Joey Gonzalez @profjoeyg Twitter profile

Last Seen Profiles

@Misscrazy199390

@SoyKarlaTorres_

@TanjaBaker1

@Myhappines49220

@_RAJF_

@GeorgesEColbert

@Dbrave_8

@lndryecosystem

@DrGregParker001

@NDKNIGHTBASE

@vedatyeler_

@briALTcookiebox

@akber_asif

@Basket1Coolidge

@JoyceLarsow8334

@wkwkwwk1233

@lauzadis_justas

@leighkeystone

@jastipeonnie

@tealcrowns

@okeanaye

@KenKrahenbuhl1

@Western_Tribune

@xb0g0

@SnowyPackel

@antonia_yamin

@uttonz

@John_aka_Alwayz

@EleazarManzano_

@KatieLusso

@MB_Farrelly

@topfan_thorsten

@cypurrpunk

@TabbieTalks

@token_summit

@WarnerMusicPH

Joey Gonzalez

@profjoeyg

3 years

I am very excited to announce that my research group just received tenure @UCBerkeley . Alright, technically I received tenure, but I could not have done this without the hard work of my amazing team of students and colleagues. Queue awards music (1/4)

40

10

541

Joey Gonzalez

@profjoeyg

11 months

Serving LLMs? My students found a way to accelerate serving by over an order-of-magnitude just by changing the way memory is managed (spoiler alert): gpu memory fragmentation = slow. Introducing vLLM with PagedAttention:

Zhuohan Li

@zhuohan123

11 months

🌟 Thrilled to introduce vLLM with @woosuk_k ! 🚀 vLLM is an open-source LLM inference and serving library that accelerates HuggingFace Transformers by 24x and powers @lmsysorg Vicuna and Chatbot Arena. Github: Blog:

20

265

1K

4

67

308

Joey Gonzalez

@profjoeyg

2 years

1/ As some of you know, I recently got tenure at @UCBerkeley . I've been thinking about my research career and the simultaneous evolution of the ML space, and I wanted to share what I've been thinking about:

How Machine Learning Became Useful

Reflecting on a Decade of Research in Machine Learning Systems

medium.com

6

37

175

Joey Gonzalez

@profjoeyg

8 months

I have always felt that RL didn't have a killer application, until now. After talking with @natolambert about the future of RL in LLMs, I think RLCF might be the next big thing. If you are working with LLMs and code you should check it out.

Generating Conversation: RLHF and LLM Evaluations with Nathan Lambert...

This week on Generating Conversation, we have Nathan Lambert with us. Nathan is a research scientist and RLHF team lead at HuggingFace. Nathan did his PhD at...

www.youtube.com

2

21

108

Joey Gonzalez

@profjoeyg

1 year

. @lmsysorg just released our rankings comparing GPT, Claude, Vicuna and others. Commercial models dominate the top spots but open models are still very good and the only option for most companies. See how @AqueductHQ makes using open LLMs easier:

2

19

71

Joey Gonzalez

@profjoeyg

2 years

. @joe_hellerstein , @vsreekanti , @cgwu0530 and I have been working on a new project (and company) to simplify infrastructure for data scientists. We're looking for feedback from data engineers who support data scientists. If you or a friend are willing to chat, we'd appreciate it!

11

21

59

Joey Gonzalez

@profjoeyg

10 months

I am excited to announce that two of the LLMs from my group (Gorilla and Vicuna) are on AI Business’s top 12 models. Congrats @shishirpatil_ , @tianjun_zhang , @xinw_ai , and the @lmsysorg team. We are looking forward to working with @Meta on Llama-v2 versions.

2

7

51

Joey Gonzalez

@profjoeyg

7 months

My students @shishirpatil_ and @tianjun_zhang and their undergrads @_royh021 and @fanjia_yan just presented some exciting new features for #GorillaLLM at Sky Camp . Now you can LoRA fine-tune and get open function calling from one place.

0

12

48

Joey Gonzalez

@profjoeyg

2 years

Congratulation @lm_zheng on receiving the prestigious Meta Phd Fellowship for your work on compilers for deep learning. Keep up the great work!

Lianmin Zheng – Meta Research | Meta Research

Lianmin Zheng is a third-year PhD student in the EECS department at UC Berkeley, advised by Ion Stoica and Joseph E. Gonzalez.

research.facebook.com

2

0

42

Joey Gonzalez

@profjoeyg

8 months

I just posted an interview with @shishirpatil_ and @tianjun_zhang , the creators of #GorillaLLM and rockstar PhD students in my lab @UCBerkeley . 🦍 #GorillaLLM enables LLMs to discover and invoke cloud APIs and command line tools.

Generating Conversation: Gorilla, An LLM for Massive APIs - Shishir...

Gorilla is an open-source LLM from the Sky Lab at UC Berkeley that generates API calls for massive APIs. Gorilla is built by fine-tuning the open-source LLM ...

www.youtube.com

1

6

42

Joey Gonzalez

@profjoeyg

1 year

We launched our epic battle (randomized trial) of the LLMs this morning. We know what ChatGPT-4 thinks of each of the models but what about you? Also let us know what you think of the interface.

lmsys.org

@lmsysorg

1 year

Introducing Chatbot Arena 🤖 ⚔️ 🤖 : We have collected the most popular open-source LLMs and need your help to determine which LLM is the best. In in this epic battle of AI versus AI, only you can decide the winner. Let the battle begin !

16

117

538

1

7

36

Joey Gonzalez

@profjoeyg

7 months

🚨New Project -- MemGPT: LLM mediated virtual context paging -- we are leveraging the function calling abilities of modern LLMs to enable direct context management just like an OS manages pages. You can try it now!

Charles Packer

@charlespacker

7 months

Introducing MemGPT 📚🦙 a method for extending LLM context windows. Inspired by OS mem management, it provides an infinite virtualized context for fixed-context LLMs. Enables perpetual chatbots & large doc QA. 🧵1/n Paper: GitHub:

9

107

465

1

6

37

Joey Gonzalez

@profjoeyg

1 year

Ever wonder how LLMs work? I tried to explain how LLMs work (the math, not the magic 🧙 ) and where things are headed 📈 to my co-founders. Let me know what you think. Did I miss anything?

Intro to LLMs (Generating Conversation, Episode 1)

Large language models have taken the world by storm, but we're still learning what they do and how they work. In this conversation, UC Berkeley Professor & A...

www.youtube.com

1

8

32

Joey Gonzalez

@profjoeyg

1 year

My students just released the results of our open crowd-sourced competition among open-source LLMs. And the winner is ...

lmsys.org

@lmsysorg

1 year

Evaluating LLMs is notoriously difficult, and academic benchmarks may fail. Inspired by chess and MOBA games, we are taking a new approach by calculating Elo ratings of models with crowdsourced battle data. - Blog: - Leaderboard:

31

277

1K

4

8

32

Joey Gonzalez

@profjoeyg

3 years

Anyone using feature stores? I am pretty excited about where the technology is headed and we wrote a short blog post about it .

Feature Stores: The Data Side of ML Pipelines

We need a principled way of managing state in real-time ML pipelines.

medium.com

1

3

31

Joey Gonzalez

@profjoeyg

7 months

MemGPT is also trending on GitHub: Well done @charlespacker , @vivianfxng , @shishirpatil_ , @nlpkevinl , and @sarahwooders ! I hope we don't have too many 🐞.

Vikram Sreekanti

@vsreekanti

7 months

. @charlespacker released MemGPT earlier this week, and it was on the front page of HackerNews for 2 days straight. 🤯 Charles joined @profjoeyg this week to talk about context 🧠, memory management 🤔, and the future of conversational AI ➡️.

0

3

7

1

5

29

Joey Gonzalez

@profjoeyg

11 months

We have a new multi-round open-ended LLM benchmark that is evaluated by LLMs. The open-source models are actually doing remarkably well but you also see more spread in the commercial models.

lmsys.org

@lmsysorg

11 months

🔥Big news from Chatbot Arena: Meet our new MT-Bench leaderboard & Vicuna-33B! We present a comprehensive, scalable, and validated leaderboard differentiating across open (Falcon, Wizard & Guanaco) and proprietary models (GPT-4, Claude & PaLM). Blog post:

14

101

436

1

7

28

Joey Gonzalez

@profjoeyg

9 months

I am teaching an AI-Systems graduate seminar this semester with @matei_zaharia . We are focusing on LLMs (obviously...) and our first required reading was the very well written and insightful T5 paper by @colinraffel et al. but I have one issue ...

1

0

28

Joey Gonzalez

@profjoeyg

9 months

I just posted a fantastic interview with @jerryjliu0 of @LlamaIndex . This is the most in-depth interview on the podcast so far and really dives into the intersection of LLM technology and data. 🧠+💽

Generating Conversation: A deep dive on Llama Index with Jerry Liu...

After a long break, we're back with episode 5 of Generating Conversation! This week, we're joined by Jerry Liu, the co-founder and CEO of Llama Index. Jerry ...

www.youtube.com

0

9

25

Joey Gonzalez

@profjoeyg

7 months

@ylecun We have been doing this with the FastChat arena as part of our @lmsysorg effort. We recently released the largest open-source chat collection along with human ratings. This is entirely open-source. Check it out:

Chat with Open Large Language Models

chat.lmsys.org

2

0

28

Joey Gonzalez

@profjoeyg

11 months

Just finished a fun talk at @mlopscommunity with @vsreekanti . Here is what has been driving our thinking (spoiler alert!). Is anyone using the Ideal LLM Stack?

2

9

24

Joey Gonzalez

@profjoeyg

11 months

We found a very simple way to extend the context length of LLMs while preserving model accuracy!

lmsys.org

@lmsysorg

11 months

🔥Introducing LongChat🤖, our new chatbots supporting 16K tokens context, and LongEval, our new benchmark for testing long context chatbots. 🤥Surprisingly, we found open LLMs often fail to achieve their promised context length. Check our blog for details:

4

106

476

1

5

22

Joey Gonzalez

@profjoeyg

9 months

It's always exciting when others outside @Berkeley_EECS contribute to our open-source projects. Thanks @morgymcg and @weights_biases for contributing to the Gorilla project! 🦍🙏

Morgan McGuire

@morgymcg

9 months

Put together a quick colab to fine-tune @OpenAI ChatGPT-3.5 on the huggingface api code from the gorilla dataset Idea being to see if something like this can help improve ChatGPT-3.5's use of tools and mimic GPT-4's `functions` capability

5

9

43

2

5

22

Joey Gonzalez

@profjoeyg

3 years

I would like to thank my @ucbrise and @berkeley_ai colleagues @joe_hellerstein , Ion Stoica, @ralucaadapopa , @KurtKeutzer , @trevordarrell , @Ken_Goldberg , @DebAtStat , and @fperez_org as well as my students including (2/4)

2

0

22

Joey Gonzalez

@profjoeyg

9 months

I started a new blog! Along with @vsreekanti and some of my students at @UCBerkeley , I'll be writing about what I'm seeing in the LLM space across research & industry + what my group is doing. First post is coming later today!

Generating Conversation | Joseph E. Gonzalez | Substack

The latest in generative AI & LLMs across research and industry. Click to read Generating Conversation, a Substack publication with thousands of subscribers.

generatingconversation.substack.com

1

6

21

Joey Gonzalez

@profjoeyg

1 year

A lot of people are talking about the build versus buy question for LLMs. For most, the answer is really easy.

LLMs: Build vs Buy? (Generating Conversation, Episode 3)

With every organization looking to adopt LLMs, whether you should build your own or buy an existing model is a topic that's come up repeatedly. In this video...

www.youtube.com

0

6

21

Joey Gonzalez

@profjoeyg

8 months

I just interviewed @ajayj_ , co-founder of @genmoai and @Berkeley_EECS PhD. @ajayj_ introduced diffusion models for image generation and 3D modeling and @genmoai he and his brother @_parasj are reimagining how we interact with generative AI.

Generating Conversation: Genmo, A Platform for Generative AI Art -...

While LLMs have been all the rage in 2023, visual models have made incredible strides in recent years. Ajay & Paras Jain, both Berkeley PhDs, have been innov...

www.youtube.com

0

4

19

Joey Gonzalez

@profjoeyg

1 year

My former students are doing some really cool working making Pandas run at scale in the data warehouse.

Ponder

@ponderdata

1 year

📣 Introducing Ponder: Run #pandas on 1TB+ DIRECTLY in your data warehouse 🚀 Learn more below! 🧵[1/N] #python #datascience #AI #database

10

40

177

1

20

Joey Gonzalez

@profjoeyg

2 years

My students have been rethinking the architecture of the cloud and found a way to make data movement 110x faster. Check out their new open-source project!

Paras Jain

@_parasj

2 years

Releasing Skyplane, a new open-source tool to move huge datasets between clouds. Skyplane is: 1. 🔥 Blazing fast (110x faster) 2. 🤑 Cheap (4x cheaper) 3. 🌐 Universal (AWS, Azure and GCP) Read more: 1/

8

57

259

0

1

20

Joey Gonzalez

@profjoeyg

11 months

We have been thinking a lot about how people can use LLMs to talk to their data and solve real problems. Next week, @vsreekanti and I will present what we have learned at the @mlopscommunity conference on LLMs in production. Check it out!

1

10

20

Joey Gonzalez

@profjoeyg

2 years

I am really excited to announce the first major release of the @AqueductHQ open-source project. It embodies almost a decade of research in prediction serving, data infrastructure, and server-less computing at @UCBerkeley . Let us know what you think.

RunLLM

@RunLLM

2 years

We just released Aqueduct v0.1! 🎉 We're on a mission to remove the complexity from getting data science & ML in production, and this release is a big step in that direction. We're also on ProductHunt today:

1

9

29

1

17

Joey Gonzalez

@profjoeyg

2 years

After running (and initially failing at) project management at my startup for the past year, I completely agree with this! Grad students, it’s time for some process. :-)

Arvind Narayanan

@random_walker

2 years

Academics would double our productivity if we learnt some basic project management skills that are bog standard in the industry. We have this myth that scholarly success is all about brilliance and creativity, but in fact 90% of it is getting sh*t done, same as any other job.

62

588

6K

0

1

18

Joey Gonzalez

@profjoeyg

8 months

Two weeks ago, our work on GraphLab received the test of time award at VLDB🎉. Funny story, our test of time paper almost wasn't published. @YuchengLow and I wrote a blog about our experience which will hopefully encourage future graduate students.

How our Test-of-Time Paper Almost Wasn’t

The non-linear path to success

generatingconversation.substack.com

1

3

18

Joey Gonzalez

@profjoeyg

8 months

Our FastChat project is killing it! Great work @lm_zheng , @infwinston , and @haozhangml ! I am looking forward to the big announcements next week. 🏟️

lmsys.org

@lmsysorg

8 months

Mistral-7B is now available at under both the "Chatbot Arena" and "Single Model" tab. Test it yourself! We are glad that our tools (FastChat/Skypilot/vLLM) helped the release of this model! Chatbot Arena now serves over 450 billion parameters for

4

36

255

1

3

18

Joey Gonzalez

@profjoeyg

1 year

@lmsysorg If you checkout the linked notebook in our @lmsysorg blog you can see the bootstrap estimates of the Elo scores. They show which models are really close and where the differences are likely more significant.

0

2

17

Joey Gonzalez

@profjoeyg

2 years

1/ Part 2! Last week, I talked about how data + compute + abstractions catalyzed the ML revolution. The natural next question is how we put those models to use (hint: it’s not testing):

The Real Challenge in (Useful) Machine Learning isn’t Learning

This blog is cross posted on the Aqueduct blog.

medium.com

1

7

17

Joey Gonzalez

@profjoeyg

7 months

@ylecun I also really want open-source LLMs to win! However, I think LLMs are more like search engines -- they require constant expensive training to maintain quality (like crawling) and it is difficult to accumulate contributions (can't merge training runs...yet).

2

17

Joey Gonzalez

@profjoeyg

8 months

Weaviate is now using our #GorillaLLM project to go from natural language to GraphQL! 🦍 Congratulations @tianjun_zhang and @shishirpatil_ ! 🎉

Connor Shorten

@CShorten30

8 months

We trained LlaMA 7B to use Weaviate!! 🦍🛠️ Presenting... Weaviate Gorilla Part 1: GraphQL! 🎉 Blog Post: YouTube: 🧵 With some more details 👇

16

76

244

2

8

18

Joey Gonzalez

@profjoeyg

3 years

@charlespacker , @pschafhalter , @sukritkalra , @TrawickNathan , @tianjun_zhang , Suzie Petryk, @GaiYu0 , @lm_zheng , @Shishir_India , @nlpkevinl , @sarahwooders , Justin Wong, and Lisa Dunlap (4/4)

1

0

16

Joey Gonzalez

@profjoeyg

8 months

Congratulations, @_parasj and @ajayj_ on your latest generative video model! It's insane!! I have to ask, how much 💸 did that cost and ... when can I read the arXiv version?

Genmo

@genmoai

8 months

Generative video models are rapidly improving in quality. Meet Replay, a new AI model that can generate stunning videos from text. Replay v0.1 is designed to create ultrasmooth HD videos with a new interface. Available today for everyone. What's New? 1. Replay understands plain

49

90

464

1

2

16

Joey Gonzalez

@profjoeyg

7 months

I am excited to announce that my @lmsysorg team is now working with @kaggle to help improve LLM evaluation. We look forward to announcing new joint challenges in the coming months.

lmsys.org

@lmsysorg

7 months

We're super excited to partner with @kaggle , welcoming the ML and data science community to Arena! Yesterday's Kaggle launch, we recorded the highest traffic to date since the Arena launch! Over 4K votes in a day🗳️ Our mission remains building an open and community-first

2

23

163

0

5

16

Joey Gonzalez

@profjoeyg

9 months

. @vsreekanti and I just wrote a fun blog about the LLM Stack and why it's so hard to build LLM applications. 🚨Spoiler alert, it's not the LLM 🚨

The Easiest Part of LLM Applications is the LLM

LLMs have brought thousands of developers into the machine learning world. The models themselves are, of course, impressive and key to the explosion of applications, but they’re only part of the...

generatingconversation.substack.com

0

6

15

Joey Gonzalez

@profjoeyg

1 year

Check out this new project from one of my students to generate short movies from text. It’s powered by really cool inference technology so anyone can try it right now.

Genmo

@genmoai

1 year

Announcing Genmo Video, a generative media platform with a new text-to-video model that can generate immersive live artwork from any prompt or any image. What will you create? 🎨▶️ Free public access: Discord: 👇1/n

15

63

253

0

3

16

Joey Gonzalez

@profjoeyg

7 months

My students decorated the Sky Computing Lab @UCBerkeley with a 🍭 candy land theme 🍬. Was this your doing @lisabdunlap ?

1

2

16

Joey Gonzalez

@profjoeyg

1 year

Open source LLMs are critical to the growth of the ML community and it is exciting to be a part of open efforts to make them easier to use. Now we just need to make them better 😛.

RunLLM

@RunLLM

1 year

Dolly v2 from @databricks is a big deal — the first commercially viable open-source LLM! Running it in the cloud — like with all foundation models — is a pain. With Aqueduct, you can do it in a single line of Python:

0

2

6

2

15

Joey Gonzalez

@profjoeyg

2 years

Google recently posted about our exciting collaboration around a new framework to easily automate model parallel training while also achieving state-of-the-art performance.

Google AI

@GoogleAI

2 years

Alpa is a framework that uses just one line of code to easily automate the complex model parallelism process for large #DeepLearning models. Learn more and check out the code.

6

99

373

0

15

Joey Gonzalez

@profjoeyg

11 months

Ever wondered if LLMs could revolutionize ... the terminal? Well we did ... and we are excited to announce the new Gorilla CLI.

Shishir Patil

@shishirpatil_

11 months

🦍Introducing the all-new gorilla-cli, now available as a pip package!✍️ With a vast collection of ~1500 🆕APIs, including 👀 Kubernetes, AWS, GCP, Azure, GitHub, Conda, Curl, Sed, and more🤩 simply state your goal, and let Gorilla CLI generate the commands for execution.

5

29

148

0

4

14

Joey Gonzalez

@profjoeyg

8 months

We all know that RAG is the killer application for LLMs but did you know that it doesn't work (out of the box)? Here are some basic steps needed to make RAG actually work:

How to Optimize Retrieval-Augmented Generation

The latest episode of our podcast is out! We had Nathan Lambert from HuggingFace on to discuss RLHF, LLM evaluations, and how to improve discussion around AI research. Check it out! Retrieval-augme...

generatingconversation.substack.com

2

5

12

Joey Gonzalez

@profjoeyg

1 year

@arankomatsuzaki Wow you are fast to find papers! We just posted this and are trying to post a demo ASAP.

0

12

Joey Gonzalez

@profjoeyg

2 years

@beenwrekt We are hosting a live/free version of the OPT-175B model () for people to study. I strongly support the need for safety measures when using large language models but how should we apply them to the research platform?

1

0

12

Joey Gonzalez

@profjoeyg

3 years

I am really excited to be part of an effort @Berkeley_EECS to introduce courses addressing social justice and technology into the core EECS curriculum. Do you know anyone who could help build and teach these important new classes?

Cathryn Carson

@CathrynCarson

3 years

. @Berkeley_EECS invites applications for a lecturer to teach "EECS for All: Social Justice in EECS" in Spring 2022. If you have teaching experience and a background at the intersection of social justice and technology, please consider applying!

0

3

2

1

2

13

Joey Gonzalez

@profjoeyg

7 months

I just posted a fun interview I did with my student @charlespacker on the challenges of conversational AI, creativity in LLM, and his exciting new work on virtual context management (MemGPT).

Generating Conversation: MemGPT, Memory Management for LLMs - Charles...

Context window management has become a critical part of every LLM application — from the basics (embeddings models, vector DBs) to more advanced techniques (...

www.youtube.com

2

3

13

Joey Gonzalez

@profjoeyg

2 years

I am excited to be a part of this new cross-campus effort using AI to help create new materials that have the potential to tackle some of the biggest challenges of climate change.

Berkeley Computing, Data Science, and Society

@BerkeleyDataSci

2 years

Imagine a technology that removes planet-warming emissions from smokestacks and turns the air's moisture into drinking water. @UCBerkeley 's new Bakar Institute of Digital Materials for the Planet will use #chemistry & #machinelearning to enact this vision.

0

11

23

0

12

Joey Gonzalez

@profjoeyg

1 year

@matei_zaharia Technically, Vicuna is constrained by the Llama license more than the data. However, it would be great to see how Dolly performs on images. You can compare the two models side by side using .

Chat with Open Large Language Models

chat.lmsys.org

0

2

12

Joey Gonzalez

@profjoeyg

4 years

Updating models is important. However, if you find that you need very frequent updates, you probably are not directly modeling the temporal variation in the underlying task. For example, don't update a CTR model with each click, use the clickstream as a feature.

Chip Huyen

@chipro

4 years

4. You won’t need to update your models as much One mindboggling fact about DevOps: Etsy deploys 50 times/day. Netflix 1000s times/day. AWS every 11.7 seconds. MLOps isn’t an exemption. For online ML systems, you want to update them as fast as humanly possible. (5/6)

8

46

489

2

0

12

Joey Gonzalez

@profjoeyg

1 year

Want to work on advancing AI to solve real problems? We are looking for postdocs to join a new collaboration with colleagues in chemistry to bring AI to the design of materials for everything from energy storage to carbon capture .

1

2

12

Joey Gonzalez

@profjoeyg

6 months

Should you be starting a GenAI company? I asked @zooie , an expert AI entrepreneur and investor, and was surprised by his answer: No, the world is full of undifferentiated picks and shovels and applications haven't yet found PMF. What do you think?

Generating Conversation: Building a Business in Generative AI - Vik...

In this episode, Vik Singh joins us to discuss his perspective on the technology, product, and business perspective on building a business in generative AI. ...

www.youtube.com

0

3

13

Joey Gonzalez

@profjoeyg

3 years

Thank you all! (5/4)

1

0

12

Joey Gonzalez

@profjoeyg

7 months

@ylecun @martin_casado I think the real immediate risk for these technologies is that: (1) we trust them — “ChatGPT told me” … “so it must be true.” (2) people use them to manipulate others — “explain why X is true to a person who believes Y.” (3) we start to rely on their opinions … see (1)

4

1

10

Joey Gonzalez

@profjoeyg

3 years

@NeerajaJY , @dan_crankshaw , Francois Belletti, @xinw_ai , @lvinwan , @richliaw , @brthananjeyan , @vsreekanti , @simon_mo_ , @_parasj , @FeinbergVlad , Eyal Sela, Devin Petersohn, @dr_othchild , I am not done yet (3/4)

1

0

12

Joey Gonzalez

@profjoeyg

1 year

In this video, we discuss what is happening with foundation models (they are everywhere). As always, I am curious what people think. Should there be chalk or crayons in my future videos?

Foundation Models (Generating Conversation, Episode 2)

LLMs are part of a larger class of models called foundation models. What are foundation models, how do they work, and how well do they generalize? Cal profes...

www.youtube.com

0

7

11

Joey Gonzalez

@profjoeyg

1 year

We just released our Google PaLM benchmarking results against OpenAI and many major open source models. I have to admit, I was a little surprised by Google’s results. However, the explanation is promising.

lmsys.org

@lmsysorg

1 year

⚔️Chatbot Arena Leaderboard Update! Exciting to welcome new entrants: - Google PaLM 2 - Claude-instant-v1 - MosaicML MPT-7B The competition is heating up🔥 Check out our analysis for all the surprising results at Remember, your vote shapes the arena.

39

194

1K

0

6

10

Joey Gonzalez

@profjoeyg

3 years

If you are interested in systems for machine learning, check out the new MLSys Conference (part of the NeurIPS foundation). In the past, registration has sold out quickly so be sure to register soon.

Alex Dimakis

@AlexGDimakis

3 years

#MLSys2021 : We are proud to announce our three keynote speakers, Bill Dally , NVIDIA, Jeannette Wing, Columbia University and Kathy Yelick, UC Berkeley. Watch them on April 6-8, 2021. Registration: @BillDally @KathyYelick @smolix

0

14

43

0

11

Joey Gonzalez

@profjoeyg

8 months

@NVIDIAAIDev I am excited to see our Paged-Attention work in the latest NVIDIA announcement. Call me academic, but where is the citation? 😉 Congratulations @woosuk_k and @zhuohan123 ! 🎉

1

11

Joey Gonzalez

@profjoeyg

1 year

We are excited to announce that we just released FastChat-T5 for public use. This is an encoder-decoder architecture (unlike Vicuna) but according to our early benchmarks it already outperforms Dolly-V2 and can be used in the same settings.

lmsys.org

@lmsysorg

1 year

We are excited to release FastChat-T5: our compact and commercial-friendly chatbot! - Fine-tuned from Flan-T5, ready for commercial usage! - Outperforms Dolly-V2 with 4x fewer parameters. Link:

30

153

742

0

2

11

Joey Gonzalez

@profjoeyg

2 years

Why are data scientists spending so much time solving the same engineering problems? I am increasingly convinced that MLOps tools are designed for large engineering teams at tech giants and not the every-day data scientists that need them.

RunLLM

@RunLLM

2 years

1/ MLOps has become increasingly popular of late as a solution to deploying and managing ML models in the cloud. But we believe MLOps is taking the data science and machine learning community in the wrong direction:

2

9

26

0

4

10

Joey Gonzalez

@profjoeyg

10 months

It's exciting to see our work on LLMs for APIs getting attention!

Lior⚡

@AlphaSignalAI

10 months

We're about to save a lot of time. The first LLM specializing in writing API calls is out. Gorilla can write your code and accurately invoke 1,600+ API calls while reducing hallucination. With a simple text input, Gorilla comes up with the semantically correct code and API to

20

135

640

2

11

Joey Gonzalez

@profjoeyg

10 months

We just released the raw conversations and user judgements from the Chatbot Arena! Hopefully this will enable other researchers to study AI safety in the wild as well as advance open-source RLHF training.

lmsys.org

@lmsysorg

10 months

We are excited to announce the first major release of the Chatbot Arena conversation dataset! - 33K conversations with pairwise human preferences - 20 SOTA models such as GPT-4, Claude, and LLaMA-based Vicuna - From 13K unique IPs in the wild - An additional 3K expert-level

14

177

731

0

1

10

Joey Gonzalez

@profjoeyg

4 years

We just posted our latest work on serving machine learning prediction pipelines in real-time.

Vikram Sreekanti

@vsreekanti

4 years

1/ Putting trained ML models in production is necessary to integrate them into real applications, but prediction serving has received relatively little attention to date. Today's solutions (e.g., AWS SageMaker) have significant shortcomings around usability and scaling.

1

5

17

0

10

Joey Gonzalez

@profjoeyg

8 months

It has been really exciting to see our work on FastChat and vLLM have real impact in the community. We are lucky to have amazing students @lmsysorg , @lm_zheng , @infwinston , @ying11231 , @woosuk_k , and @zhuohan123 leading these projects.

Hao Zhang

@haozhangml

8 months

Congrats to Mistral on the release of the best 7B model ever! Extremely exciting to see that Mistral adopted the full stack of LLM infra we built at : fastchat as the finetuning and serving infra, vllm as the inference engine, and mt-bench for evaluation!

0

1

38

0

3

10

Joey Gonzalez

@profjoeyg

1 year

Thanks @amanda_robs & @tnachen for hosting us! It was fun talking about where MLOps is headed, what is missing, and the challenges of growing a successful open-source project in the space. It was like going to therapy! 😊

Open Source Startup Podcast🎙

@OssStartup

1 year

Ep 77 of the Open Source Startup Podcast is LIVE🎙️ Check out @tnachen & @amanda_robs convo w/ @AqueductHQ Founders @vsreekanti & @profjoeyg 🎧 They discuss learnings from interviews w/ 100s of data teams, building in the competitive MLOps space & more!

1

3

12

0

2

10

Joey Gonzalez

@profjoeyg

8 months

Can GPUs in the ☁️ really drive your 🚗 and make it safer? We have been studying this question and @pschafhalter will present our findings this afternoon @ieeeiros 2023. Spoiler alert: Yes!

Self-Driving Cars Should Use the Cloud

Self-driving cars should use the cloud. The cloud provides key advantages such as cost-effective, powerful hardware. However, network latency and connectivit...

www.youtube.com

0

3

10

Joey Gonzalez

@profjoeyg

1 year

We just launched support for running open-source LLMs on @AqueductHQ with a single line of Python, and we're live on ProductHunt. Check it out!

Aqueduct - Product Information, Latest Updates, and Reviews 2024 | Product Hunt

Aqueduct automates the engineering required to take data science to production. By abstracting away low-level cloud infrastructure, Aqueduct enables data teams to run models anywhere, publish...

www.producthunt.com

0

5

9

Joey Gonzalez

@profjoeyg

1 year

As someone involved in the @lmsysorg effort, I want to see a future full of open models. However, it is worth noting that what is enabling the rapid open-source progress is: (COMPUTE) strong foundation models and (DATA) high-quality dialogue. Improving both is expensive.

Dylan Patel

@dylan522p

1 year

Google "We Have No Moat, And Neither Does OpenAI" Leaked Internal Google Document Claims Open Source AI Will Outcompete Google and OpenAI This is the opinion of one Googler, we do not agree, simply sharing. $GOOGL $MSFT $META $AI $NVDA $AMZN $AAPL

31

125

690

1

9

Joey Gonzalez

@profjoeyg

1 year

. @AqueductHQ , we're thinking hard about the best ways to orchestrate ML workflows. @vsreekanti & I have lots of questions for those of you orchestrating production ML. Let me know if you're willing to do a quick research call to discuss challenges in the space!

0

2

9

Joey Gonzalez

@profjoeyg

1 year

The launch of the @llama_index company is a big deal for anyone interested in connecting LLMs to their data and I am excited to be a part of it!

Jerry Liu

@jerryjliu0

1 year

I’m super excited to make it official: @disiok and I have started a company around @llama_index , and we’ve raised a $8.5M seed round led by @GreylockVC ! 🔥🚀 We are building the open-source data framework to unlock LLM capabilities on your private data.

96

120

1K

0

9

Joey Gonzalez

@profjoeyg

7 months

I keep getting arxiv-baited on slack: "Have you seen <random arxiv link> from today?" No, I haven't read this paper that came out an hour ago! Did we get scooped or is this just an interesting paper? I need to be emotionally prepared once this PDF loads.

Lisa Dunlap

@lisabdunlap

7 months

I feel like the academic equivalent of the iPhone alarm noise is the slack message “found this recent work on arxiv, seems similar to what you have been working on”

1

4

54

1

9

Joey Gonzalez

@profjoeyg

1 year

I'm thinking about starting an interview series on LLMs @AqueductHQ . Who would you like to see?

1

0

9

Joey Gonzalez

@profjoeyg

1 year

What if you could just tell your computer to accomplish high-level tasks spanning applications? What if you could orchestrate cloud services across clouds using just English (no shell ... no Python ... not even YAML)? My students might have the answer!

Shishir Patil

@shishirpatil_

1 year

📢 Excited to release Gorilla🦍 Gorilla picks from 1000s of APIs to complete user tasks, surpassing even GPT-4! LLMs need to interact with the world through APIs, and Gorilla teaches LLMs APIs. Presenting Gorilla-Spotlight demo🤩 Webpage:

32

206

975

0

2

9

Joey Gonzalez

@profjoeyg

1 year

This week in my video series with @vsreekanti , we had a fun discussion on what's wrong with LLMs and hallucination ... or at least I think we did?

Limitations of LLMs (Generating Conversation, Episode 4)

LLMs are incredibly powerful, but they fail in shocking and hilarious ways from time to time. What's going on? In this episode, we dive into the typical fail...

www.youtube.com

0

3

8

Joey Gonzalez

@profjoeyg

1 year

Integrating open-source LLMs into basic ML workflows is getting to be too easy. Not only can I run an LLM with 1 line of code, @AqueductHQ automates finding and managing the necessary cloud resources and GPUs from a single web interface.

RunLLM

@RunLLM

1 year

Open-source LLMs have taken off recently and have clear advantages over proprietary models (self-hosted, no vendor lock-in, etc.). But deploying and operating them — especially with existing pipelines — is a pain.

1

5

14

0

3

8

Joey Gonzalez

@profjoeyg

8 months

I have some hot takes 🔥 about the limited future of open-source LLMs and the best path forward being in domain specific applications. Checkout the full 🎥:

Jon Krohn 🇺🇦

@JonKrohnLearns

8 months

Our Podcast of the Month features Berkeley professor and brilliant LLM pioneer (behind Vicuna, Chatbot Arena & more) Dr. @profjoeyg . Does he think open-source or commercial LLMs are better? Check out today's infographic! Watch here: Joey: • Is an

0

3

7

2

8

Joey Gonzalez

@profjoeyg

6 months

It’s great to see industry leaders using our leaderboard!

Greg Brockman

@gdb

6 months

GPT-4 Turbo is top of the leaderboard on human preferences (with GPT-4 as #2 ):

72

143

2K

0

8

Joey Gonzalez

@profjoeyg

9 years

The #LearningSys2015 NIPS workshop is soliciting abstracts for research at the intersection of ML and Systems. http://t.co/DlNsYXTfUg

0

3

7

Joey Gonzalez

@profjoeyg

1 year

@matthew_d_green Thanks! We are working on improving Vicuna and releasing other open models @lmsysorg . Also check out the battle of the open LLMs. We hope to release a detailed analysis of the results this Tuesday.

Chat with Open Large Language Models

arena.lmsys.org

0

3

7

Joey Gonzalez

@profjoeyg

1 year

@AmplifyWithAI @lmsysorg We hope to release data soon! We are doing this for science.

0

8

Joey Gonzalez

@profjoeyg

2 years

I am really excited about what we have been building at Aqueduct and the future of Production Data Science (PDS) infrastructure. Let us know what you think!

GitHub - RunLLM/aqueduct: Aqueduct is no longer being maintained. Aqueduct allows you to run LLM...

Aqueduct is no longer being maintained. Aqueduct allows you to run LLM and ML workloads on any cloud infrastructure. - RunLLM/aqueduct

github.com

RunLLM

@RunLLM

2 years

We’ve been working on Aqueduct for over a year, and we’re super excited to share what we’ve been building:

1

7

18

1

7

Joey Gonzalez

@profjoeyg

3 years

Our ActNN work exploring memory-efficient training was just accepted at #ICML2021 as a long presentation! Great work @jianfei_chen , @lm_zheng , @yao_zhewei , and @DequanWang .

Jianfei Chen

@jianfei_chen

3 years

We just open sourced our ActNN library for memory efficient training. It reduces the training memory footprint by compressing the saved activations to 2 bits. It's only a few lines of code in PyTorch, try it!

0

2

5

0

3

8

Joey Gonzalez

@profjoeyg

8 months

Here is the detailed paper describing our PagedAttention research that powers vLLM (and now @nvidia 's TensorRT).

Woosuk Kwon

@woosuk_k

8 months

Exciting news! 🎉Our PagedAttention paper is now up on arXiv! Dive in to learn why it's an indispensable technique for all major LLM serving frameworks. @zhuohan123 and I will present it at @sospconf next month. Blog post: Paper:

2

34

188

2

1

7

Joey Gonzalez

@profjoeyg

1 year

@YiTayML Yeah ... wait which leading institution? 😄

1

7

Joey Gonzalez

@profjoeyg

1 year

LLMs could bring the end of `man`. 🦍 If LLMs could invoke all my shell commands for me, then I would not need to use `man` anymore. @shishirpatil_ and @tianjun_zhang can you all make invoke shell commands.

0

2

7

Joey Gonzalez

@profjoeyg

1 year

This is an excellent overview on the current legal challenges facing Generative AI from one of the leading legal experts. Very accessible. Also, my group got a shoutout from a lawyer (which is hopefully a good thing).

Generative AI Meets Copyright - Pamela Samuelson

Talk Title: “Generative AI Meets Copyright”Speaker: Pamela Samuelson, Richard M. Sherman Distinguished Professor of Law, UC BerkeleyAbstract: The question th...

www.youtube.com

0

3

7

Joey Gonzalez

@profjoeyg

9 years

Some great new GraphLab numbers! I am curious how these compare to the recent results by @Frankmcsherry .

priya joseph

@ayirpelle

9 years

@datoinc 's Yucheng Low with more blowout comparison metrics #datasmt Sgraphs n Sframe ❤️ BSD #opensource Aug http://t.co/y5Rg2xzs7m

1

2

3

2

1

7

Joey Gonzalez

@profjoeyg

2 years

Need GPUs but can't find them or can't afford them? We have a way to reduce the cost of writing ICML papers. Check it out:

Zongheng Yang

@zongheng_yang

2 years

Introducing SkyPilot: Run ML and Data Science jobs on any cloud, with massive cost savings. 🚀 Run jobs on any cloud ⏰ Get GPU/TPU/CPU in 1 click 💵 Reduce > 3x cost Read blog: 🧵1/

11

51

211

0

7

Joey Gonzalez

@profjoeyg

1 year

@Ken_Goldberg @UCBerkeley This is also an opportunity for PhDs who might have been in the tech industry for a few years to get back into fast-paced, high-impact research, and work with amazing graduate students. (Come back to academia!)

0

6

Joey Gonzalez

@profjoeyg

9 months

Check out my newest blog post with @vsreekanti on building your first LLM powered application (and what not to do):

0

2

6

Joey Gonzalez

@profjoeyg

10 months

We just released our analysis of Llama-2 using the very challenging MT-bench tests. It's surprisingly, not nearly as good as GPT-3.5 and Claude.

lmsys.org

@lmsysorg

10 months

How good is Llama 2 Chat? Key insights from our eval: 1. Llama-2 exhibits stronger instruction-following skills, yet still significantly lags behind GPT-3.5/Claude in extraction/coding/math 2. Overly sensitive to safety could cause misinterpretation on user queries 3. Comparable

13

132

534

0

1

6

Joey Gonzalez

@profjoeyg

8 months

We have a new faculty search (all levels) at UC Berkeley for people working at the interface of the computational, statistical, chemical, and physical sciences, including the development and deployment of new materials using artificial intelligence.

0

3

6

Joey Gonzalez

@profjoeyg

10 months

I am looking forward to talking about/with SuperDataScience. Also, I am honored to share this image with a Llama! 😂

Jon Krohn 🇺🇦

@JonKrohnLearns

10 months

Next week, I'm interviewing Dr. @profjoeyg — co-creator of the breakthrough open-source LLM Vicuña (pictured 😎); developer of Apache Spark, Ray, GraphLab; and Berkeley faculty — for a #SuperDataScience episode. Got Qs for him? Joey's episode will likely be #707 and be

0

7

1

0

6

Joey Gonzalez

@profjoeyg

1 year

@jerryjliu0 @gpt_index This is a really cool business application of LLMs! I could imagine this being run automatically to test for potential PII in all data pipelines. @gpt_index we are thinking about building an @AqueductHQ integration, it would be fun to chat more.

1

6