A compelling intuition is that deep learning does approximate Solomonoff induction, finding a mixture of the programs that explain the data, weighted by complexity. Finding a more precise version of this claim that's actually true would help us understand why deep learning works so well. There are a couple recent papers studying how NNs solve algorithmic tasks, which seem like exciting progress in this direction. - [[2309.02390] Explaining grokking through circuit efficiency](https://arxiv.org/abs/2309.02390) - develops a theory around when NN training learns a "memorizing" vs "generalizing" solution, which depends on each solution's "efficiency" -- how much param norm is needed to get correct & confident outputs. This theory predicts grokking phenomena - [[2310.16028] What Algorithms can Transformers Learn? A Study in Length Generalization](https://arxiv.org/abs/2310.16028) - transformers can't represent turing machines, but they can can represent a smaller class of computations, described by RASP programs. This paper finds that indeed, if data is generated by a RASP-L program, the transformer will learn exactly the right function. RASP is a assembly language for Transformers. RASP-L is a human-readable programming language which defines programs that can be compiled into Transformer weights, such that each line of RASP-L compiles into at most one Transformer layer. RASP-L lets us reason about Transformer-representable algorithms at a familiar level of abstraction— similar to standard programming languages like Python, but for programs which “run on a Transformer” instead of running on a standard von Neumann computer. "we leverage our insights to drastically improve generalization performance on traditionally hard tasks (such as parity and addition)" "we give a simple example where the "min-degree-interpolator" model of learning from Abbe et al. (2023) does not correctly predict Transformers' out-of-distribution behavior, but our conjecture does" [[2305.14699] Can Transformers Learn to Solve Problems Recursively?](https://arxiv.org/abs/2305.14699) Can Transformers Learn to Solve Problems Recursively? https://twitter.com/main_horse/status/1742373444485058949 LLM Maybe LongLM: Self-Extend LLM Context Window Without Tuning perplexity.ai uses gpt4 over google scholar [[2309.05922] A Survey of Hallucination in Large Foundation Models](https://arxiv.org/abs/2309.05922) (A Survey of Hallucination in Large Foundation Models paper) People: "LLMs hallucinate nonempirical overconfident false bullshit!" Also people: *hallucinate nonempirical overconfident false bullshit about how LLMs work without being knowledgable about results from the mechanistic interpretability field* As mechanistic interpretability, architectures AI engineering hacks progresses and hallucinations lower over time, AIs will soon get better at empirical reasoning than humans. I would argue they are already x times better than average humans at many tasks needing rational reasoning, since most humans seem to "repeat" what good enough sounding seeming "truths" people around them say, instead of developing rational empirical epistemology to critically evaluate any propositions and fact check if its possible. We can explicitly hardcode this into LLMs too I believe instead of data engineering and other smaller tricks. We have some mechanistic interpretability work scratching the surface of how truth works in LLMs (The Geometry of Truth: Emergent Linear Structure in Large Language Model Representations of True/False Datasets paper) and how internal emergent circuit generalization works (retweeted tweet). We can couple that will chain of thought and other hacks to give it time to reason etc. (AI capabilities can be significantly improved without expensive retraining paper) We could implement some internal mechanism of "I don't have this information sufficiently saved in my weights, I'm gonna find it in my context window, files, SQL/vector database, or on the internet on wikipedia, x documentation, google scholar, books,...", use Math/Physics engine (There are already attempts at doing this in ChatGPT, PerplexityAI etc.), do an evolving search process in the space of inferences with metacognitive critical evaluation, dynamic programming, use this expert subprocess,... (AGI will be made of heterogneous components, Transformer and Selective SSM blocks will be among them on LessWrong) This can be used to direct the LLM into the empirical reasoning and minimize hallucinations. All academic fields can be divided into the Crafts, the Histories, the Sciences, and Philosophy. https://twitter.com/eshear/status/1742334382223077621?t=jhTOr7A54Cws88ZozWDOuw&s=19 John Von Neumann in Computer and the Brain in 1955: "Possibility of replacing the human brain by machine which will be superior to it in any or all respects is not excluded by any natural law that we know, its therefore possible that the human race may be extinguished by machines." [- YouTube](https://youtu.be/0RknkWgd6Ck?si=eIg3fOCHiJc1p-oo) Hallucination minimization techniques https://twitter.com/omarsar0/status/1742633831234994189?t=yAWdq3gnC_p6gYPZmOgsmA&s=19 Toolformer [[2302.04761] Toolformer: Language Models Can Teach Themselves to Use Tools](https://arxiv.org/abs/2302.04761) [Physical healing as a function of perceived time | Scientific Reports](https://www.nature.com/articles/s41598-023-50009-3) time affects wound healing rate https://www.sciencedirect.com/science/authShare/S0303264723002824/20240104T004800Z/1?md5=4fe9bc08c60078b4236d99080e231c68&dgcid=coauthor From reinforcement learning to agency: Frameworks for understanding basal cognition, unifying behavior and goaldirectedness, RL with competencies Room tempature superconductor https://twitter.com/pronounced_kyle/status/1742588127628361809?t=nhlHiMMngh3F52g_BPfqdw&s=19 Robot imitation learning https://twitter.com/bindureddy/status/1742644792117600400?t=5lSJz9mCfUkl6VNwKrHvUg&s=19 𝐌𝐨𝐛𝐢𝐥𝐞 𝐀𝐋𝐎𝐇𝐀🏄 -- Learning! With 50 demos, the robot can autonomously complete complex mobile manipulation tasks: - cook and serve shrimp🦐 - call and take elevator🛗 - store a 3Ibs pot to a two-door cabinet Open-sourced! Project Website 🛜: [Mobile ALOHA](https://mobile-aloha.github.io) Code for Imitation Learning 🖥️: [GitHub - MarkFzp/act-plus-plus: Imitation learning algorithms with Co-training for Mobile ALOHA: ACT, Diffusion Policy, VINN](https://github.com/MarkFzp/act-plus-plus) Data 📊: [public_mobile_aloha_datasets – Disk Google](http://tinyurl.com/mobile-aloha-data) More about the hardware: https://x.com/tonyzzhao/status/1742603121682153852 Prompt engineering https://twitter.com/AlphaSignalAI/status/1742580813919875539?t=6WfihEYXF5xQGdStXElI3A&s=19 Custom constructed reality will blend with original reality more and more https://twitter.com/nickfloats/status/1742643438934426004?t=U8S4Dn-RNJHv8UIW_8KjoQ&s=19 [Self-Play Fine-Tuning Converts Weak Language Models to Strong Language Models](https://browse.arxiv.org/html/2401.01335v1) Self-Play Fine-Tuning Converts Weak Language Models to Strong Language Models I think that p(doom) given no AGI is greater than p(doom) given AGI, so we should pursue AGI, safely, and centralization of power is another risk here to prevent [[2305.04388] Language Models Don't Always Say What They Think: Unfaithful Explanations in Chain-of-Thought Prompting](https://arxiv.org/abs/2305.04388) Language Models Don't Always Say What They Think: Unfaithful Explanations in Chain-of-Thought Prompting Unitree robot dog: $2400 on Amazon Raspberry Pi: $57 Glock 17: $400 Ended 3D printer: $210 GPT-4 vision API: $0.01 per 1,000 tokens So ~$3100 is the cost of a robot soldier in 2024. https://twitter.com/AiBreakfast/status/1742598658758148317 new WIP list of mechinterp papers [Zotero | Groups > Mechanistic Interpretability](https://www.zotero.org/groups/5339216/mechanistic_interpretability) Nootropics https://twitter.com/nootropicguy/status/1742715531679904173?t=sL2iKmI49UAaGy1M9dCl9g&s=19 [[2401.00908] DocLLM: A layout-aware generative language model for multimodal document understanding](https://arxiv.org/abs/2401.00908) DocLLM: A layout-aware generative language model for multimodal document understanding Eacc movies https://twitter.com/npceo_/status/1742889797864259671?t=7Te9KjPQnkGLeFnFAdd8_g&s=19 Clustered graph map of ML papers https://twitter.com/MaartenGr/status/1742953286083166597?t=PU0nK_tUDWte9p6EJX0EpQ&s=19 LLM Augmented LLMs: Expanding Capabilities through Composition Proposes CALM, which introduces cross-attention between models to compose their representations and enable new capabilities arxiv.org/abs/2401.02412 https://fixupx.com/robertwiblin/status/1742595869310939641 [[2310.16028] What Algorithms can Transformers Learn? A Study in Length Generalization](https://arxiv.org/abs/2310.16028) ^ if we scale this mechanistic interpretability work then we could force big neural nets to learn and use emergent multiplication circuits and dont generalize, save every detail about concrete events, if we implemented realtime learning from experience 🤔 i wonder if one could do the same for humans tbh, would be cool, but biological neuronal populations degrade, we need more antiaging engineering Pořád si ale myslím že je v inteligenci a učení fakt nějaký fundamentální rozdíl mezi LLMs a lidmama hlavně mezi učím se algoritmem (mozek má možná forward forward místo backpropu), jinou evolucí a životem naučenou architekturou (efektivnejší s tím množství compute a memory co má, víc nelinearit a komplexity v individual neuronech, různě specializovaný části mozku, active inference matika?), odlišnýma objektivníma funkcema (i když pravděpodobně existují univerzální vzory co se jak umělý tak biologický neuronky učí dle různých analýz), tím jak mozek filtruje hodně věcí co není důležitý k přežití, tak to je možná predict next token u LLMs versus maximize chances of survival and reproduction given shared context. Individuálně nebo civilizationwise? Někteří mají za cíl vytvořit dečka, někteří chcou roboty po sobě (nerds), někteří bojují pro přežití celýho lidstva (třeba různí aktivisti), někteří vyvinou různý arbitrary instrumental (?) goals kterým dají celý život? Hmmm, zajímalo by mě jak nejlíp vidět objektivní funkci života. Přes free energy principle matematiku jdou všechny systémy vidět jako minimalizující překvapení (variational free energy v bayesian mechanice), ale to je moc obecný když chce člověk najít konkrétnosti. Je ale zajímavý jak LLMs jsou v nějakých věcech průměrný jak lidi, v jiných lepší než všichni lidi a v nějakých věcech nemají ani na batolata... Jaká matika tuto učící dynamiku a rozdíl mezi různýma umělýma a biologickýma systémama předpovídá? I want to know! Oba jsou to information processing systémy co fungují pod různou společnou a zároveň různou odlišnou matikou kterou pomalu zjišťujeme Nejsou úplně stejný ale zároveň nejsou úplně jiná věc 😄 Bude cool jak víc crackneme realtime učení u robotů, aby se to učilo jako lidi realtime z nových věcí přímo, než hacky jako daní examples do context window u už naučený sítě, ale možná to ani nebude potřeba Mozek podobně jako LLMs neměl dostatečný evoluční tlaky na hardcodnutí multiplikačního algoritmu nebo v různých kontextech úplnou memorizaci bez coarse grainingu A že když chceme humanlike intelligence benchmarky, tak ať to jsou reálné humanlike intelligence benchmarky co toto reflektují Ať benchmarky porovnávající jenom inputy a outputy, nebo porovnávající co se děje za algoritmy uvnitř Rozdíly mezi mozkem a umělýma neuronkama máme asi nejvíc prozkoumáný u vizuálního zpracovávání [Limits to visual representational correspondence between convolutional neural networks and the human brain | Nature Communications](https://www.nature.com/articles/s41467-021-22244-7) a u reasoningu teprve šlapeme na paty [Predictive coding - Wikipedia](https://en.m.wikipedia.org/wiki/Predictive_coding) [The empirical status of predictive coding and active inference - PubMed](https://pubmed.ncbi.nlm.nih.gov/38030100/) Chápu tu alergii k porovnání lidí s AI, protože všichni dělají různý porovnávání různých kvalit všude. Já se ale snažím to nedismissovat a porovnávat ty výroky s tím co empiricky máme za scientific modely v neuroscience x AI oboru o čem mám povědomí https://twitter.com/WizardLM_AI/status/1742906065359167730/photo/1 sota open source coding LLM An Animated Research Talk on: Neural-Network Quantum Field States [- YouTube](https://www.youtube.com/watch?v=rrvZDZMii-0) Variational Neural-Network Ansatz for Continuum Quantum Field Theory [[2212.00782] Variational Neural-Network Ansatz for Continuum Quantum Field Theory](https://arxiv.org/abs/2212.00782) "Here we approach this problem by introducing neural-network quantum field states, a deep learning ansatz that enables application of the variational principle to non-relativistic quantum field theories in the continuum." https://www.scientificamerican.com/article/scientists-finally-invent-heat-controlling-circuitry-that-keeps-electronics-cool1/ https://www.science.org/doi/10.1126/science.abo4297 [[1805.00899] AI safety via debate](https://arxiv.org/abs/1805.00899) ai safety via debate superintelligences debating and inferin if one is true from how it conviced watching human [[2307.15043] Universal and Transferable Adversarial Attacks on Aligned Language Models](https://arxiv.org/abs/2307.15043) Universal and Transferable Adversarial Attacks on Aligned Language Models usign gradient search to backpropagate ideal tokens to make models comply "instead of relying on manual engineering, our approach automatically produces these adversarial suffixes by a combination of greedy and gradient-based search techniques" https://www.lesswrong.com/posts/v7f8ayBxLhmMFRzpa/steering-llama-2-with-contrastive-activation-additions "generating transferrable prompt injection activation vectors without needing a complex dataset, like you just append 1 prefix or suffix to every prompt of your original dataset and then make contrast pairs that way, average over all the activations and steer via this activation addition intervention and they found a way to do interpretability with it" [GitHub - nrimsky/LM-exp: LLM experiments done during SERI MATS - focusing on activation steering / interpreting activation spaces](https://github.com/nrimsky/LM-exp/tree/main) Heygennsota deepfakes https://twitter.com/emollick/status/1743146951749533897?t=_5TcwLPQzLcAu6DqxU_tIg&s=19 Links for 2024-01-05 AI: 1. Google Deepmind introduces AutoRT, SARA-RT and RT-Trajectory to improve real-world robot data collection, speed, and generalization [Shaping the future of advanced robotics - Google DeepMind](https://deepmind.google/discover/blog/shaping-the-future-of-advanced-robotics/) 2. GPT-4V(ision) is a Generalist Web Agent, if Grounded [SeeAct](https://osu-nlp-group.github.io/SeeAct/) 3. Learning Vision from Models Rivals Learning Vision from Data — SynCLR, a novel approach for learning visual representations exclusively from synthetic images and synthetic captions, without any real data. [[2312.17742] Learning Vision from Models Rivals Learning Vision from Data](https://arxiv.org/abs/2312.17742) 4. AnyText: Multilingual Visual Text Generation And Editing [GitHub - tyxsspa/AnyText: Official implementation code of the paper <AnyText: Multilingual Visual Text Generation And Editing>](https://github.com/tyxsspa/AnyText#readme) 5. Microsoft announces Improving Text Embeddings with Large Language Models [[2401.00368] Improving Text Embeddings with Large Language Models](https://arxiv.org/abs/2401.00368) 6. FlowVid: Taming Imperfect Optical Flows for Consistent Video-to-Video Synthesis [FlowVid: Taming Imperfect Optical Flows for Consistent Video-to-Video Synthesis](https://jeff-liangf.github.io/projects/flowvid/) Science and Technology: 1. A new class of antibiotics kills one of the three most dangerous strains of drug-resistant bacteria. [A novel antibiotic class targeting the lipopolysaccharide transporter | Nature](https://www.nature.com/articles/s41586-023-06873-0) 2. Top 10 discoveries about ancient people from DNA in 2023 [Top 10 discoveries about ancient people from DNA in 2023](https://johnhawks.net/weblog/the-top-10-discoveries-about-ancient-people-from-dna-in-2023/) 3. “Science and technology have reached a level of maturity where we can begin to have real, dramatic effects on the human condition,” [The Race to Put Brain Implants in People Is Heating Up | WIRED](https://www.wired.com/story/the-race-to-put-brain-implants-in-people-is-heating-up/) [https://archive.is/wyYyv] 4. We’re Inching Closer to Answers for Why We Age and How to Slow Down the Clock [We're Inching Closer to Answers for Why We Age and How to Slow Down the Clock](https://singularityhub.com/2024/01/04/were-inching-closer-to-answers-for-why-we-age-and-how-to-slow-down-the-clock/) Psychology: 1. “levels of depressive affect in a sample of 86,138 US teenagers, broken down by sex & political leanings. Long story short, mental-health declines are found in every group,but left-leaning girls are doing worst,followed by left-leaning boys, followed by right-leaning girls & boys” [Graph of the Day: Politics and Teen Mental Health](https://stevestewartwilliams.substack.com/p/graph-of-the-day-politics-and-teen) Miscellaneous: 1. “Disappointed by the #napoleonmovie? The good news is that reality dwarfs fiction Far from being Scott's half-wit, he was a legendary tactician and a master of deception, in full display in the triumph of Austerlitz” https://twitter.com/Valen10Francois/status/1741503707701797309 The biggest survey of AI researchers ever! 2778 participants from six top AI venues answered questions from fourteen topics regarding the future of AI. 1. The expected time to human-level performance dropped 1-5 decades since the 2022 survey. 50% chance of unaided AI outperforming humans in every possible task by 2047. 2. Time to most narrow milestones has decreased since 2022, some by a lot. AI researchers are expected to be automatable ~25 years earlier than a year before. Time to NYT bestselling fiction dropped ~9 years. AI researchers give a 50% chance that AI can in 5 years: 💻 make an entire payment processing site from scratch 🎤 generate a new song that sounds like it’s by Taylor Swift 🤖 autonomously download and fine-tune a large language model 3. The median artificial-intelligence researcher believes there is a 5% chance that AI will cause human extinction. Read more: [Survey of 2,778 AI authors: six parts in pictures](https://blog.aiimpacts.org/p/2023-ai-survey-of-2778-six-things) Links for 2024-01-06 AI: 1. LLM Augmented LLMs: Expanding Capabilities through Composition — “…when PaLM2-S is augmented with a code-specific model, we see a relative improvement of 40% over the base model for code generation and explanation tasks -- on-par with fully fine-tuned counterparts.” [[2401.02412] LLM Augmented LLMs: Expanding Capabilities through Composition](https://arxiv.org/abs/2401.02412) 2. A new paper just identified 26 principles to improve the quality of LLM responses by 50%. [[2312.16171v1] Principled Instructions Are All You Need for Questioning LLaMA-1/2, GPT-3.5/4](https://arxiv.org/abs/2312.16171v1) 3. Chiral Magnets Pave the Way for Energy-Efficient Brain-Like Computing [Twisted magnets make brain-inspired computing more adaptable | UCL News - UCL – University College London](https://www.ucl.ac.uk/news/2023/nov/twisted-magnets-make-brain-inspired-computing-more-adaptable) 4. A Powerful AI Fact-Checker that Checks Content and Eliminates Hallucinations? https://www.novaspivack.com/technology/a-powerful-ai-fact-checker-that-checks-content-and-eliminates-hallucinations 5. AI’s big test: Making sense of $4 trillion in medical expenses https://www.politico.com/news/2023/12/31/ai-medical-expenses-00132557 6. Chief Justice Roberts Sees Promise and Danger of A.I. in the Courts https://www.nytimes.com/2023/12/31/us/john-roberts-supreme-court.html [https://archive.is/iNBv3] 7. OpenAI will launch its GPT “app store” next week. Users will be able to make and sell artificial-intelligence agents. [OpenAI will open its custom ChatGPT store next week - The Verge](https://www.theverge.com/2024/1/4/24025610/openai-gpt-store-ai-agent-delay) 8. “This is a completely fake video of me. The AI (HeyGen) used 30 seconds of me talking to a webcam and 30 seconds of my voice, and now I have an avatar that I can make say anything. Don't trust your own eyes.” https://twitter.com/emollick/status/1743146951749533897 Miscellaneous: 1. Jupiter Brains? Think bigger. Anders Sandberg describes an entire star cluster, collapsed into a single 10^36 kg body of iron, storing 10^61 bits of information in quark matter and running 10^85 operations per second, with 10^39 W from black hole reactors. https://www.jetpress.org/volume5/Brains2.pdf 2. How often does correlation equal causation? [How Often Does Correlation=Causality? · Gwern.net](https://gwern.net/correlation) 3. Bets, Bonds, and Kindergarteners https://www.lesswrong.com/posts/DoHcgTvyxdorAMquE/bets-bonds-and-kindergarteners 4. Life in a hospital prison for an actual genius [Lapsus$: GTA 6 hacker handed indefinite hospital order - BBC News](https://www.bbc.co.uk/news/technology-67663128) 5. Scott Alexander's Best Essays? [Links | near.blog](https://near.blog/my-favorite-links/) Psychology: 1. The average IQ of undergraduate college students has been falling since the 1940s and has now become basically the same as the population average. https://www.frontiersin.org/articles/10.3389/fpsyg.2024.1309142/abstract