can't look

I remember when creating apps was the biggest trend. Now, it seems the future lies in developing Agents, LLMs and fine-tuning models. Is this the next big evolution?

Distributed ML framework development is interesting. How should I pursue ?

What a great combo, you should do a weekly show together

I like the part with real world applications, a deeper dive on that would be valuable

Great podcast. Learned a lot.

Ravin and I worked together at sg. One of the brightest mlnds out there, nice to see him on the pod.

Hugo, Always a pleasure to talk with you. And thank you folks in the audience for all the wonderful questions

Isn't this the same as function calling in LLM?

thank you!

The course was amazing, Hamel and Dan were perfect hosts, making the experience both insightful and enjoyable.

One of the best interviews!

what interview / when is this from?

I dont hearing

what is the hamal course? 1:35:46

This is just a gem conversation. Feels like sitting in a nerd talk sofa round, during my math masters ❤

Is there a link to the report? Cheers

I 'm blessed to regularly do AI with the Missus

I really enjoyed this episode. Sébastian is a great person.

0:04:55 Prompt Engineering Fulltime

TIL: A̷t̷t̷e̷n̷t̷i̷o̷n̷ P̶y̶d̶a̶n̶t̶i̶c̶ Rizz is All you Need

Having come to CS from the humanities, I've found Jason unique in his ability to make broad associations between seemingly unrelated concepts, something I've found difficult for many engineers. He's got a knack for language, and that goes really far in working with these new systems in my experience. Love to see it.

TL; DW 00:00 Introduction: Hugo Bounde Anderson, host of Vanishing Gradients, introduces Jason Liu, an independent consultant specializing in RAG applications. The episode aims to uncover how understanding how to build terrible AI systems can lead to building better ones. Hugo promotes his and Jason's upcoming course on LLMs and his upcoming live stream on in-context learning. 04:58 Leveraging Language Models: Jason shares his journey from skepticism to embracing language models. He describes how chatGPT's conversational nature convinced him of the reasoning capabilities of LLMs, moving them beyond mere "toys". 06:25 Delivering Value with LLMs: Jason believes chat interfaces aren't scalable for delivering value. He advocates for document generation, emphasizing how reports that guide decision-making are more impactful than simple question-answering. He showcases real-world applications in fields like investment and due diligence, where the value of generated reports can be measured in millions of dollars. 15:53 Similarities Between RecSys and RAG: Jason highlights the parallels between his experience with recommendation systems and the current landscape of RAG. He sees RAG as a natural extension of his expertise, as it involves similar infrastructure and metrics. He points out that the focus remains on precision and recall, regardless of the specific LLM technology. 17:43 Consulting Playbook for RAG Systems: Jason details his playbook for building RAG systems in organizations. This involves understanding user queries through topic clustering and capability analysis. He emphasizes the need to identify areas where the system is strong and areas needing improvement, using metrics like relevance and user satisfaction. 42:42 Advice for Conducting Experiments: Jason encourages starting simple and learning by suffering the consequences of design decisions. He recommends focusing on specific datasets that pose challenges, rather than easily processed content. He advocates for continuous iteration and micro-experiments, emphasizing that the velocity of learning outweighs the need for perfectly engineered code. 46:16 The Future of AI Engineering Roles: Jason sees AI engineering as a natural evolution of data science and machine learning roles. He emphasizes the importance of quantitative skills, but believes the most valuable skill is "data sense" - the ability to critically assess a system and identify potential issues. 53:57 The Role of Charisma: Jason attributes his charisma to a conscious decision to embrace non-technical skills, particularly public speaking. He emphasizes the importance of consistent practice and believes confidence emerges from doing the work and delivering value. 58:05 Lessons from a Hand Injury: Jason shares how his hand injury challenged his sense of self-worth, which was tied to his ability to work hard. He describes his journey to embrace the concept of "being enough," a turning point that shifted his focus to maximizing leverage rather than simply coding more. 1:04:07 Maintaining Simplicity in AI Systems: Jason suggests prioritizing code that is easily deleted, recognizing that rapid technological advancements necessitate adaptability. He emphasizes the Unix-like philosophy of building modular systems with clear interfaces. He stresses the importance of comprehensive documentation, arguing that a well-documented system is more valuable than one with extensive but unclear code. 1:25:46 Synthetic Data Generation: Jason draws parallels between computer vision and the use of synthetic data in language models. He argues that generating high-quality synthetic data can significantly compensate for limited real data, especially considering the high parameter counts in LLMs. 1:30:45 Reflections on Work and Self-Worth: Jason highlights the concept of "auto-exploitation" in the entrepreneurial world, where the pressure to achieve success is often internalized. He emphasizes the importance of embracing choices and actively deciding what is truly valuable, rather than constantly striving for more. He encourages others to cultivate a sense of "being enough" to avoid self-imposed pressure and to prioritize choices aligned with their values. 1:37:43 Maintaining Mathematical Skills: Jason and Hugo discuss the importance of maintaining math skills for AI practitioners. They argue that a focus on data problems will naturally lead to the development and application of necessary math concepts. Jason highlights the importance of conceptual understanding and problem-solving skills over memorizing complex formulas. He believes practicing these skills in everyday work is sufficient.

Nice, found this video from the great LLM fine-tuning course

0:14:00 Structured Thinking

this was an excellent talk. thanks to all the participants.

Next time Hugo: don't interrupt so much!

Are they using llm for their unit tests? With llm output, how do you unit test since you are not completely in charge of the unit?

'unix like philosophy' -- I like it

Great video

Thanks to all of you for sharing your wisdom, experience and insights. Much appreciated from someone currently attempting to make decisions regarding the future direction of an AI app.

ulmfit not getting enough credit publicly! definitely the forefather of modern NLP

actually there are quite a lot have changed after LLM, although the principles are the same. Jeremy sounds like many very 'senior' AI practitioners, quite pessimistic, maybe because they feel (maybe unconsciously) left behind? it has just been just one year and we can see the unprecedented change already

Always love to listen to Hamel share his experiences. One of the best there is when it comes to practical advice!

✔️ "promosm"

promo sm

Thanks so much for these vids. I subscribe to the podcast but idkw, always find videos more engaging. Really loving your content, I found you due to DataCamp!

The best part of this is living through the cringe of realizing how many times I misspoke. It keeps me humble. <bangs head on desk>