Links for 2022-05-07
Ever wonder why a Bayesian brain should have so much trouble understanding Bayes' theorem? This preprint offers two complementary answers to this question. https://psyarxiv.com/hne9s
DeepMind’s New AI Finally Enters The Real World! https://youtu.be/zxyZSxnTrZs
This Two-Inch Diamond Disc Could Hold a Staggering Billion Blu-Ray’s Worth of Data https://gizmodo.com/quantum-computing-diamond-disc-could-store-billion-blu-1848853029
World’s Smallest Gears Measure Mere Nanometers to Power Molecular Machines https://newatlas.com/science/world-smallest-gears-molecular-machines/
“The Dream Chaser spaceplane, Tenacity, is taking shape and will be the first vehicle in our Dream Chaser fleet of orbital vehicles. Making space more accessible one innovation at a time.” https://spacenews.com/first-dream-chaser-vehicle-takes-shape/
Strong selection during the last millennium for African ancestry in the admixed population of Madagascar https://www.nature.com/articles/s41467-018-03342-5
You Can Now Ask Google to Remove Your Phone Number, Email or Address from Search Results https://krebsonsecurity.com/2022/04/you-can-now-ask-google-to-remove-your-phone-number-email-or-address-from-search-results/
Postoperative cognitive dysfunction: "Cibelli listened as the woman described how her father, a former physics professor, had shown signs of significant cognitive decline after the initial operation. Once a keen chess player, he was now unable to play the game and struggled to even do basic crosswords...At the moment, estimates suggest that the overall incidence of POCD in older patients can be as high as 50-80% at discharge, 20-50% at six weeks and 10-30% at six months post-surgery." https://www.theguardian.com/science/2022/apr/24/the-hidden-long-term-risks-of-surgery-it-give-peoples-brains-a-hard-time
An experiment conducted on hybrid matter-antimatter atoms has defied researchers’ expectations. https://www.wired.com/story/an-antimatter-experiment-shows-surprises-near-absolute-zero/ (archived version: https://archive.ph/mGjkZ)
New and Surprising Duality Discovered in Theoretical Particle Physics https://scitechdaily.com/new-and-surprising-duality-discovered-in-theoretical-particle-physics/
I remember sharing the post mentioned in the following tweet on Google+. I was also convinced that such capabilities are probably very far away. It's astonishing and scary how wrong I was. This is part of the reason I changed my mind about risks from AI.
Sadly, I haven't recorded all the updates I made towards taking AI risks more seriously after that 2012 post. But here is a rough overview:
1. The deep learning revolution and evidence that scaling a certain neural network architecture yields performance returns even in the absence of any fine-tuning or algorithmic improvement (which is entirely feasible).
2. The absence of better counter-arguments by experts. When I started criticizing the AI risk narrative I expected that experts would have much better counter-arguments than I. It turned out that their arguments are worse.
3. Foundational research seems to be progressing very well with no signs of slowing down. Algorithmic advances and synergies between neuroscience and machine learning will fasten progress even further.
4. China understands the significance of artificial intelligence. Well-resourced government-linked researchers in China want to build some *really* big neural network models. They might be throwing millions of smart people at it soon. Asians are generally taking AI increasingly seriously and have the human capital to brute force more breakthroughs.
5. Many chipmakers are now focusing on specialized AI processors. A lot of promising research in this area will give the field another boost. Even more importantly, Google has been successfully using AI to design its next generation of AI chips.
6. A bunch of important people in the field are now frequently mentioning AGI non-ironically. Google DeepMind believes that AGI might only be one or two decades away.
7. Artificial neural networks are not just getting predictably better with more training, data, and parameters but sometimes capabilities emerge in a jumpy and unpredictable way. There can be sudden phase transitions. This means that it cannot be ruled out that artificial general intelligence could appear even earlier than expected by teams such as DeepMind.