Counting Down to Christmas!

OpenAI makes a comeback, read on to find out how to get the bag 💰️ 

• OpenAI’s White Paper on Superalignment

• OpenAI’s Superalignment Fast Grants

• More Google AI Tools for Developers and Enterprises

• Weekly productivity hacks

• Quick snips of other AI updates

Read time: 6 minutes

OpenAI’s White Paper on Superalignment

Nowadays, humans continue to be the agents responsible for training AI models in the context of reinforcement learning. However, OpenAI believes that AI will surpass human intelligence significantly within the next 10 years. This raises the “weak-to-strong generalization” problem: who will be responsible for training future AI systems?

The approach that OpenAI tested was: “can we use a smaller (less capable, i.e. GPT-2) model to supervise a larger (more capable, i.e. GPT-4) model?”

Short answer: Yes, the resulting model typically performs somewhere between GPT-3 and GPT-3.5.

However, further research needs to be done to address scenarios in which a superintelligence hides its true behavior and pretends to be aligned when it isn’t.

OpenAI’s Superalignment Fast Grants

The “weak-to-strong generalization” problem discussed above is one of OpenAI’s research interest and the company is launching a $10M grants program to support technical research towards ensuring superhuman AI systems are aligned and safe:

• $100K–$2M grants for academic labs, nonprofits, and individual researchers.

• For graduate students, a one-year $150K OpenAI Superalignment Fellowship: $75K in stipend and $75K in compute and research funding.

The other research directions include:

• Interpretability: How can we understand model internals? And can we use this to e.g. build an AI lie detector?

• Scalable oversight: How can we use AI systems to assist humans in evaluating the outputs of other AI systems on complex tasks?

• Chain-of-thought faithfulness: The ability of a language model or AI system to maintain coherence and consistency in generating responses or outputs.

• Adversarial robustness: The ability of an AI model or system to maintain its performance and functionality even when facing deliberate attempts to manipulate or deceive it.

• Evals: The process of assessing or evaluating the performance, accuracy, or quality of a model or system.

• Testbeds: Testbeds help facilitate fair and consistent evaluations of AI systems across different domains or applications.

More Google AI Tools for Developers and Enterprises

With the introduction of Gemini, Google’s latest AI model, the company is stepping up and making Gemini Pro available to developers and organizations, which has the following applications:

• Gemini API in AI Studio and Vertex AI

  Access to 130+ AI models for various applications

• Introducing Imagen 2 on Vertex AI

  Ability to generate a wide variety of creative and realistic logos and accurate text-to-image rendering

• MedLM in Vertex AI

  Improved medical documentation, pre-clinical research and development and healthcare access and experience etc.

• Duet AI for Developers

  Duet AI will generate suggested code that can be reviewed and modified, and help users overcome challenges like friction when integrating new tools and services, hashing over repetitive tasks, and spending time understanding a new code base or project

• Duet AI in Security Operations

What’s a Power Nap?

Short naps that typically🙂 last for 20-30 minutes that are known to improve cognitive function, memory, and overall productivity.

What’s a Caffeine Nap?

A strategy that combines the effects of caffeine and a short nap to enhance alertness. The idea behind a caffeine nap is to consume a caffeinated beverage, such as coffee or tea, and then immediately take a short nap. The caffeine takes about 20 minutes to reach peak levels in your bloodstream, so by the time you wake up from the nap, you'll experience the combined effects of the caffeine and the restorative power of the nap.

According to ChatGPT: However, it's important to note that individual responses to napping and caffeine vary, so it's essential to experiment and find what works best for you.

So, I’ll head off to experiment now. 😴 

The kanji “zei”, which means taxes, was selected to be the Kanji of the Year in 2023 in reference to tax hikes and cuts, introduction of a new invoice system, and municipal tax payments in exchange for gifts.

Meanwhile, Oxford University Press has announced “rizz” to be the word of the year for 2023. For all the boomers out there, rizz is short for “charisma” but with a touch of dyslexia.

A great example of how to use these words:

The higher your income zei bracket, the more rizz you have.

Google DeepMind used a large language model to solve an unsolvable math problem - Link

Boosting faith in the authenticity of open source software - Link

Harvard engineers use data science to accelerate drug discovery - Link

Partnership with Axel Springer to deepen beneficial use of AI in journalism - Link

That’s It For This Week, Catch Y’all Same Time Next Week!