Charlie O'Neill (@oneill

Charlie O'Neill

851 posts

Charlie O'Neill

@oneill_c

The sea is the sea The old man is an old man The boy is a boy and the fish is a fish The sharks are all sharks no better and no worse

San Francisco

charlesponeill.com

Joined September 2016

Pinned
Charlie O'Neill
@oneill_c
Aug 28, 2025
Today, we’re launching Parsed. We are incredibly lucky to live in a world where we stand on the shoulders of giants, first in science and now in AI. Our heroes have gotten us to this point, where we have brilliant general intelligence in our pocket. But this is a local minima. We
00:00
107K
Charlie O'Neill
@oneill_c
Jan 7, 2025
New preprint! In transformers, we often describe the Q/K/V maps in an ad hoc way, but we show these linear self-attention components form a "parametric endofunctor" in a 2-category of linear maps. This takes cues from @bgavran3 et al.’s programme extended to transformers. (1/10)
33K
Charlie O'Neill
@oneill_c
Aug 5, 2024
New paper with the fantastic @christinexye (+ @jwuphysics, @kartheikiyer) : "Disentangling Dense Embeddings with Sparse Autoencoders". We present one of the first applications of SAEs to text embeddings 🧵 Paper: arxiv.org/pdf/2408.00657 Web app: huggingface.co/spaces/charlie…
7.3K
Charlie O'Neill
@oneill_c
Aug 28, 2025
Replying to @oneill_c
Parsed is the company I always wanted to build and the company I wish existed in the world. I’m so lucky to work with the smartest people I know, to be backed by some of the best investors and angels in the world (LocalGlobe, HuggingFace, DeepMind, NHS), and to have customers who
Parsed | Custom, interpretable AI systems that continuously learn
From parsed.com
3.3K
Charlie O'Neill
@oneill_c
May 19, 2025
With all the buzz of reasoning models, the questions I keep coming back to around RL for LLMs are: - how well does this do compared to SFTing on correct outputs of a specific task - (related to above) how much more data efficient is rl than sft (how much extra bang for buck do
2.8K
Charlie O'Neill
@oneill_c
Sep 2, 2024
1/7 Why are we using such simple encoders in sparse autoencoders (SAEs) for LLM interpretability? The current approach uses a basic linearity-nonlinearity, but transformers are capable of much more complex computations
4.2K
Charlie O'Neill
@oneill_c
Jan 7, 2025
Replying to @oneill_c
@bgavran3 argue in arxiv.org/abs/2402.15332 that deep learning frameworks either impose constraints (GDL) or specify tensor ops (RNNs/Transformers). We need a single theory bridging both views - hence category theory. (2/10)
arxiv.org
Position: Categorical Deep Learning is an Algebraic Theory of All...
We present our position on the elusive quest for a general-purpose framework for specifying and studying deep learning architectures. Our opinion is that the key attempts made so far lack a...
1.7K
Charlie O'Neill
@oneill_c
Feb 3, 2025
1/5 Chris Olah’s "What is a Linear Representation? What is a Multidimensional Feature?" led me to revisit the term "one-dimensional feature." At first glance I equated this with the activation dimensionality (e.g. GPT‑2's 768), but @ch402 obviously uses it in a different sense
1K
Charlie O'Neill
@oneill_c
Jan 17, 2025
Replying to @thesephist
Mathpix are pretty good, maybe slightly slow (10s per 150 pages maybe?) really good with latex and mathtex formatting
8.2K
Charlie O'Neill
@oneill_c
Dec 21, 2024
Thoughts on nihilism arising from the general pace of LM+RL research now on LessWrong, a point bolstered by release of o3 today (link below) (thanks @christinexye for reminding me of the Vonnegut passage)
3.4K
Charlie O'Neill
@oneill_c
May 5, 2025
1/🧵 Thrilled to announce our ICML‑accepted work (done in my undergrad days with @david_klindt and @gumr4n) that questions the linear‑only dogma in sparse autoencoders👇
1.4K
Charlie O'Neill
@oneill_c
Mar 27, 2023
1/ 🧵 I came across a paper shared by @jeremyphoward about parallelisable linear RNNs. The ICLR 2018 paper (openreview.net/pdf?id=HyUNwul…) raises an interesting question: Could RNNs challenge Transformers as the backbone for large language models?
9.3K
Charlie O'Neill
@oneill_c
Aug 28, 2025
Replying to @oneill_c
@parsedlabs
2.7K
Charlie O'Neill
@oneill_c
Apr 25, 2025
Thought experiment: clone a trillion copies of o3 and let them self-direct research/thinking for a decade. Would even one Einstein/Newton/Ramanujan emerge? I'd price this at ~0. Competence reaches parity; paradigm-creation still shows a human/LLM delta, either fundamental or a
6.7K