While the world is in the grasp of war and climate change the latest Stateof.ai report came like a breath of fresh air. Showing AI trends that gave us hope and restore faith in humanity - here’s three trends that we see pushing forward AI for Good.

Open research & power distribution

‍While it seemed at some point that cutting-edge AI like LLMs would be dominated by the big few - because of massive training costs and data needs - decentralized research collectives are slowly gaining “market share” in AI research. Big examples are Stable Diffusion, open-sourcing an image generating DM at a training cost of “only” 600k USD.

What it makes us think

What makes humans great has always been our ability to flexibly collaborate. Neanderthals had bigger brains, but we were more likely to share our knowledge and tools - we’re glad that humanity is again playing into their biggest strength!

What this means for you

In our bootcamp and challenges we put a strong focus on transfer learning and building on existing tools. The open sourcing of these previously closed of models allow us to use them in AI for Good. And with more startups being able to apply these cutting edge technologies (many friends of the community) you will also get the chance to work in the cutting-edge of AI.

Advances in science due to AI

AI is solving long standing problems left and right in science. Every discipline from applied math: discovering a more efficient way of matrix computation (DeepMind’s AlphaZero) to energy: using RL to stabilize plasma in fusion reactors (@Lausanne’s TCV) to medicine: protein folding (DeepMind) & treatment plan optimization to even recycling: discovering enzymes that can break down plastic (UTA). But also more “meta” topics like the use of LLMs to help AI use other tools like websites, apps and robotic equipment (don’t tell your mom).

What this means for you

No matter what discipline you’re in, you will be using AI. So, great that you’re here! But you won't perse need to be a fundamental AI researcher to reap the benefits.

What it makes us think

With the "vulnerable world hypothesis", Oxford Prof. Nick Bostrom warns for the dangers of truly destructive technologies becoming cheap and simple — and therefore exceptionally difficult to control.

While the breakthroughs highlighted show nothing but goodness, a model used for drug discovery can easily be used to develop the next generation of deadly, hard to detect nerve agents or killer viruses. Developing a weapon like this once needed a lot of resources and the leading scientists in the field, but can now be achieved by "just" transfer learning a ML model. This is different from previous destructive forces like the atomic bomb.

It also makes us think that we need to start doing applied science AI for Good challenges! We’ll get right on it :)

AI safety is booking tangible results

The infamous black box is being cracked, for example using RL acting on human feedback (RLHF) to adjust LLMs in their responses or direct human language feedback to adjust models (NYU). Since it can be quite argues to manually test models, “red teaming” has been developed by DeepMind: using LMs to provoke unsafe / power-seeking behavior in LM's and adjust them automatically.

What it means for you

AI safety is finally being taken seriously, meaning a research career or even startup in the field is feasible!

What it makes us think

While it’s great that there’s attention for AI with “malicious intent” like power-seeking behavior we’d like to see a stronger push to tackle arguably the biggest AI threat to humanity right now: recommender systems and the like influencing our behavior by optimizing for our reactiveness. Victor Frankl said: “Between stimulus and response there is a space. In that space is our power to choose our response. In our response lies our growth and our freedom.”

The most clear sign of this space reducing is political polarization, made possible due to filter bubbles (see image). The ill intent or even just negligence of companies can topple democracies. To solve this government legislation and lobbying is needed... Interestingly, we can now also use deep reinforcement learning to craft policies likely to be accepted by governments - so maybe AI offers a solution too!

‍

We hope that this short summary of what we think are the most important trends has given you the same lightness in your step as it did for us. Be sure to check out the full report of Stateof.ai and as always…

Have a fruitful day!

Your Chief Cheerleader Buster

‍