Self-driving will progress rapidly in the next few years

Waymo is opening ride-hailing service between Phoenix Sky Harbor International Airport and Downtown Phoenix to members of the public.

The service is available at the 44th Street and Washington PHX Sky Train® Station and is the only autonomous airport service of its kind in the world, available round the clock with no human driver.

Taisija Demchenko and Ninon Lizé Masclef introduced the project "reaDream" to decode dream contents from brain electrical activity into text. For this project, the subject was sleeping while wearing a Unicorn Hybrid Black BCI

2023 AI prediction: the gap between generative and predictive AI will widen.

Despite product & business model innovation in generative AI, real-world ROI will remain concentrated around predictive AI- leading to frustrated expectations.

This gap will all come down to data.

First, basic definitions:

- Generative (ie. LLMs / foundation models): Goal is to output a data point (e.g. an image)

- Predictive (or "discriminative"): Goal is to label a data point (e.g. predict whether an image contains offensive content).

A natural response would be: isn't generating data fundamentally "harder" than just labeling it? And formally the answer is (roughly) yes.

However, the widening efficacy gap comes down to how each is used in the real world.

To date, generative AI has been useful in:
- Human-in-the-loop settings (eg co-authoring code/art/text, having a cooperative dialogue); and/or
- "Fuzzy success measure" settings (e.g. an experimental chatbot with no correctness guarantees.

These types of settings have lower / less well-defined performance bars. The proverbial "80%" demo-ready accuracy will be enough to ship- and create some value as product/UX/business models get worked out.

So highly-visible generative AI progress will continue to accelerate.

The point of most predictive AI systems, on the other hand, is to *automate* some process at high accuracy so that human oversight is not needed. Eg. triage a medical image, read a loan document, etc.

In these use cases: success is very measurable, and the bar is very high.

Reaching the performance needed to deploy predictive AI requires labeling (and re-labeling) training data for each task and setting.

Building predictive AI models on top of generative AI "foundation models" will help, but not solve this- foundations are just foundations.

Because of this data gap, predictive AI will seem stuck while generative AI accelerates in 2023.

Most high-value AI will still be predictive- so there will significant frustrations around AI ROI.

However- it will be an exciting time to work on bridging this data gap.

OpenAI just dropped a prototype of 3D DALLE (called “Point-E”)

It isn’t as good as Google’s DreamFusion, but blazing fast! Like ~600x faster to generate. 2D DALLE has already turned the creative world upside down. How will 3D DALLE disrupt games, VR, metaverse, …?

Bonus: it’s fully open-source! Both the code and all model weights (for 40M, 300M, 1B parameters) are available at https://t.co/thj7kSAfuv

Arxiv paper: arxiv.org/abs/2212.08751
2023 will be a year of powerful 3D generative models, and 2D will simply be their planar projection.

The Metaverse is expected to increase demand for even more data: the average VR Metaverse user requires up to 40x more data than HD video streaming.

The forecasted evolution of global data traffic between 2021-2027.

This figure presents the number of open source AI papers that cite the use of specific AI chips according to analysis by Zeta Alpha. The 2022 figure represents the annual sum extrapolated from 4 Dec 2022.

Yesterday was published
V2 of the State of AI Report Compute Index

A few notes:

- Usage in AI research papers (early adopters) as a leading indicator of industry usage.
- Papers using AI semi startup chips almost all have authors from the startup.
- FY2022 is extrapolated from 4 Dec '22.

You'll find live counts of AI research papers using chips from NVIDIA, TPUs, ASICs, FPGAs, and AI semi startups.

First, here's an updated count of papers using:
- any NVIDIA chip
- Google's TPU
- ASICs
- FPGAs
- chips from graphcoreai SambaNovaAI CerebrasSystems Cambricon and habanalabs
- Huawei's Ascend 910

You'll see that NVIDIA is just *leagues* ahead...by 2 orders of magnitude!

For now here are the AI research paper counts for the 5 major AI semiconductor startup contenders.

Combined, they're used 125x less than NVIDIA's chips.

After the explosive growth of the last few years, Amazon now has 7-8% of US retail sales, and Amazon’s own retailing (i.e excluding Marketplace) is 3-4%
All ‘online-only’ retailers combined have 12-13% of US retail.

China's Zhejiang province to build $28.7 billion metaverse industry by 2025

The Zhejiang Province of China is setting in motion plans to construct a metaverse industry worth more than 200 billion Yuan by 2025. The South China Morning Post reported that the home of Alibaba is eyeing the development that would be worth $28.5 billion.

The western province is the latest Chinese region to extend plans for a metaverse industry. The report noted that the concept was met with an enthusiastic reception by local authorities. All of which are participating in the developmental plans.

Zhejiang Embracing Metaverse

The Zhejiang province is the home of Alibaba Group Holding headquarters, the driving force behind the new plans to develop a metaverse industry worth billions. The proposed plan is aiming to, “develop metaverse-related industries valued at more than 200 billion yuan.”

The report from South China Morning Post stated that the news comes from an official development plan. Furthermore, that plan was published by the provincial development and reform commission last week.

Artificial intelligence could help cool buildings more efficiently. DeepMind trained an AI to minimise energy usage by cooling systems under different weather conditions

Stablecoins have settled more than $7 trillion in value, a record-breaking value compared to previous years — link

PwC just released their Global Crypto Regulation Report 2023

Below is their Crypto Regulation at a Glance table which provides a nice concise jurisdiction by jurisdiction comparison.

The full report is 68 pages long and you can access it here

- Where crypto regulation is heading
- Views from global standard setters
- EU single market for digital assets (nice pictorial for upcoming #mica regulatory framework and requirements for crypto businesses)
- Summaries from different jurisdictions

Brain computer interface to distinguish between self and other related errors in human agent collaboration

When a human and machine collaborate on a shared task, ambiguous events might occur that could be perceived as an error by the human partner.

In such events, spontaneous error-related potentials (ErrPs) are evoked in the human brain. Knowing whom the human perceived as responsible for the error would help a machine in co-adaptation and shared control paradigms to better adapt to human preferences.

Therefore, we ask whether self- and agent-related errors evoke different ErrPs. Eleven subjects participated in an electroencephalography human-agent collaboration experiment with a collaborative trajectory-following task on two collaboration levels, where movement errors occurred as trajectory deviations.

Independently of the collaboration level, we observed a higher amplitude of the responses on the midline central Cz electrode for self-related errors compared to observed errors made by the agent.

On average, Support Vector Machines classified self- and agent-related errors with 72.64% accuracy using subject-specific features.

These results demonstrate that ErrPs can tell if a person relates an error to themselves or an external autonomous agent during collaboration. Thus, the collaborative machine will receive more informed feedback for the error attribution that allows appropriate error identification, a possibility for correction, and avoidance in future actions.

A new speech-to-speech translation architecture, UnitY

UnitY translates source language speech to the target text and the corresponding discrete acoustic units in the two-pass way.

Animoca Brands Unveils $2B Metaverse Fund

Hong Kong-based blockchain gaming giant Animoca Brands is all set to roll out a massive $2 billion fund, dubbed “Animoca Capital,” to invest in metaverse businesses, according to co-founder and executive chairman Yat Siu.

While speaking to Nikkei Asia, the exec revealed that the upcoming metaverse fund, which will make its first investment in 2023, will focus on digital property rights. It also seeks to provide opportunities to access Web3 companies.

Meanwhile, its portfolio boasts over 380 investments, including Colossal, Axie Infinity, OpenSea, Dapper Labs (NBA Top Shot), Alien Worlds, and Star Atlas, among others.

Crypto trends for 2023 by Messari - must read

Someone implemented a framework to allow you to train PaLM (Google’s 540B parameter LM) using the same reinforcement learning strategy as ChatGPT!

Open source strikes again.