Видео

ChatGPT summary: This live chat during an AI/ML discussion reflects a global audience sharing their backgrounds and interests in machine learning, data science, and AI applications. Participants from various cities, including Sydney, Santiago, Auckland, Chicago, and others, greeted each other and briefly introduced their professional roles, such as ML engineers, data scientists, and students. Key discussion points include: 1. ML and AI Applications: Participants discussed using AI for practical applications like identifying waste streams, forecasting, and structured data extraction. There was interest in integrating AI/ML models with client data and improving model monitoring and retraining processes. 2. Challenges in AI/ML: The chat touched on issues such as the difficulty of solving nonlinear problems with current AI/ML models and the limitations of large language models (LLMs). Some expressed concerns about the hype around AI/ML and the need for a deeper understanding of the mathematical and engineering fundamentals behind these technologies. 3. LLM and MLOps: The conversation covered the roles of LLMs in AI, with some viewing them as advanced librarians that help summarize and connect knowledge. There was a request to spend more time discussing MLOps and LLM Ops, indicating a need for practical insights into managing and optimizing ML operations. 4. Learning and Career Development: Several participants sought advice on learning AI/ML from end to end, checking their knowledge, and exploring the benefits of freelancing versus full-time employment in the field. Resources and course recommendations were shared, including links to relevant materials and platforms for further education. 5. Neurodiversity and Creativity: One participant highlighted the correlation between creativity and neurodiversity, suggesting that real breakthroughs in AI may come from individuals who think differently, rather than following typical patterns. Overall, the chat was a dynamic exchange of ideas, with participants expressing both enthusiasm and caution about the current state of AI/ML and its future directions. Thanks to All!

Fantastic chat. I always enjoy listening about and reading Chip's work.

Incredible chat, folks! Lots to learn. Thanks

thanks guys wasnt expecting seeing Hugo here :D

It would be helpful to start by explaining why we need a runner

Is there any where we can see the repo for the ai-examples with pixi? Whisper is very interesting and we would love to see how that was set up as a good starting point for us if that is something that is public of course

Great Data/ML Platform recipe Jon and Eric! Can't wait to bring it to my kitchen to try it on :)

where is the link ?

I love listening to Simon. So dear host, please let the guest talk and find a way to drive the conversation without interrupting them so much.

Hello Hugo, is this event in-person ONLY? If yes, where exactly?

Will registrants get access to a video copy after the fact?

why the link is not working ? :(

Way to go Eric!!

Every time I hear Simon talk I get super excited and sometimes I don't even know what I'm excited about. But, I just know something has happened and that my spare time over the next week is accounted. Simon rocks!!!

George Hotz is looking to build a machine + software that you can buy for the price of a small car, run off house electricity and use to train something like Llama from scratch. TinyGrad and TinyBox.

In the past two months - Mark Zuckerburg said the next Llama will be state of the art. GPU Sales also show Meta in the lead, tied with MS.

Why every start up losses ideas, 100 % commerce, return my money from dalle 2, if not go f**k yourself closedAI

Totally closed org, they're threatened that someone others will achieve the same level of AI development. The revolution is coming soon. I highly prob that my opinion here will vanish, but revolution is coming.

❄️💙🎄😊

Emo-Robo-Whiney-Country-Data-Core 🤖

This was fantastic. A lot to learn from Jeremy

1:02:00 I like this argument: some people love having more power than others...they really dedicate themselves to it...Ensuring openness and democracy allow the rest of us to invest in defense...cos its much easier to work on defense than attack...and stop the attack path when such people go crazy...

You can't have an LLM just randomly generating text about a product, there are laws that govern that

As of the last few weeks ChatGPT with GPT4 has been garbage. It just like when you google a question and get a vague, off topic list of possibilities. It used to be specific.

the guy just hates open ai

Of course certain people want to lobotomize AI. They cannot let the general population have such a truth oracle. AI could set is free, and that can never be. When they say AI will destroy the world, they mean THEIR world. It's their status-quo they are worried about.

Could i run mistral 7B on raspberry pie ?

37:37 mistral 7B

was there a name for that technique jeremy mentioned for the good model training the "shit" model?

Q*anon I'm dying 😂

Share Spotify link, it's trouble to hear while traveling

what a dystopian vision

I listened with interest, and Jeremy is an actual expert. But I am not persuaded that superintelligent AI is far away, or that an AI on every desktop offers much protection. If anything, AI for all is going to accelerate the discovery of architectures for efficient learning and agentic AI, and that's just getting closer and closer to the conditions for successfully self-enhancing AI. And superintelligent AI seems likely to overwhelm humanity even if we are assisted by AIs of human intelligence. So we still need something like "superalignment". Futhermore, even if superintelligent AI that surpasses humans-with-AI is somehow chimerical, won't humans-with-AI overwhelm humans-without-AI? And I also note that democratic processes seem to have a very small part in determining the outcome of humanity's AI adventure. AI is driven forward by money and power now, decentralization depends on the boldness and ingenuity of maverick techies, and either way, the whole thing is an enormous experiment in creating nonbiological thinking beings, potentially giving away humanity's chief evolutionary advantage to entities that aren't even DNA-based.

Hi, really enjoyed the talk. Does Jeremy has an X account I could follow? Thanks

Is that a Tim tam Hugo was eating in the beginning 😂?

Great talk mostly though the bellyaching was a bit cringe….

Starting my own AAA, any hot tips?

I got lost from where you found the Jupyter notebook page. It suddenly appeared in the video from somewhere. Didn't find it so the rest of the video was mute for me.

Very good workshop! Appreciate y'all uploading this

🎯 Key Takeaways for quick navigation: 00:00 🎉 *Introduction and Background* - Ville Tuulos introduces the topic of building production systems with generative AI and data science. - He discusses the evolution of online shopping experiences, emphasizing the need for more human-like interactions in e-commerce. - Ville highlights the potential of AI and generative models in transforming user experiences. 05:10 🛋️ *Reimagining Online Shopping* - The shift towards more immersive and human-like online shopping experiences. - Exploring the possibilities of interactive product visualization and personalization. - The importance of leveraging generative AI to create unique and engaging customer interactions. 09:03 🤖 *Building Sophisticated AI Systems* - The technical challenges involved in creating AI-driven e-commerce experiences. - Emphasis on the need for custom AI solutions rather than relying on off-the-shelf models. - The importance of data governance and quality for AI applications. 12:05 📈 *Maximizing Revenue and User Experience* - Strategies for optimizing revenue through AI-driven product recommendations and marketing. - The shift of content generation responsibility from non-technical departments to AI engineers. - The growing complexity of business operations and the role of AI in different domains. 16:47 🛠️ *Challenges and Infrastructure* - The realization that AI and generative systems pose complex challenges for various projects. - The need for common infrastructure and tooling to support multiple AI initiatives. - Practical considerations for scaling AI development across different domains and industries. 19:05 📜 *Building Production Systems within Existing Frameworks* - Adapting to privacy laws and compliance requirements. - Navigating the challenge of increased computational demands in modern systems. - The fragmentation and heterogeneity of compute resources in today's landscape. 20:00 🔄 *Adapting to Diverse Compute Needs* - The absence of a one-size-fits-all approach for compute resources. - The need for CPUs, GPUs, distributed training, and specialized hardware. - The evolving complexity of compute requirements. 21:11 🛠️ *Orchestrating Interrelated Components* - The importance of orchestrating multiple interrelated components in real-world systems. - Examples of component orchestration in tasks like inventory forecasting. - Managing complex pipelines involving data processing, feature engineering, training, and more. 22:10 🔄 *Iterative Development and Tracking* - Emphasizing the need for iterative development in building production systems. - Tracking code, models, and data changes over time. - The importance of continuous improvement in AI and ML projects. 23:18 🚀 *Operating AI Systems in Production* - Challenges of operating large-scale AI systems in production. - Connecting systems to upstream and downstream services and databases. - The need for robust observability layers for distributed systems. 24:17 📦 *Managing Software Supply Chain Complexity* - Addressing the increasing complexity of the software supply chain in ML and AI. - Challenges with dependency management in a rapidly evolving library ecosystem. - The limitations of traditional containerization in handling dynamic libraries. 25:26 🔄 *Unifying Concerns with Frameworks* - The need for unified frameworks to address concerns in ML and AI system development. - Drawing parallels with the evolution of full-stack engineering in the JavaScript ecosystem. - The potential for frameworks to simplify the development of AI-driven systems. 26:49 🌐 *Resources and Announcements* - Encouraging exploration of Metaflow and its relevance to AI system development. - Mentioning upcoming announcements related to distributed training and integration. - Highlighting the active open-source community around Metaflow. 27:33 📢 *Call to Action* - Inviting the audience to start building AI-driven systems. - Offering assistance and access to open-source resources. - Emphasizing the role of AI in optimizing business processes across various departments. Made with HARPA AI

This is a fantastic session. I would love to see Bindu Reddy take part in upcoming panel discussions.

Great! My reaction was: this would work great if a run only produced one model, but if I remember correctly, the recommended pattern for hyper parameter tuning is to have one flow that fans out and trains many models. At that point you might be the sort of group that goes for a dedicated model registry tool like MLFlow, CometML or SageMaker registry, but is there a way to make tagging in multi model runs can work natively in Metaflow?

"Promosm" 💦

from Austin Texas

Incredible conversation, folks! Thanks

I am Olaniyan remilekun from Nigeria

Hey is there any javascript integration

Looks great, can you share please a video how to configure everything? For me thats currently a big hussle, and I don't know which parameter to set. Will be a big help for using metaflow.

Nice please let me know the VS Code theme that is used

Great tutorial, links and program worked fine for me. Thanks for sharing.