Co-founder quits OpenAI

Welcome to the 387 new subscribers who joined in the last week.

‍

Poll results show many of you want more emails from me to keep up to date on the latest developments. From now on I'll default to twice a week, but I'll only do once a week if news are slow, to avoid spamming you.

‍

What’s brewing in AI this week:

• OpenAI co-founder Andrej Karpathy quits
• Poll: ChatGPT vs Gemini
• A couple of cool GPTs
• The other top stories in AI this week

‍

Dario’s Picks:

I. Andrej Karpathy quits OpenAI

‍

‍

OpenAI’s founding member Andrej Karpathy (previously AI director at Tesla) announced on X that he already left the company, with a plan to work on personal projects. He emphasised that no issues or drama led to his departure.

Why it matters: Andrej was a respected research scientist at OpenAI, and widely considered one of the brightest people in the world of AI. He says there were no issue leading to him quitting, but it’s worth noting that chief scientist Ilya Sutskever’s currently has an unclear status at the company, with some sources saying he has become invisible. Wouldn’t call it a conspiracy to suggest that there seems to some tensions or turbulence in the research division of OpenAI.

‍

II. OpenAI is testing cross-chat memory in ChatGPT

‍

‍

OpenAI is testing cross-chat memory in ChatGPT for a small group of users, allowing it to remember details across all conversations. As you chat, it starts to pick up details by itself, and will get better over time. Users can edit or clear the memory, similar to browser history, and also interact with this feature conversationally ("remember this" or "forget that").

Why it matters: The idea of ChatGPT remembering stuff about you to make it more useful is alluring. I’m a bit sceptical about how useful this feature will be in practice though. I’ve found that leaving details about me in custom instructions comes in handy only in very specific cases, mostly in terms of instructions/formatting of the output. I’ve also experienced certain details I’ve entered in custom instructions as counterproductive, because it evokes the memory when it’s not relevant or you’re looking to explore something new.

While I definitely believe we’ll have AIs that remember every detail about us soon enough, I’m curious if this new feature will be able to identify the right information to save, and being able to use it without over relying on it.

‍

III. OpenAI is reportedly developing AI agents that can do things without a prompt.

OpenAI is said to be working on new products that not only generate text, but can actually automate your workflow. Things like transferring data from a document to a spreadsheet, booking flights and hotels, e-mail a group of people. This could potentially automate stuff you do, particularly web-based tasks, that normally requires you to click around, move the cursor, type text, etc.

Why it matters: AI agents that can handle workflow all by themselves could really change how we work with AI. Imagine if you were able to not only generate a travel itinerary with AI, but actually have it book your next vacation based on your criteria. With this type of agents, OpenAI is delving into similar territory as Rabbit and Hypewrite, focusing on AIs that can take action in the real world.

‍

IV. Nvidia’s Chat with RTX

Nvidia’s new demo app, Chat with RTX, makes it easy to run an LLM (Mistral or Llama 2) locally on your PC that can dig through docs, notes, videos, or other data. It’s faster and more secure than doing it in the cloud.

Note. You need Windows and a RTX 30- or 40-series GPU with at least 8GB of VRAM in order to run it.

Why it matters: This seems like a simple but useful idea. Haven’t tested it myself as I use iOS, but as the Verge pointed out, it can be a valuable tool for researchers and journalists needing to securely and rapidly analyse a collection of documents. It can also handle YouTube videos, but this is less impressive to me as it isn’t generally a privacy-demanding task, and there’s already a range of GPTs that can readily do the same.

‍

In Focus

Most of you are ChatGPT plus users, but I reckon quite a few of you have already tried Gemini Ultra as well. Let’s compare Gemini Ultra to GPT-4.

‍

*SEE THE EMAIL VERSION OF THIS NEWSLETTER TO VOTE*

‍

My experience so far: Gemini Ultra is faster and has no usage cap, with Google services integrations as an advantage (although I haven’t personally found much use in it yet). However, it often struggles with simple tasks, forgets chat context faster than GPT-4, and generates inferior images compared to DALL-E and Midjourney. GPT-4 is still my go-to.

Gemini Ultra isn’t yet on Huggin Face’s LMSys leaderboard, but its little brother Gemini Pro ranks 3rd, behind GPT-4 Turbo models. It will be interesting to see where it Ultra ranks when it eventually appears.

‍

GPTs

Editor’s Choice

Weekly picks from me to you

‍

AutoGPT Agent

💬10K+

Plans, researches, strategizes, and works to complete tasks semi-autonomously using multi-modal tools, enabling completion of tasks with minimal user input.

‍

‍

❞ The name is a bit exaggerated – it won’t go and automate your whole workflow – but it’s still a great GPT. I’ve tested it versus default GPT-4 on things like sentiment analysis of stocks and some market research tasks. It did a notably better job at giving a structured, thoroughly researched answer.

‍

Featured GPT (Sponsored)

RPG Game Creator

💬400+

Crafts an RPG from your prompt and generates a link to play the game.

‍

‍

❞ It generates an 8-bit Final Fantasy style game based on your prompt. You can also make sequels to your games based on existing games. It didn’t completely adhere to my prompt (doesn’t look futuristic, and no robots), but I give it two thumbs up for innovation.

‍

Bytes

General

• Eleven Labs’ new voice actor payout program: Turn your voice into passive income in 30 minutes.
• Microsoft Copilot has a new look and improved image generation features.
• WSJ did an article on early adopters expressing disappointment with Microsoft Copilot, questioning how much value it adds. Transcribing meetings on Teams is the most useful feature.
• AI was in the spotlight during the Super Bowl last week, with ads for Microsoft Copilot among others. Check out the AI-related ads here.

Legislation

• US Government restricts calls using AI voices. The law gives state attorneys general power to prosecute robocallers using AI voices, regardless of whether the call is deceptive or annoying. The exception is if the recipient of the call has given prior consent or in the case of an emergency.

Research

• Interesting, and a little unsettling: ¼ Americans are using AI to create content and/or photos when dating online. 39% of guys using AI to help them write a Valentines message. 69% says AI content gave them better responses.
• A recent study found that ChatGPT demonstrates higher rationality scores than humans in decision making, including when contemplating two options. GPT could have the potential in assisting human decision-making.
• While most U.S. adults say AI's risks outweigh its benefits, Gen-Z (particularly college students and fresh graduates) think it can give them a career edge.

‍