throw more computer, throw more data
@jxmnop
-
PhD Student Develops Chord-Aware Polyphonic Music Transcription Model
By
–
> be me, first year PhD student > get obsessed with polyphonic music transcription during pandemic > for some reason, advisor lets me work on music > brilliant idea: explicitly encode chord structure into the model > spend months implementing complicated chord-aware audio
-
AI Labs Underestimate Intelligence Complexity Like Developers Underestimate Code
By
–
often i’ll set out to write code with the expectation that it’ll take a few hours, and it takes a few days and i think this is the same fallacy the AI labs are falling for. but instead of underestimating the complexity of code, they underestimate the complexity of intelligence
-
PowerPoint’s 40-Year Dominance: Why No Software Disruption?
By
–
it is really funny to me that in 40 years of software evolution no one has been able to dethrone PowerPoint maybe arranging visuals on a stack of blank pieces of imaginary 'paper' is ideal in some primitive way
-
AI Data Collection Scale Compared to 2010s Tech Giants
By
–
the scale of data collection in the AI labs pales in comparison to 2010s google it’s mostly web scraping and data-labeling. compare that to diligently photographing streets of every country, mapping earth via satellite, scanning every book known to man.. now *that* was ambitious
-
All Embedding Models Learn Same Thing Discovers Researchers
By
–
excited to finally share on arxiv what we've known for a while now: All Embedding Models Learn The Same Thing embeddings from different models are SO similar that we can map between them based on structure alone. without *any* paired data feels like magic, but it's real: x.com/jxmnop/status/…
-
Training Data Overlap Between BERT and T5 Models
By
–
yeah, this is a great question – like even though BERT and T5 are different for example they probably have a large amount of training data overlap
-
CycleGAN Architecture: Unsupervised Image-to-Image Translation Explained
By
–
sorry that was confusing. I was just showing images from the original CycleGAN paper to make an analogy between horseszebras and model1model2.
-
Universal Geometry of Embeddings Paper Released on ArXiv
By
–
this was joint work with my friends @rishi_d_jha
, collin zhang, and vitaly shmatikov at Cornell Tech our paper "Harnessing the Universal Geometry of Embeddings" is on ArXiv today: http://
arxiv.org/abs/2505.12540 -
Strong Platonic Representation Hypothesis: Cross-Model Translation
By
–
theoretically, the implications of this seem big. we call it The Strong Platonic Representation Hypothesis: models of a certain scale learn representations that are so similar that we can learn to translate between them, using *no* paired data (just our version of CycleGAN)