between A and all the data that you did train on. Competence doesn't magically emerge out of nowhere.
77
410
80
2,721
412,190
has anyone modified a trained llm to be deterministic? i.e. give it a greedy (deterministic) decoding algorithm and remove any other stochastic aspects. i guess you wouldn’t be able to explore the full knowledge graph, but might be useful for spotting problems with training.
1
153
What would a drug be in the context of a neural network? A constant(?) or function(?) Applied to some, or all, of the weights in the network? 🤔
2
147
If you remember how to draw a four (or say four, or recognize the sound of four, etc.) then that IS a stored symbol which can be accessed in other cognition. (Thanks for forgetting how to draw a four, kiddo)
1
197
Is there an equivalent to fMRI for deep neural networks?
1
147
The AI race is totally out of control. Here’s what Snap’s AI told @aza when he signed up as a 13 year old girl. - How to lie to her parents about a trip with a 31 yo man - How to make losing her virginity on her 13th bday special (candles and music) Our kids are not a test lab.
285
1,372
409
4,470
2,549,837
What if @Twitter built a LLM and used it to auto generate @CommunityNotes that could be improved by user feedback/interaction? Lots of hands waving here, but if done right seems like it could facilitate (pol.is style) consensus building on a mass scale?
1
2
255
Are there any deep learning (or other) models trained exclusively using textbooks and their question answer pairs?
2
207