The Memo - 2/Feb/2023

100M ChatGPT users, Baidu's ChatGPT via ERNIE, Microsoft FLAME, GPT-4 + Bing, and much more!

Feb 02, 2023

FOR IMMEDIATE RELEASE: 2/Feb/2023

Welcome back to The Memo.

Who else is reading this alongside you? Well, as of Feb/2023, The Memo includes paid subscribers from these places:

Accenture
Amazon
Automattic
Baird
Bakersfield College
Boston Consulting Group
Brock University
Century 21
Cirque du Soleil
Citi
Comcast
Cornell University
DeepMind
Deloitte
Duke University
Ericsson
Georgia Tech
Google
Harvard
IBM
IEEE
Indiana University
Infosys
Ipsos
Lockheed Martin
Mastercard
Microsoft
MIT
Moodle
Nasdaq
Naver
Oracle Cerner
Pearson (Wechsler)
PepsiCo
Princeton University
PwC
RAND Corporation
Rogers
Siemens
Sony
Stanford University
Steam Labs
Telekom
Telstra
Tencent
Tesla AI
Uber
University of Washington
VMware
Yandex
…and a few thousand more. Thanks to Ili & Joanna for some hand-picked selections.

Did you know: Many readers use The Memo subscription as a tax deduction for their business or organization. You can generate an invoice here.

In the Toys to play with section, we look at the latest human-like text-to-speech generator including a free option to clone your own voice into AI, a new GPT to SQL tool, and a free, open-source ChatGPT clone using a distributed network…

The BIG Stuff

ChatGPT with >10M daily users (24/Jan/2023)

I’ve been running lots of expert calls with Fortune 500s throughout the beginning of 2023, and each time I do, I ‘guesstimate’ that ChatGPT has 10 million concurrent users.

According to ARK Invest, it turns out I was pretty close! (‘chatGPT at >10 million daily users in 40 days Instagram took 355 days to get to 10 million registered users‘…)

Brett Winton @wintonARK

chatGPT at >10 million daily users in 40 days Instagram took 355 days to get to 10 million registered users

Further, the Swiss investment bank UBS yesterday reported that the number is actually 13 million daily users as of Jan/2023, and that ChatGPT had 100 million active users in Jan/2023.

ChatGPT is averaging more than 13 million visitors a day, more than double the level in December.
…venture-capital investors speculate that the market for generative AI applications could be as large as $1 trillion… the world has over one billion knowledge workers; OpenAI charges $42 a month for the professional version of ChatGPT. If you assume every one of those people gets two accounts—one general, and one specialized—you get close $1 trillion.
Walmsley notes: “This may seem like a crazy bullish scenario, and pricing could fall, but our core point is that the market has potential to be exceedingly large.”

Read the cleanest source via Barron’s with quotes: https://archive.is/XRl0R

The Interesting Stuff

OpenAI are urgently looking to scale their models (31/Jan/2022)

13 million daily users drawing on that kind of heavy computing power is a big ask, and OpenAI are looking for the smartest engineers in San Francisco to help them out here! (‘Seeking software engineers across frontend/backend/infra to help us scale & improve ChatGPT who can join ASAP.’)

Greg Brockman @gdb

Seeking software engineers across frontend/backend/infra to help us scale & improve ChatGPT who can join ASAP. Looking for track record of building great products with a team. Describe an exceptional accomplishment, area(s) of expertise, & ideal start date to gdb@openai.com.

OpenAI advising the US government (30/Jan/2023)

Sam Altman, creator of ChatGPT and CEO of OpenAI, is making the rounds in D.C. this week, aiming to demystify the artificial intelligence tool that has captivated the world… In the meetings, Altman told policymakers that OpenAI is on the path to creating “artificial general intelligence,” a term used to describe an artificial intelligence that can think and understand on the level of the human brain.

Read the full article: https://www.semafor.com/article/01/26/2023/chatgpt-creator-sam-altman-visits-washington-to-meet-lawmakers

Baidu building ChatGPT into their search engine in Mar/2023 (30/Jan/2023)

China’s largest search engine company plans to debut a ChatGPT-style application in March, initially embedding it into its main search services… The tool, whose name hasn’t been decided, will allow users to get conversation-style search results much like OpenAI’s popular platform.
Baidu has spent billions of dollars researching AI in a years-long effort to transition from online marketing to deeper technology. Its Ernie system — a large-scale machine-learning model that’s been trained on data over several years — will be the foundation of its upcoming ChatGPT-like tool…

Read via Bloomberg: https://archive.is/jOrZJ

ChatGPT updated with better maths and facts (30/Jan/2023)

Read the change log: https://help.openai.com/en/articles/6825453-chatgpt-release-notes

Microsoft integrating GPT into Teams (2/Feb/2023)

We’re bringing the power of large language models, including OpenAI’s GPT, to Teams Premium, as we make meetings more intelligent, personalized, and protected. - Microsoft CEO, 1/Feb/2023

Read the release: https://www.microsoft.com/en-us/microsoft-365/blog/2023/02/01/microsoft-teams-premium-cut-costs-and-add-ai-powered-productivity/

GPT-4 and Bing (1/Feb/2023)

An OpenAI insider has noted (via Semafor):

Microsoft’s second-place search engine Bing is poised to incorporate a faster and richer version of ChatGPT, known as GPT-4, into its product in the coming weeks, marking rapid progress in the booming field of generative AI and a long-awaited challenge to Google’s dominance of search.
OpenAI’s latest software responds much faster than the current version, and the replies sound more human and are more detailed, according to people familiar with the product and rollout plans.

First LLM for 2023: Microsoft FLAME 60M (1/Feb/2023)

This is possibly the first large language model release for 2023.

We present FLAME, a T5-based model trained on Excel formulas that leverages domain insights to achieve competitive performance with a substantially smaller model (60M parameters) and two orders of magnitude less training data. We curate a training dataset using sketch deduplication, introduce an Excel-specific formula tokenizer for our model, and use domain-specific versions of masked span prediction and noisy auto-encoding as pretraining objectives. We evaluate FLAME on formula repair, formula auto-completion, and a novel task called syntax reconstruction. FLAME (60M) can outperform much larger models, such as Codex-Davinci (175B), Codex-Cushman (12B), and CodeT5 (220M), in 6 out of 10 settings.

It is also aligned with Chinchilla for data, using ~9B tokens for 60M params ("We start from a dataset of 927M formulas" estimating 10x multiplier for a total of 9B tokens).

Read the paper: https://arxiv.org/abs/2301.13779

See it on my timeline: https://lifearchitect.ai/timeline/

See it on my shared Sheet.

Hyper + ChatGPT (23/Jan/2023)

Aaron Ng @localghost

Hyper + ChatGPT + AI Animation = personal interactive avatar teachers on any topic.

Flawless AI replacing movie dialogue (Feb/2023)

This is really useful for dubs, and even making an “R” film into a “G” film!

Toys to Play With

PetalsChat using BLOOMZ-176B (Jan/2023)

As reported back in The Memo edition 11/Nov, the BigScience Workshop BLOOMZ-176B instruction/fine-tuned model is a newer LLM with more languages, that is similar in size to GPT-3.

A chatbot version was made available this month for free via Petals, a distributed network that enables anyone with a computer to contribute their idle processing power to train and run text-generating AI models.

I had fun playing with this, and the results are better than expected. No other model really comes close to OpenAI’s GPT family (see my Model sizes visualization), but the fine-tuned BLOOMZ seems to be nice and clean.

Try it (warning: it can be very slow): http://chat.petals.ml/

View the repo: https://github.com/borzunov/chat.petals.ml

ElevenLabs text-to-speech engine

Stop whatever it is you’re doing right now.

This thing is better than Synthesia, better than Amazon Polly, even better than Sonantic.io.

You can also clone your own voice (also for free) by uploading a sample of more than 1-min, but less than 10MB.

You can try it for free by typing in any text. I’ll be leaning on this thing a lot in 2023!

Try it: https://beta.elevenlabs.io/

Update about malicious actors and new membership mandates (31/Jan/2023).

Watch my video:

And here’s a funny parody of Daniel Day Lewis—voice by ElevenLabs AI—teaching us about oil, with images generated by text-to-image AI, Stable Diffusion: