Links for 2023-05-26
A Paralyzed Man Can Walk Naturally Again With Brain and Spine Implants
It is a “digital bridge” between the brain and the spinal cord, bypassing injured sections. They use a recursive exponentially weighted Markov-switching multilinear model algorithm to decode the intention to perform lower limb movements.
New York Times: https://archive.is/PMnEi
Nature: https://archive.is/98uMj
Paper: https://www.nature.com/articles/s41586-023-06094-5
New superbug-killing antibiotic discovered using AI https://news.mit.edu/2023/using-ai-scientists-combat-drug-resistant-infections-0525
Improving Factuality and Reasoning in Language Models through Multiagent Debate https://composable-models.github.io/llm_debate/
Repurposing LLMs As Both World Model & Reasoning Agent: “On plan generation, numerical reasoning, and logical reasoning tasks, RAP (our framework) outperforms previous methods including CoT/Least-to-Most + self-consistency. Specifically, RAP (LLaMA-33B) even outperforms GPT-4 in a setting of Blocksworld.” https://arxiv.org/abs/2305.14992
An efficient finetuning approach that reduces memory usage enough to finetune a 65B model on a single 48GB GPU while preserving full 16-bit finetuning task performance. https://arxiv.org/abs/2305.14314
Goat: Fine-tuned LLaMA Outperforms GPT-4 on Arithmetic Tasks https://arxiv.org/abs/2305.14201
Resolving code review comments with ML: “…we describe applying recent advances of large sequence models in a real-world setting to automatically resolve code review comments in the day-to-day development workflow at Google…” https://ai.googleblog.com/2023/05/resolving-code-review-comments-with-ml.html?m=1
“GPT-4/PaLM-2 have both shown almost perfect performance on existing grade school math dataset. What about more challenging STEM questions, especially the ones which require specific theorems, like Stoke's theorem, Wiener Process, etc? ... We found that GPT-4 is performing surprisingly well on TheoremQA. It achieves 51% accuracy with PoT and 43% with CoT. All the existing open-source LLMs are only getting 10-14% accuracy.” https://twitter.com/WenhuChen/status/1660832837715611648
“Introducing LeTI, a new LM finetuning paradigm that explores LMs' potential to learn from textual interactions & feedback, allowing LMs to understand not just if they were wrong, but why.” https://twitter.com/xingyaow_/status/1659246169950650391
Can we use LLM to Speedup LLM inference? https://zhengzangw.github.io/blogs/seqsch/
A Generalist Dynamics Model for Control https://arxiv.org/abs/2305.10912
Apple’s upcoming Personal Voice feature can clone your voice in 15 minutes using AI https://www.macworld.com/article/1919264/ios-17-macos-14-accessibility-live-speech-personal-voice.html
"While all attention was focused on Neuralink, Paradromics quietly developed a breakthrough device and technology" https://www.prnewswire.com/news-releases/paradromics-raises-33-million-in-funding-achieves-breakthrough-medical-device-designation-from-fda-301827969.html
Once again, ideology distorts science: The editor-in-chief of Scientific American flubs big time, wrongly asserting that sparrows have four sexes. https://whyevolutionistrue.com/2023/05/18/once-again-ideology-distorts-science-the-editor-in-chief-of-scientific-american-flubs-big-time-wrongly-asserting-that-sparrows-have-four-sexes/
The Oldest Known Blueprints Depict Stone Age ‘Megastructures’ https://www.vice.com/en/article/pkagbb/the-oldest-known-blueprints-depict-stone-age-megastructures
Videos showing Ukrainian maritime drones attacking the Ivan Khurs SIGINT intelligence collection ship 400 km from the Ukrainian shore: 1. https://twitter.com/NOELreports/status/1661720194673156097 2. https://twitter.com/kromark/status/1661382232190140416
“Desperate air strike attempts by Russian Su-34 fighter jet on Ukrainian-backed forces who had settled at Grayvoron checkpoint in Belgorod region. Su-34 flies at ultra-low altitude to drop high-explosive free-falling bombs. The target was clearly not hit.” https://twitter.com/clashreport/status/1661819276439724046
Here's a reminder that there's really no need to panic about climate change. As technology improves, we will increasingly be able to mitigate its negative effects while decoupling growth from greenhouse gas emissions.
Of course, as with everything we do, there is a small chance that the consequences will be catastrophic. But that can be offset by spending money to develop and build new nuclear reactors. And if things really go wrong, we always have the option of geoengineering, such as pumping sulfur dioxide into the stratosphere.
In any case, curbing economic growth is not a viable answer because it cannot be enforced, and it would mean denying people massive amounts of wealth and the benefits of technological progress. Imagine if we had stopped economic growth in 1850. Would we really be better off with a world that was 1.1 °C cooler, but economically and technologically at 1850 levels?
P.S. Without space rockets, we wouldn't even be able to launch climate-monitoring satellites that allow us to understand and verify our environmental impact. But you cannot build rockets and satellites without an advanced technological civilization, which exists only because of economic growth.
The Effects of Immigration in Denmark
https://inquisitivebird.substack.com/p/the-effects-of-immigration-in-denmark