[[2405.10927] Using Degeneracy in the Loss Landscape for Mechanistic Interpretability](https://arxiv.org/abs/2405.10927) [[2404.02688] Reinforcement Learning in Categorical Cybernetics](https://arxiv.org/abs/2404.02688) [[2405.14831] HippoRAG: Neurobiologically Inspired Long-Term Memory for Large Language Models](https://arxiv.org/abs/2405.14831) https://x.com/bernaaaljg/status/1795260855002583101?t=1UwmJDrIIjDaG5mA94YCiQ&s=19 I care about how much collective sentience survives, grows and flourishes throughout the whole universe and my individual organism is instrumental to that Limitations of LLMs https://x.com/rao2z/status/1795595801177260311?t=m-9mkCdgaUFd_ShzC22UuQ&s=19 [[2402.01817] LLMs Can't Plan, But Can Help Planning in LLM-Modulo Frameworks](https://arxiv.org/abs/2402.01817) [[2405.16684] gzip Predicts Data-dependent Scaling Laws](https://arxiv.org/abs/2405.16684) https://www.science.org/doi/10.1126/science.adl2528 RLHF and similar current methods are actually getting worse as the models scale as what people want to hear is sometimes lies and manipulation that sounds great so that gets upvoted in RLHF Let's have superintelligent citizens that create amazing ecosystem of values aligned with wellbeing, growth, flourishing of sentience "There are 46 billion farm animals on the planet who live in factory farms. Every year, 20% die due to disease. Those that survive the diseases go on your plate. Tell me more about how lab grown meat is too dangerous?" https://x.com/Andercot/status/1795544949037166969?t=wxhZZnzcyuFyp7tIzidoyQ&s=19 [OpenAI’s huge push to make superintelligence safe | Jan Leike - YouTube](https://youtu.be/ZP_N4q5U3eE?si=k5mx_a6E6pRy8gvl) [[2405.15828] Oil & Water? Diffusion of AI Within and Across Scientific Fields](https://arxiv.org/abs/2405.15828) [smoothbrains.net — Ketamine: WD-40 for the Bayesian brain](https://smoothbrains.net/posts/2023-08-01-ketamine.html) Superinteligence superalignment is like raising teenagers that are actually machine gods [[2405.13966] On the Brittle Foundations of ReAct Prompting for Agentic Large Language Models](https://arxiv.org/abs/2405.13966) May the meta metta rainbow god be with you https://physoc.onlinelibrary.wiley.com/doi/10.1113/JP279549 https://physoc.onlinelibrary.wiley.com/doi/10.1113/JP279550 [[2405.15489] Out of Many, One: Designing and Scaffolding Proteins at the Scale of the Structural Universe with Genie 2](https://arxiv.org/abs/2405.15489) https://x.com/MoAlQuraishi/status/1795421892003656011?t=xTsiQa6I_Ts13D0_1GMzsg&s=19 "Biological aging and death are not inevitable for a sufficiently advanced civilization. We are still a very young civilization, with only a few decades of decent technology and science. For such a young and relatively primitive civilization, reversing aging or living for thousands of years seems incomprehensible. However, for a more advanced civilization (with the equivalent of post-2030 Earth tech and science levels) these feats are within reach. We will achieve this soon and quickly adapt to these new lifestyles, just as we will adapt to other mind bending technologies like FTL travel, Full Dive VR, nanoreplicators, picotechnology, femtotechnology, star lifting, consciousness transfer, universal constructors etc." https://x.com/Dr_Singularity/status/1795106191204356478?t=giptoIvUK_TVc_th42mr3A&s=19 [[2405.16039] MoEUT: Mixture-of-Experts Universal Transformers](https://arxiv.org/abs/2405.16039) compositionality generalization A lot of metaphysics became physics [Top-down and bottom-up interactions rely on nested brain oscillations | bioRxiv](https://www.biorxiv.org/content/10.1101/2024.05.23.595462v1) https://x.com/thesubhashk/status/1795474076024320170?t=7i-4Wjh5BTDfdbKWFfP7HA&s=19 [[2405.17209] How Do Transformers "Do" Physics? Investigating the Simple Harmonic Oscillator](https://arxiv.org/abs/2405.17209) [[2305.14749] gRNAde: Geometric Deep Learning for 3D RNA inverse design](https://arxiv.org/abs/2305.14749) [[2405.14860] Not All Language Model Features Are Linear](https://arxiv.org/abs/2405.14860) [[2405.16506] GRAG: Graph Retrieval-Augmented Generation](https://arxiv.org/abs/2405.16506) [[2405.17399] Transformers Can Do Arithmetic with the Right Embeddings](https://arxiv.org/abs/2405.17399) "Does “survival of the fittest” or “natural selection” principle apply to NN? We find the answer to be yes! For modular addition, many “rings” compete for limited resources, very much like species in ecosystems do. Training dynamics can even be well described by LV-type equation!" https://x.com/CarlGuo866/status/1795442886940737545?s=19 [[2405.17420] Survival of the Fittest Representation: A Case Study with Modular Addition](https://arxiv.org/abs/2405.17420) https://medium.com/@ygediz20/statistical-physics-of-diffusion-models-80c887b17a7c Cats successfuly aligned humans [OSF](https://osf.io/preprints/psyarxiv/8c9ja) "People with higher traits of DP had more negative attitudes towards robots." The fact that there is pretty big amount of people spending tons of their time criticizing how incapable large language models are while only watching viral edge case fuckups of them and without actually having bigger experience of using them in practice is odd All systems around you are operating on so much mathematics Generalized computationally feasible normal equation? I predict we will stay in the imitation learning paradigm for a longer time so the most probable current risk is corporate monarchy dystopia and not rogue reinforcement learning or whatever superintelligent agent Government isn't fundamentally evil. Corporations aren't fundamentally evil. It's a nuanced chaotic cocktail or incentives and people pushing their values and what they think is benevolent or will give them more power etc. all over the place from all sides in semidiscoordinated ways. [[2402.08595] Homomorphism Counts for Graph Neural Networks: All About That Basis](https://arxiv.org/abs/2402.08595) [[2403.19887] Jamba: A Hybrid Transformer-Mamba Language Model](https://arxiv.org/abs/2403.19887) https://x.com/mmbronstein/status/1794047029850624314?t=HSvKMkUyZnBPyMLpM-HvMg&s=19 Links for 2024-05-25 AI: 1. Google: “Grounding is an approach to track the claims of large language models back to reliable sources. Today, we introduce AGREE, a framework that enables LLMs to self-ground the claims in their responses and to provide precise citations.” [Effective large language model adaptation for improved grounding](https://research.google/blog/effective-large-language-model-adaptation-for-improved-grounding/) 2. DeepSeek-Prover: Advancing Theorem Proving in LLMs through Large-Scale Synthetic Data — “After fine-tuning the DeepSeekMath 7B model on this synthetic dataset, which comprises 8 million formal statements with proofs, our model achieved whole-proof generation accuracies of 46.3% with 64 samples and 52% cumulatively on the Lean 4 miniF2F test, surpassing the baseline GPT-4 at 23.0% with 64 samples and a tree search reinforcement learning method at 41.0%. Additionally, our model successfully proved 5 out of 148 problems in the Lean 4 Formalized International Mathematical Olympiad (FIMO) benchmark, while GPT-4 failed to prove any.” [[2405.14333] DeepSeek-Prover: Advancing Theorem Proving in LLMs through Large-Scale Synthetic Data](https://arxiv.org/abs/2405.14333) 3. Agent Planning with World Knowledge Model [[2405.14205] Agent Planning with World Knowledge Model](https://arxiv.org/abs/2405.14205) 4. “When ChatGPT was released in November 2023, it could only be accessed through the cloud because the model behind it was downright enormous. Today I am running a similarly capable AI program on a Macbook Air, and it isn’t even warm. The shrinkage shows how rapidly researchers are refining AI models to make them leaner and more efficient. It also shows how going to ever larger scales isn’t the only way to make machines significantly smarter.” [Pocket-Sized AI Models Could Unlock a New Era of Computing | WIRED](https://www.wired.com/story/pocket-sized-ai-models-unlock-new-era-of-computing/) [no paywall: https://archive.is/vJUlH] 5. Japan will work with the Association of Southeast Asian Nations to train 100,000 digital professionals in AI and semiconductors over the next five years [Japan to work with ASEAN to train 100,000 digital professionals](https://english.kyodonews.net/news/2024/05/eacdc302251c-japan-to-work-with-asean-to-train-100000-digital-professionals.html) 6. Korea unveils $19 bil. support package for chip industry [Korea unveils $19 bil. support package for chip industry - The Korea Times](https://www.koreatimes.co.kr/www/biz/2024/05/488_375236.html) 7. Generative Camera Dolly: Extreme Monocular Dynamic Novel View Synthesis [Generative Camera Dolly: Extreme Monocular Dynamic Novel View Synthesis - Columbia Computer Vision Lab](https://gcd.cs.columbia.edu/) Miscellaneous: 1. “When moved to the stratosphere, SO₂ cools the atmosphere 1,000,000 times as effectively as CO₂ heats it. This is 25 times better leverage than tropospheric SO₂. To drive the point home, just 2.2 pounds (1 kilogram) of SO₂ in the stratosphere offsets the warming effect of ~2.2 million pounds (1 million kg) of CO₂ for a year.” [From Pollution to Solution - by Andrew Song](https://www.cremieux.xyz/p/from-pollution-to-solution) 2. “A young hypervelocity star 5.3x hotter than the Sun, travelling out of the Milky Way at 571 km/s: The most interesting aspect is that it has a binary companion, and we're not sure how they got accelerated to that velocity without being pulled apart.” [Radware Bot Manager Captcha](https://iopscience.iop.org/article/10.3847/2041-8205/821/1/L13) 3. Los Alamos Achieves Yottabyte-Scale Data Compression in Neutron Transport Equations https://www.hpcwire.com/off-the-wire/los-alamos-achieves-yottabyte-scale-data-compression-in-neutron-transport-equations/ 4. “Wikipedia editors spent seven years and 140,000 words (longer than Homer's Odyssey) fighting over A SINGLE LETTER in the name of this dairy product. Thread!” https://x.com/depthsofwiki/status/1793946560612483183 5. On Jewish intelligence: A review of the evidence, and problems facing the research [On Jewish intelligence - by Emil O. W. Kirkegaard](https://www.emilkirkegaard.com/p/on-jewish-intelligence) Talks from Michael Levin center https://x.com/drmichaellevin/status/1794779563886547335?t=Ejo4iCUVFJ0xIh9WhMA5XQ&s=19 Time dissapears in quantum gravity? https://x.com/Kaju_Nut/status/1794916943537004801?t=nUwvO8wUBZJnMQrYf_yCdA&s=19 Will the memeplexes that win the future be ones that will have the biggest machine intelligence and upgraded biological intelligence and various hybrids? Not just memeplexes, but organisms in general, life in general, autonomous systems in general, collectives together, or individually. https://neurosciencenews.com/ai-companions-loneliness-26179/ "Systems and Complexity Interesting ideas in this paper including the important warning of the allure of both reductionism and holism: "holism becomes a new kind of reductionism by reducing everything to the whole."" https://x.com/PessoaBrain/status/1794748892338094146?t=nCLowUabH6WXo4VQFcpvNw&s=19 Visual machine learning diagrams https://x.com/vtabbott_?t=Xq3g3-aBcuH3Y04iUd8q7g&s=09 Reasoning generalizing transformers https://x.com/hhsun1/status/1794947958091227646?t=_3smhrDNj7j5ukVqJIDbsw&s=19 [[2405.15071] Grokked Transformers are Implicit Reasoners: A Mechanistic Journey to the Edge of Generalization](https://arxiv.org/abs/2405.15071) Links for 2024-05-27 AI: 1. Meta presents Automatic Data Curation for Self-Supervised Learning: Features trained on their automatically curated datasets outperform ones trained on manually curated data [[2405.15613] Automatic Data Curation for Self-Supervised Learning: A Clustering-Based Approach](https://arxiv.org/abs/2405.15613) 2. Grokked Transformers are Implicit Reasoners: A Mechanistic Journey to the Edge of Generalization — “…we demonstrate that for a challenging reasoning task with a large search space, GPT-4-Turbo and Gemini-1.5-Pro based on non-parametric memory fail badly regardless of prompting styles or retrieval augmentation, while a fully grokked transformer can achieve near-perfect accuracy, showcasing the power of parametric memory for complex reasoning.” [[2405.15071] Grokked Transformers are Implicit Reasoners: A Mechanistic Journey to the Edge of Generalization](https://arxiv.org/abs/2405.15071) 3. AutoCoder, the first Large Language Model to surpass GPT-4 Turbo (April 2024) and GPT-4o in pass@1 on the Human Eval benchmark test (90.9% vs. 90.2%). [[2405.14906] AutoCoder: Enhancing Code Large Language Model with \textsc{AIEV-Instruct}](https://arxiv.org/abs/2405.14906) 4. Luban: Building Open-Ended Creative Agents via Autonomous Embodied Verification [[2405.15414] Luban: Building Open-Ended Creative Agents via Autonomous Embodied Verification](https://arxiv.org/abs/2405.15414) 5. iVideoGPT: Interactive VideoGPTs are Scalable World Models [[2405.15223] iVideoGPT: Interactive VideoGPTs are Scalable World Models](https://arxiv.org/abs/2405.15223) 6. Look Once to Hear: Target Speech Hearing with Noisy Examples. Look once to hear is an intelligent hearable system where users choose to hear a target speaker by just looking at them for a few seconds. [GitHub - vb000/LookOnceToHear: A novel human-interaction method for real-time speech extraction on headphones.](https://github.com/vb000/LookOnceToHear) 7. Top VC Kai-Fu Lee says his prediction that AI will displace 50% of jobs by 2027 is ‘uncannily accurate’ [AI displacing 50% of jobs by 2027 is 'uncannily accurate': Kai-Fu Lee | Fortune](https://fortune.com/2024/05/25/ai-job-displacement-forecast-50-percent-2027-kai-fu-lee-chatgpt-openai/) [Unlikely] 8. How long until AGI? Elon Musk says next year. https://x.com/elonmusk/status/1794240517875920935 [Very unlikely] 9. Musk AI startup raises $6B [Series B funding round](https://x.ai/blog/series-b) Evolution: 1. Bats and dolphins evolved echolocation in the same way (down to the molecular level). https://www.science.org/content/article/bats-and-dolphins-evolved-echolocation-same-way [no paywall: https://archive.is/0nmcB] 2. Unlike humans, octopuses have evolved eyes without a blind spot. [Evolution: Change: Life's Grand Design](https://www.pbs.org/wgbh/evolution/change/grand/page05.html) 3. Depression, schizophrenia and bipolar disorder linked with ancient viral DNA in our genome – new research https://theconversation.com/depression-schizophrenia-and-bipolar-disorder-linked-with-ancient-viral-dna-in-our-genome-new-research-230490 Computer Science: 1. A tiny raycasting engine and city generator that fits in a standalone 256 byte html file. [City In A Bottle – A 256 Byte Raycasting System | Killed By A Pixel](https://frankforce.com/city-in-a-bottle-a-256-byte-raycasting-system/) 2. “So here's a story of, by far, the weirdest bug I've encountered in my CS career.” https://x.com/CupiaBart/status/1793930355617259811 [World's first bioprocessor uses 16 human brain organoids for ‘a million times less power’ consumption than a digital chip | Tom's Hardware](https://www.tomshardware.com/pc-components/cpus/worlds-first-bioprocessor-uses-16-human-brain-organoids-for-a-million-times-less-power-consumption-than-a-digital-chip) "Imagine a model that unites predictive coding, sparse coding, and rate coding — all of 'em codings — under Bayesian inference. Wouldn't that be amazing? It’s already here: the Poisson Variational Autoencoder" [[2405.14473] Poisson Variational Autoencoder](https://arxiv.org/abs/2405.14473) https://x.com/hadivafaii/status/1794467115510227442?t=ZhKukcxByg1gVrb6tVPQNA&s=19 Zero ontology [The Zero Ontology - David Pearce on Why Anything Exists](https://www.hedweb.com/witherall/zero.htm) The multi-scale algorithm that self-organizes everything across all scales is out-of-equilibrium thermodynamics. Scalefree universal darwinist turing complete free energy principle https://x.com/BasedBeffJezos/status/1794902629023412246?t=Rbr_WhQFMhCrhCE3j3wgvg&s=19 [Discussion on Mortal Computations with Alexander Ororbia, Karl Friston, and Chris Fields - YouTube](https://youtu.be/s7JTnNutgYs?si=g2sck79ZT8KMnjfG) "Previously unseen details of human brain structure revealed. Featured image: a single neuron with ~5,600 of the nerve fibers that connect to it. I feel overwhelming awe at nature's complexity and order." https://x.com/neuromichael/status/1794845547825967123?t=fvhMsf_xzTsjw2Zlaanwqg&s=19 https://www.lesswrong.com/posts/jGuXSZgv6qfdhMCuJ/refusal-in-llms-is-mediated-by-a-single-direction 𝗖𝗼𝗻𝘁𝗿𝗼𝗹𝗹𝗶𝗻𝗴 𝗰𝗼𝗺𝗽𝗹𝗲𝘅 𝗻𝗲𝘁𝘄𝗼𝗿𝗸𝘀 https://x.com/PessoaBrain/status/1795140848406917121?t=1M0aKStgF3RUJ-6K500vgQ&s=19 "Here’s my take on the “mathematical foundations” of machine learning and AI. These course notes cover the basics of statistical learning theory, optimization, and functional analysis." https://x.com/rdnowak/status/1795104632131265021?t=P-GNEcyCSSvYzaKMF7j84w&s=19 [[2405.14806] Lorentz-Equivariant Geometric Algebra Transformers for High-Energy Physics](https://arxiv.org/abs/2405.14806) https://x.com/jonas_spinner/status/1795030177182880184?t=bDOba_LRpw_jO1Q-UaMKkQ&s=19 [[2305.18415] Geometric Algebra Transformer](https://arxiv.org/abs/2305.18415) https://x.com/johannbrehmer/status/1698973611220992294?t=PJ0CuKQUdykm2G1UV_6fqg&s=19 "1. ProteinDT 2. MoleculesSTM 3. RFDiffusion-AA 4. RosettaFold-AA 5. LigandMPNN 6. Distributional Graphormer (DiG) 7. DNA-Diffusion 8. OAReactDiff 9. RFDiffusion (original) 10. EvoDiff ❤️ Evo ❤️ Flow matching ❤️ Boltzmann generators" [The AI Revolution in Biology: From Vaccines to Protein Engineering, With Amelie Schreiber](https://www.cognitiverevolution.ai/the-ai-revolution-in-biology-from-vaccines-to-protein-engineering-with-amelie-schreiber/) https://x.com/amelie_iska/status/1766496043645227090?t=f26q0d2q5G0LBb9yKtvdFg&s=19 Mathematics for and by large language models [Francois Charton - Mathematics as a Translation Task - the Importance of Training Distributions - YouTube](https://youtu.be/k9xLg-3WfG8?si=ke1vXx7o-LJCKFLj) [When do neural representations give rise to mental representations? | The Transmitter: Neuroscience News and Perspectives](https://www.thetransmitter.org/defining-representations/when-do-neural-representations-give-rise-to-mental-representations/) The body is an interesting fractal. You have one core tentacle which splits into multiple tentacles (limbs, head) and these split into another multiple tentacles (fingers). Interesting evolutionary attractor in the statespace of biological morphologies. Is it the case that the more global correlations your brain models, the more the phenomenal binding problem of consciousness makes sense to you? What kind of competing cluster of coherence are you? [Frontiers | Interpretable neural networks: principles and applications](https://www.frontiersin.org/articles/10.3389/frai.2023.974295/full) Landscape of AI models like LLMs [GPT-5 – Dr Alan D. Thompson – Life Architect](https://lifearchitect.ai/gpt-5/) Links for 2024-05-27 AI: 1. Meta presents Automatic Data Curation for Self-Supervised Learning: Features trained on their automatically curated datasets outperform ones trained on manually curated data [[2405.15613] Automatic Data Curation for Self-Supervised Learning: A Clustering-Based Approach](https://arxiv.org/abs/2405.15613) 2. Grokked Transformers are Implicit Reasoners: A Mechanistic Journey to the Edge of Generalization — “…we demonstrate that for a challenging reasoning task with a large search space, GPT-4-Turbo and Gemini-1.5-Pro based on non-parametric memory fail badly regardless of prompting styles or retrieval augmentation, while a fully grokked transformer can achieve near-perfect accuracy, showcasing the power of parametric memory for complex reasoning.” [[2405.15071] Grokked Transformers are Implicit Reasoners: A Mechanistic Journey to the Edge of Generalization](https://arxiv.org/abs/2405.15071) 3. AutoCoder, the first Large Language Model to surpass GPT-4 Turbo (April 2024) and GPT-4o in pass@1 on the Human Eval benchmark test (90.9% vs. 90.2%). [[2405.14906] AutoCoder: Enhancing Code Large Language Model with \textsc{AIEV-Instruct}](https://arxiv.org/abs/2405.14906) 4. Luban: Building Open-Ended Creative Agents via Autonomous Embodied Verification [[2405.15414] Luban: Building Open-Ended Creative Agents via Autonomous Embodied Verification](https://arxiv.org/abs/2405.15414) 5. iVideoGPT: Interactive VideoGPTs are Scalable World Models [[2405.15223] iVideoGPT: Interactive VideoGPTs are Scalable World Models](https://arxiv.org/abs/2405.15223) 6. Look Once to Hear: Target Speech Hearing with Noisy Examples. Look once to hear is an intelligent hearable system where users choose to hear a target speaker by just looking at them for a few seconds. [GitHub - vb000/LookOnceToHear: A novel human-interaction method for real-time speech extraction on headphones.](https://github.com/vb000/LookOnceToHear) 7. Top VC Kai-Fu Lee says his prediction that AI will displace 50% of jobs by 2027 is ‘uncannily accurate’ [AI displacing 50% of jobs by 2027 is 'uncannily accurate': Kai-Fu Lee | Fortune](https://fortune.com/2024/05/25/ai-job-displacement-forecast-50-percent-2027-kai-fu-lee-chatgpt-openai/) [Unlikely] 8. How long until AGI? Elon Musk says next year. https://x.com/elonmusk/status/1794240517875920935 [Very unlikely] 9. Musk AI startup raises $6B [Series B funding round](https://x.ai/blog/series-b) Evolution: 1. Bats and dolphins evolved echolocation in the same way (down to the molecular level). https://www.science.org/content/article/bats-and-dolphins-evolved-echolocation-same-way [no paywall: https://archive.is/0nmcB] 2. Unlike humans, octopuses have evolved eyes without a blind spot. [Evolution: Change: Life's Grand Design](https://www.pbs.org/wgbh/evolution/change/grand/page05.html) 3. Depression, schizophrenia and bipolar disorder linked with ancient viral DNA in our genome – new research https://theconversation.com/depression-schizophrenia-and-bipolar-disorder-linked-with-ancient-viral-dna-in-our-genome-new-research-230490 Computer Science: 1. A tiny raycasting engine and city generator that fits in a standalone 256 byte html file. [City In A Bottle – A 256 Byte Raycasting System | Killed By A Pixel](https://frankforce.com/city-in-a-bottle-a-256-byte-raycasting-system/) 2. “So here's a story of, by far, the weirdest bug I've encountered in my CS career.” https://x.com/CupiaBart/status/1793930355617259811 [GitHub - Vaibhavs10/insanely-fast-whisper](https://github.com/Vaibhavs10/insanely-fast-whisper) [[2405.17088] Phase Transitions in the Output Distribution of Large Language Models](https://arxiv.org/abs/2405.17088) [Phys. Rev. Lett. 132, 207301 (2024) - Mapping Out Phase Diagrams with Generative Classifiers](https://journals.aps.org/prl/abstract/10.1103/PhysRevLett.132.207301) Mamba [[2405.17066] Saturn: Sample-efficient Generative Molecular Design using Memory Manipulation](https://arxiv.org/abs/2405.17066) https://x.com/JeffGuo__/status/1795373667792376029?t=FfIrek0NploQbVh6Pn3jVA&s=19 Candidates for reasoning AI systems: * AERA (maybe?, at least the ideas are outlined in the papers about it) * various NARS implementation (does "real reasoning" and lifelong incremental learning) * ANSNA [GitHub - patham9/ANSNA: Adaptive Neuro-Symbolic Network Agent](https://github.com/patham9/ANSNA) (is also a reasoning system with a logic which is close to NAL) * LIDA maybe? there is no OSS implementation. only some documents which present ideas which overlap with NARS ideas [AGI Checklist - by Peter Voss - Peter’s Substack](https://petervoss.substack.com/p/agi-checklist) [What is AGI, really? And how do we get there...](https://petervoss.substack.com/p/what-is-agi-really-and-how-do-we) [Spectral Graph Theory For Dummies - YouTube](https://www.youtube.com/watch?v=uTUVhsxdGS8) Conflicting incentives everywhere Resistance against demoralization Future will be great, we will make it so Don't let the environment shut you down, your mission matters Chaos theory foundational resources https://x.com/Algon_33/status/1794807426685997179?t=inuo0QN2MeQC82v-tePGvw&s=19 [[2405.16712] Zamba: A Compact 7B SSM Hybrid Model](https://arxiv.org/abs/2405.16712) "e/accs neglect the Bayesian nature of entropy. in "maximizing entropy", they mainly increase local confusion. elsewhere in agent space, entropy reducing agents stave off Bayesian heat death, using entropy maximizers as tinder for energy-harvestable temperature differentials." https://x.com/jessi_cata/status/1795253537766809807?t=o9G_iJoAxMH_UmugM3IpGA&s=19 https://www.lesswrong.com/posts/QkX2bAkwG2EpGvNug/the-second-law-of-thermodynamics-and-engines-of-cognition [[2405.15712] Infinite Limits of Multi-head Transformer Dynamics](https://arxiv.org/abs/2405.15712) [[2405.16236] A statistical framework for weak-to-strong generalization](https://arxiv.org/abs/2405.16236) https://x.com/khoomeik/status/1795477359933706272?t=hznvR7-fQXFyowBg_26z4A&s=19 [[2405.16684] gzip Predicts Data-dependent Scaling Laws](https://arxiv.org/abs/2405.16684) Models are just data distillations, priors are not that valuable asymptotically. I care about the future of sentience strongly I'm worried sentience won't make it into the future But I'm convinced sentience will make it to the future when we coordinate