Links for 2023-04-12
Remember Ted Chiang's dismissive article in The New Yorker,[1] in which he wrote that one criterion for measuring the quality of a large language model is that it be good enough to be used to train new models? He thought that such a result would require a major breakthrough. That was only two months ago!
Two months later, not only is it clear that the output of LLMs can be used to train new models,[2][3][4][5] but we're living in a world where LLMs are the foundation of goal-directed agents with access to thousands of tools.[6][7][8]
People still haven't grasped how fast progress is moving.[9] Not even hard science fiction writers.
References:
https://www.newyorker.com/tech/annals-of-technology/chatgpt-is-a-blurry-jpeg-of-the-web
https://www.lesswrong.com/posts/dcoxvEhAfYcov2LA6/agentized-llms-will-change-the-alignment-landscape
“Here was the Rosetta stone of neuroscience—a single paper and a single idea that united all the diverse and wondrous capabilities of the human mind...What brains are telling us is once you have AGI with common sense intuitions, integrating analytic reasoning on top is straightforward.” https://praxtime.com/2023/04/08/intelligence-is-prediction/
“The Segment Anything Model (SAM) by Meta AI is a step toward the first foundation model for image segmentation. SAM is capable of one-click segmentation of any object from photos or videos + zero-shot transfer to other segmentation tasks.” https://segment-anything.com/
A survey of LLM hallucinations and how to reduce them https://arxiv.org/abs/2202.03629
“To advance the state of the art of instruction-tuning for LLMs, we present the first attempt to use GPT-4 to generate instruction-following data for LLM finetuning.” https://instruction-tuning-with-gpt-4.github.io/
“We wrapped up the first LLM hackathon for applications in materials and chemistry last week. The results to me were astounding. We are at the point now where some tasks that took years can now be completed in days. Here is a list of the fantastic submissions!” https://twitter.com/BenBlaiszik/status/1644069117656444928
TagGPT: Large Language Models are Zero-shot Multimodal Taggers — “…through elaborate prompt engineering, LLMs are able to extract and reason about proper tags given textual clues of multimodal data, e.g., OCR, ASR, title, etc.” https://arxiv.org/abs/2304.03022
“Evaluating GPT-4 and ChatGPT on Japanese Medical Licensing Examinations...highlighting LLMs' potential in a language that is typologically distant from English.” https://arxiv.org/abs/2303.18027
“The first method using diffusion to close the sim2real gap for vision-based tactile sensing. Tactile Diffusion generates synthetic tactile images from sim data, capturing the complex illumination of the gel deformation.” https://arxiv.org/abs/2304.01182
Sounds emitted by plants under stress are airborne and informative https://www.cell.com/cell/fulltext/S0092-8674(23)00262-3
“Muon colliders may be the next big thing in high-energy particle physics. But they could create a dangerous beam of high-energy neutrinos!” https://mathstodon.xyz/@johncarlosbaez/110152791510451952
~100 Interesting Questions https://www.lesswrong.com/posts/pypsxBmdLpWnDErjD/100-interesting-questions
The genius of mathematician James Glimm. https://www.palladiummag.com/2023/03/02/what-genius-looks-like/
The Employment Effects of Generous and Unconditional Cash Support https://marginalrevolution.com/marginalrevolution/2023/04/the-employment-effects-of-generous-and-unconditional-cash-support.html
“Based on the available information, it appears the project is reasonably anticipated to yield a lab-generated monkeypox virus that is 1,000 times more lethal in mice than the monkeypox virus currently circulating in humans and that transmits as efficiently as the monkeypox virus currently circulating in humans.” https://republicans-energycommerce.house.gov/posts/chairs-rodgers-guthrie-and-griffith-demand-answers-on-nih-planned-experiments-using-more-lethal-group-of-monkeypox-virus
“The number of labs continues to soar. Of the most restrictive, known as Biosafety Level (BSL) 4 labs, some 69 exist or are under construction in 27 countries. Half were built in the past decade, including 10 last year.” [Washington Post] https://archive.is/rWjhD
See also:
Approximately 116 first-generation immigrants have founded American Fortune 500 companies. 106 come from Europe or the anglosphere (Canada/Australia/SA) and 6 are East Asian.
England has 21 entries, Germany 17, Canada 14, Russia 9, Ireland 8, Scotland 7, Switzerland 6, France 4, Poland 4, Taiwan 4, Sweden 3, Hungary 3, Brazil 1, Italy 1, Greece 1, Cuba 1, Zambia 1, Norway 1, Australia 1, Netherlands 1, Turkey 1, South Africa 1, Lithuania 1, Bosnia 1, Croatia 1, Finland 1, China 1, Hong Kong 1.
The Brazilian is a white Jew. The Cuban is also white. The “Turkish immigrant” is Armenian.
Source: http://startupsusa.org/fortune500/
More: The Demographics of Innovation in the United States, Page 27 http://www2.itif.org/2016-demographics-of-innovation.pdf?_ga=2.44113956.1263383430.1513639962-1733797525.1513639962