Ashish @ashishnh

Like good spirited and informative discussions! Will block people arguing in bad faith or using personal attacks! Joined August 2009

Tweets

382K
Followers

914
Following

5K
Likes

378K

Sam Stein @samstein

3 hours ago

SCOTUS has basically given Trump whatever he has wanted, save for that very early abrego garcia order

64 228 1K 42K 41

Bill Pulte's father and stepmother claimed tax exemptions on primary residences in two states, leading local tax officials in Michigan to revoke the one on a property in Bloomfield Township after inquiries from Reuters.

(((Charles Fishman))) 💧 @cfishman

3 days ago

232 2K 7K 643K 612

113 373 2K 207K 121

Richard Hanania @RichardHanania

23 hours ago

79% now think immigration is a good thing, a new record. Next time they win, Democrats will have a mandate to dismantle ICE. And Trump has set the precedent that a president can basically eliminate agencies he doesn’t like. There will be no excuse not to dismantle this system.

63 78 566 36K 51

Download Image

Matt Darling 🌐🏗️ @besttrousers

4 hours ago

The median household in the United States has $8,000 in their checking account, owns their home, and has a net worth of $193,000.

Dan Hill @dandinohill

23 hours ago

The median household in the United States has $8,000 in their checking account, owns their home, and has a net worth of $193,000.

263 2K 12K 193K 112

85 71 1K 125K 168

Smita Prakash @smitaprakash

13 hours ago

Seems like the US is in a race with itself on how many friendly nations it can alienate before the year ends. South Korea is the latest one.

Raphael Rashid @koryodynasty

16 hours ago

Seems like the US is in a race with itself on how many friendly nations it can alienate before the year ends. South Korea is the latest one.

205 3K 12K 1.7M 2K

17 97 695 19K 15

James Surowiecki @JamesSurowiecki

7 hours ago

Rittenhouse and Penny are not in prison because they got due process.

steb @StebDaVee

22 hours ago

Rittenhouse and Penny are not in prison because they got due process.

47 1 36 75K 0

34 89 1K 31K 11

Akshay 🚀 @akshay_pachaar

6 hours ago

Build agents that can actually do real-world tasks! Agent Reinforcement Trainer (ART) is a framework to train multi-step LLM agents for real-world tasks using GRPO. You just need a few lines of code. No manual rewards needed! ✨ 100% open-source.

6 36 255 20K 383

Download Image

Rohan Paul @rohanpaul_ai

9 hours ago

The paper shows LLM agents look steady at first, but they break under basic behavioral checks. New @GoogleDeepMind paper. Even when 2 agents begin with opposite preferences, like one liking something and the other disliking it, their conversation almost never ends in open…

8 16 100 10K 64

Download Image

James Surowiecki @JamesSurowiecki

4 hours ago

Trump attended the US Open as a "guest of Rolex" - which just happens to be facing 39% tariffs on its watches that Trump imposed a few weeks ago. Open corruption continues.

42 793 2K 42K 93

Download Image

Pramod Goyal @goyal__pramod

20 hours ago

The greatest document of all time about LLMs. Remember the madlad, senior Google director who released the greatest document on AI Agents. He is back, and he created one for LLMs this time. > Architecture > Pre-training > Reasoning > RL

10 111 949 80K 2K

Download Image

TuringPost @TheTuringPost

11 hours ago

KV cache compression techniques ▪️KV caching (basic) – stores previously computed Keys and Values in memory and calculates attention only for new tokens. ▪️ Quantization – represents KV cache with fewer bits. ▪️ Low-rank decomposition – compresses the KV cache into smaller…

8 84 680 34K 537

Download Image

Richard Hanania @RichardHanania

4 hours ago

Financial Times reports that it's an open secret that Korean companies cut corners on visas. But it's because American administrations have been demanding the impossible from them. A senior South Korean official said the companies had been placed in an “impossible position”,…

34 180 1K 73K 199

Download Image

Google DeepMind @GoogleDeepMind

4 days ago

EmbeddingGemma is our new best-in-class open embedding model designed for on-device AI. 📱 At just 308M parameters, it delivers state-of-the-art performance while being small and efficient enough to run anywhere - even without an internet connection.

44 302 2K 364K 598

Download Video

Avi Chawla @_avichawla

13 hours ago

2) LoRA-FA While LoRA significantly decreases the total trainable parameters, it requires substantial activation memory to update the low-rank weights. LoRA-FA (FA stands for Frozen-A) freezes matrix A and only updates matrix B. Check this 👇

1 2 7 2K 4

Download Image

Avi Chawla @_avichawla

13 hours ago

1) LoRA - Add two low-rank trainable matrices, A and B, alongside weight matrices. - Instead of fine-tuning W, update the low-rank matrices instead. Even for the largest of LLMs, LoRA matrices take up a few MBs of memory. QLoRA is a quantized variant of LoRA. Check this👇

1 2 9 2K 12

Download Image

Avi Chawla @_avichawla

13 hours ago

Some background! LLM weights are matrices of numbers adjusted during finetuning. Most PEFT techniques involve finding a lower-rank adaptation of these matrices, which is a smaller-dimensional matrix that can still represent the information stored in the original.

1 2 16 3K 11

Download Image

Avi Chawla @_avichawla

13 hours ago

Traditional fine‑tuning is impractical for LLMs (billions of params; 100s GB). Since this kind of computing isn't accessible to everyone, parameter-efficient finetuning (PEFT) is extensively used. Today, we’ll cover the top 5 PEFT techniques, step by step.

3 7 57 4K 75

Download Image

Avi Chawla @_avichawla

13 hours ago

I have been fine-tuning LLMs for over two years now! Here are the top 5 LLM fine-tuning techniques, explained visually:

13 59 593 57K 1K

Sabeer Bhatia @sabeer

4 hours ago

I believe policies like GST and demonetization are deliberate attempts to keep Indians “poor”.

162 232 1K 22K 24

elvis @omarsar0

4 hours ago

Another banger paper on reasoning LLMs! They train models to "think wider" to explore multiple ideas that produce better responses. It's called native thought parallelism and proves superior to sequential reasoning. Great read for AI devs! Here are the technical details: