honestly, lots of big stuff

Oh boy...what a week it was for AI yesterday! There are quite a few links in today's email so don't let it scare you. Like pacing yourself for a Thanksgiving meal.

Welcome to the 153 folks joining us since yesterday.

Let's get to it.

Prompt: List the topics in today’s email

🫦 BEN'S BITES

Hugging Face Tasks. Tasks cover all you need to know about machine learning. Click on a task e.g. image classification and it’ll show you a summary of what that is (including a visual), resources to learn about it, any datasets, spaces and models that are popular. It’s going to be the go-to documentation for anyone building ML apps.

Another update from HF: You can now duplicate Spaces on Hugging Face. Clone and remix, edit the code, customise the look or run it on different hardware.

An AI POV of “value investing” vs “growth investing”.

Assembly AI is running a virtual AI hackathon, with $50k in cash prizes.

Suhail is now all in on AI and released all the liked images and their prompts on playgroundai.

Notion now has AI baked in (includes my link to move up the waitlist)! It basically has GPT-3 functionality now* (*on a waitlist of course). This is a good reminder that some features are that, they’re features, not products. If you were working on an AI writing tool for Notion, you’re sure to be disheartened by this but it's the future we will see. Most platforms will have AI as a feature. You can imagine; Figma with text-to-image, Databases with text-to-formula, Zapier with text-to-automation and so on. It’s an exciting time but brings up the ‘what is the moat in AI?’ question even more. And when we get GPT-4, all tools will get a capability upgrade by moving from GPT-3. You can read more about the release from The Verge.

Runway released another AI Magic Tool: generate subtitles and transcripts automatically. It’ll be a fun head to head for them vs Descript. Runway are the best marketers around right now. Their videos are 👌.

Ben’s Bites Early Access: Mind Maps in Miro, if you sign up you’ll get access first. This tool uses GPT-3 in Miro, the whiteboard tool, to produce mind-maps and more.

An interesting thread by Peter Diamandis on how AI -> human collaboration. Again, people still think AI will take people’s jobs and for some, it will. But a lot of jobs will have a human in the loop. AI will make humans much more productive.

Can a model trained on a *plausible* amount of *raw* speech explain both intelligent behavior and its brain bases?

Typefully, the Twitter tool for thread, growth and more, is adding AI (see my point above) called Vesper. A writing tool to give tweet ideas, continue tweets and gives comments to improve your tweets. Elon will have to add another badge ‘Written by AI’.

Mixture-of-Experts with Expert Choice Routing by Google - which is essentially a new way to route traffic in a system. It’s a way of dramatically increasing model capacity without a proportional increase in computation. You can view it on Hugging Face - yes, HF can hold huge models (1.6T parameters!)

Galactica, which I mentioned yesterday, explores scientific papers and provides answers to questions you ask it. However, as with most AI experiences, there are instances of incorrect information. This user asked it about some things he knows about and it was wrong or biased, sounded right and authoritative. Galactica does have a ‘not verified’ badge on answers that aren’t, well, verified. But like wikipedia or any media, people do believe what they read. It’s a tricky situation to navigate. Gary wrote a piece on it’s inaccuracy and there’s also this example on ‘Bears living in Space’ which we can all confidently say is false.

A thread on how diffusion models work.

A proposal for a shift in focus from "learning from data" to "learning what data to learn from" in AI research. The authors argue this exploration is essential to all learning systems, including supervised learning, and proposes the problem of generalized exploration to conceptually unify exploration-driven learning between supervised learning and reinforcement learning.

AI-powered software that can extract human motion from videos and automatically retarget the data to a character rig. You know in movies like Avatar, the actors have all those bobbles to trace their movements for CGI. This is kinda that, minus the bobbles!

A multiplayer AI creation canvas.

Universal Q&A - A chatbot for any website on the internet. You can go to a homepage for OpenAI and ask ‘do you offer maternity leave?’ and it will return the answer from crawling through the site and finding relevant info! Love this.

a16z published a post on why AI models will replace artists long before they’ll replace programmers…because it is easier to automate creative tasks than programming tasks. (h/t Matt from a16z, who contributed + is a fellow reader)

Tome, a beautiful way to craft any story–instead of boring slide decks, has integrated DALL-E into their tool. Tome’s website is also beautiful.

LLMs have the potential to be used for artificial general intelligence (AGI) and the technology is not 10 years away. An interesting take!

Sam Lessin wrote 10 companies to start in AI today.

DeepMind, in partnership with, ahem, PartnershipAI, developed a set of best practices for responsible data collection.

Readwise lets you save, read, and highlight anything -- from articles to PDFs to tweets to newsletters. And they’ve just announced ‘Ghostreader’, it’s AI feature “copilot for reading”. It’ll define words in context, lookup people/places, simplify language, you know, all the usual GPT-3 bells and whistles. The announcement thread gives some great context and examples.

Using brain recordings to reconstruct what people are seeing at the time. UMMM This seems fairly HUGE!

After my very basic explanation of emergence yesterday, I saw this thread on how emergence has been happening for decades and in very small networks.

Evaluating the factual consistency of LLMs through summarization. The benchmark involves comparing the scores an LLM assigns to a factually consistent versus a factually inconsistent summary for an input news article.

Token Turing Machine (TTM), a sequential, autoregressive Transformer model with memory for real-world sequential visual understanding.

The HELM benchmark for language models aims to improve the transparency of language models.

Investigating the ability of large language models (LLMs) to translate between languages. While their performance is impressive, it still lags behind that of state-of-the-art supervised systems.

MAsked Generative Encoder (MAGE), the first framework to unify SOTA image generation and self-supervised representation learning.

Nitro Diffusion - the first Multi-Style Model trained from scratch.

Cradle raised $5.5m from Index, Kindred and 30+ angels. Cradle helps biologists design improved proteins in record time using powerful prediction algorithms and AI design suggestions.

Microsoft and Nvidia team up to build new Azure-hosted AI supercomputer.

🧑‍💻 WHO'S HIRING IN AI

VEED.IO - Simple online video editing. They are building out their AI team and hiring engineers.

Buildspace - where builders, build! They're looking for an ML/AI instructor to build their new course.

Add your company here | Join the Talent pool

🖼 AI IMAGES OF THE DAY

Cyperpunk Biker.

🤗 SHARE BENS BITES

Send this with 1 AI-curious friend and receive my AI project tracker database!

or copy/paste this link: https://bensbites.beehiiv.com/subscribe?ref=PLACEHOLDER

👋 SEE YA

⭐️ HOW DID WE DO?

⭐️ REAL REVIEWS