@addyosmani The best self-check of whether you understand something well enough is when you have to explain it to someone else. If you can't explain it, you don't understand it. Having to teach is a good guard against cognitive surrender. One of the reasons I love it.
@ylecun@musedivision@geoffreyhinton Is there evidence that there is a fundamental difference between discrete and continuous reasoning spaces other than the efficiency gains? If we allowed AIs to invent their token space instead of using human language, would the two converge to the same thing?
@Plinz Physics is half Math and half observation and experimentation, which involves interaction with the physical world. There will be speedups, but I'd say their job is safe for some time.
@swombat Evolution is good at stumbling upon new discoveries by chance, and LLMs are good judges. Thatโs a recipe for automated discovery, though not RSI. (2/2)
@swombat Current models are not capable of producing novel, out-of-distribution ideas. Until that becomes a thing, RSI is not possible. That said, as far as I know, DeepMind has been working on a hybrid of LLMs and evolutionary algorithms. (1/2)
"To be a scientist is the ultimate act of charity."
โLove this framing from @Ada_Palmer on the @dwarkesh_sp podcast. Sheโs paraphrasing Francis Baconโs radical 17th-century idea: that the true purpose of science isn't just "truth," but the active relief of human suffering.
@rasbt @redtachyon @techeconomyana People often overlook the importance of open-source and weighted models in applications involving sensitive data. Policies in most US-based academic institutions prohibit researchers from uploading sensitive data to the cloud for training/inference/fine-tuning.
@MatthewBerman Unlikely, but one could argue that is an ideal outcome for humanity. We have a highly capable and helpful assistant, but we always remain in charge.
The implications of this paper seem big. There is something inherently problematic in how modern neural networks learn and form internal representations.
https://t.co/639R5OC6iX
What does humanity need more right now, a super-intelligent AI or a super-wise AI? It seems to me we are in a more dire need for the latter. Why is nobody talking about it?
Looking forward to the event! Our team will present our poster titled "A Framework for End-to-End Machine Translation of American Sign Language to English for Improved Telehealth Services for the Hearing Impaired."
Health Scholar Showcase is just around the corner! Donโt miss your chance to join us for an afternoon of engaging research at the largest #txst faculty research showcase.
RSVP here and secure your spot for the 2025 #thrcHSS: https://t.co/OTNZORhCFY
@AmandaAskell I find Claude a bit overzealous to spit out Python code when brainstorming a technical idea without being asked to do so. Often I have to explicitly tell it to not give me code as an answer to keep the conversation going.
@Dr_Singularity Knowledge acquisition is bound not just by the number of intelligent agents that operate in an environment but also by the laws of physics and the speed by which matter and energy can be transformed. Chemical reactions take time, energy is limited, etc.
@NielsHoven This is a nice parallel to pre-training versus fine tuning in terms of ML. Letting kids free to learn from random experiences but never training them to follow norms could make them super creative, but not very useful as part of a society.