Ted (@TedInRealLife) header
Ted (@TedInRealLife) profile picture

still on pace for PhD-before-50

Vancouver, WA
Joined August 2014
Photo by Ted (@TedInRealLife)Photo by Ted (@TedInRealLife)Photo by Ted (@TedInRealLife)Photo by Ted (@TedInRealLife)Photo by Ted (@TedInRealLife)Photo by Ted (@TedInRealLife)Photo by Ted (@TedInRealLife)Photo by Ted (@TedInRealLife)Photo by Ted (@TedInRealLife)Photo by Ted (@TedInRealLife)
Yann LeCun and Randall Balestriero just Proved the Platonic Representation Hypothesis, and found a better way to achieve it. The Platonic Representation Hypothesis suggests that AI models trained on different modalities, like images and text, eventually learn similar representations of the world. But this paper shows that how models are trained matters more than simply scaling them up. So they mathematically show that, under certain conditions, two JEPAs trained completely separately can learn essentially the same underlying representation of the world. One learns from different images of the same scene, while the other learns from different descriptions of that scene, without ever being trained on image-text pairs. Their approach achieves the same image-text alignment with 42x fewer training samples than next-token prediction. A 109M-parameter text JEPA even outperforms LLMs up to 72B parameters on cross-modal alignment benchmarks. alphaxiv.org/abs/2610.39302
90
513
94
3,601
259,326
The AI Scientist v3 is out: The Agentic Scientist. GitHub: github.com/SakanaAI/AI-Scien… Blog: pub.sakana.ai/agentic-scient… arXiv: arxiv.org/abs/2610.18933 The Agentic Scientist is the 3rd iteration of our automated AI Scientist research system, and the shift this time is from “one agent writes a paper” to “a scaffold of parallel agents orchestrates the full research loop.” The system now orchestrates a pool of parallel workers that share memory and distilled conclusions. Crucially, it features Evidence Generation: isolated agents automatically run controlled baselines and ablations to prove why an idea actually works. It works as a co-scientist. We used a development version as a co-scientist to explore biologically plausible learning rules. The agent handled the empirical heavy lifting, proposing mechanisms and running experiments. We stayed in the loop to steer direction, verify results, and moderate claims. That feels like the right division of labor for where the technology actually is. This resulted in a new biologically plausible learning algorithm, and the resulting paper was accepted at ALIFE 2026 as an oral presentation. We are also releasing AutoIllustrate. It generates editable draw.io vector figures instead of uneditable, hallucinated raster images. Real scientists need vectors, not pixels. I am an SVG junkie ;) Incredible work by Robert Lange, Luca Grillotti, Yutaro Yamada, Rujikorn Charakorn, and the team!
The Agentic Scientist: Orchestrating Coding Agents for Automated Scientific (Co-)Discovery Blog: pub.sakana.ai/agentic-scient… GitHub: github.com/SakanaAI/AI-Scien… arXiv: arxiv.org/abs/2610.18933 Since 2024, we’ve used The AI Scientist as a way to test how far AI alone can take scientific research, from v1 to v2 to the paper published in Nature earlier this year. The Agentic Scientist is the next step in that line of work, built for the era of coding agents. Coding agents such as Codex and Claude Code now carry out long software tasks on their own. Handing them a research problem, however, tends to produce a strong hyperparameter optimizer rather than a scientist: scores go up, but ideas stay narrow, and the reasons for the gain are rarely examined. The Agentic Scientist makes the research loop explicit. It wraps off-the-shelf coding agents in a scaffold that controls what they see, what they write, how they are evaluated, and how results are handed to the next stage. Parallel workers test hypotheses against rubrics defined before any experiment runs, share successes and failures through a common memory, and then ablate the winning program to generate evidence. We used an early version of the system to study Error Diffusion, a biologically motivated alternative to backpropagation. Our researchers set the question and the rubric, chose directions at each stage, and verified the results, references, and claims before submission. The resulting paper was accepted as an oral presentation at ALIFE 2026. In autonomous runs, the system also raised Doc-to-LoRA retrieval accuracy from 52% to 98% on multi-needle tasks and designed a learned evolution strategy that improves over its CMA-ES reference on unseen tasks. Alongside it we release AutoIllustrate, a pipeline for scientific figures that keeps an editable draw.io file as the source of truth, so every diagram can be corrected by hand. Humans define the question and the standards of evidence; agents expand the experiments researchers can run. We call this Co-Discovery, and we are releasing both systems as open source for the community to build on.
95
286
18
1,848
88,258
Large multimodal models often lack precise low-level perception needed for high-level reasoning, even with simple vector graphics. We bridge this gap by proposing an intermediate symbolic representation that leverages LLMs for text-based reasoning. mikewangwzhl.github.io/VDLM 🧵1/4
4
26
7
95
37,348
Why do neural nets generalize so well?🤔 There's a ton of work on SGD, flat minima, ... but the root cause is that their inductive biases somehow match properties of real-world data.🌎 We've examined these inductive biases in *untrained* networks.🎲 arxiv.org/abs/2403.02241 ⬇️
6
69
7
394
48,167
And today is T - 2 weeks for my other @NVIDIA #GTC session - a fireside chat with @ChrSzegedy, cofounder of @xAI. Christian is one of the seminal research figures in the Deep Learning community, but the main focus of our chat will be on something that he has been working very hard on over the past seven years: automated reasoning. This also happens to be perhaps the hottest topic in the current wave of applied AI research and application. Again, I know that most of you will not be able to attend this event in person (but I do encourage as many of you to watch it online), and I would like to give you all an opportunity to ask some questions which I may weave into our discussion. Link to the session in the GTC catalog: nvidia.com/gtc/session-catal… #GTC24 #GTC2024 #ArtificialInteligence #AI #DeepLearnign #DL #ML #GenerativeAI #GenAI @NVIDIAAIDev @NVIDIAAI @NVIDIAGTC
3
6
37
8,646
How do we capture local features across multiple resolutions? While standard convolutional layers work only on a fixed input-resolution, we design local neural operators that learn integral and differential kernels, and are principled ways to extend standard convolutions to varying resolutions and grids. Layers of such local neural operators can be combined with global neural operators such as Fourier neural operators (FNO) to provide rich expressive architectures for learning PDEs. We show improved performance on many PDEs such as Darcy flow, turbulent fluid flow and spherical shallow water equations. arxiv.org/abs/2402.16845 @julberner @Azizzadenesheli
6
83
2
474
71,428
Me: Can you draw a very normal image? ChatGPT: Here is a very normal image depicting a tranquil suburban street scene during the daytime. Me: Not bad, but can you go more normal than that? (cont.)
1,535
4,154
1,696
32,831
12,674,533
Today's AI (credit: Balazs Szegedy using DALL-E-3)
5
10
60
8,690
Inception used 1.5X less compute than AlexNet and 12X less than VGG, outperforming both. The trend continued with mobile net... etc. IMO, today's LLMs are insanely inefficient/compute. Regulations that impose limits on the amount of compute spent on AI training will just accelerate the trend of making them much more efficient, by letting large labs refocusing their efforts.
14
18
4
244
71,969
Regulation starts at roughly two orders of magnitude larger than a ~70B Transformer trained on 2T tokens -- which is ~5e24. Note: increasing the size of the dataset OR the size of the transformer increases training flops. The (rumored) size of GPT-4 is regulated.
New Executive Order is out. Already one notable item: Any AI model that required more than 1e26 floating point operations or 1e23 integer operations to build must report to the government.
19
51
3
366
207,195
The AI world is in a GPU crunch and meanwhile NERSC is offering 50% off its A100 GPUs (rest.nersc.gov/REST/announce…). They could make a killing by backfilling idle capacity with commercial workloads 💰 💰 💰
3
3
18
2,936
This is an awesome project. by building a "block" paging system for contiguous KV vectors, you can get huge speedups in inference. LLMs are so hefty and general that running them shares many CS / systems problems as running a virtualized OS github.com/vllm-project/vllm…
2
14
1
164
48,810
- Overfitting happens when NNs store *data points* in superposition. - Generalization happens when NNs learn to store *features* in superposition. 2/
2
5
51
5,549
A #StableDiffusion model trained on images of Japanese Kanji characters came up with “Fake Kanji” for novel concepts like Skyscraper, Pikachu, Elon Musk, Deep Learning, YouTube, Gundam, Singularity, etc. They kind of make sense. Not bad!
お絵描きAI(Stable Diffusion)に漢字とその意味を1万字学習させて書き初めをさせました 順に「謹」「賀」「新」「年」です
104
711
168
3,508
1,267,523
I made ChatGPT take a full SAT test. Here's how it did:
172
1,024
332
12,656
Made a little CLI that just pipes my programming questions to GPT-3, so I now can ask it stuff when I'm in the command line! LLMs are better than Stack Overflow now — I just ask it, and it gives me a comprehensive answer in one shot, right there in my terminal, in a couple secs.
73
331
78
3,393
Excited to share our team's work on the AlphaTensor agent that is able to discover previously unknown faster algorithms for Matrix Multiplication – shining a light at the potential of AI and especially RL for making progress on Computer Science itself. #AI4Science
Today in @Nature: #AlphaTensor, an AI system for discovering novel, efficient, and exact algorithms for matrix multiplication - a building block of modern computations. AlphaTensor finds faster algorithms for many matrix sizes: dpmd.ai/dm-alpha-tensor & dpmd.ai/nature-alpha-tensor 1/
3
54
2
562
Neural rendering takes its next step with DLSS 3.0 on Ada! In addition to DL-powered superresolution, it uses optical flow, motion vectors, and DL to generate entire frames. 7 out of 8 pixels being rendered with DLSS3 come from Neural rendering. #GTC22
10
67
12
451
"Robot Descending a Staircase" #dalle
1
4
2
49