šŸ˜ŗ OpenAI is so back

PLUS: this chatbot can summarize 150 PAGES of context

Welcome, humans.

We hope you had a wonderful Thanksgiving. We're immensely grateful for each of you, and also for ChatGPT, which helped us skillfully sidestep every political debate that arose while we stuffed our faces with stuffing.

Hereā€™s what you need to know about AI today:

  • Sam Altman is BACK at OpenAI and two board members were replaced.

  • Claude 2.1 is here with a 200K context windows and 2x less hallucinations.

  • Inflection announced an upgraded model to power Pi.

  • Grok will soon be available to Twitter Premium+ subscribers.

Sam & Greg are BACK at OpenAI.

when Helen Toner gets locked out of the OpenAI office

Itā€™s official: Sam Altman is rejoining OpenAI as CEO, with a revamped board feat. Bret Taylor (former CTO of Facebook, co-CEO of Salesforce), Larry Summer (former Secretary of the Treasury/Harvard President), and Adam Dā€™Angelo (CEO, Quora/initial insurgent).

(the new boardā€™s first job is to select a new board of 9 people).

Thus concludes 5 intense days of Twitter drama, late-night boba orders, and heinous amounts of chicken McNuggets.

Recently, some insights dropped about potential sources of the internal strife:

The board conflict was partially due to ex-board member Helen Toner publishing a paper that painted OpenAI in a less-than-favorable light.

But the deeper cause likely centers around how to navigate Q*, a new OpenAI project that could be a scary big leap in AI.

We donā€™t know the technical details of Q* (nobody does), but it could be a huge unlock:

  • Todayā€™s AI models are like 2nd graders telling a story. Chatty, but they donā€™t think ahead and only stop when they feel like it. Itā€™s a big part of why AI chatbots lie and canā€™t do math.

  • Q* could allow models like ChatGPT to think ahead and plan what theyā€™re going to say like a thoughtful high schooler. This makes AI acting on autopilot (for the first time) much more likely.

If Q* is real, an AI chatbot that writes human-like sentences isnā€™t even close to what weā€™ll see in the next couple of years. Buckle up!

FROM OUR PARTNERS

Create AI-generated images, now <1 second.  

We all know what Midjourney's like: type in a prompt, wait 2 minutes, and maybe get a usable image.

A big new innovation called LCM-LoRAs can generate images in real-time (i.e. no waiting). Plus, it lets you modify your images just by moving your cursor or applying simple shapes.

Itā€™s like an instant AI filter for images (and videos, 3D and more).

Wanna build using LCM-LoRAs? Check out a walkthrough from our friends at OctoML, featuring a real-world e-commerce example (and adorable octopi!).

Claude can now handle 150-page PDFs (with an important caveat)ā€¦

Claude 2.1 hallucinates less

Hereā€™s the scoop on Claude v2.1:

  • Now supports 200K context windows = 150K words = >500 pages of text.

  • 2x fewer hallucinations, but itā€™s more prone to responding with ā€œI decline to answerā€.

  • Easier to test and use Claudeā€™s API within existing processes.

The 200K context window is key (150 pages!), but a caveat worth noting via Greg Kamradtā€™s assessment: the more you give Claude, the worse it is at recalling info.

Itā€™s also noticeably better at recalling info from the very top or bottom of a document, rather than info lost in the middle.

Claude is still our go-to for summarizing/dissecting lengthy PDFs, codebases, research papers, books, and more. Just remember to fact-check its outputs before they hit the workplace.

Around the Horn.

  • Elon Muskā€™s chatbot ā€œGrokā€ is set to be available for Twitter Premium+ subscribers this week.

  • OpenAI is still on track for employee share sales at an $86B valuation.

  • Stability released Stable Video Diffusion, its first image-to-video model.

  • Bard can now understand some YouTube video content.

Treat To Try.

We describe Pi as the most ā€œhumanā€ chatbotā€”on par with our most emotionally intelligent friends.

Its creator, Inflection AI, just announced a new-and-improved model thatā€™s actually quite good at math and coding.

It still doesnā€™t match up with GPT-4 (and probably Claude 2.1), but a great option for tackling workplace/home situations that require a personal touch.

We curated the top 27 tools you need for work in our ā€œTop Tools For Businessā€!

Monday Meme.

Here are some of our favorite memes from the past week!

A Cat's Commentary.

Thatā€™s all for today, for more AI treats, check out our website.

Get your brand in front of 180,000+ professionals here.

See you cool cats on Twitter: @nonmayorpete & @noahedelman02

What'd you think of today's email?

Login or Subscribe to participate in polls.