🔥 ThursdAI Sunday special - Deep dives into Crew AI with Joao then a tasty Bagel discussion with Jon Durbin

ThursdAI - Sunday special deep dive, interviews with Joao, and Jon, AI agent Crews and Bagel Merges.

Happy Sunday dear reader,

As you know by now, ThursdAI pod is not a standard interview based podcast, we don't focus on a 1:1 guest/host conversation, but from time to time we do! And this week I was very lucky to have one invited guest and one surprise guest, and I'm very happy to bring you both these conversations today.

Get your Crew together - interview with João Moura, creator of CrewAI

We'll first hear from João Moura, the creator of Crew AI, the latest agent framework. João is a director of AI eng. at Clearbit (acquired by Hubspot recently) and created Crew AI for himself, to automate many of the things he didn't want to keep doing, for example, post more on Linkedin.

Crew has been getting a lot of engagement lately, and we go into the conversation about it with João, it's been trending #1 on Github, and received #2 product of the day when Chris Messina hunted this (to João's complete surprise) on Product Hunt.

CrewAI is built on top of Langchain, and is an agent framework, focusing on Orchestration or role-playing, autonomous agents.

In our chat with João we go into the inspiration, the technical challenges and the success of CrewAI so far, how maintenance for crew is now partly a family effort and what's next for crew

Merges and Bagels - chat with Jon Durbin about Bagel, DPO and merging

The second part of today's pod was a conversation with Jon Durbin, a self described AI tinkerer and software engineer. Jon is a Sr. applied AI researcher at Convai, and is well known in our AI circles as a master finetuner and dataset curator.

This interview was not scheduled, but I'm very happy it happened! If you've been following along with the AI / Finetuning space, Jon's Airoboros dataset and set of models have been often mentioned, and cited, and Jon's latest work on the Bagel models took the lead on HuggingFace open LLM leaderboard

So when I mentioned on X (as I often do) that I'm going to mention this on ThursdAI, Jon came up to the space and we had a great conversation, in which he shared a LOT of deep insights into finetuning, DPO (Direct Preference Optimizations) and merging.

The series of Bagel dataset and models, was inspired by the Everything Everywhere All at Once movie (which is a great movie, watch it if you haven't!) and is alluding to, Jon trying to throw as many datasets together as he could, but not only datasets!

There has been a lot of interest in merging models recently, specifically many folks are using MergeKit to merge models with other models (and often a model with itself) to create larger/better models, without additional training or GPU requirements. This is solely an engineering thing, some call it frankensteining, some frankenmerging.

If you want to learn about Merging, Maxime Labonne (the author of Phixtral) has co-authored a great deep-dive on Huggingface blog, it's a great resource to quickly get up to speed

So given the merging excitement, Jon has set out to create a model that can be an incredible merge base, many models are using different prompt techniques, and Jon has tried to cover as many as possible. Jon also released a few versions of Bagel models, DPO and non DPO, that and we had a brief conversation about why the DPO versions are more factual and better at math, but not great for Role Playing (which is unsurprisingly what many agents are using these models for) or creative writing. The answer is, as always, dataset mix!

I learned a TON from this brief conversation with Jon, and if you're interested in the incredible range of techniques in the Open Source LLM world, DPO and Merging are definitely at the forefront of this space right now, and Jon is just at the cross-roads of them, so definitely worth a listen and I hope to get Jon to say more and learn more in future episodes so stay tuned!

So I'm in San Francisco, again...

As I've mentioned on the previous newsletter, I was invited to step in for a colleauge and fly to SF to help co-host a hack-a-thon with friends from TogetherCompute, Langchain, in AGI house in Hillsborough CA. The Hackathon was under the Finetune VS RAG theme, because, well, we don't know what works better, and for what purpose.

The keynote speaker was Tri Dao, Chief Scientist @ Together and the creator of Flash Attention, who talked about SSM, State space models and Mamba.

Harrison from Langchain gave a talk with a deepdive into 5 techniques for knowledge assistants, starting with basic RAG and going all the way to agents 👏

I also gave a talk, but, I couldn't record a cool gif like this for myself, but thanks to Lizzy I got a pic as well 🙂 Here is the link to my slides if interesting (SLIDES)

More than 150 hackers got together to try and find this out, and it was quite a blast for me to participate and meet many of the folks hacking, hear what they worked on, what worked, what didn't, and how they used WandB, Together and Langchain to achieve some of the incredible results they hacked together in a very short time.

The projects showcased a range of creative applications leveraging RAG, finetuning, and other large language models. Several projects like Magic RAG, CareerNavigator-AI, and CompetitionAI used RAG for document retrieval and knowledge enhancement. Others like rags2pizza and Naturalist DALL-E focused more on finetuning models for specific domains. Some projects compared finetuning and RAG, finding that combining both gives superior performance over using either alone but that result wasn't conclusive.

My vote as a judge (which I did not expect to be) eventually went to the team that built the OptiMUS project, they had generated a systentic dataset, cleaned it up, finetuned a model on it, and showed that they want to optimize AI agents. They used WandB to track their work and I hope they take this project forward and keep making advancements in AI. Congrats for the win Ali and Shayan, hope you enjoy the WandB branded Airpods (even I don't have those) and the Meta Quest, well deserved!

Thank you for tuning in! See you next week!

Full Transcription :

[00:00:00] Alex Volkov: Hi. Welcome back to Thursday. The Sunday special episode. This is Alex Volkov. And I'm recording this in. A gorgeous space. In San Francisco. Where I was. Invited to judge hackathon. And now I'm hanging out with a few friends from cerebral valley. So thank you. Valley folks. For letting me use this place for recording and Today, we have a special episode for you. As If you hear this on Sunday. Today's not a Thursday. We often times have special guests on the pod. Where conversations. Or deeper.

[00:00:45] Alex Volkov: And usually I reserve that slot for a Sunday special release. So this is what you're hearing now. In today's episode, we actually have two conversations. Although I only planned on one. And the first part is the planned part that you hear from Joao Maura. He is a director of AI in Clearbit, and now acquired by HubSpot. And he's also the creator of Crew AI and the Gentek AI framework that can run. By orchestrating.

[00:01:14] Alex Volkov: c

[00:01:15] Alex Volkov: The digital AI agents and have them work together.

[00:01:19] Alex Volkov: And I think you'll hear from, Joao why this peaked interest. For many folks. Specifically. Because as we caught up with. Wow.

[00:01:29] Alex Volkov: Crew AI was trending on GitHub and getting number two on product hunt at the same time. And it's a really cool framework. And I think the underlying. Power of this is that it can use open source, local models. A lot of previous agent attempts used GPT4 For example, and the crew AI can use things like Mistral or Mixtral running in LM studio or Ollama on your Mac, which I think is super cool.

[00:01:55] Alex Volkov: And I think on device AI, plus something like this framework is going to be very, very powerful. It was a great conversation was wow. And surprising to me, the second guest was not planned. However you may have heard from the …