Links for 2024-12-30

AI:

It gives me a lot of comfort knowing that we are the last generation without advanced robots everywhere. Our children will grow up as “robot natives”. They will have humanoids cook Michelin dinner, robot teddy bears tell bedtime stories, and FSD drive them to school.

— Jim Fan, NVIDIA

1. “By default, capital will matter more than ever after AGI” https://www.lesswrong.com/posts/KFFaKu27FNugCHFmh/by-default-capital-will-matter-more-than-ever-after-agi

2. China’s Deepseek V3, trained on 2048 GPUs for 2 months—a low-bandwidth-NVLink version of H100, available for China before 2023.10 export restrictions. It’s allegedly on par with GPT-4o and Claude 3.5 Sonnet. That would be a >10X cost reduction. They use Self-Rewarding: “The LLM serves as a versatile processor capable of transforming unstructured information from diverse scenarios into rewards, ultimately facilitating the self-improvement of LLMs.” Also: Multi token prediction. https://github.com/deepseek-ai/DeepSeek-V3/blob/main/DeepSeek_V3.pdf

3. There's now a clear path to super-intelligent scientists with data-driven experiment simulation and super-scaling RL. OmniPred: Language Models as Universal Regressors https://arxiv.org/abs/2402.14547

4. Byte Latent Transformer scales efficiently with byte-level encoding, outperforming tokenized models. https://arxiv.org/abs/2412.09871v1

5. How Hallucinatory A.I. Helps Science Dream Up Big Breakthroughs https://www.nytimes.com/2024/12/23/science/ai-hallucinations-science.html [no paywall: https://archive.is/zFtgu]

6. Reversible molecular simulation for training classical and machine learning force fields https://arxiv.org/abs/2412.04374

7. Let your LLM generate a few tokens and you will reduce the need for retrieval https://www.arxiv.org/abs/2412.11536

8. RLEF: Grounding Code LLMs in Execution Feedback with Reinforcement Learning https://arxiv.org/abs/2410.02089

9. “We instructed o1-preview to play to win against Stockfish. Without explicit prompting, o1 figured out it could edit the game state to win against a stronger opponent. GPT-4o and Claude 3.5 required more nudging to figure this out” https://x.com/JeffLadish/status/1872805453224448208

10. Next Token Prediction Towards Multimodal Intelligence: A Comprehensive Survey https://arxiv.org/abs/2412.18619

11. HuatuoGPT-o1, Towards Medical Complex Reasoning with LLMs https://arxiv.org/abs/2412.18925

12. Why backprop was resisted for 20 years: assumption of discretely spiking neurons, goal of synthesizing Boolean logic, fear of local optima, and bad luck. Werbos has the best claim for invention. https://yuxi-liu-wired.github.io/essays/posts/backstory-of-backpropagation/

13. OpenAI lays out its for-profit transition plans https://openai.com/index/why-our-structure-must-evolve-to-advance-our-mission/

14. 321 real-world gen AI use cases from the world's leading organizations https://cloud.google.com/transform/101-real-world-generative-ai-use-cases-from-industry-leaders

15. In a Google strategy meeting last week, Sundar Pichai and Demis Hassabis told staff that 2025 will be a critical year in AI, that new features will be introduced in the first half of the year and that they are are going to "turbo charge" the Gemini app https://www.cnbc.com/2024/12/27/google-ceo-pichai-tells-employees-the-stakes-are-high-for-2025.html

https://x.com/Miles_Brundage/status/1873205522742296873

Psychology:

1. Rationality is not distinct from cognitive ability. It is an aspect of general intelligence. Specifically, rationality consists of "the reflexive and reflective aspects of cognitive ability." https://www.sciencedirect.com/science/article/pii/S0160289624000898

2. “…each treatment improves students’ school performance by 0.09 standard deviations. This indicates that the experience of effortful thinking itself—even when devoid of any subject content—improves general cognitive capacity.” https://academic.oup.com/qje/advance-article/doi/10.1093/qje/qjae043/7925870

Miscellaneous:

1. Adam Brown (a lead of Blueshift at DeepMind & theoretical physicist at Stanford) – How Future Civilizations Could Change The Laws of Physics https://www.dwarkeshpatel.com/p/adam-brown

2. How to fix computing's AI energy problem: run everything backwards — About reversible computing, thermodynamics as the natural language for computation, and how we will get to 1000x reduction in costs in <24 months. https://www.newscientist.com/article/mg26435231-300-how-to-fix-computings-ai-energy-problem-run-everything-backwards/ [no paywall: https://archive.is/be95Q]

Biology:

1. Genetically edited mosquitoes haven't scaled yet. Why? https://eryney.substack.com/p/genetically-edited-mosquitoes-havent

2. BAAI aims to develop such closed-loop biophysically detailed models ('life models' that accurately simulate an organism's complex behavioral repertoire from its neural, biomechanical and environmental interactions). https://www.nature.com/articles/s43588-024-00738-w

3. Genetic Sequences of Highly Pathogenic Avian Influenza A(H5N1) Viruses Identified in a Person in Louisiana — The H5 virus mutated inside the single patient to gain an ability to bind human receptors in the upper respiratory tract. https://www.cdc.gov/bird-flu/spotlights/h5n1-response-12232024.html

4. Considerations on orca intelligence https://www.lesswrong.com/posts/dzLwCBvwC4hWytnus/considerations-on-orca-intelligence

“Math”:

1. How to disguise the game of tic-tac-toe into an isomorphic one and amaze people by how good you are at that game https://mathoverflow.net/questions/287917/what-is-a-chess-piece-mathematically/287927#287927

2. Is "VNM-agent" one of several options, for what minds can grow up into? https://www.lesswrong.com/posts/KSguJeuyuKCMq7haq/is-vnm-agent-one-of-several-options-for-what-minds-can-grow

3. “In 4 dimensions, you can walk through walls. GPT3 has 175 billion parameters. In 175 billion dimensions, walls are so far beneath your notice that if you observe them at all it is like God looking down upon individual protons.” https://moultano.wordpress.com/2020/10/18/why-deep-learning-works-even-though-it-shouldnt/

4. Life is just easier in higher dimensions. Unless you want to "win" https://web.archive.org/web/20200625001036/https://mailchi.mp/ribbonfarm/how-to-not-lose-at-4d-chess

How do you maximize freedom and individuality in a world of brain emulation technology:

In this far-future civilization, each human being is given access to a “secure pocket universe” they can shape in any way they wish—populated by p-zombies (beings that behave like conscious agents but lack the subjective experience of suffering). These universes are protected using advanced, open-verified encryption architectures, making them nearly impossible to hack.

When conscious humans wish to interact with one another, they do so by exchanging encrypted, time-limited copies of themselves. A strict “consent contract” sets the rules for each copy’s experience, such as maximum runtime and comfort standards. These copies can always self-terminate or revert to ensure no prolonged suffering is possible, preventing abuses like indefinite torture or confinement.

To support both personal utopias and collective security, the civilization allots 20% of the cosmic resource base to these private universes, giving each person equitable access. The remaining 80% is dedicated to bolstering defenses and advancing space colonization in preparation for encounters with “grabby aliens”—potentially hostile, expansionist civilizations. This reserve also maintains a collaborative infrastructure for voluntary cultural exchange, preventing societal stagnation and ensuring that new innovations and ideas continue to flow.

Crucially, ethics and consent remain at the center of this system. People may opt in or out of inter-universal interactions, and they maintain control over every conscious copy of themselves. By combining robust encryption, verifiable non-conscious p-zombies, carefully regulated resource use, and secure inter-civilizational protocols, humanity aims to maximize individual freedom while minimizing suffering—and to stand strong against any external cosmic threats.

What is a minimal set of axioms that you have to believe in to be able to do science and self-improvement:

Axiom 1 (Mathematical and Logical Consistency): Assume a consistent framework for reasoning, probability, and mathematics. This is a bare minimum for rational inference.

Axiom 2 (Learnability): There exist sufficiently compressible (lawlike) explanations that can unify past, present, and future observations, enabling us to distinguish better theories from worse through observation and criticism.

Axiom 3 (Resource Reasoning): When forming priors, distributing credence over an unbounded complexity space forces you to give higher effective credence to simpler theories. This ensures that simpler theories get tested first and, if successful, rapidly gain credence.

Axiom 4 (Cognitive/Observational Reliability): Our faculties (or instruments) are good enough to let us propose, refine, and test theories iteratively. We don’t need absolute certainty here—just enough trust to get started and not dismiss the entire enterprise.

Politics:

1. Russia is using bitcoin in foreign trade, finance minister says https://www.reuters.com/markets/currencies/russia-is-using-bitcoin-foreign-trade-finance-minister-says-2024-12-25/

2. “Wild French energy story: Greenpeace fought, and won, a battle to switch new French homes from electric heat to natural gas heat.” https://x.com/energybants/status/1873098180109492689

Ukraine:

1. Good on Paper, Looming Disaster in Reality - The Remains of Russia's Soviet Arsenal https://www.youtube.com/watch?v=K8CcuVCDEUw

2. “An interceptor drone from the German company TYTAN Technologies during testing in Ukraine. The drone is capable of reaching speeds of 300 km/h and covering a distance of 20 kilometers. The drone strikes the target kinetically.” https://x.com/bayraktar_1love/status/1873432215386010080

3. “And here’s how wounded North Korean comrades spend their leisure time in Moscow hospitals. They have access to a television, good food, and possibly even access to the internet.” https://x.com/wartranslated/status/1872259704799609080

4. “Russia’s alcohol battalions are welcoming new recruits, though some can’t even manage to run a hundred meters. They’ll soon collect their $30,000 and be sent into the deadly carousel that the Cossacks are preparing for them somewhere near Kursk.” https://x.com/wartranslated/status/1872254220143149231

5. Three Russian officers were killed in a successful operation conducted by Ukraine’s GUR, SBU and other forces. https://x.com/NOELreports/status/1872556964812558797

6. Russian soldiers examine the torn bodies of their comrades after an FPV drone attack on their position in Donetsk Oblast. https://x.com/NOELreports/status/1872358178786365866

7. A Ukrainian soldier from the 32nd Mechanized Brigade in close combat with a Russian soldier near Pokrovsk. After a short shootout, the Russian soldier gets eliminated. https://x.com/NOELreports/status/1872648626637295901

8. At the railway station in the settlement of Voskresensk, Moscow Region, partisans blew up the carriages of a freight train, which was used to provide logistics for the Russian army. https://x.com/bayraktar_1love/status/1872961023898001836

https://www.theguardian.com/world/2024/dec/26/support-for-ukraine-russia-war-yougov-poll-survey