Sagar @sagarpath

Master of My Domain #AutonomousSystems #AgenticAI #Optimization #LLMEngineering Wondering about wondering... All views are my own. SpaceTime Joined December 2016

Tweets

2K
Followers

398
Following

1K
Likes

37K

Hieu Pham @hyhieu226

a day ago

It is easier for an LLMs to write a chess engine than to learn to play chess. What are similar examples to humans?

240 66 3K 314K 524

Towards a Unified View of LLM Post-Training This work proposes Hybrid Post-Training, which switches between RL and SFT using simple performance feedback to balance exploration and exploitation. More below:

5 27 164 25K 176

Download Image

Boaz Barak @boazbaraktcs

2 days ago

Giving first lecture in my AI safety course tomorrow. I plan to record and post lectures online on the course webpage boazbk.github.io/mltheorysemina…

15 65 619 49K 537

Google DeepMind @GoogleDeepMind

a day ago

We’re helping to unlock the mysteries of the universe with AI. 🌌 Our novel Deep Loop Shaping method published in @ScienceMagazine could help astronomers observe more events like collisions and mergers of black holes in greater detail, and gather more data about rare space…

56 213 1K 553K 262

Download Gif

Richard Sutton @RichardSSutton

2 days ago

Dwarkesh Patel is 100% right on this: AI's utility is very strongly dependent on continual learning. youtu.be/nyvmYnz6EAg?si…

48 128 1K 388K 716

Tanishq Mathew Abraham, Ph.D. @iScienceLuvr

3 days ago

Fantastic Pretraining Optimizers and Where to Find Them "we conduct a systematic study of ten deep learning optimizers across four model scales (0.1B-1.2B parameters) and data-to-model ratios (1–8× the Chinchilla optimum)." "we find that all the fastest optimizers such as Muon…

9 50 259 40K 202

Download Image

Wes Roth @WesRothMoney

3 days ago

Sundar Pichai explained that AGI means human-level expertise across disciplines, while ASI is a self-improving version far beyond humans. He noted some researchers use the term "AJI" to describe the current jagged state of AI progress. He said AI shows both impressive…

35 46 270 34K 110

Download Video

François Chollet @fchollet

a week ago

With enough compute, all approaches start looking alike. Compute is the great equalizer.

112 155 1K 377K 147

Ben MacLaurin @ben__maclaurin

a week ago

everytime i feel uninspired i re-read @ryolu_'s writing

19 86 1K 111K 1K

Download Image

Xander Chin @XanderChin

a week ago

wrote a piece on how we do tiling and gradient descent on TinyTPU included some animations of the weight and bias gradient descent

9 63 659 30K 458

Download Video

Andrej Karpathy @karpathy

a week ago

In era of pretraining, what mattered was internet text. You'd primarily want a large, diverse, high quality collection of internet documents to learn from. In era of supervised finetuning, it was conversations. Contract workers are hired to create answers for questions, a bit…

Prime Intellect @PrimeIntellect

a week ago

117 398 3K 1.2M 2K

Download Video

261 851 7K 847K 5K

rank decomposition @rankdim

a week ago

1. why RL didn't work before? 2. why it works now?

93 65 1K 291K 806

Jackson Atkins @JacksonAtkinsX

2 weeks ago

NVIDIA research just made LLMs 53x faster. 🤯 Imagine slashing your AI inference budget by 98%. This breakthrough doesn't require training a new model from scratch; it upgrades your existing ones for hyper-speed while matching or beating SOTA accuracy. Here's how it works:…

91 697 4K 429K 4K

Download Image

hardmaru @hardmaru

2 weeks ago

Our new GECCO paper builds on our past work, showing how AI models can be evolved like organisms. By letting models evolve their own merging boundaries, compete to specialize, and find ‘attractive’ partners to merge with, we can create adaptive, robust and scalable AI ecosystems.

Sakana AI @SakanaAILabs

2 weeks ago

38 153 824 169K 497

Download Image

19 53 404 63K 173

hardmaru @hardmaru

2 weeks ago

The high cost required to train large foundation models makes them the fastest depreciating asset in human history.

François Chollet @fchollet

2 weeks ago

The high cost required to train large foundation models makes them the fastest depreciating asset in human history.

99 359 2K 227K 312

40 91 1K 118K 179

Yann LeCun @ylecun

2 weeks ago

My previous meeting room at Meta was named after the title of this paper.

Samee Ur Rehman @sameeurehman

2 weeks ago

My previous meeting room at Meta was named after the title of this paper.

3 12 114 186K 84

Download Image

74 69 1K 183K 300

Demis Hassabis @demishassabis

2 weeks ago

Simulations are the future, & one of the main tools we’ll ultimately use to understand and predict things about the universe. This is why I’m so excited about Genie 3, our latest interactive world simulator - here are some insanely cool things you might have missed about it 🧵:

161 601 4K 393K 2K

Download Video

Mario Sieg @_mario_neo_

2 weeks ago

my matrix multiplication kernels for magnetron now beat PyTorch's performance on my CPU: Magnetron matmul: 1955.2 GFLOP/s Torch matmul: 1752.3 GFLOP/s check out the kernel code: github.com/MarioSieg/magn… magnetron detects cache sizes and uses state of the art block tuning…

16 64 628 47K 392

Guangxuan Xiao @Guangxuan_Xiao

2 weeks ago

Just wrote a post on my understanding of the statistics behind block sparse attention. My take is that it works by using the "learned similarity gap," which creates a simple SNR formula connecting retrieval quality with model architecture. Read more: guangxuanx.com/blog/block-spa…

5 46 385 33K 398

Bonnie Li @bonniesjli

2 weeks ago

We can now train AI inside the mind of another AI. 🤯 🌍 Our world model, Genie 3, imagines and generates new worlds on the fly. 🤖 Our embodied agent, Sima, is dropped in and learns to navigate them autonomously. The entire loop—from the environment to the action—is generated…