Greg Diamos @GregoryDiamos Twitter profile

Pinned Tweet

Greg Diamos

8 months

If you are an AI startup blocked on GPUs, send me a note. At Lamini, we figured out how to use AMD GPUs, which gives us a relatively large supply compared to the rest of the market.

28

120

1K

Last Seen Profiles

@__RC__h_

@I_psychledlic

@_TheImpactAD_

@_m3ganfoxx

@coachcobb_

@JulianaAnsari1

@Muslimah0331

@JohannaKir8459

@alphagrey96

@riri_coeota1616

@AliceMarga_Bot

@Sulimanaltamim3

@CWJediMaster

@BrophyJed

@TokayevKZ

@MonaOnyxxx

@JaloniCambridge

@ProConnArtist

@bunny_haruno

@NON_DAY_

@SYAMROCK47

@ccg224eth

@Abdul_naji77

@Schizointel

@Compton4KY2024

@V5b5gB

@Onallee

@carmine_san_

@thechiliphilly

@_Toorauz

@PAOtoreka_yoko

@zeonchriss

@MasciScout

@TwilCanni

@pubity

@asht0nAB

Greg Diamos

@GregoryDiamos

9 months

Lamini now supports JSON output at full speed. We guarantee that your LLM produces a valid JSON object according to your spec. Works for any LLM we support, e.g. Mistral 7b show below. API Docs:

4

18

220

Greg Diamos

@GregoryDiamos

8 months

@remusrisnov IMO, ROCm is 90% of the way there, but the market treats it like it is 0% of the way there. We filled in the 90%->100% gaps for LLM finetuning and inference with a team of a few good hackers.

5

13

195

Greg Diamos

@GregoryDiamos

7 months

New LLMs for new years! On my vacation I streamed a 5 hour walkthrough of building an optimized medical LLM from scratch using an AMD supercomputer from a beach in Hawaii. All of the code is open source! via @YouTube

ICD11 LLM Finetuning on AMD GPU Supercomputer

Learn how to finetune your own LLM on the ICD11 standard from the world health organization (or any dataset) using an AMD supercomputer. Fully open source a...

www.youtube.com

2

25

139

Greg Diamos

@GregoryDiamos

3 years

I'm so excited to release two giant speech datasets today: with clean CC licenses with academic and commercial use! Special shout out to Daniel Galvez, Mark Mazumder, and the whole team who put in a huge effort to create these.

Multilingual Spoken Words Dataset | MLCommons Datasets

MLCommons Multilingual Spoken Words Dataset Corpus is a large and growing audio dataset of spoken words in 50 languages for machine learning.

mlcommons.org

3

28

132

Greg Diamos

@GregoryDiamos

5 months

Here's a new earnings call Q&A dataset I made with Lamini. It has millions of questions and answers generating using a Lamini LLM reading earnings call transcripts. Still uploading - 10k finished so far - should have about 1M by tonight

lamini/earnings-calls-qa · Datasets at Hugging Face

huggingface.co

6

15

112

Greg Diamos

@GregoryDiamos

1 year

Excited to release 1.109 billion times faster LLM switching using a PEFT cache in HBM. See the blog for details of how it works. Finding an opportunity for a 1 billion times speedup suggests that we are just scratching the surface of fine tuning custom LLMs.

Lamini

@LaminiAI

1 year

Training multiple LLMs taking forever? 😤 Costing you a fortune?💸 Enter PEFT! Get ready to multiply!! 🚀 1000 models, just 1 machine! 🤖 3 months of training -> 3 milliseconds ⚡️ Just one API call, load and train with Lamini! 👉 👀

1

7

42

5

6

107

Greg Diamos

@GregoryDiamos

4 months

Hiring a world class ML Engineer at Lamini. Drop me a note if you want to push model accuracy to the limit on unlimited AMD compute.

9

15

96

Greg Diamos

@GregoryDiamos

3 months

This is the paper that convinced me - Showing that a Frankenstein CUDA cluster could beat a 10,000 cpu map reduce cluster

Andrej Karpathy

@karpathy

3 months

# CUDA/C++ origins of Deep Learning Fun fact many people might have heard about the ImageNet / AlexNet moment of 2012, and the deep learning revolution it started. What's maybe a bit less known is that the code backing this winning submission to the

166

901

7K

1

12

89

Greg Diamos

@GregoryDiamos

9 months

Nice work replicating this. The recipe is out. AMD GPUs work.

Lamini LLM Finetuning on AMD ROCm™: A Technical Recipe | Lamini - Enterprise LLM Platform

www.lamini.ai

Cognitive Computations

@cognitivecompai

9 months

I've been working on my AI servers in my garage. (8x AMD Instinct mi100) Yesterday I got the first server running and inferencing with oobabooga. Now duplicating the os drive to get the 2nd server running. Then once infiniband is setup I will start working on getting

42

34

474

3

12

84

Greg Diamos

@GregoryDiamos

5 months

We are hiring an HPC (MPI / OpenAI Triton) Engineer at Lamini. Apply here: We are inventing and building the largest AMD LLM training system in the world. Join us in strongly scaling to 1000s of GPUs and beyond.

2

16

85

Greg Diamos

@GregoryDiamos

10 months

Hit me up if you need GPUs for finetuning LLMs. We are bringing online more capacity together with AMD. Start training on AMD in 3 lines of code. pip install lamini from lamini import LlamaV2Runner model = LlamaV2Runner() model.load_data(path=...) model.train(args=...)

Lisa Su

@LisaSu

10 months

Love working with @LaminiAI and @realSharonZhou making LLMs easy and accessible for all on @AMD @AMDInstinct GPUs! So cool what can be done with @LaminiAI LLM Superstations!!

18

58

270

2

13

72

Greg Diamos

@GregoryDiamos

8 months

I remember the same time, it wasn’t luck. My first exposure to the CUDA vision was from John Nickolls . He very clearly saw it as becoming the dominant form of computing.

Bryan Catanzaro

@ctnzr

8 months

I worked at Intel on Larrabee applications in 2007. Then I went to NVIDIA to work on ML in 2008. So I was there at both places at that time and I can say: NVIDIA's dominance didn't come from luck. It came from vision and execution. Which Intel lacked.

44

195

1K

3

5

67

Greg Diamos

@GregoryDiamos

7 years

We used a supercomputer to perform the largest study to date of how deep learning scales up with more data and faster computers. It turns out to be simple and predictable across diverse applications.

0

27

53

Greg Diamos

@GregoryDiamos

5 months

Training LLMs in the Wild Wild West - or how I learned to stop worrying and do multi-node training on AMD GPUs.

Multi-node LLM Training on AMD GPUs | Lamini - Enterprise LLM Platform

www.lamini.ai

0

6

52

Greg Diamos

@GregoryDiamos

6 years

Our latest enhancement to text to speech makes knowledge distillation much easier to train and results in high quality models that are also extremely efficient to deploy on GPUs.

0

15

48

Greg Diamos

@GregoryDiamos

5 months

Now over 1 million Q&As - I'm going to stop now. Generating this dataset using Claude 3 would have cost about $50,000. Enjoy it for free.

Greg Diamos

@GregoryDiamos

5 months

Here's a new earnings call Q&A dataset I made with Lamini. It has millions of questions and answers generating using a Lamini LLM reading earnings call transcripts. Still uploading - 10k finished so far - should have about 1M by tonight

6

15

112

3

7

46

Greg Diamos

@GregoryDiamos

8 months

@belacquant What info do you need? AMD CDNA has a whitepaper: ROCm is open source: Some of our SW is open and documented: Good docs take time, we will keep publishing.

Lamini AI

Lamini AI has 31 repositories available. Follow their code on GitHub.

github.com

2

6

49

Greg Diamos

@GregoryDiamos

5 years

I'm excited to pursue a new adventure.

6

1

46

Greg Diamos

@GregoryDiamos

8 years

I'm excited to announce the release of the persistent rnn library @ICML :

GitHub - baidu-research/persistent-rnn: Fast Recurrent Networks Library

Fast Recurrent Networks Library. Contribute to baidu-research/persistent-rnn development by creating an account on GitHub.

github.com

1

20

45

Greg Diamos

@GregoryDiamos

5 months

Lamini startup program has unlimited GPUs available - with a full LLM stack. Mi250X - $1.15/hour Mi300X - $3.30/hour

Lamini Startup Program Application

Thank you for your interest in Lamini's LLM Compute Startup Program! Lamini offers an all-in-one LLM pods for every developer to build LLM experts efficiently. We'll review your application and get...

docs.google.com

4

20

44

Greg Diamos

@GregoryDiamos

6 years

Check out , a community supported benchmark for measuring deep neural network performance on software frameworks and hardware accelerators.

0

17

41

Greg Diamos

@GregoryDiamos

7 months

First MI300 live in production.

Sharon Zhou

@realSharonZhou

7 months

> rocm-smi like freshly baked bread, 8x MI300X is online if you're building on open LLMs and you're blocked on compute, lmk. Everyone should have access to this wizard technology called LLMs. that is to say, the next batch of @LaminiAI LLM pods are here.

16

36

228

1

8

39

Greg Diamos

@GregoryDiamos

10 months

Love deploying LLMs with ROCm on AMD Instinct GPUs.

Lisa Su

@LisaSu

10 months

Love working with @LaminiAI and @realSharonZhou making LLMs easy and accessible for all on @AMD @AMDInstinct GPUs! So cool what can be done with @LaminiAI LLM Superstations!!

18

58

270

2

3

37

Greg Diamos

@GregoryDiamos

8 months

@remusrisnov @realGeorgeHotz There are oh so many. IOMMU, I'm looking at you! It was essential to use MI class servers, e.g. MI210/MI250/MI300. I read that ROCm on Radeon/RDNA has gotten better recently, but last year it was not pretty and we have moved on completely to CNDA by now.

1

0

36

Greg Diamos

@GregoryDiamos

5 months

An LLM is a database that speaks English instead of SQL and runs on a GPU instead of a disk.

7

5

32

Greg Diamos

@GregoryDiamos

3 months

Announcing Lamini series A to build more powerful supercomputers to enable enterprises to build their own Expert LLMs

Sharon Zhou

@realSharonZhou

3 months

Super excited to announce our Series A!! ✨ @LaminiAI Raises $25M For Enterprises To Develop Top AI Capabilities In-House ▫️ We have incredible enterprise customers who are able to build LLMs with capabilities that exceed general LLMs, e.g. don't hallucinate on their revenue

43

35

438

4

2

29

Greg Diamos

@GregoryDiamos

10 months

Using Lamini, ROCm finally matches CUDA in SW compatibility for LLMs. It took years to get this to work, but it finally does. This blog begins to explain how it works, it was very unintuitive. More technical details to come...

This link will take you to a page that’s not on LinkedIn

lnkd.in

Sharon Zhou

@realSharonZhou

10 months

Excited to announce a HUGE secret with @LisaSu : @LaminiAI has been building LLMs on @AMD GPUs *in production* for over the past year! We’ve made running LLMs on AMD super easy and a highly competitive option through our LLM Superstation, available now at ~10x lower cost than

36

103

782

0

7

26

Greg Diamos

@GregoryDiamos

5 months

The entire data pipeline used to create this dataset is open source at:

GitHub - lamini-ai/lamini-earnings-calls

Contribute to lamini-ai/lamini-earnings-calls development by creating an account on GitHub.

github.com

2

4

27

Greg Diamos

@GregoryDiamos

11 months

I enjoyed Song Han's brain dump of 400+ slides & videos worth of content on MLSys methods. Interesting mix of practical methods like data/model parallelism, and bleeding edge research topics from automl to quantum ML

0

4

25

Greg Diamos

@GregoryDiamos

3 months

GPUs are currently disrupting cloud - if you walk into a datacenter you can see it happening

7

0

24

Greg Diamos

@GregoryDiamos

5 months

Excited to partner with @Meta to bring Llama finetuning to every enterprise.

Sharon Zhou

@realSharonZhou

5 months

Excited to partner with @Meta ! We have a ton of great F500 enterprises, building on open foundation models like Llama 2: ▫️ Deploying Lamini Stack on premise ▫️ Scaling on Lamini Instances in the cloud Let me know if you want to chat 🙌

5

12

96

2

23

Greg Diamos

@GregoryDiamos

5 years

My son opened his first GPU with TensorCores. Which one will learn faster?

4

0

23

Greg Diamos

@GregoryDiamos

3 months

Jensen said the same thing to us internally in 2009 - stand on the shoulders of giants “on CPUs and C” - and use the GPU to reach higher In hindsight it’s amazing to see CUDA enabled deep learning, which enabled LLMs

Jon Erlichman

@JonErlichman

3 months

2024 valuations: Nvidia: $2.2 trillion Intel: $132 billion 2009 valuations: Intel: $90 billion Nvidia: $5 billion

190

4K

24K

0

2

23

Greg Diamos

@GregoryDiamos

10 months

@DylanOnChips @LaminiAI @AMD @nvidia @CRN It took years to get this to work, but ROCm finally can run a full training and inference stack at hundred GPU scale.

1

4

24

Greg Diamos

@GregoryDiamos

10 months

@LisaSu @LaminiAI @realSharonZhou @AMD @AMDInstinct We have been running on hundreds of @AMDInstinct GPUs in production for over one year. The hard part of this space has been software, but we are now at the point where ROCm works for LLMs.

1

19

Greg Diamos

@GregoryDiamos

4 months

Awni is one of my favorite people to work with ever

Awni Hannun

@awnihannun

4 months

We’re hiring people to work with us on MLX. If you’re interested, can write fast GPU kernels, and have machine learning experience, reach out. More here:

18

91

765

1

19

Greg Diamos

@GregoryDiamos

6 months

What worked for me was asking what I really cared about, and then just start building it.

Andrej Karpathy

@karpathy

6 months

Hi everyone yes, I left OpenAI yesterday. First of all nothing "happened" and it’s not a result of any particular event, issue or drama (but please keep the conspiracy theories coming as they are highly entertaining :)). Actually, being at OpenAI over the last ~year has been

2K

1K

23K

0

1

19

Greg Diamos

@GregoryDiamos

10 months

@DylanOnChips @jaygoldberg @AMD @nvidia I mean that we can run LLM training and inference for models that we use in production like Llama 2 (70B, Code, etc). It needs to work (usability) and be fast (e.g. close to peak TOPs). It's not easy to do this. Only ROCm and CUDA work out of platforms I've tried.

1

3

17

Greg Diamos

@GregoryDiamos

4 months

Tons of code. If you expect someone else to write your code while you work on algorithms, get out. No salary. If you want a $1 million comp package. No way.

27

0

17

Greg Diamos

@GregoryDiamos

1 year

Check out @LaminiAI , the LLM engine that gives every developer the superpowers that took the world from GPT-3 to ChatGPT! Excited to work with @realSharonZhou to build this!

Sharon Zhou

@realSharonZhou

1 year

I’m super excited to announce @LaminiAI , the LLM engine that gives every developer the superpowers that took the world from GPT-3 to ChatGPT! We make it easy to rapidly train custom LLMs from @OpenAI @EleutherAI @Cerebras @Databricks @HuggingFace @Meta 🧵

94

605

3K

0

2

18

Greg Diamos

@GregoryDiamos

4 months

You can call Llama 3 on Lamini now Check out the docs:

0

3

16

Greg Diamos

@GregoryDiamos

6 months

More AMD GPUs on the way

Lamini

@LaminiAI

6 months

Our 2024 first startup cohort is working hard at building LLMs on Lamini 💪 🌶️ We are now accepting applications for our next batch in March. If you are an early-stage startup building LLM applications and needing compute, please apply now! 🙌 🥳

0

3

15

0

1

15

Greg Diamos

@GregoryDiamos

8 months

@kadokaelan Thanks, opened

0

13

Greg Diamos

@GregoryDiamos

5 years

The MLPerf training paper is live! Can we improve ML training speed by more than 50% per year?

0

6

15

Greg Diamos

@GregoryDiamos

11 months

I didn’t realized how unique the NVIDIA flat culture was until I left. Jensen and team got it right and more of the industry should pay attention to how much faster you can innovate if you drop the hierarchy.

Dan Hockenmaier

@danhockenmaier

11 months

@petergyang Interview where he talked about his management style:

10

202

2K

2

0

14

Greg Diamos

@GregoryDiamos

7 years

I’m excited to present our recent work showing that deep learning scaling is simple and predictable at 1:50pm at the Deep Learning at Supercomputer Scale NIPS workshop today.

0

2

13

Greg Diamos

@GregoryDiamos

5 years

Einstein was an amazing research project that succeeded in having impact on Volta in a surprisingly short amount of time. Industry researchers will know how hard this is to pull off. It wouldn’t have happened without an amazing development team and visionary leadership.

Bryan Catanzaro

@ctnzr

5 years

Indeed, it was. And several important ideas from Einstein’s vision made it into Volta’s design. I think Volta’s improvements over prior GPUs are generally underappreciated. Volta is really a huge change in the way GPUs are built. And now those ideas are in Turing as well.

2

18

53

0

4

14

Greg Diamos

@GregoryDiamos

6 months

It's insane that you can almost fit a 200B parameter FP8 LLM in one MI300 GPU. (192GB of HBM)

Greg Diamos

@GregoryDiamos

7 months

First MI300 live in production.

1

8

39

0

3

14

Greg Diamos

@GregoryDiamos

5 months

I asked 20 engineers if they knew how allreduce works inside of frameworks like Megatron-LM. 0 / 20 explained it correctly. This tech is 10 years old, but it is so hidden in the guts of libraries that few people see it or implement it. Maybe it's time to visit this topic again?

Greg Diamos

@GregoryDiamos

5 months

Training LLMs in the Wild Wild West - or how I learned to stop worrying and do multi-node training on AMD GPUs.

0

6

52

2

0

13

Greg Diamos

@GregoryDiamos

9 months

Now with code snippets in our Llama 2 70B playground.

1

2

13

Greg Diamos

@GregoryDiamos

4 months

No egos. If your goal is personal fame over building amazing models, this is not the right place for you. No research. If your goal is to publish and rack up citations, this is not the right place.

2

0

13

Greg Diamos

@GregoryDiamos

5 years

SysML hasn’t quite reached NeurIPS level of attendance, but wow, there are a lot of people here!

0

12

Greg Diamos

@GregoryDiamos

8 months

@yusufg We have optimized training and inference for LLMs. So we can offer a higher level interface. Just submit 10k LLM requests to a cluster of 1000s of GPUs, we will handle it for you. You can also drop to a lower level, eg PyTorch, BLAS, etc. that would work fine

0

1

10

Greg Diamos

@GregoryDiamos

2 months

Lamini is up -

Lamini - Enterprise LLM Platform

Lamini is the enterprise LLM platform for existing software teams to quickly develop and control their own LLMs. Lamini has built-in best practices for specializing LLMs on billions of proprietary...

www.lamini.ai

Benjamin De Kraker 🏴‍☠️

@BenjaminDEKR

2 months

ChatGPT, Claude, and Perplexity -- three totally different AI services -- are all down right now. What the actual F is going on?

825

273

3K

1

2

12

Greg Diamos

@GregoryDiamos

9 months

This was a good paper. One of the first clear demonstrations of what you can do with 40 GB of on-chip SRAM.

Cerebras

@CerebrasSystems

9 months

(1/2) Argonne National Laboratory researchers demonstrate the Cerebras CS-2 is 130x faster than NVidia A100 on a nuclear physics simulation workload. Read the paper here:

6

8

44

0

2

12

Greg Diamos

@GregoryDiamos

1 year

Let's democratize LLMs. Thank you @realSharonZhou and @AndrewYNg for creating a simple and accessible 1-hour course.

Sharon Zhou

@realSharonZhou

1 year

Super excited to announce Finetuning LLMs, a short-course made with my friend @AndrewYNg !! 🎉 By importing open-source @LaminiAI core & @HuggingFace & @PyTorch & @Weights_Biases , you can: ✅ Gain an expert's intuition behind finetuning* ✅ Understand how finetuning fits in vs.

18

122

749

1

2

12

Greg Diamos

@GregoryDiamos

3 years

@IanCutress Wow, MI200 is a monster.

1

0

12

Greg Diamos

@GregoryDiamos

5 years

The MLPerf inference benchmark is launched!

0

5

11

Greg Diamos

@GregoryDiamos

6 months

Open source Lamini software development kits are out:

0

2

11

Greg Diamos

@GregoryDiamos

1 year

This is how a tiny diffusion LLM generates a sentence.

0

11

Greg Diamos

@GregoryDiamos

4 months

Incredible return to open source!

Philipp Schmid

@_philschmid

4 months

Casual Easter Monday with a huge gift from @OpenAI !🤯 They just released an old GPT-3.5 version. 😍 👉

117

199

1K

2

0

11

Greg Diamos

@GregoryDiamos

9 months

Learn how we built an optimized LLM finetuning system on AMD's ROCm AI stack. Leveraging AMD Instinct GPUs & optimizations for major speedups with more to come! 🚀 👉 Read more in-depth technical details:

Lamini LLM Finetuning on AMD ROCm™: A Technical Recipe | Lamini - Enterprise LLM Platform

www.lamini.ai

0

2

10

Greg Diamos

@GregoryDiamos

9 months

Get ready for bigger LLMs! As more AMD GPU capacity comes online we are opening up big model finetuning and inference to a bigger audience. How did we do it?

0

2

10

Greg Diamos

@GregoryDiamos

6 months

We open sourced the Lamini python package here: Apache 2 License Our philosophy is to make our framework code as open source and transparent as possible because we believe that this is the best way to ensure high quality, simplicity, and security.

GitHub - lamini-ai/lamini

Contribute to lamini-ai/lamini development by creating an account on GitHub.

github.com

0

2

10

Greg Diamos

@GregoryDiamos

7 years

Excited to present our successful validation of mixed precision deep neural network training at @reworkAI !

1

10

Greg Diamos

@GregoryDiamos

5 months

A video walkthrough of how to implement RAG in 80 lines of python code -

ICD11 LLM Finetuning on AMD GPU Supercomputer

Learn how to finetune your own LLM on the ICD11 standard from the world health organization (or any dataset) using an AMD supercomputer. Fully open source a...

www.youtube.com

1

4

10

Greg Diamos

@GregoryDiamos

10 months

Much lower barrier to entry for running LLMs on AMD.

AMD extends support for PyTorch Machine Learning development on select RDNA™ 3 GPUs with ROCm™ 5.7

Researchers and developers working with Machine Learning (ML) models and algorithms using PyTorch can now use AMD ROCm 5.7 on Ubuntu® Linux® to tap into the parallel computing power of the Radeon™ RX...

community.amd.com

0

2

10

Greg Diamos

@GregoryDiamos

9 months

We had to write a custom inference server from scratch. We do license it.

Cognitive Computations

@cognitivecompai

9 months

Anyone know how to inference on AMD? tgi and vllm don't work

10

0

13

2

1

8

Greg Diamos

@GregoryDiamos

1 year

Databricks built a high performance and secure data system. I'm excited to work together to enable every business to build and own LLMs trained on their own data. @LaminiAI @databricks . Thank you @matei_zaharia

Matei Zaharia

@matei_zaharia

1 year

This is a big addition to the Databricks platform, letting software vendors reach our 10,000+ customers much more easily! Thanks to our early dev partners @retool , @posit_pbc , @Kumo_ai_team and @LaminiAI !

1

9

51

0

1

9

Greg Diamos

@GregoryDiamos

9 months

These tiny models are so useful for LLM development. The unit tests of LLMs.

Stas Bekman

@StasBekman

9 months

If this is useful for your work, I have just created a <1MB tiny random llama2 model including a tiny 3k tokenizer. This is crucial for extremely fast testing/development. You can easily adapt the tiny model maker script to any other model

15

74

589

1

2

8

Greg Diamos

@GregoryDiamos

7 months

Mixtral paper is out. The experts specialize a bit per domain, eg math, code

Devendra Chaplot

@dchaplot

7 months

We just released Mixtral 8x7B paper on Arxiv:

47

495

3K

0

9

Greg Diamos

@GregoryDiamos

7 years

Check out our recent work. Deep neural network generalization error scaling is simple and predictable, even for large complex architectures on hard open problems:

0

4

8

Greg Diamos

@GregoryDiamos

8 months

@belacquant We have some docs for LLM training and inference on AMD here: We also wrote a blog post about the AMD tech stack here:

Lamini LLM Finetuning on AMD ROCm™: A Technical Recipe | Lamini - Enterprise LLM Platform

www.lamini.ai

1

0

9

Greg Diamos

@GregoryDiamos

5 years

The MLPerf community has made amazing progress in a short amount of time. I’m really excited to see the inference effort coming together.

0

1

9

Greg Diamos

@GregoryDiamos

8 months

Good old fashion ML is dying LLMs popularized zero-shot learning, or “prompt engineering” which is drastically easier to use and more effective than labeling data. IMO, it’s a short matter of time before this takes over all of what used to be called “deep learning”.

4

1

8

Greg Diamos

@GregoryDiamos

6 years

MLPerf training results are out!

1

6

9

Greg Diamos

@GregoryDiamos

5 years

If you want to partner with our team at Landing AI, we are open to working with you, and I would be excited to explore using cutting-edge AI technology to make your products even better and to level up the AI capability inside your organization.

Why I Joined Landing AI

AI is at the heart of some of the world's most exciting products, from smart speakers to self-driving cars. Yet most companies don't know how to take advantage of the opportunity.

www.linkedin.com

1

9

Greg Diamos

@GregoryDiamos

3 months

We put Llama3 and Phi3 in the Lamini playground. You can submit tons of traffic to them and finetune them now. Playground: API Docs:

2

8

Greg Diamos

@GregoryDiamos

3 months

Appreciate your contributions getting us this far - you accelerated the scaling law timeline by years Can’t wait to see what you do next

Ilya Sutskever

@ilyasut

3 months

After almost a decade, I have made the decision to leave OpenAI. The company’s trajectory has been nothing short of miraculous, and I’m confident that OpenAI will build AGI that is both safe and beneficial under the leadership of @sama , @gdb , @miramurati and now, under the

2K

3K

27K

0

2

8

Greg Diamos

@GregoryDiamos

5 years

Im experimenting with short videos about systems and machine learning. Let me know what you think.

AI Datacenters - Seven Lessons Learned

Lessons learned while building computers that are very good at machine learning.

www.youtube.com

0

1

8

Greg Diamos

@GregoryDiamos

5 months

Without Infiniband and NVLink - we needed to optimize the software in our own collectives and models library to scale. See the section on Lamini collectives in our blog.

Sharon Zhou

@realSharonZhou

5 months

Excited to share how we’re scaling to thousands of GPUs in production! …with multi-node LLM training, on not just Nvidia but @AMD GPUs Details 👉 Great blog by our team, led by Ayushi 💅 tl;dr - Push the limits of training LLMs on enterprise data

6

15

163

1

0

8

Greg Diamos

@GregoryDiamos

5 months

Generate millions of questions and answers from 💰💰💰 earnings call transcripts 💰💰💰 with this simple open source Lamini library. It's easy to make your LLM an expert in the most up to date information. 🚀 🚀🚀

GitHub - lamini-ai/lamini-earnings-calls

Contribute to lamini-ai/lamini-earnings-calls development by creating an account on GitHub.

github.com

0

1

8

Greg Diamos

@GregoryDiamos

4 months

@nbrempel No “ai salary” - we pay above market rate for similar stage startups

1

0

4

Greg Diamos

@GregoryDiamos

7 years

Getting ready to survey progress on AI Chips and AI Datacenters at #OReillyAI

0

8

Greg Diamos

@GregoryDiamos

9 months

Why tokenize when you can just predict pixels. Reading papers is more fun than Twitter. I went back over Here's a fun one.

0

2

8

Greg Diamos

@GregoryDiamos

5 years

Are you an outstanding engineer who wants to break into AI? Reach out to me to learn more.

0

7

Greg Diamos

@GregoryDiamos

1 year

Public alpha is here: HPC LLM fine-tuning!

Sharon Zhou

@realSharonZhou

1 year

Excited to announce @LaminiAI ’s new public alpha: free, fast & furious LLM finetuning! Finetuning was complex & expensive before. Now, anyone can do it in: ⏳10 mins 🧑‍💻3-5 lines of code 💸$0 free 👉🏻 🧸Free tier: toy ~400M LLMs. This is just the start 🧵

12

43

193

1

0

7

Greg Diamos

@GregoryDiamos

7 years

We finally have evidence that low precision training works reliably, even for our biggest models!

Baidu Research

@BaiduResearch

7 years

Check out our paper on training deep neural networks using half precision FP numbers, a joint work w/ @NvidiaAI

1

51

115

0

1

7

Greg Diamos

@GregoryDiamos

8 months

I spent some time this week playing with Intel's new Neural Chat LLM which is at the top of the Huggingface LLM leaderboard. Nice work Intel. The pace of new LLMs is relentless. On Lamini, we support new models immediately. Try it out now at:

0

7

Greg Diamos

@GregoryDiamos

6 months

Over the last few months I’ve talked to many top AI researchers and software engineers leaving big tech. I was curious to hear where they want to go next. Two answers rise to the top: 1) OpenAI 2) Found an AI startup I wonder what that says about the job market

0

1

7

Greg Diamos

@GregoryDiamos

2 months

We got hit with a DDOS after this post. Maybe that had something to do with the outages last week. As the LLM space matures it is helpful to think about security. What would happen if there was a data breach that leaked all conversation history?

1

0

2

Greg Diamos

@GregoryDiamos

3 months

@ylecun @npinto @karpathy @AndrewYNg I remember looking at the shader assembly (SASS) of one of Bryan’s early conv kernels with David Tarjan - they were very close to matrix multiplication - one of the most compute bound kernels we had in CUDA

1

0

6

Greg Diamos

@GregoryDiamos

6 months

Scaling laws: improve AI by sucking up more data

0

6

Greg Diamos

@GregoryDiamos

5 months

It includes scores from a reward model for DPO or RLHF training.

0

1

6

Greg Diamos

@GregoryDiamos

11 months

Interesting perspective from @petewarden Why Nvidia's AI Supremacy is Only Temporary I agree with a lot of this, but not that CPUs are good enough for inference. Scaling laws give an advantage to accelerators, but the SW also matters.

Why Nvidia’s AI Supremacy is Only Temporary

Nvidia is an amazing company that has executed a contrarian vision for decades, and has rightly become one of the most valuable corporations on the planet thanks to its central role in the AI revol…

petewarden.com

0

7

Greg Diamos

@GregoryDiamos

2 months

We have some good SW to go with that HW.

Sharon Zhou

@realSharonZhou

2 months

It’s official: Satya announced AMD GPUs on Azure! 🎉 If you want to run LLMs on the most cost-effective AMD GPUs on Azure, please contact us @LaminiAI :) Contact:

4

7

65

1

7

Greg Diamos

@GregoryDiamos

5 years

Even though AI is starting to work well, many companies still struggle to land it in mass production. After seeing many examples, we wrote down these best practices. I expect that many of them can be applied to your AI deployments.

Taking Your AI Projects from Pilot to Production

Six tips to keep from getting stuck in the proof of concept stage.

www.industryweek.com

0

3

7

Greg Diamos

@GregoryDiamos

3 months

Excited to see that LLMs with photographic memory are starting to work really well internally We would have never found this with a pure research perspective and no user input

0

6

Greg Diamos

@GregoryDiamos

2 months

@IanCutress @LaminiAI @realSharonZhou It was fun reminiscing about the early days of CUDA @IanCutress - and what comes next. Drop by anytime.

0

1

3