OpenAI's new safety committee

Howdy, wizards.

‍

PSA: 🎈whatplugin.ai is 1 year old this week. A special shout out to everyone who’s been sticking around since the beginning – I truly appreciate you.

‍

Productivity tip: I know ChatGPT’s new voice mode is all the rage, but have you tried the dictation mode on the desktop app? It’s like voice-typing into ChatGPT. I’m using it all the time, and I rarely type to it using my keyboard anymore.

‍

Here’s what’s brewing in AI this week:

1. OpenAI’s new safety committee consists of Altman and other company directors. OpenAI is basically saying capability > safety.
2. Ex OpenAI researcher Jan Leike joins Anthropic. That’s understandable.
3. A new LLM leaderboard has entered the chat. Which LLM habla mejor español?
4. Custom GPTs are now available for free. I’m ditching one of my ChatGPT plus subscriptions.
5. Perplexity introduces Pages. One-click “Wikipedia-pages” about anything.
6. GPTs: top newcomers on whatplugin.ai

‍

Dario’s Picks

1. OpenAI’s new safety committee consists of Altman and other company directors

OpenAI announced, in a short blog post that they’ve formed a new safety and security committee, responsible for making recommendations on critical safety decisions. The new team is led by OpenAI directors Bret Taylor, Adam D’Angelo, Nicole Seligman, and, Sam Altman (surprise surprise). There’s also several other members from within OpenAI.

The company’s “superalignment” team which had a mission to ensure AI doesn’t go haywire was recently abolished after prominent researchers left, essentially citing the company’s disregard for safety as key reasons.

‍

Why it matters

OpenAI’s best safety researchers recently quit saying the company doesn’t take safety seriously. And now the CEO is on the safety committee. I think even GPT-3 would agree that’s a clear conflict of interest.. why wouldn’t they bring in a neutral third-party? It’s pretty clear that OpenAI is taking a firm position as prioritising capability (and profit) over safety concerns.

‍

2. Ex OpenAI researcher Jan Leike joins Anthropic

Ex superalignment lead for OpenAI, Jan Leike, has just joined Anthropic to continue his work with superalignment there. He resigned from OpenAI a couple of weeks back citing disagreement with leadership about the company’s lack of safety focus.

‍

Why it matters

Anthropic was started by Dario Amodei and other ex-OpenAI employees to create an AI company with stronger safety focus. Contrary to OpenAI, that’s a narrative Anthropic has been able to uphold, greatly helped by their emphasis on transparency, way more safety research contributions and not trying to steal Scarlett Johansson’s voice to make Claude sound more seductive.

‍

3. A new LLM leaderboard has entered the chat

‍

There’s a new LLM leaderboard in town by Scale – it’s called SEAL and ranks models based on “private datasets that can’t be gamed”.

SEAL (Safety, Evaluations, and Alignment Lab) was made to address the common problem of biased data in LLM evaluations. Currently, can see the top LLMs for coding, math, instruction following and Spanish.

‍

Why it matters

If you want to know what the best LLM is, you’ve so far mostly had the chatbot arena leaderboard to rely on. They also have leaderboards for different categories now, but now you can cross-check between the two. I admire the simplicity of Scale’s presentation and I especially like that you can click on the task names (e.g. Coding) and you get more background on the data and how they derive the measurement.

‍

4. Custom GPTs are now available for free in ChatGPT

OpenAI rolled out the ability to use custom GPTs for free users last week, as well as advanced data analysis, vision and memory. This means some of the biggest features previously paywalled by OpenAI are now free for all.

The main difference between free and paid is now (until the new voice mode rolls out to everyone) is that you have 5x higher usage caps and you can create custom GPTs. And the desktop app! As I said in the start of the newsletter, the dictate functionality has quickly become indispensable in my daily work.

You can find GPTs to try through the official GPT store or whatplugin.ai.

‍

Why it matters

This is amazing news for all the casual ChatGPT users out there. This week I even cancelled one of my two ChatGPT Plus subscriptions, since you get most of the good stuff for free anyway.

(I immediately spent that extra money on credits for Originality.ai, an AI detection tool. As a bit of an SEO geek, I’ve been keeping tabs on some top publishers on Google for over a year to see if they use AI in their content, and if this changes over time. I’m really happy with the accuracy so far – maybe I’ll share some of my findings later.)

‍

5. Perplexity introduces Pages

‍

Perplexity got a neat new feature this week that’s a bit futuristic. You can now automatically generate “pages” about any subject (they remind me a bit of Wikipedia pages).

Here’s how it works: You input a topic, and it scans the web for relevant info and images, then puts it all together in an nice layout. Citations for info are included, and you can easily edit and add new content.

You can use it to easily share anything you make with perplexity – research, how-to guides, learning material..

‍

Why it matters

Haven’t tried this feature myself yet, but it looks like a great learning tool, you can basically input any subject and instead of dense layers of text you now get a nicely laid out blog article, with pictures and all.. The stuff we can make with a single prompt is getting more advanced by the day.

‍

GPTs

Top newcomers on whatplugin.ai

Highly rated GPTs that made it into whatplugin.ai’s database in the last week. How rankings work.

‍

‍