Francois Chollet's mental model for LLMs: LLMs are stores of programs. Querying LLMs involves selecting a program from the latent program space and running it on your data. And the ability of LLMs to interpolate between these programs is what makes LLMs so flexible. [x.com](https://twitter.com/PicoPaco17/status/1783983717037342793?t=PuNtVE7Qu5prckE6dgwv2A&s=19) [How I think about LLM prompt engineering](https://fchollet.substack.com/p/how-i-think-about-llm-prompt-engineering) [François Chollet - Creating Keras 3 - YouTube](https://youtu.be/oe6fuxhGVRE?si=5rRkf8xvvsegv1Tc) [What is GitHub Copilot Workspace? Sneak peek into your new developer environment - YouTube](https://youtu.be/pkotufZchjE?si=XvPxrOATqB_ywwgQ) https://arxiv.org/abs/2404.11018 https://arxiv.org/abs/2404.18930 [Scientists Say New Material Can Suck Carbon Out of Atmosphere Faster Than Trees](https://futurism.com/the-byte/new-material-carbon-atmosphere) [What if we simulated biology using physics - YouTube](https://www.youtube.com/watch?v=ncC-GMzF9RY) [The Unity of Consciousness (Stanford Encyclopedia of Philosophy)](https://plato.stanford.edu/entries/consciousness-unity/) https://arxiv.org/abs/2302.00843 [Sequential predictive learning is a unifying theory for hippocampal representation and replay | bioRxiv](https://www.biorxiv.org/content/10.1101/2024.04.28.591528v1) [x.com](https://twitter.com/dlevenstein/status/1785311847928713579?t=9ryTq9_E8Q8Wd89Vk_4b0g&s=19) We need to find ways to make our biological neural networks more compatible with our engineered neural networks and other information processing systems for data communication and merging. [Watch These 39 Minutes If You Want To Live To 200+ - YouTube](https://youtu.be/hkS1Eww5jTc?si=h6x4ZXPelKrI4VZ_) Illya next token prediction is enough for AGI [x.com](https://twitter.com/ns123abc/status/1785504804619608367?t=j74nJZlDvMJasxvreNzAew&s=19) https://arxiv.org/abs/2310.13018 MLPs are so foundational, but are there alternatives? MLPs place activation functions on neurons, but can we instead place (learnable) activation functions on weights? Yes, we KAN! We propose Kolmogorov-Arnold Networks (KAN), which are more accurate and interpretable than MLPs. [x.com](https://twitter.com/ZimingLiu11/status/1785483967719981538?t=08imquqmzvPs2sMuDzAtaA&s=19) https://arxiv.org/abs/2404.19756 LLM that turns any problem into a Turing Machine problem and solve it using a Turing machine LLM that turns any problem into a python code and solves it bc running that Python code Med Gemini [x.com](https://twitter.com/alan_karthi/status/1785117444383588823?t=nehPygbWzxnoqw4OXnV5mw&s=19) https://arxiv.org/abs/2404.18416 [A Simulated Universe | David Chalmers and Scott Aaronson | Part 3 - YouTube](https://m.youtube.com/watch?v=7PlmOXQ18jk&t=14s&fbclid=IwZXh0bgNhZW0CMTEAAR1InlecrUjLZf2p_cBDJQEib49zoOwuBBA9-zxghnKD03cYAW9kmtN1oeY_aem_AchR4ujWkEvtWJXm_f6cKJCGFrzd6JFU8Gs8d2WjNTiRpILzjcNkQpJBwGwiAFYhEHBcb28wUE73ffhfBPGXDroI) [How AI Is Unlocking the Secrets of Nature and the Universe | Demis Hassabis | TED - YouTube](https://youtu.be/0_M_syPuFos?si=Zz4DgvxC1A9j5Peh) [x.com](https://twitter.com/Dr_Singularity/status/1785403555525837287?t=TU0gVBq13pM-KFxfkDpdbw&s=19) ASI is born, quickly leads to unimaginable abundance. Soon, you will be a billionaire. ASI is born, quickly takes over, exterminates or enslaves or mind-controls all of humanity. Soon, you will be a slave, or a pet, or dead. Reality might be dystopia, or anarchy, or some combination or in the middle. It might also be a gradual process instead of a single point in time that is already unfolding. Pick your predictions, or build them into reality. Caught a civilization trying to access forbidden time technologies. Had a nice chat about the ethics of temporal manipulation. Staying optimistic doesn't just feel good, it will help you live longer. We are still early, but because progress in AI and robotics in exponential, soon (2030's) there will be billions of them working 24/7. Global wealth creation will be enormous by today's standards. Rogue ASI isn't by default as many assume [Coding the Cosmos: Does Reality Emerge From Simple Computations? - YouTube](https://www.youtube.com/live/ITJ3AF3TK5M?si=c0R9_lYcQW4vps_I) [What is GitHub Copilot Workspace? Sneak peek into your new developer environment - YouTube](https://youtu.be/pkotufZchjE?si=XvPxrOATqB_ywwgQ) [Robot bee swarms fly collision-free in close formation](https://newatlas.com/robotics/festo-bionicbee/) [Scientists Find a Surprising Way to Transform A and B Blood Types Into Universal Blood](https://singularityhub.com/2024/04/29/scientists-find-a-surprising-way-to-convert-a-and-b-blood-types-to-universal-blood/) If we solve/cure aging, we will have the time to solve everything else. Ethics of imprisoning the immortal humans. How do prison systems adapt to a civilisation that doesn't age (post ASI humans)? What does a life sentence mean in an ageless society? [x.com](https://twitter.com/iScienceLuvr/status/1785135037379199162?t=97c7yB2ecBXQombJFCzq0g&s=19) https://arxiv.org/abs/2404.18021 [What Supplements does Ray Kurzweil take and why? — TRANSCEND](https://transcend.me/blogs/supplementation/what-supplements-does-ray-kurzweil-take-and-why) [Verity - Will AI Replace All Jobs?](https://www.improvethenews.org/controversy/ai-replace-jobs) US AI regulation [x.com](https://twitter.com/nearcyan/status/1784864119491100784?t=0hIccKIeyMdfl7bW5Sa2VQ&s=19) [#65 Prof. PEDRO DOMINGOS [Unplugged] - YouTube](https://youtu.be/IUngGy9P3kE?si=ZcAU31ituuG7lySq) [The Master Algorithm - Wikipedia](https://en.wikipedia.org/wiki/The_Master_Algorithm) LLMs history [x.com](https://twitter.com/jannchie/status/1784621770018058651?t=kvpbtPXJxnGOnab9Lz3VOw&s=19) [HRL Laboratories | News | HRL Demonstrates the Potential to Enhance the Human Intellect's Existing Capacity to Learn New Skills](https://www.hrl.com/news/2016/02/10/hrl-demonstrates-the-potential-to-enhance-the-human-intellects-existing-capacity-to-learn-new-skills) https://www.scientificamerican.com/article/we-can-now-send-thoughts-directly-between-brains/ [Brain–computer interface - Wikipedia](https://en.wikipedia.org/wiki/Brain%E2%80%93computer_interface?wprov=sfla1) Do you compare things by distance between points in geometric space or by symbolic structural similarity,... By geometrical or topological analysis? [#51 FRANCOIS CHOLLET - Intelligence and Generalisation - YouTube](https://youtu.be/J0p_thJJnoo?si=N1y08o2tcMpNUaCy) 44:00 "You can't have infinite growth on a planet with finite resources!" After all, why shouldn't we try to solve nuclear fusion, print atoms and molecules, build Dyson spheres, mine other planets, and travel to other solar systems and galaxies [x.com](https://twitter.com/burny_tech/status/1784291567210988019?t=yuy3mjhATUjWuroNgEf4mg&s=19) But could grounded math and coding selfplay generalize to other reasoning modalities for general superintelligent systems though? [x.com](https://twitter.com/teortaxesTex/status/1784202972559298895?t=G1qmVwpBXkuZpFO7BAoLoA&s=19) Or [x.com](https://twitter.com/cheng_pengyu/status/1780965366531006887?t=N-fYiFyMiORTBCPqsyT1lA&s=19) [Pioneering Quantum Physicists Win Nobel Prize in Physics | Quanta Magazine](https://www.quantamagazine.org/pioneering-quantum-physicists-win-nobel-prize-in-physics-20221004/) [Reinforcement Learning By the Book - YouTube](https://youtube.com/playlist?list=PLzvYlJMoZ02Dxtwe-MmH4nOB5jYlMGBjr&si=mM06QeUJmmuQ_yze) [The future of the food industry: Food tech explained](https://www.techtarget.com/whatis/feature/The-future-of-the-food-industry-Food-tech-explained) [Sir Michael Atiyah - From Algebraic Geometry to Physics - a Personal Perspective [2010] - YouTube](https://m.youtube.com/watch?si=8HGj7FoaeynhygKo&t=139&fbclid=IwZXh0bgNhZW0CMTEAAR1CIpzfJSRocftvlkknZY85aM1nuEJYHRc_MHVM-pRHdfQIyQTMA2TVAdQ_aem_AbKrcSdtjWqauljEgYffySnHXV0izkxNMQTudAk2mBgIEtNlDMXWozbtlUCHADBTE6csi1mE1Yy_HQ9x9oLc82RZ&v=wvpNhZEIlN4&feature=youtu.be) I think generating fast cheap quality food is in majority a technological issue Tím nalepkuju celej ten proces produkce v supply chainu S asociovánýma technologiema [The future of the food industry: Food tech explained](https://www.techtarget.com/whatis/feature/The-future-of-the-food-industry-Food-tech-explained) Vidím dost budoucnost v levnějším a levnějším naškálovaným 3D printed jídle [Revo Foods launches ‘industrial-scale’ food 3D printer - 3D Printing Industry](https://3dprintingindustry.com/news/revo-foods-launches-industrial-scale-food-3d-printer-227699/) Ale souhlasím že největší problém současného systému je zvětšující se divide mezi rich a poor, mezi powerful a powerless, a často přemýšlím jak by to šlo destabilizovat I want to redistribute the benefits of technology more [x.com](https://twitter.com/getjonwithit/status/1784258756202688675?t=jb3V1gUPSXsOJJKB0mBGrg&s=19) here's a fun "coincidence" (maybe...): There are exactly four known fundamental forces (gravitational, electromagnetic, weak, strong), and exactly four normed division algebras (real numbers, complex numbers, quaternions, octonions). The set of complex numbers of magnitude 1 under complex multiplication forms the gauge group of electromagnetism: U(1). The set of quaternions of magnitude 1 under quaternionic multiplication forms the gauge group of the weak force: SU(2). The automorphism subgroup of the octonions preserving a magnitude-1 imaginary element i^2=-1 forms the gauge group of the strong force: SU(3). And the isometry group of a (flat) Lorentzian manifold with real-valued coordinates forms the gauge group of gravity: ISO(1, 3). Clearly, the relationship is most direct in the case of the electromagnetic and weak forces, which were unified first (the Weinberg-Salam model). Clearly, the relationship is least direct in the case of gravity, which remains un-unified with the other three. [Informal QFT 1 - Classical Gauge Field Theory - YouTube](https://youtu.be/PueK5qkHMmc?si=yp7NCPpV92bZw6Q-) [The Biggest Ideas in the Universe: Introduction - YouTube](https://youtu.be/HI09kat_GeI?si=HjeUPvuMEivjnN8B) AI agent landscape AI infrastructure landscape [x.com](https://twitter.com/chiefaioffice/status/1783932905355362745?t=B2mEpTPiwurIRv6fJuPBDA&s=19) [Amazon could run out of workers in US in two years, internal memo suggests | Amazon | The Guardian](https://www.theguardian.com/technology/2022/jun/22/amazon-workers-shortage-leaked-memo-warehouse) https://finance.yahoo.com/news/amazon-grows-over-750-000-153000967.html?guccounter=1 Claude 3 Opus can simulate a Turing Machine. The ability to be a (universal) Turing machine could, in principle, be the foundation of the ability to reliably perform complex rigorous calculation and cognition - the kind of tasks where there is an exact right answer, or exact constraints on what is a valid next step, and so the ability to pattern-match plausibly is not enough. And that is what people always say is missing from LLMs. [x.com](https://twitter.com/ctjlewis/status/1779740038852690393?t=qmuJ2foWJD3lSA5SYP2dPQ&s=19) "If we use the framework derived from Kolmogorov and Martin-Löf, randomness *just is* unpredictability or incompressibility. Is a process "really" random/noisy? The only negative evidence is prediction. I elaborate on the mathematics of randomness here:" https://3quarksdaily.com/3quarksdaily/2014/10/randomness-the-ghost-in-the-machine.html [TechScape: How cheap, outsourced labour in Africa is shaping AI English | Technology | The Guardian](https://www.theguardian.com/technology/2024/apr/16/techscape-ai-gadgest-humane-ai-pin-chatgpt) [‘It’s destroyed me completely’: Kenyan moderators decry toll of training of AI models | Artificial intelligence (AI) | The Guardian](https://www.theguardian.com/technology/2023/aug/02/ai-chatbot-training-human-toll-content-moderator-meta-openai) [Thermal computing is heating up | New Scientist](https://www.newscientist.com/article/dn16512-thermal-computing-is-heating-up/) Extropic opinion [x.com](https://twitter.com/0xKyon/status/1784591427462389822?t=NawuhkRNQVRQn-WETT64Nw&s=19) Science of consciousness conference [x.com](https://twitter.com/Caldwbr/status/1784347239294390389?t=LSfQWUHIAF5N0I0B6rcocg&s=19) Computronium optimized for hedonium. [x.com](https://twitter.com/algekalipso/status/1784359087423291518?t=bjDVd_XMaeMfqUyhFkYwAw&s=19) Become one with the math Statespace of minds [x.com](https://twitter.com/ukc10014/status/1784497737485955279?t=7KpSzwPUDNQOVlsKIBOztQ&s=19) Let's map out the space of information processing systems: biological, nonbiological and hybrids. Existing and those that could exist. Where each degree of freedom (discrete or continuous) corresponding to the mathematics by which it's governed: hardcoded, emergent, rigid and flexible. The data it's processing, the architecture it's running on, it's morphology, the algorithms it implements, the various scales of fundamental substrates it's embedded in, it's developmental stages, it's cognitive patterns, it's behavioral patterns, and so on. Consciousness in LLMs https://arxiv.org/abs/2402.12422 [Connor Leahy - e/acc, AGI and the future. - YouTube](https://youtu.be/m459AQ1o_60?si=NhC2NzCM1bdXykBo) [OSF](https://osf.io/preprints/psyarxiv/4cbuv) Maxx out agency creativity meaning [Vědci vytvořili umělé buňky z programovatelné DNA. Jsou odolnější než ty přírodní — ČT24 — Česká televize](https://ct24.ceskatelevize.cz/clanek/veda/vedci-vytvorili-umele-bunky-z-programovatelne-dna-jsou-odolnejsi-nez-ty-prirodni-348642?fbclid=IwZXh0bgNhZW0CMTEAAR0ptnk5j5NYkHCAShr3ErqNAiHzgq96AL4wI48UCoklqFFXv-3kRDWaCbc_aem_ASZFLRbe8E0CsQfoydfmqo6XubUs9IYInptIPZ8CWgxP8CU8DVqgtciCMgyQnFO1jHIW6nnlV5Dmr-MF_185x6Jf) [Designer peptide–DNA cytoskeletons regulate the function of synthetic cells | Nature Chemistry](https://www.nature.com/articles/s41557-024-01509-w) https://arxiv.org/abs/2402.03175 [Exponential Quantum Speedup for the Traveling Salesman Problem](https://eprint.iacr.org/2024/626) https://www.lesswrong.com/posts/y9tnz27oLmtLxcrEF/plainly-coded-ai-may-be-feasible-by-using-gpt-6 OpenELM - a new open language model that employs a layer-wise scaling strategy to efficiently allocate parameters and leading to better efficiency and accuracy; comes with different sizes such as 270M, 450M, 1.1B, and 3B. [x.com](https://twitter.com/dair_ai/status/1784608604093292860) Arctic - an open-source LLM (Apache 2.0 license.) that uses a unique Dense-MoE Hybrid transformer architecture; performs on par with Llama3 70B in enterprise metrics like coding (HumanEval+ & MBPP+), SQL (Spider) and instruction following (IFEval). [x.com](https://twitter.com/dair_ai/status/1784608605821370625) Make Your LLM Fully Utilize the Context - presents an approach to overcome the lost-in-the-middle challenge common in LLMs. It applies an explicit "information-intensive" training procedure on Mistral-7B to enable the LLM to fully utilize the context. [x.com](https://twitter.com/dair_ai/status/1784608607536848903) Self-Evolution of LLMs - provides a comprehensive survey on self-evolution approaches in LLMs. [x.com](https://twitter.com/dair_ai/status/1784608616210706916) Naturalized Execution Tuning (NExT) - trains an LLM to have the ability to inspect the execution traced of programs and reason about run-time behavior via synthetic chain-of-thought rationales; improves the fix rate of a PaLM 2 model on MBPP and Human by 26.1% and 14.3%... [x.com](https://twitter.com/AnsongNi/status/1783311827390070941) https://www.consciousentities.com/mogi.htm https://arxiv.org/abs/2403.15796 [x.com](https://twitter.com/dmvaldman/status/1784699985642307660?t=JsyaaHR2P94RJjvUVRqazA&s=19) "Great paper, arguing emergent abilities are only a function of pre training loss and not model/dataset size. ie, if you (inefficiently) overtrain a small model to the loss of GPT4, you'd get all the abilities of GPT4." "Thank you, to the people who have the strength and the moral resolve to look at that darkness in the world. I know it's hard, but we can only fix the problems we can look at. Let's try to make sure that the superintelligent AI doesn't learn from how most humans treat animals. Let's try to make sure that the AI cares about all beings, no matter how intelligent, no matter how cute, no matter how emotionally salient. Let's make a superbenevolent AI, that is not just smarter than any human, but more kind than any human." [x.com](https://twitter.com/Kat__Woods/status/1784759081280082067?t=vMmks-MziMZ72ocMFXOtkA&s=19) https://direct.mit.edu/jocn/article-abstract/doi/10.1162/jocn_a_02146/120312/The-Spiraling-Cognitive-Emotional-Brain?redirectedFrom=fulltext [Quantum Computing Meets Genomics: The Dawn of Hyper-Fast DNA Analysis](https://scitechdaily.com/quantum-computing-meets-genomics-the-dawn-of-hyper-fast-dna-analysis/) https://www.pcmag.com/news/china-creates-neucyber-its-version-of-neuralink-brain-chip?fbclid=IwZXh0bgNhZW0CMTEAAR3ugPMcuBVEdI7tLzFtHTkwJ_jB4kPsYeyZ_6dQeAFGlMJUrnqwzVASOQo_aem_AaKlkwp_6SfcTOCwWwnoi5DBB7Fbe_2PnXPqq-NV4RwOkeIKJI1LwytBSEjDXaZWcL2W6FPteBRtrbbcM5gQLg3G [The biggest science breakthroughs in 2023 - YouTube](https://www.youtube.com/watch?v=mtBgoic_ogQ) [Machine Learning: The Great Stagnation - by Mark Saroufim](https://marksaroufim.substack.com/p/machine-learning-the-great-stagnation) Maskology is getting better in LLMs [x.com](https://twitter.com/teortaxesTex/status/1782910115919675778) https://www.promptingguide.ai/ https://arxiv.org/abs/2402.07927 https://towardsdatascience.com/using-self-organizing-map-to-bolster-retrieval-augmented-generation-in-large-language-models-5d739ce21e9c loop quantum gravity x string theory [String Theory Meets Loop Quantum Gravity | Quanta Magazine](https://www.quantamagazine.org/string-theory-meets-loop-quantum-gravity-20160112/) I'm waiting when any of the LLM critics create at least smaller model that vastly outperforms smaller LLMs not in single small toy problems. I think lots of the criticisms are valid, but I wanna see results in practice which I don't see yet! [x.com](https://twitter.com/burny_tech/status/1785320274130243746?t=BLc_1DiahXmERJugrJAYpg&s=19) https://arxiv.org/abs/2303.14617 https://www.sciencedirect.com/book/9780128215456/conceptual-breakthroughs-in-the-evolutionary-biology-of-aging Dreaming is where the brain nightly fine-tunes itself on the day’s chat transcripts, but mixed in with enough stochastically synthesized data so as to minimize catastrophic forgetting of pretrained concepts No reason why you couldn’t also do this in an LLM system In a world of AI doomers, foom-ers and skeptics, aim for Reasonable Optimism [x.com](https://twitter.com/KompendiumProj/status/1773411105429447002?t=I0YWfA3oM1WNoWwzXe2fyw&s=19) Bayesian backpropagation alternative [TAGI | Tractable Approximate Gaussian Inference for Bayesian Neural Networks - YouTube](https://youtu.be/jqd3Bj0q2Sc?si=BCprHZsGK_efrNz6) Types of antiaging papers https://imgur.com/a/ZpIihwU https://www.cell.com/trends/cognitive-sciences/fulltext/S1364-6613(24)00075-5 [StoryDiffusion: Consistent Self-Attention for Long-Range Image and Video Generation](https://storydiffusion.github.io/) [Paper page - Better & Faster Large Language Models via Multi-token Prediction](https://huggingface.co/papers/2404.19737) [MIT Claims Superconducting Breakthrough Means Fusion Power Can Be Practical](https://futurism.com/the-byte/mit-magnets-ready-fusion?fbclid=IwZXh0bgNhZW0CMTEAAR2Bv-2g3Ebfc8DfUIQHOPjW1MIyEb4hnIVtzDog3RiSE3yTbmap75ibj8Q_aem_Ad4p18w6sBesxqCF22egtDel6bjiqZiUXG7vdthl_mzNstoCe1WXzyZrk6ztNyCgF0F0Kn_IP6biLF_6tISVEGp8) [Desalination system could produce freshwater that is cheaper than tap water](https://www.freethink.com/futurology/desalination?fbclid=IwZXh0bgNhZW0CMTEAAR15Mj3suhQEs6Eao7r3uZutFnSEL00VVbEorP0UZqNFqo7rxEWamKsHcFM_aem_ATy19jHhm-Q9NN1NAGAHOfUlKb5KOc9Awr_M7J7jxZv5h0b2mBNgs-IhHPBd6n4VToltc09ePprBCAOMSR04eugg) https://www.marktechpost.com/2024/04/24/researchers-at-mit-propose-maia-an-artificial-intelligence-system-that-uses-neural-network-models-to-automate-neural-model-understanding-tasks/ https://arxiv.org/abs/2404.14394 Mamba from scratch [MAMBA from Scratch: Neural Nets Better and Faster than Transformers - YouTube](https://youtu.be/N6Piou4oYx8?si=Nv7UDWxzWvL2pNBh) Beaming knowing Microsoft CEO Satya Nadella: we are in Year 2 of the Intelligence Revolution and scaling laws will bring greater reasoning, planning and memory, leading to a new phase of economic growth [x.com](https://twitter.com/tsarnick/status/1785415907713429967) Demis Hassabis: if humanity can get through the bottleneck of safe AGI, we could be in a new era of radical abundance, curing all diseases, spreading consciousness to the stars and maximum human flourishing Let's build this future [x.com](https://twitter.com/tsarnick/status/1785243160589021656) Mathematics is the queen of sciences https://imgur.com/AiyjQr4 Machine learning books [x.com](https://twitter.com/DionysianAgent/status/1785395743039062368) https://imgur.com/ZzpTIt9 [Why a Forefather of AI Fears the Future - YouTube](https://www.youtube.com/watch?v=KcbTbTxPMLc) [Something Strange Happens When You Follow Einstein's Math - YouTube](https://www.youtube.com/watch?v=6akmv1bsz1M) Andrew Mack (my MATS mentee) found an *unsupervised* method to elicit latent model capabilities, find backdoored outputs (without knowing how to activate the backdoor!), and override safety training. [Mechanistically Eliciting Latent Behaviors in Language Models — AI Alignment Forum](https://www.alignmentforum.org/posts/ioPnHKFyy4Cw2Gr2x/mechanistically-eliciting-latent-behaviors-in-language-1)