I-JEPA: Efficient method for Self-Supervised Learning of image features.

No need for data augmentation, just masking.

Joint embedding predictive architecture, not generative.
And it's open source.

Paper: arxiv.org/abs/2301.08243

Code & models: https://github.com/facebookresearch/ijepa

World Economic Forum report on decentralised identity

The report provides tools, frameworks and recommendations for policy makers, government officials, and others looking to engage with dID tech.

Why YouTube could give Google an Edge in AI

Google last month upgraded its Bard chatbot with a new machine-learning model that can better understand conversational language and compete with OpenAI’s ChatGPT.

As Google develops a sequel to that model, it may hold a trump card: YouTube.

YouTube is the single biggest and richest source of imagery, audio and text transcripts on the internet. And Google’s researchers have been using YouTube to develop its next large-language model, Gemini, according to a person with knowledge of the situation.

The value of YouTube hasn’t been lost on OpenAI, either: The startup has secretly used data from the site to train some of its artificial intelligence models, said one person with direct knowledge of the effort.

Hugging Face and AMD partner on accelerating state-of-the-art models for CPU & GPU platforms

For years, brilliant folks have failed to make deep learning work well on AMD hardware, giving NVIDIA a lead.

Quantification of information processing capacity in living brain as physical reservoir

Delve into the ways scientists measure the brain's information processing capability, a discovery that might contribute to the development of more efficient AI systems.

The introduction of Orca by Microsoft AI, a 13 billion parameter model that learns explanation traces from GPT4, represents a significant breakthrough in advancing instruction-tuned models

Orca may surpass existing models through explanation tuning, scaling tasks and instructions, and rigorous #evaluation, marking a substantial leap forward in AI system capabilities.

Incorporating step-by-step explanations in training processes holds promise for fully unlocking the potential of large foundation models and driving progress in natural language processing.

At its annual Google I/O conference, Google’s CEO, Sundar Pichai, demonstrated how their large language model (LLM) called Bard could describe X-rays.

You just need to ask the AI system the right question like “What is on this picture” or “Can you write me a report analyzing this chest x-ray?”.

Later this summer, Med-PaLM 2, an LLM with 540 billion parameters and knowledge from scientific papers and websites, will be made available to a select group of customers using Google’s Cloud.

This development highlights the importance of physicians mastering the ability to formulate commands (prompts) to communicate seamlessly with artificial intelligence. Prompts refer to the AI language that allows us to ask AI to perform specific tasks, such as describing a mammogram or generating a creative image in a chosen style.

LLMs are not AGIs and lack moral agency, but seem to be good at predicting human moral judgments ... In Mindplex article reviews this informally.

Wysa launched its AI-powered chatbot that helps people manage their mental health long before ChatGPT fueled enthusiasm for technologies that seem to think and talk like humans

Wysa’s interactive bot uses techniques from cognitive behavioral therapy to help people manage anxiety, stress, and other common issues.

But under the hood it doesn’t share ChatGPT’s DNA: The bot uses natural language processing to interpret input from users, but it always delivers one of its pre-written and vetted responses. No generative responses means no potentially unsafe content.

It's a formula that’s been working so far for Wysa, which announced a Series B funding round last year and says 6 million people have tried its app. Wysa is freely available to consumers with paid content options, and is also used by the U.K.'s National Health Service and U.S. employer groups and insurers.

Stanford researchers has been evaluated by organizations that build language models.

In the same way that model scores drive model improvement, so model scores will drive improvements in development and deployment practices.

They then found substantial overlap with the EU AI Act, and thus we initially scoring based on it. But this is just the beginning - there are many aspects not covered by the Act.

A fresh review on the generative AI in brain imaging, with some nice infographics explaining different approaches.

Meta wants companies to make money off Its open-source AI, in challenge to Google

Meta Platforms CEO Mark Zuckerberg and his deputies want other companies to freely use and profit from new artificial intelligence software Meta is developing, a decision that could have big implications for other AI developers and businesses that are increasingly adopting it.

Meta is working on ways to make the next version of its open-source large-language model—technology that can power chatbots like ChatGPT—available for commercial use.

The move could prompt a feeding frenzy among AI developers eager for alternatives to proprietary software sold by rivals Google and OpenAI. It would also indirectly benefit Meta’s own AI development.

Terence Tao reflecting on GPT-4 in the AI Anthology coordinated by Eric Horvitz:

"I expect, say, 2026-level AI, when used properly, will be a trustworthy co-author in mathematical research, and in many other fields as well."

Deutsche Bank applies for digital asset license amid growth push.

Voice commanded essay copilot

Inspiring demo! Sit back and talk to your computer with high-level instructions, collaborating on a larger document.

OpenAI is considering launching a marketplace in which customers could sell AI models they customize for their own needs to other businesses

Such an appstore could be a version of the OpenAI app store, offering businesses a way to access advanced large-scale models that can, for example, detect financial fraud in online trading transactions or answer questions about specific markets with up-to-date information.

Creating such an app store can also be a hedge against future competitors that is not dominated by any single AI model.

It is not clear if OpenAI will charge a commission on these sales or otherwise generate revenue from the market.

Stat Health brings in $5.1M for its in-ear wearable to track cerebral blood flow

J2 Ventures, BonAngels Venture Partners and a diverse group of angel investors backed the company through the seed funding.

Stat Health also received grant funding from the U.S. Air Force.

Stat Health designed its wearable to help understand symptoms like dizziness, brain fog, headaches, fainting and fatigue upon standing.

These common symptoms can indicate illnesses like long COVID and postural orthostatic tachycardia syndrome (POTS). They also may signal myalgic encephalomyelitis/chronic fatigue syndrome (ME/CFS), and other orthostatic syndromes.
According to Stat Health, reduced blood flow to the brain upon standing causes the symptoms for these illnesses.

The company clinically tested its offering at Johns Hopkins, and it was peer-reviewed in the March 2023 issue of the Journal of the American College of Cardiology (JACC). Stat Health said it demonstrated the ability to predict fainting minutes before it happens.

Carve out a few hours to learn streamlit.io
Powerful for rapid prototyping, interactive visualization.
It's a hammer and you'll start seeing a lot of nails.

Are we at the beginning of a new era of small models? Here is newest LLM trained fully at Microsoft Research

phi-1 achieves 51% on HumanEval w. only 1.3B parameters & 7B tokens training dataset.

Any other >50% HumanEval model is >1000x bigger (e.g., WizardCoder from last week is 10x in model size and 100x in dataset size).

Corporates Using NFTs - Are They More Than a Passing Fad - Deloitte

NFTs, and the distributed blockchain networks behind them, represent a breakthrough in digital rights management as well as digital representations of assets. Here are just some of the other possible breakthroughs NFTs could drive: 

1 Building a bridge between the digital and physical worlds to authenticate and provide evidence of a transfer

2 Democratising ownership of digital collectibles—for example, the creation of new ways to monetise art, photographs, music, intellectual property (IP), and more

3 Selling digital items—homes, high-end sneakers, streetwear, and more—for use with avatars in gaming and online worlds

4 Developing “super wallets” that allow an NFT owner to keep a verified record of all licenses and rights, along with product warranties, event tickets, access passes for secure locations for work or leisure, and more 

5 Securing the ticketing industry against fraud, providing a percentage of secondary sales revenue to performers or venues, and creating unique keepsakes

6 Extending and monetising brands in new ways for both existing and new customer bases

7 Offering utility services—for instance, serving as a VIP card granting access to a secret concert, membership in an exclusive community, or special discounts on products.

CVPR 2023 announced the Best Paper Awards! It's the world's most prominent computer vision conference, with citation impact just below Nature & Science.

- Best Paper 1: VisProg uses GPT to generate executable code that parses an image and does effective visual reasoning, even though GPT itself is blind. Similar principle as Voyager (executable code to play Minecraft).

VisProg: prior.allenai.org/projects/vispr…
Paper: arxiv.org/abs/2211.11559
Another paper with similar high-level idea is called ViperGPT, worth checking out: arxiv.org/abs/2303.08128

- Best Paper 2: Unified Autonomous Driving (UniAD), a comprehensive framework that incorporates full-stack driving tasks in one network.

Planning-oriented Autonomous Driving: opendrivelab.github.io/UniAD/
Paper: arxiv.org/abs/2212.10156

- Best Paper Honorable Mention: DynIBaR, a new state-of-the-art on synthesizing novel views from a monocular video of a complex dynamic scene.

DynIBaR, Neural Dynamic Image-Based Rendering: dynibar.github.io
Paper: arxiv.org/abs/2211.11082

- Best Student Paper: a new 3D point cloud registration technique that finds the optimal pose to align a pair of point clouds.

3D Registration with Maximal Cliques: arxiv.org/abs/2305.10854

- Best Student Paper Honorable Mention: a diffusion model that can be customized to a particular subject with only 3-5 example images.

DreamBooth: dreambooth.github.io
Paper: arxiv.org/abs/2208.12242