Links for 2023-11-23

1. "With Q*, OpenAI have likely solved planning/agentic behavior for small models. Scale this up to a very large model and you can start planning for increasingly abstract goals. It is a fundamental breakthrough that is the crux of agentic behavior." https://www.lesswrong.com/posts/JnM3EHegiBePeKkLc/possible-openai-s-q-breakthrough-and-deepmind-s-alphago-type

2. How much to update on recent AI governance moves? https://www.lesswrong.com/posts/HAmH8bzDNrHDwmcPb/how-much-to-update-on-recent-ai-governance-moves

3. After closed door OpenAI drama, open source AI backed by Nvidia, Eric Schmidt, could emerge winner https://www.cnbc.com/2023/11/22/in-openai-fallout-open-source-ai-could-be-among-the-big-tech-winners.html

4. Stable Video Diffusion: Scaling Latent Video Diffusion Models to Large Datasets https://stability.ai/research/stable-video-diffusion-scaling-latent-video-diffusion-models-to-large-datasets

5. GPT-4V(ision) for Robotics: Multimodal Task Planning from Human Demonstration https://microsoft.github.io/GPT4Vision-Robot-Manipulation-Prompts/

6. GPT4Motion: Scripting Physical Motions in Text-to-Video Generation via Blender-Oriented GPT Planning https://arxiv.org/abs/2311.12631

7. “Can we boost chain-of-thought reasoning by guiding decoding toward correct solutions? Excited to present 🌟GRACE🌟: a guided decoding method that outperforms greedy decoding, self-consistency, and verifiers on math and symbolic reasoning.” https://arxiv.org/abs/2305.14934

8. UNcommonsense Reasoning: Abductive Reasoning about Uncommon Situations https://arxiv.org/abs/2311.08469

9. LLM System 2 Attention: S2A uses the full reasoning power of LLMs via generation to make complex attention decisions when soft attention fails. https://arxiv.org/abs/2311.11829

10. The Chosen One: Consistent Characters in Text-to-Image Diffusion Models https://omriavrahami.com/the-chosen-one/

11. “If you only get about five words to convey an idea, what will someone extrapolate from those five words? Rather than guess, you can use LLMs to experimentally discover what people are likely think those five words mean. You can use this to iterate on what five words you want to say in order to best convey your intended meaning.” https://www.lesswrong.com/posts/HD8kLyYcSuYRi4vzP/extrapolating-from-five-words

12. The Pentagon is moving toward letting AI weapons autonomously decide to kill humans: "We'll counter the PLA's mass with mass of our own, but ours will be harder to plan for, harder to hit, harder to beat," https://www.businessinsider.com/us-closer-ai-drones-autonomously-decide-kill-humans-artifical-intelligence-2023-11

13. Why and How to Dollarize Argentina https://cpsi.media/p/why-and-how-to-dollarize-argentina

14. Löwenheim–Skolem theorem: “It means that there's no set of global statements about local relations that pins down which infinity you're talking about. Either the continuous numbers of physics are an illusion, or physics runs on second-order logic, or ???” — Eliezer Yudkowsky https://plato.stanford.edu/entries/paradox-skolem/

Fact-Check ✅

Links for 2023-11-24

1. [1hr Talk] Intro to Large Language Models https://www.youtube.com/watch?v=zjkBMFhNj_g

2. The mind’s eye of a neural network system: “The tool we’ve developed helps you find places where the network is saying, ‘Hey, I need more information to do what you’ve asked.’ I would advise people to use this tool on any high-stakes neural network decision scenarios or image prediction task.” https://www.purdue.edu/newsroom/releases/2023/Q4/the-minds-eye-of-a-neural-network-system.html

3. Is neuromorphic computing the game changer we've been waiting for? https://www.eetimes.com/podcasts/arm-inventor-steve-furber-on-spinnaker-1-2-and-beyond/

4. 5 questions for Scott Aaronson https://www.politico.com/newsletters/digital-future-daily/2023/11/17/5-questions-for-scott-aaronson-00127828

5. Search algorithm reveals nearly 200 new kinds of CRISPR systems https://news.mit.edu/2023/search-algorithm-reveals-nearly-200-new-kinds-crispr-systems-1123

6. Study finds male-female brain differences and is able to predict "individual IQ scores for males and females separately using whole-brain functional connectivity." https://academic.oup.com/cercor/article/30/3/888/5540233

7. There’s a 1,200-year-old Phone in the Smithsonian Collections https://www.smithsonianmag.com/smithsonian-institution/theres-a-1200-year-old-phone-in-the-smithsonian-collections-180947641/

8. PCP theorem: “Given a formal proof of any mathematical statements. You can verify its proof with probability 1-epsilon, by just looking at a constant number of the random bits of the proof (in a suitable encoding), where that number depends only on epsilon.” https://en.wikipedia.org/wiki/PCP_theorem

9. What's the Geometry of Numbers? - Minkowski's Theorem https://www.youtube.com/watch?v=RquUIXUMLcc

10. A tale of two problem solvers https://www.youtube.com/watch?v=ltLUadnCyi0

The price of an NVIDIA A100 and a 155 mm artillery shell are now in the same ballpark.

Try explaining to someone from World War I why a machine that can perform 312,000,000,000,000 calculations per second is not much more expensive than an artillery shell that they fire by the millions over their muddy trenches.

Links for 2023-11-25

1. GPT-4 can solve a wide variety of problems, and doesn't look very much like a goal-directed optimizer; certainly it doesn't look like a very coherent one. Are there any kinds of cognition, then, that do require "goal-directedness"? Nate Soares responds in a new post. https://www.lesswrong.com/posts/AWoZBzxdm4DoGgiSj/ability-to-solve-long-horizon-tasks-correlates-with-wanting

2. Combining AI with traditional wet lab work creates a virtuous circle from lab to data and back to the lab. https://proto.life/2023/11/perspective-the-rise-of-wet-artificial-intelligence/

3. Watermarks in the Sand: Impossibility of Strong Watermarking for Generative Models https://arxiv.org/abs/2311.04378

4. Kyutai is a French AI research lab with a $330 million budget that will make everything open source https://techcrunch.com/2023/11/17/kyutai-is-an-french-ai-research-lab-with-a-330-million-budget-that-will-make-everything-open-source/

5. “access to lots of capital, computing power and data has helped AI researchers in Abu Dhabi train up Falcon, an open-source LLM that in some ways beats Meta’s … the UAE may well be the third-most-important country for AI, after America and China.” https://www.economist.com/leaders/2023/11/23/how-to-thrive-in-a-fractured-world [https://archive.is/dYw0k]

6. Experimental brain-like computing system more accurate with custom algorithm https://newsroom.ucla.edu/releases/experimental-brain-like-computing-system-custom-algorithm

7. “Terry Tao's blog post on the PFR project is essential reading for anybody interested in formal mathematics with Lean.” https://terrytao.wordpress.com/2023/11/18/formalizing-the-proof-of-pfr-in-lean4-using-blueprint-a-short-tour/

8. Wedderburn's little theorem: a finite division ring is a field. https://en.wikipedia.org/wiki/Wedderburn%27s_little_theorem

9. In middle-aged mice: a low isoleucine diet led to longer lifespan than either a control or low-protein diet. Less cancer. The protocol was free feeding and the low isoleucine group ate more and yet was leaner. https://news.wisc.edu/mice-eating-less-of-specific-amino-acid-overrepresented-in-diet-of-obese-people-live-longer-healthier/

Links for 2023-11-26

1. “There's another whole route to scaling LLM capabilities. LLMs themselves might not need to scale up much more at all to enable AGI. Scaffolding LLMs to them to make language model cognitive architectures can fill at least some of the gaps in their abilities. Many groups are working on this.” https://www.lesswrong.com/posts/hg9SvxBQw6KFNEzrH/what-s-the-evidence-that-llms-will-scale-up-efficiently

2. “These results suggest that without any additional training, LLMs can serve as general sequence modelers, driven by in-context learning.” https://arxiv.org/abs/2307.04721

3. DROC: Distilling and Retrieving Generalizable Knowledge for Robot Manipulation via Language Corrections https://sites.google.com/stanford.edu/droc

4. LucidDreamer: Domain-free Generation of 3D Gaussian Splatting Scenes https://luciddreamer-cvlab.github.io/

5. MetaDreamer: Efficient Text-to-3D Creation With Disentangling Geometry and Texture https://arxiv.org/abs/2311.10123

6. Jorge: Approximate Preconditioning for GPU-efficient Second-order Optimization https://arxiv.org/abs/2310.12298

7. “In this work, we propose a distributed optimization algorithm, Distributed Low-Communication (DiLoCo), that enables training of language models on islands of devices that are poorly connected.” https://arxiv.org/abs/2311.08105

8. Rethinking Attention: Exploring Shallow Feed-Forward Neural Networks as an Alternative to Attention Layers in Transformers https://arxiv.org/abs/2311.10642

9. “…thanks to recent advances in artificial intelligence (AI), the latest electronic noses - high-tech sensors that can detect and report specific smells - are quickly improving their levels of speed and accuracy.” https://www.bbc.com/news/business-67354443

10. How Laws Evolved by Natural Selection https://www.psychologytoday.com/us/blog/202308/how-laws-evolved-by-natural-selection

11. “Cosmological natural selection… of universes capable of prolific reproduction. It was originally proposed that reproduction occurs through singularities resulting from supernovae, and subsequently argued that life may facilitate the production of the singularities that become offspring universes. Here I argue technology is necessary for production of singularities by living beings, and ask whether the physics of our universe has been selected to simultaneously enable stars, intelligent life, and technology capable of creating progeny. Specific technologies appear implausibly equipped to perform tasks necessary for production of singularities…” https://iopscience.iop.org/article/10.1088/1367-2630/ab9d9d/pdf

12. “How do elite groups form? The truth is that, in many cases, elite groups are genetically different from the rest of the population.” https://www.aporiamagazine.com/p/how-do-elite-groups-form

Computer scientist who had his bike stolen tried to explain binary search to the cops.

Source: https://archive.is/TgrKu

Links for 2023-11-27

1. A framework for robotic excavation and dry stone construction using on-site materials https://www.youtube.com/watch?v=TxpE5yryCTU

2. AI can figure out sewing patterns from a single photo of clothing https://www.newscientist.com/article/2404358-ai-can-figure-out-sewing-patterns-from-a-single-photo-of-clothing/ [https://archive.is/T2q7u]

3. Cell-free quest for new antibiotics: A new method combines synthetic biology with artificial intelligence https://www.nature.com/articles/s41467-023-42434-9

4. Introducing EUGENe: An Easy-to-Use Deep Learning Genomics Software https://today.ucsd.edu/story/introducing-eugene-an-easy-to-use-deep-learning-genomics-software

5. Does creatine supplementation make you smarter? The largest RCT to date. Weak Bayesian evidence for a small effect (not significant), still uncertain. Veg vs omni made no difference. Previous studies likely too small. https://www.lesswrong.com/posts/CbaznRo9fKpriw2mi/paper-out-now-on-creatine-and-cognitive-performance

6. Aaron Silverbook on anti-cavity bacteria https://www.lesswrong.com/posts/CdxdMLbBwHgEcGYCv/aaron-silverbook-on-anti-cavity-bacteria

7. Researchers 3D printed and deployed a communication satellite in 90 min https://www.uoc.edu/portal/en/news/actualitat/2023/279-nanosatellite-for-emergency-broadband.html

Links for 2023-11-28

1. GPQA, a graduate-level “Google-proof” Q&A benchmark designed for scalable oversight! GPQA is a dataset of *really hard* questions that PhDs with full access to Google can’t answer. Highly skilled non-expert humans spending 30+ mins with unrestricted web search: 34% accuracy; GPT-4: 39%; PhDs in corresponding domains: 65%. https://arxiv.org/abs/2311.12022

2. Rankitect: Ranking Architecture Search Battling World-class Engineers at Meta Scale https://arxiv.org/abs/2311.08430

3. Simplifying Transformer Blocks https://arxiv.org/abs/2311.01906

4. PhysGaussian: Physics-Integrated 3D Gaussians for Generative Dynamics https://xpandora.github.io/PhysGaussian/

5. Baidu Inc. reported a better-than-expected 6% increase in revenue, partly due to the early stages of monetizing its ChatGPT-style AI, Ernie bots https://www.bnnbloomberg.ca/baidu-s-revenue-beats-as-chatgpt-style-ai-starts-monetization-1.2001478

6. Doug Lenat's source code for AM and EURISKO (+Traveller?) found in public archives https://white-flame.com/am-eurisko.html

7. Northvolt in new sodium-ion battery breakthrough https://www.ft.com/content/826fb9bb-b27f-4c68-ad4b-a54ad74713c9 [https://archive.is/HBjrC]

8. Electric car battery prices are going back down faster than expected https://electrek.co/2023/11/20/electric-car-battery-prices-are-going-back-down-faster/

9. “Individuals with lower intelligence levels (typically measured as IQ) tend to be more likely to engage in criminal behavior," and there is "consistent evidence of a mostly linear" association between low IQ and criminality. https://psycnet.apa.org/record/2015-31412-011

10. “Caring about race differences is not a fetish: Critics accuse race realists of obsessing over race. But refuting widespread lies about race is not an intellectual failure or a fetish. It is a duty.” https://www.aporiamagazine.com/p/caring-about-race-differences-is

All SpaceX launches from Cape Canaveral in 13 years:

https://twitter.com/AlexSvanArt/status/1630896377176702977

Pika 1.0, the idea-to-video platform that brings your creativity to life: https://pika.art/

"You know how image generation went from blurry 32x32 texture patches to high-resolution images that are difficult to distinguish from real in roughly a snap of a finger? The same is now happening along the time axis (extending to video) and the repercussions boggle the mind just a bit. Every human becomes a director of multi-modal dreams, like the architect in Inception." -- Andrej Karpathy

Links for 2023-11-29

1. GPT-4 for radiology. Far from perfect, but state-of-the-art performance on some tasks: “Surprisingly, we found radiology report summaries generated by GPT-4 to be comparable and, in some cases, even preferred over those written by experienced radiologists” https://www.microsoft.com/en-us/research/blog/gpt-4s-potential-in-shaping-the-future-of-radiology/

2. NVIDIA Powers Training for Some of the Largest Amazon Titan Foundation Models https://blogs.nvidia.com/blog/nemo-amazon-titan/

3. An Embodied Generalist Agent in 3D World https://embodied-generalist.github.io/

4. Bard can now watch YouTube videos for you https://www.theverge.com/2023/11/22/23972636/bard-youtube-extension-update-search-video-content

5. Vitalik Buterin: "AI risk 1: existential risk. My p(doom) is around 0.1." https://vitalik.eth.limo/general/2023/11/27/techno_optimism.html (thread: https://twitter.com/VitalikButerin/status/1729251822391447904)

6. Apocalypse insurance, and the hardline libertarian take on AI risk https://www.lesswrong.com/posts/mSeesg7i4d9scWAet/apocalypse-insurance-and-the-hardline-libertarian-take-on-ai

7. 5 more brain-computer interface companies you need to know https://www.massdevice.com/brain-computer-interface-companies-need-to-know-2023/

8. “Hence at the macroevolutionary scale, there can be said to have been a “ten-million-year explosion” in primate G leading up to modern humans.” https://www.sciencedirect.com/science/article/abs/pii/S0160289623000764

9. The prefrontal cortex: from monkey to man https://academic.oup.com/brain/advance-article/doi/10.1093/brain/awad389/7424860

10. First of a 3-part series on advances in the aging biotech space. Stay tuned for future posts focusing on reprogramming and cell/organ replacement! https://medium.com/prime-movers-lab/aging-biotech-moving-beyond-the-hallmarks-b2c417ec59e8 [https://archive.is/wip/ICXyz]

11. No Digging Required: Space Mining on the Moon and Beyond May Be Solar Powered https://www.space.com/moon-asteroid-space-mining-with-concentrated-sunlight.html

Links for 2023-11-30

1. AI tool GNoME finds 2.2 million new crystals, including 380,000 stable materials that could power future technologies 💎 https://deepmind.google/discover/blog/millions-of-new-materials-discovered-with-deep-learning/

2. “We trained a small transformer (100M params) for basic arithmetic. W. the right training data it nails 12x12 digits multiplication w/o CoT (that's 10^24 possibilities, so no it's not memorization🤣).” https://arxiv.org/abs/2311.14737

3. ToolChain*: Efficient Action Space Navigation in Large Language Models with A* Search — Uses algorithms like A* to improve LLM answers, improving sota on both planning and reasoning tasks https://arxiv.org/abs/2310.13227

4. “Google DeepMind work creating agents that can mimic experts in goal-driven behaviour in real-time 3D environments. What sets their AI agent apart is its ability to learn from humans in real-time, adopting a third-person perspective. It mirrors the way we learn by observing others – quick and efficient!” [Note: The preprint was already published in March 2022] https://www.nature.com/articles/s41467-023-42875-2

5. Goodhart's law Example: Training Verifiers to Solve Math Word Problems https://www.lesswrong.com/posts/euwMMxwuBeS5QZkC4/goodheart-s-law-example-training-verifiers-to-solve-math

6. The Surprising Genius of Sewing Machines https://www.youtube.com/watch?v=RQYuyHNLPTQ

7. Nano-Sized Powerhouses: Ultrafast Laser Technology Miniaturized on Tiny Photonic Chips https://www.caltech.edu/about/news/ultrafast-lasers-on-ultra-tiny-chips

8. Antares is “building micro-sized nuclear reactors to provide power to remote off-grid locations,” with a vision of “abundant clean energy for all, from Earth to the asteroid belt” https://twitter.com/juliadewahl/status/1719448793027109114

9. FDA has agreed that Loyal’s first drug supports a reasonable expectation of large dog lifespan extension. https://loyalfordogs.com/posts/loyal-announces-historic-fda-milestone-for-large-dog-lifespan-extension-drug

10. In Continued Defense Of Effective Altruism https://www.astralcodexten.com/p/in-continued-defense-of-effective

11. In a sample of 259 extremely high IQ adolescents (from the top 0.01% of the population in intelligence), 37% had doctorates, 9% had patents, 39% had peer-reviewed publications, along with various other accomplishments. https://journals.sagepub.com/doi/abs/10.1177/0956797616644735

Links for 2023-12-01

1. Large-scale pancreatic cancer detection via non-contrast CT and deep learning https://www.nature.com/articles/s41591-023-02640-w

2. “Welcome to the future. ‘Over 17 days of continuous operation, the A-Lab realized 41 novel compounds from a set of 58 targets… identified using large-scale ab initio phase-stability data from the Materials Project and Google DeepMind.” https://www.nature.com/articles/s41586-023-06734-w

3. “Machine learning for neuroscience: We build interpretable biological network reconstructions from electrode recordings with ML and optimal transport.” https://arxiv.org/abs/2306.03603

4. Q: “The reports about the Q* model breakthrough that you all recently made, what’s going on there?” Altman: “No particular comment on that unfortunate leak. But what we have been saying — two weeks ago, what we are saying today, what we’ve been saying a year ago, what we were saying earlier on — is that we expect progress in this technology to continue to be rapid...You can always hit a wall, but we expect that progress will continue to be significant.” https://www.theverge.com/2023/11/29/23982046/sam-altman-interview-openai-ceo-rehired

5. Concept Sliders: LoRA Adaptors for Precise Control in Diffusion Models https://sliders.baulab.info/

6. P(doom) roundup: what probability do people put on AI killing everyone? https://twitter.com/AISafetyMemes/status/1729892336782524676

7. "E. coli bacteria use iron levels as a way to store information about different behaviors that can then be activated in response to certain stimuli... These iron memories persist for at least four generations and disappear by the seventh generation." https://news.utexas.edu/2023/11/21/bacteria-store-memories-and-pass-them-on-for-generations/

8. A combinatorially complete fitness landscape with >260.000 E.coli genotypes shows unexpected 🤯 properties. Despite ~500 fitness peaks in this landscape, adaptive evolution can easily navigate to highest fitness peaks https://www.science.org/doi/10.1126/science.adh3860

9. Genes that increase the chances of getting addicted to alcohol increase the risk of feeling lonely. Genes that increase the risk for autism increase IQ. Genes that increase the risk for anorexia increase the chances of getting a higher education 🧬🧠 [published in 2021] https://bigthink.com/health/tiny-genetic-differences-behavioral-effects/

10. Births in South Korea are down by 11.5% in Q3 2023 compared to Q3 2022. Seoul TFR: 0.54 https://twitter.com/nonebusinesshey/status/1729710339598569513

Astronomers have detected over 5,300 planets near other stars.

Here's a beautiful visualization of some of them: https://www.halcyonmaps.com/terrestrial-exoplanets/