Q&A with a "cultist"? And APAC local models

What a week! 😉

This newsletter has-
Podcast link
Updates/Resources

wanna stay together?

Teaser clip for new YT video
Weekly standup

Speed-round Interview with Devansh

Video on fine-tuning

What we’ve learned🤓

The recent podcast with Alex Atallah, Founder of OpenRouter  The pod is just 10 minutes long, making it an easy and quick listen (and even easier to follow and rate it ;)

Also available on Youtube

New things

Video coming soon on using Jan - why locally run models are neat.

There’s now a living doc/resource page (soon to be integrated with a form/Notion). Troubleshooting guides, software recommendations and links to helpful folks for local AI. (Call local LLM 101).

There’s also a repo list for useful local ai links.

We kicked off a weekly standup over on Linkedin of all places

(Check in on Mondays to catch up with what folks are working on/ need help with/can help with)

Quick news🗞️(feel up to date, without pressure)

New tool to tackle hallucination in multi-modal  (meaning more accurate results)

OpenAI is adding watermarks to DALLE

Q&A with Devansh

Last week, I mentioned one of my favourite writers- Devansh.

Here’s a speed round Q&A with him.

Renee-What does multimodality mean for “normal people” (your tiktok on embeddings was great)

Devansh-Are you asking me for the definitions or what implications it has? Either way, the answer is the same for both normal people and AI Folk.

Definition- Most AI systems are limited by the data they can ingest. An image classification algorithm will only be able to process images, text gen only gets text etc. This reduces the potential for a system to be deployed autonomously in the world. Self-driving, automatic delivery, and other robotics use cases for eg would need to be able to consider multiple facets to make a decision. Multi-modality is a way to train your AI to consider multiple data streams/formats to make decisions

What it implies- Multi-modality enables the deployment of systems that are much more capable than their ancestors. Having more data modalities is like adding new new features to traditional ML (but on crack).

PS- You might be the first person ever to watch my TikToks. Thank you <3

Renee- What’s something people get wrong about the AI space in your opinion?

Devansh- The entire risk conversation is so skewed (as I’ve covered on multiple occasions). Most recently, it’s my observation that people are treating AI as magic which makes things worse for everyone.

Renee- Who would you interview in OS or AI if you could (they can be dead, that’s allowed)

Devansh- Depending on how you’re willing to stretch AI, it would have to be Fermat. I have to know what his proof for the last theorem was.

In AI- Sebastian Risi. His work on self-organizing AI is beautiful.

Renee-What's the backstory of the "chocolate milk cult"?

Devansh-I wanted to get a PhD, but I didn’t to get a Master’s Degree. I decided that writing about research would help me make a connection with professors. Turns out that I really don’t like academia (this dislike is mutual) and that certain people love my writing. So I decided to hitch my horses to this wagon.

If you’re asking about the name it comes from two or three things. In university, I was accused of starting a mini-cult (This is not true, I was trying to start a religion for tax benefits + some other things). That “cult” is one of the few true positives from my college life. Hence the names, and the occasional reference to “being woke” etc.

Chocolate Milk just because I love it. I drink mostly plain milk. The chocolate milk is a treat that I have for myself. As a way of getting myself more consistent with the writing, I would drink it while writing. Hence my writing is the chocolate milk cult.

Find Devansh here

A theme we’ll be exploring in the coming episodes is training your own personal GPT- here’s a video from “AI Jason”

AI Jason (I speak about him on the pod- another Australian!)

Wanna keep the party going?

Thank you for reading 🤘 (un)supervised learning — your support allows me to keep doing this work.

How to participate

Share 🤘 (un)supervised learning. When you use the referral link below, or the “Share” button on any post, you'll get credit for any new subscribers.

Earn benefits. When more friends (or enemies 😈) use your referral link to subscribe (free or paid), you’ll be on your way to…

• Shoutout on the pod for 5 referrals

• Get Wall of Love for 10 referrals

• Get an(un)Supervised Learning Tshirt for 20 referrals (wow)

  

  Tune in to the podcast next week for a chat with ML Engineer from Hugging Face on developer empathy.

  
  PS- what’s your best Dad joke?