Nadeesha Amarasinghe @nadeesha99 Twitter profile

Last Seen Profiles

@ely_byro

@HienNguyen65277

@PLatrisha24702

@KCheslie85320

@Shaqwana180Leon

@PasutriAsli2

@PACIFICADO77836

@Joe19466457

@Ridge795261

@ChrismaJoa52834

@bokeplokalmalam

@thibaudderuelle

@Lose2Liliana

@MicalWills57662

@SaturninSolie

@DeirdraA57468

@rixon

@FunGames713784

@here_turn28456

@GraceBrown78531

@AnnabelMal27981

@EliteCrypt79860

@EShamicka1169

@owner_duck

@zuzakrajewska

@GloryHole_VIP

@NoemiSaw93080

@jandakembangstw

@SegMintIO

@gazorpiaan

@DrpcIn73996

@Jy0uWDrc3upI79E

@lane_wadle

@hma2200033

@lingjiujiu

@AjayRare53561

Nadeesha Amarasinghe

@nadeesha99

19 days

As someone who’s worked on the infrastructure to accelerate autonomy. It was still surreal to see everything come together in one coherent and clear vision of the future. Looking forward to bringing this closer to reality!

Ashok Elluswamy

@aelluswamy

19 days

Reflecting on the We, Robot event, we had: - 19 Cybercabs and 29 Model Ys driving themselves - 1,300 trips transporting over 2,000 guests - Continuous operation of over the 3.5 hours - And every trip was perfectly safe!

532

1K

10K

15

33

884

Nadeesha Amarasinghe

@nadeesha99

7 months

Give it a try! Expect steady improvements

Tesla

@Tesla

7 months

You can now subscribe to FSD (Supervised) for $99/month in the US Upgrades > Software Upgrades > Subscribe

2K

13K

14

12

421

Nadeesha Amarasinghe

@nadeesha99

7 months

“Quasi-infinite” data…

5

12

359

Nadeesha Amarasinghe

@nadeesha99

7 months

Very high compute:engineer ratio… need help

Elon Musk

@elonmusk

7 months

@thetechbrother This is not accurate. Tesla would be second highest and X/xAI would be third if measured correctly.

147

223

2K

4

26

304

Nadeesha Amarasinghe

@nadeesha99

6 months

Future is already here. just not evenly distributed… yet. Impressive work from the Optimus team!

Tesla Optimus

@Tesla_Optimus

6 months

Trying to be useful lately!

3K

6K

33K

3

11

294

Nadeesha Amarasinghe

@nadeesha99

1 year

Amazing work by the team! Also come help run these nets on the bot!

Tesla Optimus

@Tesla_Optimus

1 year

Optimus can now sort objects autonomously 🤖 Its neural network is trained fully end-to-end: video in, controls out. Come join to help develop Optimus (& improve its yoga routine 🧘) →

3K

8K

35K

10

9

282

Nadeesha Amarasinghe

@nadeesha99

11 months

Come work on foundation models and foundational infra! One of the most challenging/interesting models from an inference perspective.

Tesla AI

@Tesla_AI

11 months

Tesla AI is building next-generation autonomy on a single foundation video network that directly drives the car Join the team and build state-of-the-art end-to-end models using massive fleet data on one of the world's largest training clusters

484

1K

7K

7

13

189

Nadeesha Amarasinghe

@nadeesha99

11 months

Optimus got a face & body lift!

Tesla Optimus

@Tesla_Optimus

11 months

There’s a new bot in town 🤖 Check this out (until the very end)!

3K

7K

32K

8

2

182

Nadeesha Amarasinghe

@nadeesha99

10 months

@Tesla prolly only company with all the pieces from AI training infra, efficient edge inference, software, hardware and batteries “under one roof” to make humanoid robots happen. Excited for the future!

Milan Kovac

@_milankovac_

10 months

2023 has been awesome for Optimus. We’ve moved from an exploratory prototype (Bumblebee/cee) to a more stable, Tesla-designed platform (Optimus Gen-1). We’ve improved our locomotion stack, frequently walking off-gantry without falls and with a faster, increasingly more

428

1K

8K

3

14

142

Nadeesha Amarasinghe

@nadeesha99

7 months

The variability and complexity of driving in the real world needs lots of data, compute and a powerful model. Come build the infrastructure for scaling the data, compute and model!

Tesla

@Tesla

7 months

Reply with your best FSD video that you’d show to someone who hasn’t experienced V12 yet 👀

2K

1K

8K

1

15

137

Nadeesha Amarasinghe

@nadeesha99

3 months

Great work @tonyduan_ and team! Simplicity and scale prevails in the end.

Tony Duan

@tonyduan_

3 months

Lots of hard work and many late nights went into the making of FSD 12.5, from across the team. Many ideas were simplified and re-worked from first principles. Hope everyone has a chance to try it out. It's a release we're proud of.

304

551

6K

1

4

132

Nadeesha Amarasinghe

@nadeesha99

11 months

MLX from @Apple is an interesting mix of ideas/learnings from numpy/Jax/pytorch. Things that stood out to me: - very simple and clean implementation, can easily look under the hood and see how things work - lazy eval of the compute graph -> leaves the door open for all kinds of

6

18

122

Nadeesha Amarasinghe

@nadeesha99

11 months

Our cars are as good as our videos

Elon Musk

@elonmusk

11 months

Beats a Porsche 911 while towing a 911

20K

55K

487K

1

4

118

Nadeesha Amarasinghe

@nadeesha99

24 days

Get hyped!

Tesla

@Tesla

26 days

The future will be streamed live 10/10, 7pm PT

7K

15K

58K

3

2

116

Nadeesha Amarasinghe

@nadeesha99

11 months

Some dope 3D reconstruction

Tesla

@Tesla

11 months

The 2023 Holiday Update rolls out next week Here’s what’s coming... Custom Lock Sounds Replace the horn lock sound of your vehicle with another sound—like a screaming goat 🐐 LAN Party on Wheels Play your favorite games on the rear touchscreen 🎮 Rear Screen Bluetooth

1K

3K

17K

2

8

104

Nadeesha Amarasinghe

@nadeesha99

11 months

Also come work with me! “Platonic Ideal” of Machine Learning, Software Engineering and AI Accelerators. We dive deep on all layers of the stack to run networks on car and bot with high accuracy and low latency.

Tesla AI

@Tesla_AI

11 months

Tesla AI is building next-generation autonomy on a single foundation video network that directly drives the car Join the team and build state-of-the-art end-to-end models using massive fleet data on one of the world's largest training clusters

484

1K

7K

2

9

99

Nadeesha Amarasinghe

@nadeesha99

11 months

Quite impressive the rapid progress we’ve seen in neural reconstruction! Good job @xiuming_zhang @philduan and team, one of the coolest feature shipped this year!

Ashok Elluswamy

@aelluswamy

11 months

High-fidelity park assist is shipping this weekend to Tesla customers without ultrasonic sensors as part of the holiday release!

300

415

5K

2

96

Nadeesha Amarasinghe

@nadeesha99

23 days

Lots of exciting work to be done towards a bright and autonomous future! If you’re excited about the intersection of ML, Systems and Infrastructure, consider applying @Tesla_AI

Marc Benioff

@Benioff

23 days

@elonmusk One of the most inspiring visions of the future I have ever seen! Everyone should watch this. ❤️🤘

1K

4K

19K

3

6

87

Nadeesha Amarasinghe

@nadeesha99

10 months

Lots of tangible progress! I sometimes forget @Tesla manufacturing prowess.

Tesla

@Tesla

10 months

Tesla 2023 recap Made possible by the hard work of our amazing teams around the world, and each of our owners & supporters. Thank you for helping us continue to accelerate the transition to sustainable energy!

1K

4K

21K

2

83

Nadeesha Amarasinghe

@nadeesha99

1 year

We put a lot of effort to accurately and efficiently model the quantization of our int8 accelerator during fp16 training. We also develop novel techniques/tools/evals to identify and mitigate the sensitivity of our networks to inference time quantization. If you’re excited by

Elon Musk

@elonmusk

1 year

@Scobleizer An accurate assessment. What is also mindblowing is that the inference compute power needed for 8 cameras running at 36FPS is only about 100W on the Tesla-designed AI computer. This puny amount of power is enough to achieve superhuman driving! It makes a big difference that we

308

716

5K

0

7

64

Nadeesha Amarasinghe

@nadeesha99

10 months

3 points in the AI Accelerator design space: GPU, TPU-like and spatial dataflow. My best attempt to compare them: GPU ( @nvidia @AMD ): most common, lots of HW threads + HW thread scheduler for zero cost context switching. High HW complexity which enables lots of

5

60

Nadeesha Amarasinghe

@nadeesha99

8 months

Some nice clean code!

Toby Pohlen

@TobyPhln

8 months

Grok-1 is out in the open.

83

127

918

3

1

58

Nadeesha Amarasinghe

@nadeesha99

7 months

One of the most interesting and high signal podcast episodes I’ve seen in a while. Quite inspiring!

Dwarkesh Patel

@dwarkesh_sp

7 months

Had so much fun chatting with my friends @TrentonBricken and @_sholtodouglas . No way to summarize it, except: This is the best context dump out there on how LLMs are trained, what capabilities they're likely to soon have, and what exactly is going on inside them. You would be

39

124

1K

1

51

Nadeesha Amarasinghe

@nadeesha99

11 months

@tim_zaman Thanks for everything Tim! You took AI Infra to new heights at Tesla. It was a pleasure working with you, I certainly learned a lot. Looking forward to what you do next!

1

47

Nadeesha Amarasinghe

@nadeesha99

6 months

Cool to see model visualization and debugging tools being open sourced! We (I suspect many others) have built similar tools for better understanding quantization/numerics, latency, memory and many other details of our models as part of the deployment process.

Google AI

@GoogleAI

6 months

Presenting Model Explorer, a novel graph visualization tool that streamlines the deployment of large models to on-device platforms, such as mobile phones and browsers, where visualizing conversion, quantization, and optimization data is especially useful.

19

204

881

2

1

17

Nadeesha Amarasinghe

@nadeesha99

10 months

@_jasonwei @hwchung27 Also thankful for @hwchung27 and this great talk :)

Large Language Models (in 2023)

I gave a talk at Seoul National University.I titled the talk “Large Language Models (in 2023)”. This was an ambitious attempt to summarize our exploding fiel...

www.youtube.com

0

5

36

Nadeesha Amarasinghe

@nadeesha99

3 months

Need more talented humans to build world class AI Infrastructure and accelerate real-world AI.

AI & Robotics | Tesla

Apply now to work on Tesla Artificial Intelligence & Autopilot and join our mission to accelerate the world’s transition to sustainable energy.

www.tesla.com

Elon Musk

@elonmusk

3 months

But definitely looking to increase our human talent advantage, so please apply at @xAI , as well as @Tesla and @X

2K

4K

30K

0

3

34

Nadeesha Amarasinghe

@nadeesha99

10 months

Zero-sum behavior starts when growth stalls

2

3

28

Nadeesha Amarasinghe

@nadeesha99

10 months

A fermi estimate of AI Inference compute deployed by @Apple in a year: Around 150M iPhones (13/14/pro/pro-max) with A15/A16 chip with an average 16TOPs of (FP16?) compute = 2400 Exa-Flops inference compute. For comparison 2M H100s (sales estimate for 2024) = 4000 BF16

4

0

27

Nadeesha Amarasinghe

@nadeesha99

3 months

Examining neural net design/architecture through the lens of information and compression (bits in/bits out, size of feature bottlenecks, bits per param, bits per sample) is often insightful and helps eliminate un-desired bottlenecks.

0

22

Nadeesha Amarasinghe

@nadeesha99

8 months

Truly impressed by Jensen’s energy, breadth and depth. One of the 🐐’s

3

0

21

Nadeesha Amarasinghe

@nadeesha99

1 year

@tim_zaman is a great manager and AI Infra is extremely critical. Great opportunity to learn from the best and have outsized impact!

Tim Zaman

@tim_zaman

1 year

Managing a single supercomputer is hard. Now imagine having a lot - we're looking for fullstack engineers to help tie everything together into our machine learning platform (gui, tooling, job/model observability, health, etc). js+py. Interested? DMs open

11

27

360

0

18

Nadeesha Amarasinghe

@nadeesha99

3 months

To build great AI infrastructure you have to do some AI. Occasionally going up abstraction layers to build datasets, train models and construct evals leads to better infrastructure.

0

1

19

Nadeesha Amarasinghe

@nadeesha99

2 months

💯 Incentive structure of the Canadian startup and tech ecosystem need to change significantly to retain even a fraction of the talent that leaves.

kache

@yacineMTB

2 months

canada's biggest export is waterloo interns. we need to put a tariff on them

44

38

1K

2

1

19

Nadeesha Amarasinghe

@nadeesha99

1 year

Uncle @satyanadella don’t miss

0

2

17

Nadeesha Amarasinghe

@nadeesha99

10 months

MCM (Multi-chip Module) GPU: 2017 paper from @nvidia on scaling up gpu with MCM. We’re only now (4-6 years later) starting to see MCM based gpus in production. - reticle limits and silicon yield upper-bound the practical size of a single die - must

3

1

16

Nadeesha Amarasinghe

@nadeesha99

3 months

Would be interesting to see the distribution and magnitude of “bits” ingested/processed by sensor modality and how this has shifted and increased over HW generations.

Waymo

@Waymo

3 months

We recently introduced our 6th-generation Waymo Driver. Today, our Vice President of Engineering, Satish Jeyachandran, shares insights into how this next-gen system will drive Waymo's business forward.

18

78

447

2

0

16

Nadeesha Amarasinghe

@nadeesha99

10 months

When in doubt simplify!

2

0

14

Nadeesha Amarasinghe

@nadeesha99

11 months

Very well written blog post on LLM inference + optimizations

How to make LLMs go fast

Blog about linguistics, programming, and my projects

vgel.me

0

3

14

Nadeesha Amarasinghe

@nadeesha99

1 year

@__eknight__ and the team are super talented! can’t miss opportunity

Ethan Knight

@__eknight__

1 year

So excited that FSD customers were able to catch a glimpse of what we’ve been working on. Sincere thanks to my colleagues for the many long nights and hard work that led up to this point. Want to build end to end with us? Join the team @Tesla_AI !

520

276

1K

1

0

13

Nadeesha Amarasinghe

@nadeesha99

10 months

Dug up my thesis from a “past life” @UofT We did research at the “bottom of the stack” on high speed die-to-die IO. Particularly relevant for scaling AI Accelerators using advanced packaging (MCM or Silicon Interposer) like @nvidia grace hopper, @AMD

2

0

12

Nadeesha Amarasinghe

@nadeesha99

10 months

@itsclivetime @itsclivetime it was a pleasure working together the last 2 years, on data pipelines, perf optimizations, quantization, low precision training…. You’re certainly one of the smartest and most creative people I’ve ever worked with. I will miss our chats about ML/HW and everything

1

0

12

Nadeesha Amarasinghe

@nadeesha99

1 year

@soumithchintala Agree with this, everyone serious about AI is working hard to decouple themselves from $NVDA and 70-80% margins. Expect good returns in the near term but things will look different 2-3+ years out, as alternative SW stacks mature and internal AI Accelerator efforts mature.

1

11

Nadeesha Amarasinghe

@nadeesha99

10 months

There is no need to attend math camp

Google DeepMind

@GoogleDeepMind

10 months

Introducing AlphaGeometry: an AI system that solves Olympiad geometry problems at a level approaching a human gold-medalist. 📐 It was trained solely on synthetic data and marks a breakthrough for AI in mathematical reasoning. 🧵

126

1K

4K

1

0

11

Nadeesha Amarasinghe

@nadeesha99

1 year

Rewrite it with #MLIR is the “Rewrite it in Rust” for compilers

1

0

10

Nadeesha Amarasinghe

@nadeesha99

10 months

Cross-entropy is deep

0

10

Nadeesha Amarasinghe

@nadeesha99

10 months

More cool and creative methods for making large models fit in smaller compute!

AK

@_akhaliq

10 months

The LLM Surgeon paper page: State-of-the-art language models are becoming increasingly large in an effort to achieve the highest performance on large corpora of available textual data. However, the sheer size of the Transformer architectures makes it

1

26

151

1

0

10

Nadeesha Amarasinghe

@nadeesha99

11 months

@DrJimFan Amazing work! Always felt like this was one of NVIDIA’s biggest edges. Which other company making a HW accelerator has a world-class research org.

0

10

Nadeesha Amarasinghe

@nadeesha99

1 year

Anyone interested in learning about the nuts and bolts of an AI inference framework should read @ggerganov Single file (15K+ LOC 😮‍💨) with everything: tensor memory layout, quantization, vector instructions, op implementations, dispatching and more

0

2

9

Nadeesha Amarasinghe

@nadeesha99

1 year

Awesome work by the Tesla team! Real world deployment of learning based approaches is very hard. Requires tight integration and excellence across all parts of the stack both HW and SW. Tesla is in a unique position to make this dream a reality. Join the team!

Tesla Optimus

@Tesla_Optimus

1 year

Multiple fully Tesla-made Bots now walking around & learning about the real world 🤖 Join the Tesla AI team →

3K

5K

26K

0

2

10

Nadeesha Amarasinghe

@nadeesha99

10 months

Language appears to be humanity’s best collective invention and the current best compressor of reality

1

0

9

Nadeesha Amarasinghe

@nadeesha99

5 months

patches or tokens or both

1

0

7

Nadeesha Amarasinghe

@nadeesha99

1 year

So many interesting systems problems to be solved in LLM inference. Very reminiscent of all the distributed systems that needed to be built to run current web and search infrastructure.

1

0

7

Nadeesha Amarasinghe

@nadeesha99

7 months

Lots of exciting work in long context!

Chunting Zhou

@violet_zct

7 months

How to enjoy the best of both worlds of efficient training (less communication and computation) and inference (constant KV-cache)? We introduce a new efficient architecture for long-context modeling – Megalodon that supports unlimited context length. In a controlled head-to-head

4

51

230

0

1

9

Nadeesha Amarasinghe

@nadeesha99

7 months

Increasingly apparent AGI at test time will be one file and one BLOB

llama3/llama/model.py at main · meta-llama/llama3

The official Meta Llama 3 GitHub site. Contribute to meta-llama/llama3 development by creating an account on GitHub.

github.com

0

1

9

Nadeesha Amarasinghe

@nadeesha99

9 months

When stuck, inject some randomness

2

0

9

Nadeesha Amarasinghe

@nadeesha99

2 months

Always had a blast working with @micaelccarvalho Highly recommend getting on this 🚀

0

7

Nadeesha Amarasinghe

@nadeesha99

11 months

This whole book is awesome! Well detailed and super practical.

Stas Bekman

@StasBekman

11 months

This is the first pass on the new chapter for ML Engineering: The AI Battlefield Engineering - What You Need To Know This a WIP and your feedback for improvement is always welcome.

13

107

745

0

7

Nadeesha Amarasinghe

@nadeesha99

1 year

💯

Andrej Karpathy

@karpathy

1 year

With many 🧩 dropping recently, a more complete picture is emerging of LLMs not as a chatbot, but the kernel process of a new Operating System. E.g. today it orchestrates: - Input & Output across modalities (text, audio, vision) - Code interpreter, ability to write & run

314

2K

9K

0

6

Nadeesha Amarasinghe

@nadeesha99

10 months

They found a cheat code!

AK

@_akhaliq

10 months

Zero Bubble Pipeline Parallelism paper page: Pipeline parallelism is one of the key components for large-scale distributed training, yet its efficiency suffers from pipeline bubbles which were deemed inevitable. In this work, we introduce a scheduling

7

70

324

1

0

8

Nadeesha Amarasinghe

@nadeesha99

9 months

Hyped! Device mesh sounds familiar 🤔

PyTorch

@PyTorch

9 months

PyTorch 2.2 is here 🎉 Featuring: - SDPA support of FlashAttention-2 - New ahead-of-time extension of TorchInductor - device_mesh, a new abstraction for initializing and representing ProcessGroups - A standardized, configurable logging mechanism called TORCH_LOGS

12

191

1K

0

8

Nadeesha Amarasinghe

@nadeesha99

7 months

@tim_zaman Distillation and inference optimized silicon should help fit on “edge”. Gemini nano is a compelling example

0

8

Nadeesha Amarasinghe

@nadeesha99

1 year

When is @Apple gonna put an LLM on the iPhone periodically fine tuned locally (on device) with some retrieval. Would be the ultimate personal assistant + personal search engine.

0

7

Nadeesha Amarasinghe

@nadeesha99

10 months

@RemiCadene It was a pleasure working with you @RemiCadene and I’m gonna miss it! Very excited for what you do next!

1

0

7

Nadeesha Amarasinghe

@nadeesha99

1 year

@TheGregYang @xai Inspired me to crack this open again @TheGregYang

1

0

6

Nadeesha Amarasinghe

@nadeesha99

11 months

🤯 we can distill these massive models to run effectively on edge

Sundar Pichai

@sundarpichai

11 months

Gemini Nano is super efficient for tasks that are on-device. Android developers can sign up for an early access program for Gemini Nano via Android AICore and Pixel 8 Pro users can already see it rolling out in features like Summarize in Recorder and Smart Reply in Gboard + much

67

155

2K

1

0

6

Nadeesha Amarasinghe

@nadeesha99

2 years

Most social media (i.e @Twitter / @Reddit ) news and content platforms should be training/fine-tuning LLMs in-house. Replacing the now decades old database technologies and indexes that currently power their search.

0

6

Nadeesha Amarasinghe

@nadeesha99

11 months

Great overview! I’m waiting for a true LLM inference optimized chip + system to emerge. There’s sufficient specialization and large enough demand.

Raja Koduri

@RajaXg

11 months

Very encouraging to see the steady increase of viable hardware options that can handle various AI models. At the beginning of the year, there was only one practical option - nVidia. Now we see at-least 3 vendors providing reasonable options. Apple, AMD and Intel. We have been

18

81

424

1

0

6

Nadeesha Amarasinghe

@nadeesha99

10 months

AI has reached your local realtor… The economy can begin healing now.

0

6

Nadeesha Amarasinghe

@nadeesha99

1 year

@zoink Then they also may not have the time or focus to make AGI happen. Works either way!

0

5

Nadeesha Amarasinghe

@nadeesha99

6 months

@pariljain @kvogt @lukeholoubek Congrats @pariljain go big!

0

5

Nadeesha Amarasinghe

@nadeesha99

11 months

Great job @awnihannun and co!

0

5

Nadeesha Amarasinghe

@nadeesha99

2 years

🤩

Runway

@runwayml

2 years

New year. New magic. New ways to Make the Impossible. What will you make in 2023? Get started now:

38

243

2K

0

5

Nadeesha Amarasinghe

@nadeesha99

3 months

@ArmenAgha Whoa! Excited to see what you do next. Really enjoyed following chameleon and your multimodal work.

0

5

Nadeesha Amarasinghe

@nadeesha99

10 months

Curious what @ezhang887 @itsclivetime think 😉

0

5

Nadeesha Amarasinghe

@nadeesha99

3 months

@ysmulki did awesome work with us! He pushed several parts of our stack forward and laid the foundation for some new and exciting areas of exploration. A lot of impact over a short period of time 👏

yash

@ysmulki

6 months

excited to share that I’ll be interning at @Tesla_AI working on Autopilot ML infra this summer! HMU if you’re in SF or the Bay Area and want to grab a coffee

4

0

56

1

0

4

Nadeesha Amarasinghe

@nadeesha99

10 months

@YunTaTsai1 I’m getting those vibes too

1

0

4

Nadeesha Amarasinghe

@nadeesha99

1 year

Bard is much better for querying and summarizing recent research than chatgpt. Partly due to the 2021 cutoff in training data and possibly due to better retrieval in bard. Maybe this is more even with browsing enabled.

0

4

Nadeesha Amarasinghe

@nadeesha99

1 year

New number format just dropped

Open Compute Project

@OpenComputePrj

1 year

AMD, Arm, Intel, Meta, Microsoft, NVIDIA, and Qualcomm Standardize Next-Generation Narrow Precision Data Formats for AI! The MX Alliance has released the Microscaling Formats (MX) Specification v1.0 in an open, license-free format. Read the story here:

0

27

64

1

0

3

Nadeesha Amarasinghe

@nadeesha99

6 months

@farkhora Congrats farzad!! Soon to be fardad :p (dad/bad-pun)

0

4

Nadeesha Amarasinghe

@nadeesha99

7 months

@dwarkesh_sp @dwarkesh_sp you are killing it! I haven’t seen a host with this level of depth and insightfulness.

0

4

Nadeesha Amarasinghe

@nadeesha99

10 months

@itsclivetime @nvidia @AMD Nice way to look at it

0

3

Nadeesha Amarasinghe

@nadeesha99

10 months

@ezhang887 @itsclivetime @nvidia @AMD Hmm yesss 🙄

0

2

Nadeesha Amarasinghe

@nadeesha99

1 year

@thoefler @thoefler are there more details available on the chip architecture?

2

0

2

Nadeesha Amarasinghe

@nadeesha99

10 months

@YunTaTsai1 @Apple has late mover advantage, mlx has many of the good ideas from PyTorch/Jax/Numpy

1

0

3

Nadeesha Amarasinghe

@nadeesha99

10 months

@YunTaTsai1 Haha, I’m surprised as well ;p

0

3

Nadeesha Amarasinghe

@nadeesha99

4 months

Just @carlosalcaraz things

Wimbledon

@Wimbledon

4 months

WE'RE GOING TO A FIFTH SET! #Wimbledon

69

307

4K

0

3

Nadeesha Amarasinghe

@nadeesha99

1 year

😂

Andreas Kirsch 🇺🇦

@BlackHC

1 year

Jensen's inequality == not everyone has access to the latest @nvidia GPUs 🤓

8

49

448

0

3

Nadeesha Amarasinghe

@nadeesha99

2 years

@karpathy Can’t wait for most DL optimizations and perf gains to happen automagically in the compiler. The current state of affairs for cpu programs. Feels like we’re still in the early days.

0

2

Nadeesha Amarasinghe

@nadeesha99

1 year

🤯 30min game 32pts to go up 4-1 @carlosalcaraz @DjokerNole

0

2

Nadeesha Amarasinghe

@nadeesha99

1 year

@ylecun First in-house AI accelerator from @MetaAI interesting design point. Seems like it will immediately provide value on business critical workloads while establishing the chip design and compiler foundations needed internally. Expecting fast follow ups here.

0

2

Nadeesha Amarasinghe

@nadeesha99

1 year

Lonnie walker activated

0

2

Nadeesha Amarasinghe

@nadeesha99

2 years

Truth

Mihir Patel

@mvpatel2000

2 years

Modern AI really built on a stack from NVIDIA that almost no one really understands. Spend too long peering into all the drivers and distributed libraries and you get swallowed. Anyone know the magic env vars to chant to get sharp working ofed? 🥲

1

0

11

0

2

Nadeesha Amarasinghe

@nadeesha99

2 years

All roads lead to compilers

PyTorch

@PyTorch

2 years

We just introduced PyTorch 2.0 at the #PyTorchConference , introducing torch.compile! Available in the nightlies today, stable release Early March 2023. Read the full post: 🧵below! 1/5

23

520

2K

0

1

2

Nadeesha Amarasinghe

@nadeesha99

2 years

@tim_zaman @realGeorgeHotz @farkhora 😂

0

2

Nadeesha Amarasinghe

@nadeesha99

10 months

@itsclivetime @Apple Ya conservative estimate, also A17 is 2x more compute at 35TOPs! Also even for on-device inference Apple has yet to make the Neural Engine easy to run models on. Most current inference frameworks still use the CPU/GPU. If I was Apple I would push hard on building a great/easy

0

2

Nadeesha Amarasinghe

@nadeesha99

2 years

Academia2Real Transfer

0

2

Nadeesha Amarasinghe

@nadeesha99

1 year

@ezyang Two things that come to mind for me: - ML workloads are synchronous and and read only whereas traditional distributed systems must critically support async updates to shared state (i.e distributed databases). - there isn’t a significant understanding yet of gpu/accelerator

0

2

Nadeesha Amarasinghe

@nadeesha99

3 months

@starskyreverie @AnthropicAI Congrats Nikhil! Best of luck in this new adventure

0

1