Framing this for my grandchildren.

RT @richardcsuwandi: Most AI systems today follow the same predictable pattern: they're built for specific tasks and optimized for objectiv….

Highly recommend applying to Neo Scholars if you're in college. DM me if you are thinking about it. There's been so many instances where I see something cool on Twitter and find out its creator is a part of the Neo community.

Cool to see a company putting this into a real product. This was my vision behind building Memory Lane in 2023!. .

if you enjoyed this thread, you might also enjoy my deep dive thread on formalizing scientific discovery.

@kenneth0stanley @jeffclune @_rockt @jennyzhangzt Some things I'll be watching/reading next:. "Novel Opportunities in Open-Endedness" -- this shows how open-endedness is different from existing hillclimbing vai toy example "Open-Endedness, World Models, and the Automation of Innovation".

For further reading, I recommend following these people and their work:. @kenneth0stanley @jeffclune @_rockt @jennyzhangzt.

The rest of the paper talks about how open-ended ASI could remain safe and aligned. Open-ended systems need to maintain a two-way street with humans so that they remain aligned to human goals while providing us a new understanding of the world.

The other 3 proposed paths are similar to RL in that they are ways for an agentic AI to adapt to its environment. - self improvement: model evaluates its own performance and tweaks itself to do better next time. - task generation: model sets goals for tasks to achieve in its env.

AlphaGo and others demonstrate how reinforcement learning can achieve superhuman open-ended systems. Exploration techniques induce novelty and the use of critics shows that these outputs can be learnable as well. LLMs are showing lots of promise right now with RL too, even

The authors believe we need open-endedness for ASI, and propose a couple paths for reaching open-endedness by altering how we train foundation models. The obvious first idea is using RL.

today, we have systems that are either: .- general and not open-ended.- open-ended but not general. AlphaGo is open-ended but only in the domain of Go. It generates novel strategies, which a human is able to learn from, making it open-ended. Modern LLMs are general but not.

practical systems have limitations to this abstract definition. humans are the most common observer for these open-ended systems, but we have limited memories. so the learnability of anything we observe plateaus when we reach those limits. for example, if you started reading.

this definition incorporates both novelty and learnability--both of which require an observer that judges these qualities. Novelty is a system's ability to keep creating things that the observer didn't predict. Learnability is when something becomes more predictable the more

first, what is open-endedness?. informally, it's giving AI creativity: the ability to continuously generate novel outputs. formally, the paper defines it as:

how do we get truly novel outputs from AI? . major labs including DeepMind and OpenAI have had teams dedicated to open-endedness. makes sense given the impact it could have on enabling discovery and self-improvement. doing today's deep dive thread on this field:

More context on the speed upgrade.

Speed matters!. Last week, I added a feature that significantly reduced load times on soarXiv, and it's translated to almost 3x conversion rate (% of people who open a paper).

RT @tomtumiel: I made an infinite, personalized feed for arxiv papers:

Look for this button at the bottom center of the screen