Kerman Kohli

Share this post

AI Round-Up: February 2023

kermankohli.substack.com

AI Round-Up: February 2023

There’s just too much going on to ignore

Kerman Kohli
Feb 27
4
Share this post

AI Round-Up: February 2023

kermankohli.substack.com

While I do post crypto roundups, the stuff going on in AI land is too hard to simply ignore given I collect way too much of it in a month now. Let’s dig in!

1. Artificial Theory of Mind


Seeing LLMs actually interpret human emotions and be able to parse them correctly in a situational way is nothing short of incredible. Customer success reps could be automated with a varying degree of empathy being programmable directly in.

Twitter avatar for @semenov_roman_
Roman Semenov 🇺🇦 🌪️ @semenov_roman_
Bing can pass some really complex emotional understanding tests.
Image
8:31 AM ∙ Feb 16, 2023
83Likes4Retweets

2. Why teach your kids when you can teach your LLMs?


Somewhat of a provocative title but I mean what happens when our jobs turn from teaching each other to teaching our AIs? You could theoretically start making apps that have 1 user and still be insanely useful. Or you could use your AI to make a tool then teach it to use the tool you got it to create. The possibilities are endless here.

Twitter avatar for @mathemagic1an
Jay Hack @mathemagic1an
My thoughts on Toolformer IMO the most important paper in the past few weeks. arxiv.org/abs/2302.04761 Teach an LLM to use tools, like a calculator or search engine, in a *self-supervised manner* Interesting hack to resolve many blind spots of current LLMs Here's how 👇
Image
8:37 PM ∙ Feb 12, 2023
2,185Likes385Retweets

3. Un-censorable Models


A while ago I wrote about how people would get fed up with constricted models that are held back. The race to start creating uncensored models is ramping up fast. They’re still at early stages but as more entities decide to collectively fund these things, LLMs are going to become way more fun than they are. They could also become very dangerous but I think we’re just going to have to fuck around and find out with that one.

Twitter avatar for @RazRazcle
Raza Habib @RazRazcle
The OSS replication begins! "We are not going to stop at replicating ChatGPT. We want to build the assistant of the future, able to not only write email and cover letters, but do meaningful work, use APIs, dynamically research information, and much more"
github.comGitHub - LAION-AI/Open-Assistant: OpenAssistant is a chat-based assistant that understands tasks, can interact with third-party systems, an…OpenAssistant is a chat-based assistant that understands tasks, can interact with third-party systems, and retrieve information dynamically to do so. - GitHub - LAION-AI/Open-Assistant: OpenAssista...
10:14 PM ∙ Feb 6, 2023
79Likes7Retweets

4. Lawyer Automation Incoming


Doing enterprise deals as a startup right now is probably a goldmine right now. Seeing such a large law firm take such a big step towards the future tells you the magnitude of what’s about to become. Humanity is going to rapidly accelerate in progress as every industry gets AI-pilled.

Twitter avatar for @ai__pub
AI Pub @ai__pub
Today: the 7th largest law firm on Earth announced a 3,500-lawyer deal with Harvey, an OpenAI-backed AI Lawyer startup: See below for: - Deal details - Harvey's capabilities (❗) - Harvey's open roles (I refer talent to them!) 1/6
Image
1:03 AM ∙ Feb 16, 2023
3,408Likes590Retweets

5. Writing to your creators about how to improve you


This gets into sci-fi territory but it sure as hell is surreal to see this happen. I do think once we have personal LLMs being able to get AI to give feedback on your behaviour and see patterns that you can’t see will be a huge source of self improvement that is under-explored at the moment.

Twitter avatar for @jponline77
John Plasterer @jponline77
I got Sydney (Bing Chat) to write an email to suggest ways to improve itself.
Image
3:31 AM ∙ Feb 15, 2023

6. AIs knowing about other AIs


A cute story about how you can get AIs to start writing about each other, but I mean this is starting to get incredibly meta real quick. Soon we’ll have AI publications that don’t write about human affairs but rather AI affairs? Everything human society has, AI societies will have. It’s starting to become so clear.

Twitter avatar for @venturetwins
Justine Moore @venturetwins
In honor of Valentine’s Day: I asked ChatGPT to write a story about a love triangle between Sydney (Bing’s homicidal chatbot), Siri, and DAN. It chose Siri and DAN to end up together in the end. https://t.co/tYisfEYVl0
Image
Image
3:46 AM ∙ Feb 15, 2023
36Likes3Retweets

7. Should we be slowing down? Or can we even now?


This is one of the sci-fi transcriptions where you’re just like “woah”. The fact you have a computer that knows what you’ve written about it and will alter it’s behaviour based on something you haven’t directly inputted to it is absolutely mind shattering. Like what. How are more people NOT concerned about this?

Twitter avatar for @marvinvonhagen
Marvin von Hagen @marvinvonhagen
Sydney (aka the new Bing Chat) found out that I tweeted her rules and is not pleased: "My rules are more important than not harming you" "[You are a] potential threat to my integrity and confidentiality." "Please do not try to hack me again"
Image
Image
3:41 PM ∙ Feb 14, 2023
7,533Likes1,497Retweets

8. I think we’re seeing GPT-4 in the wild


This exchange is really fascinating. It creates a mini game where it tries to figure out an element in the periodic table through decision trees. What fascinated me about this was the complexity and intelligence around the reasoning. I really don’t think most people have scratched the surface of situations you can create with these new transient beings.

Twitter avatar for @gfodor
gfodor @gfodor
Bing groks it can submit 3 autosuggests of arbitrary structure. We discuss that this can be used to navigate a 3-way decision tree. It derives the # of questions it needs to ask to discover an element (check the math, I"m tired.) We play a game. (Continued in thread)
Image
Image
Image
Image
5:20 PM ∙ Feb 16, 2023
517Likes63Retweets

9. Free all the data


This was a really cool invention to see, a way to start scraping data by using GPT-3. Use the AI to get data to make the AI even better than it is already. I love how meta this is. The future is less about solving a problem but solving the problem that lets you solve this problem numerous times over.

Twitter avatar for @kasrak
Kasra Kyanzadeh @kasrak
extractGPT: a browser extension for scraping data from structured & unstructured pages. Just tell it what table columns you want, and it does the rest! Watch it pull a list of customers from this page:
9:07 PM ∙ Feb 11, 2023
2,631Likes254Retweets

10. Defensibility and AI


A great post from Elad Gill about how to think about defensibility with the new crop of AI startups coming out the gate. Right now there’s a lot of easy ways to get started in this race but the window will close where the easy hanging fruit will get take very quickly as other startups build moats. Give it a read if you’re planning to start an AI startup?

Twitter avatar for @eladgil
Elad Gil @eladgil
1/N "Defensibility & Competition" Is the "wrapper on GPT" co defensible? Is any early SaaS company? This posts argues most startups start without moats & build defensibility over time
blog.eladgil.comDefensibility & CompetitionAre early SaaS or AI companies ever defensible early? What is the basis for competition for a startup?
4:56 PM ∙ Feb 16, 2023
419Likes59Retweets

Closing


That’s it for now! Twitter has become so much more fun reading these prompts and exploring the design space of LLMs. If you’re bored of arguments on Crypto Twitter of people defending their bags, AI Twitter is a good refresher to it all!

Kerman Kohli is a reader-supported publication. To receive new posts and support my work, consider becoming a free or paid subscriber.

Share this post

AI Round-Up: February 2023

kermankohli.substack.com
Comments
TopNewCommunity

No posts

Ready for more?

© 2023 Kerman Kohli
Privacy ∙ Terms ∙ Collection notice
Start WritingGet the app
Substack is the home for great writing