This is an ape Kanzi playing Minecraft! A fascinating experiment on non-human biological neural networks

Researchers are from the Ape Initiative, a non-profit org.

There're so many similar techniques that the ape trainers used:

1. In-context reinforcement learning: Kanzi gets a fruit or peanut whenever he hits a marked milestone in the game, incentivizing him to follow the in-game guides.

2. RLHF: Kanzi doesn't understand much language, but he can see the trainers cheering him on, and he occasionally cheers them back! That gives him a strong signal that he's on the right track.

3. Imitation learning: the trainers show Kanzi just 1 demonstration of how to do a task, and he immediately grasps the concepts. It's much more efficient than using rewards alone.

4. Curriculum learning: they start with very simple environments to gradually teach Kanzi the controls. At the end, Kanzi is able to navigate complex caves, mazes, and the Nether.

It also amazes how strong the ape's vision system is. Kanzi never saw Minecraft in his life, and for sure his ancestors didn't either. Yet he rapidly adapts to Minecraft's texture and physics, which are dramatically different from the natural world.

This level of generalization is far beyond what most powerful vision models can do today. AIs are approaching human level on understanding language, but far behind animals on parsing pixels.

⚡️ Google DeepMind’s CEO says their new AI model Gemini will far surpass the capabilities GPT-4

Gemini (still several months from release) is a multi-modal AI system that combines their Go-winning program, AlphaGo, with LLM capabilities.

Smart Contracts. The Bank for International Settlements has released a paper that examines the oracle issue with respect to decentralised systems.

Summary

Smart contracts are self-executing programmable contracts between two or more parties. Smart contracts do not require a vetting authority because their legitimacy relies on decentralised ledger technologies. However, the implementation of many potentially useful smart contract applications depends upon verifying that some real-world event has taken place. This is a problem. Given its fully decentralised nature, how does a smart contract select what the true state of the world is?

Contribution

This paper considers a situation where multiple individuals seek to enter into agreements based on the outcome of a real-world event, but where there is no trusted party (ie contractible source) that can be used to determine payoffs.

In this case, payoffs must be based on some form of collective agreement on the true state of the world. A truth-telling problem arises in this environment, as agents report the state of the world that is most beneficial for them.

Findings

The general result, which applies to simultaneous voting games, is that the only way that individuals are willing to vote according to the true state is if they are completely indifferent as to what the true state should be.

That is, their payoffs cannot depend on their actions or their individual reports.

This general result suggests that, in the absence of some additional motivation (eg an inherent preference toward truth-telling) which links individual payoffs to the truth, there is no way to implement contracts that pay out based on an observed state without a trusted source.

Researchers of the Human Brain Project have coupled the measurements of brain waves associated with disorders of consciousness (DoC) with glucose usage in specific brain areas, identifying where in the brain the waves might be generated.

The study for the first time adds the dimension of location and energy consumption to the reading of brain waves in DoC, and reveals the important role of subcortical areas in driving cortical activity associated with consciousness.

Blockchain Node Engine is now available to everyone! You can get fully synced Ethereum archive nodes in less than a day for a flat hourly rate.

Neurolangelo: the demo here is a 3D virtual reconstruction of NVIDIA’s headquarters in Santa Clara.

Transporting reality back to sim has huge applications in VR/AR. It’s also a potential way to multiply synthetic training data for robotics in sim.

Open source.

TSMC will ship 550,000 of the latest Nvidia H100 GPUs in 2023

Saudi Arabia and UAE squeezing the global supply by purchasing 1000s of the GPUs, priced US$40,000 each, to join the global AI race.

How a California health system puts VR into practice

Hoag Healthcare, an Orange County health system with almost 600,000 patients and almost 2,000 doctors, on their use of the tech.

They're using Penumbra’s REAL System, which is designed for rehab care for issues like strokes or brain tumors, and to help patients recovering from substance use disorder when they experience cravings.

More than 1,000 Hoag patients have used the tech. It's also helped the health system manage its own staff's stress levels; clinicians are using the headsets to take breaks between patients.

Hoag has about 70 headsets that it loans out. 
For acute and outpatient rehab, VR offers a way for patients to start therapy faster than if they'd have to wait for fully in-person care, and it keeps them motivated.

The Monetary Authority of Singapore (MAS) has announced a new regulatory framework for stablecoins in which issuers must meet key requirements regarding value stability, capital, redemption and transparent disclosure in order to be recognized as a “MAS Regulated Stablecoin.”

American Association for the Advancement of Science (AAAS) just released The Decision Tree for Responsible AI that could help "operationalizing a broad set of principles that AAAS has identified as core components of an ethical approach to developing and implementing AI".

Its goal is to assist those who may be contemplating AI solutions in asking appropriate questions and identifying practical steps that will help them progress towards implementing the technology in an effective, ethical way.

Bytedance, the company behind TikTok, uses Ray for fast & cheap offline inference with multi-modal LLMs.

They generate embeddings for a staggering 200 TB of image and text data using a model with >10B parameters.

ByteDance chose Ray Data over Spark for large model scaling due to its:

1. Flexible API for connecting LLM stages
2. Cost reduction via sharding & heterogeneous resource scheduling (CPUs for preprocessing, GPUs for inference)
3. In-memory data transfers between stages, no disk.

Ray Data's streaming execution is super efficient for large datasets and hybrid CPU+GPU workloads.

Check out these blog posts to learn more:

1. Ray vs. Spark comparison:
2. Exploring Ray's streaming execution engine:

DeepLearning AI released a new course "Large Language Models with Semantic Search”, built with Cohere

This course teaches the techniques needed to leverage LLMs into search.

Course access is free for a limited time during the DeepLearning.AI learning platform beta!

New breakthrough from Meta AI: Instruction Backtranslation.

Applied to LLama this method outperformed Claude, Guanaco, LIMA, and Falcon-Instruct.

It can automatically label text with instructions and picks ideal ones for LLM finetuning. No humans.

How it works:
1. Instruction Generation: Using LLaMa, instructions are self-generated from web documents.

2. Quality Control: LLaMa self-curates the high-quality prompts from the generated candidates.

3. Model Finetuning: The newly-curated instructions are used to finetune a more robust model

Tether has announced that it will stop minting and issuing USDT on Polkadot test chain Kusama, Bitcoin Omni, and BCH. Users can swap USD₮-Omni, USD₮-Kusama, and USD₮-BCH-SLP to other chains on platforms that support these assets (such as Bitfinex).

Samsung Game launching an open beta today for the Samsung Game Launcher, which lets you play mobile games on Samsung devices using Samsung’s mobile cloud gaming service.

The new beta service of the Samsung mobile cloud gaming platform will be available in the U.S. and Canada, and it can be included as an update for an existing game launcher already used by as many as 150 million monthly active users on Samsung Galaxy devices.

DeepMind released AlphaZeroᵈᵇ, a team of diverse AlphaZero agents that collaborate to solve chess puzzles and demonstrate increased creativity.

AlphaZeroᵈᵇ discovers a group of strong chess players with distinct playing styles. It models the players with a single, player-conditioned architecture and encourages them to play differently using behavioural and response diversity techniques.

Experiments show that players in AlphaZeroᵈᵇ play chess differently, solve more puzzles together, outperform a more homogeneous team and specialize in different openings. Remarkably, some of the players solve the challenging penrose positions.

The newly formed CPC Central Science and Technology Commission (CSTC) has recently held its first meeting.

Although only minimally reported, the revelation of the first meeting suggests that China's industrial policy has been taken over by the top Party leadership in order to concentrate national resources on achieving tech breakthroughs.

The first digital twin of the human immune system highlights the vast societal opportunities in physical-digital fusion.

Virtual replicas of ‘real life’ give humanity exponential insights.

Much of the real world – like the seabed, the innards of cars, or the workings of our bodies – is hidden.

Digital-physical fusion creates dynamic representations of real-world objects, systems, and processes in the digital world and can make these things become clearer. And the human immune system is, in many ways, the ultimate example.

Why do we need a digital twin of the human immune system?

Digital twins provide a raft of advantages in all walks of society and business. The immune system project could mean:

1. A clear model of something very complicated

2. The chance to see – and understand – connections that were previously opaque

3. Immediate support for tackling immune-dependent conditions like organ transplants, cancer, and autoimmune diseases

4. Quicker and cheaper drug discovery

5. A doorway to drugs with less harmful side effects

6. The route to personalized medicine