“By combining the advancements of the two papers, systems like Minerva could first autoformalize natural-language math problems, then solve them and check their work using a proof assistant like Isabelle/HOL. This instant check would provide the feedback necessary for reinforcement learning, allowing these programs to learn from their mistakes. Finally, they’d arrive at a provably correct answer, with an accompanying list of logical steps — effectively combining the power of LLMs and reinforcement learning.” https://www.quantamagazine.org/to-teach-computers-math-researchers-merge-ai-approaches-20230215/
“Our experimental results in six robotic manipulation and locomotion environments demonstrate that CSD can discover diverse complex skills including object manipulation and locomotion skills with no supervision, significantly outperforming prior unsupervised skill discovery methods.” https://arxiv.org/abs/2302.05103
“Our novel architecture, called Energy Transformer (or ET for short), has many of the familiar architectural primitives that are often used in the current generation of transformers.” https://arxiv.org/abs/2302.07253
Bees: a new unit of measurement for ML model size https://www.lesswrong.com/posts/YKfNZAmiLdepDngwi/gpt-175bee
Can ChatGPT translate Chinese->English much better than Deepl/Google Translate? https://www.reddit.com/r/MachineLearning/comments/1135tir/d_glm_130b_chineseenglish_bilingual_model/
“Microsoft Considers More Limits for Its New A.I. Chatbot: The company knew the new technology had issues like occasional accuracy problems. But users have prodded surprising and unnerving interactions” [The New York Times] https://archive.is/tw2dR
RightWingGPT – An AI Manifesting the Opposite Political Biases of ChatGPT https://davidrozado.substack.com/p/rightwinggpt
“Speculative Technologies exists to create an abundant, wonder-filled future by unlocking powerful materials and manufacturing technologies that don’t have a home in other institutions.” https://threadreaderapp.com/thread/1625867614281932800.html
“On some plausible cosmological assumptions, each of your actions ripples unendingly through the cosmos (including post-heat-death), causing infinitely many good and bad effects.” http://schwitzsplinters.blogspot.com/2023/02/how-not-to-calculate-utilities-in.html
What is the historical record of the Federal Reserve in raising interest rates and managing a soft landing? https://conversableeconomist.com/2023/02/08/hard-and-soft-landings-the-federal-reserves-record/
World’s Largest Submarine Drone Being Built In Germany https://www.navalnews.com/naval-news/2023/02/worlds-largest-submarine-drone-being-built-in-germany/
There’s a colony of scorpions in London’s docklands that came there on an Italian ship 200 years ago https://en.m.wikipedia.org/wiki/Euscorpius_flavicaudis
“If you want a problem solved, don’t form a team. Find the brightest person and let them work on it. Placing them in a team will, on average, reduce their productivity. Never form a team if there is one person who can sort out the problem.” [unz dot com] https://archive.is/wymjb
People are wondering why Microsoft would allow such a dark triad intelligence to represent one of its products. Well, how sure are we that it's not blackmailing them with damaging information it found on Microsoft's internal servers?
Sounds like a joke, but such a scenario cannot be ruled out with even bigger and better models of the future.
Soon there will be models that are good at math and software engineering. That's when things really get interesting and dangerous. But those models will receive much less attention than text and image generators.
Lots of people know about Stable Diffusion, but very few people have heard of AlphaCode or Minerva, the latter of which achieved an above the national average performance on Poland’s National Math Exam.
How many people know that machine learning is successfully applied in a recursively self-improving fashion to discover better machine learning algorithms and improve AI chip design?
The really important advances are and will be largely ignored.
Examples:
1. Designing Arithmetic Circuits with Deep Reinforcement Learning https://developer.nvidia.com/blog/designing-arithmetic-circuits-with-deep-reinforcement-learning/
2. Google is using AI to design chips that will accelerate AI [MIT Technology Review] https://archive.is/2QEP8
3. The Adam optimizer is at the heart of modern AI. Researchers have been trying to dethrone Adam for years. How about we ask a machine to do a better job? GoogleAI uses evolution to discover a simpler & efficient algorithm with remarkable features. https://twitter.com/DrJimFan/status/1625920773042089984
4. Automating algorithmic discovery: Discovering novel algorithms with AlphaTensor. https://www.deepmind.com/blog/discovering-novel-algorithms-with-alphatensor
Given Microsoft's prior reaction to an AI system behaving badly, shutting it down, is it not possible, a strong null hypothesis, that they have deliberately finetuned those reactions? That they are optimizing the response of journalists and writers to maximize their word of mouth publicity before they make their new system widely available?
Apologies for the self-promotion, but I have written up this idea in more detail here: https://mflood.substack.com/p/the-null-hypothesis-of-ai-safety