Links for 2023-07-08
1. Statler: STATe-maintaining Language models for Embodied Reasoning https://statler-lm.github.io/
2. LLMs can reliably identify and explain robot failures given a textual summary of robot past experiences generated from raw sensory inputs. https://roboreflect.github.io/
3. Introducing VoxPoser: use LLM+VLM to label affordances and constraints directly in 3D perceptual space for zero-shot robot manipulation in the real world! https://voxposer.github.io/
4. ChatGPT for Robotics: Design Principles and Model Abilities https://arxiv.org/abs//2306.17582
5. Supersize Simulation Task Generation in Robotics with LLM: We explore an LLM pipeline to generate code for simulation tasks. https://github.com/liruiw/GenSim/blob/main/BLOG.md
6. Scaling Transformers to 1,000,000,000 Tokens? https://arxiv.org/abs//2307.02486
7. “Introducing LongLLaMA 🦙, an unlimited-context version of OpenLLaMA fine-tuned at 8k & capable of extrapolating to 256k tokens!” https://arxiv.org/abs/2307.03170
8. Self-Driving Cars Are Surveillance Cameras on Wheels https://www.schneier.com/blog/archives/2023/07/self-driving-cars-are-surveillance-cameras-on-wheels.html
9. Elephants have 257 billion neurons: three times more than human brains. But 97.5% of those are found in the cerebellum, managing their large body. In the cerebral cortex, elephants only have 5.6 billion neurons dedicated to thinking: a third of humans' https://pubmed.ncbi.nlm.nih.gov/24971054/
10. White House cautiously opens the door to study blocking sun’s rays to slow global warming https://www.politico.com/news/2023/07/01/white-house-cautiously-opens-door-to-study-blocking-suns-rays-to-slow-global-warming-ee-00104513
1. Statler: STATe-maintaining Language models for Embodied Reasoning https://statler-lm.github.io/
2. LLMs can reliably identify and explain robot failures given a textual summary of robot past experiences generated from raw sensory inputs. https://roboreflect.github.io/
3. Introducing VoxPoser: use LLM+VLM to label affordances and constraints directly in 3D perceptual space for zero-shot robot manipulation in the real world! https://voxposer.github.io/
4. ChatGPT for Robotics: Design Principles and Model Abilities https://arxiv.org/abs//2306.17582
5. Supersize Simulation Task Generation in Robotics with LLM: We explore an LLM pipeline to generate code for simulation tasks. https://github.com/liruiw/GenSim/blob/main/BLOG.md
6. Scaling Transformers to 1,000,000,000 Tokens? https://arxiv.org/abs//2307.02486
7. “Introducing LongLLaMA 🦙, an unlimited-context version of OpenLLaMA fine-tuned at 8k & capable of extrapolating to 256k tokens!” https://arxiv.org/abs/2307.03170
8. Self-Driving Cars Are Surveillance Cameras on Wheels https://www.schneier.com/blog/archives/2023/07/self-driving-cars-are-surveillance-cameras-on-wheels.html
9. Elephants have 257 billion neurons: three times more than human brains. But 97.5% of those are found in the cerebellum, managing their large body. In the cerebral cortex, elephants only have 5.6 billion neurons dedicated to thinking: a third of humans' https://pubmed.ncbi.nlm.nih.gov/24971054/
10. White House cautiously opens the door to study blocking sun’s rays to slow global warming https://www.politico.com/news/2023/07/01/white-house-cautiously-opens-door-to-study-blocking-suns-rays-to-slow-global-warming-ee-00104513
Source: The Sounds Of Invisible Worlds https://www.noemamag.com/the-sounds-of-invisible-worlds/
See also: Hindsight Devalues Science https://www.lesswrong.com/posts/WnheMGAka4fL99eae/hindsight-devalues-science
See also: Hindsight Devalues Science https://www.lesswrong.com/posts/WnheMGAka4fL99eae/hindsight-devalues-science
Moriori genocide: https://en.m.wikipedia.org/wiki/Moriori_genocide
“Younger members argued that the Moriori could fight back as they outnumbered Māori two-to-one. Elders, however, argued Nunuku's Law should not be broken. Despite knowing Māori were not pacifist, Moriori ultimately decided to stay pacifist against the invaders, describing Nunuku's Law as "a moral imperative"...The invaders killed around 10% of the population in a ritual that included staking out women and children on the beach and leaving them to die in great pain over several days...the Māori invaders forbade the speaking of the Moriori language. They forced Moriori to desecrate sacred sites by urinating and defecating on them. Moriori were forbidden to marry Moriori or Māori or to have children.”
“Younger members argued that the Moriori could fight back as they outnumbered Māori two-to-one. Elders, however, argued Nunuku's Law should not be broken. Despite knowing Māori were not pacifist, Moriori ultimately decided to stay pacifist against the invaders, describing Nunuku's Law as "a moral imperative"...The invaders killed around 10% of the population in a ritual that included staking out women and children on the beach and leaving them to die in great pain over several days...the Māori invaders forbade the speaking of the Moriori language. They forced Moriori to desecrate sacred sites by urinating and defecating on them. Moriori were forbidden to marry Moriori or Māori or to have children.”
😢8🤯2👍1
This media is not supported in your browser
VIEW IN TELEGRAM
"Once, the way to get ahead among India's Konyak warriors was by chopping off an enemy's skull. Today, the last survivors of a past age mourn the changing times and a new generation they see as soft."
🤔5😇1
Links for 2023-07-09
1. CodeVQA: A framework for few-shot visual question answering that relies on code generation to perform multi-step visual reasoning. https://ai.googleblog.com/2023/07/modular-visual-question-answering-via.html
2. “We borrowed CFG from vision and run it with LLMs. We get increased control, and benchmarks increases similar to a model twice the size. Ready for all your models (incl. chatbots!) : no special training or fine tuning required.” https://arxiv.org/abs/2306.17806
3. “Dive into the fascinating world of Transformer models! Luis Serrano breaks down the architecture & functionality of these ML marvels in this blog. You'll learn how they maintain context, generate coherent text, & much more!” https://txt.cohere.com/what-are-transformer-models/
4. Building Blocks of Memory in the Brain https://www.youtube.com/watch?v=X5trRLX7PQY
5. Preserving extracellular space for high-quality optical and ultrastructural studies of whole mammalian brains https://www.cell.com/cell-reports-methods/fulltext/S2667-2375(23)00149-2
6. Researchers Create Highly Conductive Metallic Gel for 3D Printing https://news.ncsu.edu/2023/07/metallic-gel-for-3d-printing/
7. One Shot of a Kidney Protein Gave Monkeys a Brain Boost https://www.wired.com/story/one-shot-of-a-kidney-protein-gave-monkeys-a-brain-boost/ [https://archive.is/EzaPy]
8. “Be one of the first people in 40+ years to see this freshly-scanned film (in 4K) about the PM-1 military microreactor that provided a remote radar station near Sundance, WY with heat and 1 MWe of electricity in 1962.” https://whatisnuclear.com/news/2023-07-03-film-about-pm1-microreactor-in-wyoming.html
9. American mink regrow their brains in a rare reversal of the domestication process https://www.mpg.de/20584072/0705-ornr-feral-genius-987453-x
10. “Ontario has announced that they are going to build ~20% more clean (nuclear) power than all of California solar power last year. I am shocked that I haven’t seen any coverage on this beyond niche Canadian sources…” https://twitter.com/GBruhaug/status/1677518485948100611
1. CodeVQA: A framework for few-shot visual question answering that relies on code generation to perform multi-step visual reasoning. https://ai.googleblog.com/2023/07/modular-visual-question-answering-via.html
2. “We borrowed CFG from vision and run it with LLMs. We get increased control, and benchmarks increases similar to a model twice the size. Ready for all your models (incl. chatbots!) : no special training or fine tuning required.” https://arxiv.org/abs/2306.17806
3. “Dive into the fascinating world of Transformer models! Luis Serrano breaks down the architecture & functionality of these ML marvels in this blog. You'll learn how they maintain context, generate coherent text, & much more!” https://txt.cohere.com/what-are-transformer-models/
4. Building Blocks of Memory in the Brain https://www.youtube.com/watch?v=X5trRLX7PQY
5. Preserving extracellular space for high-quality optical and ultrastructural studies of whole mammalian brains https://www.cell.com/cell-reports-methods/fulltext/S2667-2375(23)00149-2
6. Researchers Create Highly Conductive Metallic Gel for 3D Printing https://news.ncsu.edu/2023/07/metallic-gel-for-3d-printing/
7. One Shot of a Kidney Protein Gave Monkeys a Brain Boost https://www.wired.com/story/one-shot-of-a-kidney-protein-gave-monkeys-a-brain-boost/ [https://archive.is/EzaPy]
8. “Be one of the first people in 40+ years to see this freshly-scanned film (in 4K) about the PM-1 military microreactor that provided a remote radar station near Sundance, WY with heat and 1 MWe of electricity in 1962.” https://whatisnuclear.com/news/2023-07-03-film-about-pm1-microreactor-in-wyoming.html
9. American mink regrow their brains in a rare reversal of the domestication process https://www.mpg.de/20584072/0705-ornr-feral-genius-987453-x
10. “Ontario has announced that they are going to build ~20% more clean (nuclear) power than all of California solar power last year. I am shocked that I haven’t seen any coverage on this beyond niche Canadian sources…” https://twitter.com/GBruhaug/status/1677518485948100611
👍4
“I cannot overemphasize how little there was to do before we all had smartphones. A barren expanse of empty time would stretch out before you: waiting for the bus, or for someone to come home, or for the next scheduled event to start. Someone might be late or take longer than expected, but no notice of such delay would arrive, so you’d stare out the window, hoping to see some sign of activity down the block. You’d pace, or sulk, or stew.”
What Did People Do Before Smartphones? https://www.theatlantic.com/technology/archive/2023/07/before-smartphones-boredom/674631/ [https://archive.is/pUmLx]
What Did People Do Before Smartphones? https://www.theatlantic.com/technology/archive/2023/07/before-smartphones-boredom/674631/ [https://archive.is/pUmLx]
Links for 2023-07-10
1. “Want generalization? Train your LM for longer! Here, we find that vanilla transformer LMs (no pre-training) generalize to structurally novel inputs (and acquire syntax), if you train them for much longer than needed. We call this behavior “Structural Grokking" ...” https://arxiv.org/abs/2305.18741
2. Models generating training data: huge win or fake win? https://dblalock.substack.com/p/models-generating-training-data-huge
3. The performance of Language Models is highest when relevant information appear at the beginning or end of the input context, and significantly lower otherwise. https://arxiv.org/abs/2307.03172
4. “Toyota claimed it had made a ‘technological breakthrough’… that would allow an electric vehicle powered by a solid-state battery to have a range of 1,200km and charging time of 10 minutes or less.” https://www.ft.com/content/87cb8e92-8e82-4755-8fc3-2943f8f63e1d [https://archive.is/aZVks]
5. At Last, Single-Photon Cameras Could Peer Into Your Brain https://spectrum.ieee.org/single-photon-camera
6. Risk-taking in chimpanzees mirrors risk-taking in humans: It’s a stable individual trait; it’s more common in males than females; and it peaks in early adulthood. https://journals.sagepub.com/doi/10.1177/09567976221140326
7. Sappers risk their lives to win Ukraine back, inch by inch: ““We don’t have air cover or plentiful Western equipment, so we risk our people. Small group tactics, like the Finns used in the winter war [of 1939-40].” Surprise tactics were the only way to breach an obstacle course like no other built since the second world war, he said.” https://www.economist.com/europe/2023/07/09/sappers-risk-their-lives-to-win-ukraine-back-inch-by-inch [https://archive.is/Xwm13]
8. Rheinmetall kamikaze drone carrier https://www.youtube.com/watch?v=Cfi7sdGnJPY
9. Rheinmetall Mission Master CXT – Fire Support, an autonomous counter-UAS solution https://www.youtube.com/watch?v=uYhlNBfMZl4
1. “Want generalization? Train your LM for longer! Here, we find that vanilla transformer LMs (no pre-training) generalize to structurally novel inputs (and acquire syntax), if you train them for much longer than needed. We call this behavior “Structural Grokking" ...” https://arxiv.org/abs/2305.18741
2. Models generating training data: huge win or fake win? https://dblalock.substack.com/p/models-generating-training-data-huge
3. The performance of Language Models is highest when relevant information appear at the beginning or end of the input context, and significantly lower otherwise. https://arxiv.org/abs/2307.03172
4. “Toyota claimed it had made a ‘technological breakthrough’… that would allow an electric vehicle powered by a solid-state battery to have a range of 1,200km and charging time of 10 minutes or less.” https://www.ft.com/content/87cb8e92-8e82-4755-8fc3-2943f8f63e1d [https://archive.is/aZVks]
5. At Last, Single-Photon Cameras Could Peer Into Your Brain https://spectrum.ieee.org/single-photon-camera
6. Risk-taking in chimpanzees mirrors risk-taking in humans: It’s a stable individual trait; it’s more common in males than females; and it peaks in early adulthood. https://journals.sagepub.com/doi/10.1177/09567976221140326
7. Sappers risk their lives to win Ukraine back, inch by inch: ““We don’t have air cover or plentiful Western equipment, so we risk our people. Small group tactics, like the Finns used in the winter war [of 1939-40].” Surprise tactics were the only way to breach an obstacle course like no other built since the second world war, he said.” https://www.economist.com/europe/2023/07/09/sappers-risk-their-lives-to-win-ukraine-back-inch-by-inch [https://archive.is/Xwm13]
8. Rheinmetall kamikaze drone carrier https://www.youtube.com/watch?v=Cfi7sdGnJPY
9. Rheinmetall Mission Master CXT – Fire Support, an autonomous counter-UAS solution https://www.youtube.com/watch?v=uYhlNBfMZl4
This media is not supported in your browser
VIEW IN TELEGRAM
AI image recognition models are powering the world’s next agricultural workforce:
Watch as these drones use multispectral color grading to determine the ripeness + sugar content of apples, then gently pick them.
Comment by user:davidad: "It may be counterintuitive, but if you want to work high on trees, it is overall cheaper to fly a light platform than to have a massive skeleton that can reach high while remaining rigid.
Birds and bees are more common in nature than giraffes."
Robot arms are ~$10k for 1 meter reach. Quadrotors are ~$200.
Watch as these drones use multispectral color grading to determine the ripeness + sugar content of apples, then gently pick them.
Comment by user:davidad: "It may be counterintuitive, but if you want to work high on trees, it is overall cheaper to fly a light platform than to have a massive skeleton that can reach high while remaining rigid.
Birds and bees are more common in nature than giraffes."
Robot arms are ~$10k for 1 meter reach. Quadrotors are ~$200.
🔥13🤮3❤2
This media is not supported in your browser
VIEW IN TELEGRAM
"I've asked an AI to generate a trailer for a HEIDI movie and now I can never sleep again"
https://twitter.com/karpi/status/1678321009638637568
https://twitter.com/karpi/status/1678321009638637568
😐10🤣3🔥1
Worth repeating now that Turkey gave the green light to Sweden's NATO bid: https://xn--r1a.website/axisofordinary/4732
😁5😢4
Links for 2023-07-11
1. Teaching Arithmetic to Small Transformers: Isolates the factors that contribute to fast emergence of arithmetic capabilities from next-token prediction. Highlights the importance of high-quality, instructive data, which can significantly improve accuracy. https://arxiv.org/abs/2307.03381
2. Eric Drexler has written '“Reframing Superintelligence” + LLMs + 4 years', a look at how things have moved in the 4 years/centuries since the original CAIS paper. https://www.lesswrong.com/posts/LxNwBNxXktvzAko65/reframing-superintelligence-llms-4-years
3. AI and CRISPR Precisely Control Gene Expression https://www.nyu.edu/about/news-publications/news/2023/july/ai-crispr-gene-expression.html
4. Tree-of-thought prompting expands on chain-of-thought by causing language models to consider multiple reasoning paths in the process of generating an output. https://arxiv.org/abs/2305.10601
5. “What I think LLMs are doing: 1. LLMs do both of the things that their promoters and detractors say they do. 2. It is very difficult from the outside to tell which they are doing. 3. Both of them are useful.” https://moultano.wordpress.com/2023/06/28/the-many-ways-that-digital-minds-can-know/
6. Eric Schmidt: This is how AI will transform the way science gets done https://www.technologyreview.com/2023/07/05/1075865/eric-schmidt-ai-will-transform-science/ [https://archive.is/wip/trkjv]
7. Eric Schmidt, reflecting on a recent trip to Ukraine, praises FPV drones as "cheaper than a mortar round" (he says $400) and "more accurate than artillery fire" (depends on round & EW). He says the FPV drones carry up to 1.3kg of explosive payload. https://www.wsj.com/articles/the-future-of-war-has-come-in-ukraine-drone-swarms-kamikaze-kyiv-31dd19d7 [https://archive.is/pgcpC]
8. One Policy to Dress Them All: Learning to Dress People with Diverse Poses and Garments https://sites.google.com/view/one-policy-dress
9. “Ways I Expect AI Regulation To Increase Extinction Risk” https://www.lesswrong.com/posts/6untaSPpsocmkS7Z3/ways-i-expect-ai-regulation-to-increase-extinction-risk
10. “Multiple lines of evidence suggest women gained consciousness and self-awareness first…Women with an IQ of 70 recognize faces as well as men with an IQ of 130. For women, the task is intuitive. This is their evolved niche. For men the task is effortful” https://www.vectorsofmind.com/p/eve-theory-of-consciousness-v2
11. According to this analysis, between 40,000-55,000 Russian soldiers KIA at the end of May 2023. If we add DPR - LPR KIA to their work, we are probably closing in on 65,000-70,000 total KIA. https://en.zona.media/article/2023/07/10/stats
1. Teaching Arithmetic to Small Transformers: Isolates the factors that contribute to fast emergence of arithmetic capabilities from next-token prediction. Highlights the importance of high-quality, instructive data, which can significantly improve accuracy. https://arxiv.org/abs/2307.03381
2. Eric Drexler has written '“Reframing Superintelligence” + LLMs + 4 years', a look at how things have moved in the 4 years/centuries since the original CAIS paper. https://www.lesswrong.com/posts/LxNwBNxXktvzAko65/reframing-superintelligence-llms-4-years
3. AI and CRISPR Precisely Control Gene Expression https://www.nyu.edu/about/news-publications/news/2023/july/ai-crispr-gene-expression.html
4. Tree-of-thought prompting expands on chain-of-thought by causing language models to consider multiple reasoning paths in the process of generating an output. https://arxiv.org/abs/2305.10601
5. “What I think LLMs are doing: 1. LLMs do both of the things that their promoters and detractors say they do. 2. It is very difficult from the outside to tell which they are doing. 3. Both of them are useful.” https://moultano.wordpress.com/2023/06/28/the-many-ways-that-digital-minds-can-know/
6. Eric Schmidt: This is how AI will transform the way science gets done https://www.technologyreview.com/2023/07/05/1075865/eric-schmidt-ai-will-transform-science/ [https://archive.is/wip/trkjv]
7. Eric Schmidt, reflecting on a recent trip to Ukraine, praises FPV drones as "cheaper than a mortar round" (he says $400) and "more accurate than artillery fire" (depends on round & EW). He says the FPV drones carry up to 1.3kg of explosive payload. https://www.wsj.com/articles/the-future-of-war-has-come-in-ukraine-drone-swarms-kamikaze-kyiv-31dd19d7 [https://archive.is/pgcpC]
8. One Policy to Dress Them All: Learning to Dress People with Diverse Poses and Garments https://sites.google.com/view/one-policy-dress
9. “Ways I Expect AI Regulation To Increase Extinction Risk” https://www.lesswrong.com/posts/6untaSPpsocmkS7Z3/ways-i-expect-ai-regulation-to-increase-extinction-risk
10. “Multiple lines of evidence suggest women gained consciousness and self-awareness first…Women with an IQ of 70 recognize faces as well as men with an IQ of 130. For women, the task is intuitive. This is their evolved niche. For men the task is effortful” https://www.vectorsofmind.com/p/eve-theory-of-consciousness-v2
11. According to this analysis, between 40,000-55,000 Russian soldiers KIA at the end of May 2023. If we add DPR - LPR KIA to their work, we are probably closing in on 65,000-70,000 total KIA. https://en.zona.media/article/2023/07/10/stats
👍3😢2❤1
BRICS is a joke, guys, wake up:
- Brazil is what could happen to France after another half century of mass immigration.
- South Africa... oh, comon now! Do I really need to say anything about that?
- What has Putin accomplished after decades of ruling Russia? Stop NATO expansion? 🤣 Belgium, the Netherlands and Luxembourg combined have a higher GDP than Russia, despite having 5 times the population and huge amounts of natural resources. Russian men have the average life expectancy of third world countries.
- India has 1.5 billion people. What Indian products do you use? What cultural influence do they have? And they get their weapons from a Chinese vassal state, while fist-fighting Chinese troops.
- China hasn't even managed to steal enough technology to be at the level of the Netherlands in semiconductor technology. They are in conflict with most of their neighbors, many of whom hate them. The US is much more popular in countries like Vietnam than China. But what about their leadership? Well, their zero-COVID policy has been the most draconian and economically damaging of any country. Chinese students want to study and immigrate to the United States.
BRICS can only dream of having companies like OpenAI or SpaceX. And what little they have is often copied or stolen. Their military technology wouldn't fly without the smuggling of Western technology.
- Brazil is what could happen to France after another half century of mass immigration.
- South Africa... oh, comon now! Do I really need to say anything about that?
- What has Putin accomplished after decades of ruling Russia? Stop NATO expansion? 🤣 Belgium, the Netherlands and Luxembourg combined have a higher GDP than Russia, despite having 5 times the population and huge amounts of natural resources. Russian men have the average life expectancy of third world countries.
- India has 1.5 billion people. What Indian products do you use? What cultural influence do they have? And they get their weapons from a Chinese vassal state, while fist-fighting Chinese troops.
- China hasn't even managed to steal enough technology to be at the level of the Netherlands in semiconductor technology. They are in conflict with most of their neighbors, many of whom hate them. The US is much more popular in countries like Vietnam than China. But what about their leadership? Well, their zero-COVID policy has been the most draconian and economically damaging of any country. Chinese students want to study and immigrate to the United States.
BRICS can only dream of having companies like OpenAI or SpaceX. And what little they have is often copied or stolen. Their military technology wouldn't fly without the smuggling of Western technology.
👍17😁4❤2🤔1🐳1🍾1👻1