Accelerate knowledge Mechanistic interpretability acceleration Next generation of capable AI systems will be neurosymbolic https://www.lesswrong.com/posts/C5guLAx7ieQoowv3d/lecun-s-a-path-towards-autonomous-machine-intelligence-has-1 Those that will control the most intelligent systems in the future will control the future https://www.perplexity.ai/search/arrow-impossibility-theorem-JdFD6..xT56FXlrEuKCm7Q [[2206.07697] MACE: Higher Order Equivariant Message Passing Neural Networks for Fast and Accurate Force Fields](https://arxiv.org/abs/2206.07697) https://www.pnas.org/doi/10.1073/pnas.2310223120 [[2406.04127] Are We Done with MMLU?](https://arxiv.org/abs/2406.04127) [[2406.04303] Vision-LSTM: xLSTM as Generic Vision Backbone](https://arxiv.org/abs/2406.04303) [Neural Networks, Manifolds, and Topology -- colah's blog](https://colah.github.io/posts/2014-03-NN-Manifolds-Topology/) [[2406.04268] Open-Endedness is Essential for Artificial Superhuman Intelligence](https://arxiv.org/abs/2406.04268) Nobody actually knows if scaling will continue to drastically increase AI capabilities or not. Everyone is just guessing, or predict using limited currently known theory, but it's fundamentally an empirical problem. [How Does the Screen of Consciousness Emerge? - Competing Clusters, Blanketing, Rastering, Filling-In - YouTube](https://youtu.be/Pua9gI15KCs?si=aLBzotUhb11XGd0e) https://openai.com/index/extracting-concepts-from-gpt-4/ mechanistic interpretability [[2406.00244] Controlling Large Language Model Agents with Entropic Activation Steering](https://arxiv.org/abs/2406.00244) https://x.com/proceduralia/status/1798085571337683371?t=2OwpgSBJYKo1DL70xIUmGQ&s=19 We're finally inventing more steering wheels for AI systems [New theory suggests time is an illusion created by quantum entanglement](https://bgr.com/science/new-theory-suggests-time-is-an-illusion-created-by-quantum-entanglement/?fbclid=IwZXh0bgNhZW0CMTEAAR0e-R96IRXM4H6lhegQbrnsls1dpB1G-VQSSbuHb_HDw9nwHzQ6Gb-FAwM_aem_ZmFrZWR1bW15MTZieXRlcw) Reinforcing tokens for selfcorrection, system two thinking, planning, search, agentic behavior etc. could implicitly emerge neurosymbolic circuits that are not as hardcoded [Leopold Aschenbrenner - 2027 AGI, China/US Super-Intelligence Race, & The Return of History - YouTube](https://youtu.be/zdbVtZIn9IM?si=maVa6Nb7-MUzDttu) https://x.com/skdh/status/1798322646540529957?t=7MWa3VJw1V7eBecKugEICg&s=19 "AI becoming intelligent and AI developing its own values and being able to act on them are two different problems. The former -- intelligence -- is likely to precede the latter. I don't think it will be long and I think not enough attention is being paid to the most obvious immediate problem: Any individual, company, or government in possession of a device that vastly outperforms average human intelligence will be incredibly powerful. The potential for abuse is enormous. The human brain doesn't work with magic. It's a big wobbly thing with a lot of connections that can process a lot of information -- sometimes well, sometimes not so well. It is not an optimal device for intelligence because it's not what it evolved to be. The human brain evolved to keep us alive. This means among other things it needs to be energy efficient, which indeed it is quite good at. But artificial brains have no such limits so of course they will eventually outperform humans. If you don't understand why superior intelligence is scary, you've read too many touchy sci-fi stories in which human irrationality saves the day. But humans didn't come to dominate this planet because they're somtimes irrational, they came to dominate despite of it." Yes, humans aren't special and we evolved not to be the most intelligent, our fitness function is to survive in our evolutionary environment. (if you don't define intelligence as the ability to survive) The space of possible machine learning architectures and algorithms and hardware that it could be running on is so enormous, and what is currently the most popular, which is large language models, which is deep learning, training models through curve-fitting differentiable parametric curves on transistor-based digital computers with their current transistor-based Von Neumann architecture, GPUs with specialized circuits for matrix multiplication, TPUs, is just a tiny subspace of all possible machine learning architectures, algorithms and hardware. AI field is full of different types of algorithms and architectures that are under the radar, such as statistical learning with regression and clustering, various alternative deep learning architectures to Transformer and Diffusion (which AlphaFold combines), such as Mamba, LSTMx, RWKW, Jamba, convolutional neural networks, graph neural networks (TacticAI), etc., then symbolic, neurosymbolic systems (AlphaZero with reinforcement learning, AlphaGeometry, DreamCoder), energy based models (Yann Lecun), bayesian AI (Active Inference, VERSES), forward forward algorithm (Hinton), neural cellular automata and selforganizing AI (Joscha Bach), evolutionary AI, causal learning, and so on. Also various AI accelerators, whether classical digital processors supporting different algorithms (matrix multiplication) and architectures (e.g., specialized hardware for artificial neural networks and Transformers), quantum processors (quantum machine learning), analog processors (optical), stochastic processors (Extropic) (thermodynamic), or bio processors (organoids) (FinalSpark) etc., implementing different machine learning algorithms that make more use of the hardware level as accelerators for higher computational efficiency, are becoming more and more common. I am interested in what other paradigms and hybrids will emerge at both the hardware and software levels! We are predicting the world into existence and editing it by actions "🧠/acc, safe-by-design/acc, math/acc, biosecurity/acc, importantly **Pareto-topia/acc** (see link below — and therefore, nano/acc, cryo/acc, 🚀/acc), cooperation/acc, dangerous-geopolitical-conflict/decel Overall, d/acc" [Paretotopian Goal Alignment | Effective Altruism](https://www.effectivealtruism.org/articles/ea-global-2018-paretotopian-goal-alignment) https://x.com/AdamMarblestone/status/1798306865287508421?t=JrloRVDmYr5Ji2IBY4Nafw&s=19 [[2406.01705] Dark Matter](https://arxiv.org/abs/2406.01705) [[2405.15059] Message-Passing Monte Carlo: Generating low-discrepancy point sets via Graph Neural Networks](https://arxiv.org/abs/2405.15059) https://x.com/tk_rusch/status/1798053622703665541?t=tCdMFjRh6InKRsDFVyIH8w&s=19 [Researchers Find That Higher Intelligence Is Correlated With Left-Wing Beliefs](https://futurism.com/neoscope/left-wing-beliefs-intelligence?fbclid=IwY2xjawDGePhleHRuA2FlbQIxMQABHfRkGIvYOlljVpr_XJR7r6NoivcXrio0rDL9102fokDRmihF-h9IJcw0xQ_aem_AYa5B5DEB9_UnuhR3pGSv253CM_GnaOXL_w3UjaauujDdrEDRsgWK597lCGe1MelHe6a9jTk4h-IYDrLBMe89ngr) [[2406.00104] Scalable Bayesian Learning with posteriors](https://arxiv.org/abs/2406.00104) https://x.com/Sam_Duffield/status/1797992718473261559?t=5co8yKgvBK1_tgETqKEAvw&s=19 [[2406.02543] To Believe or Not to Believe Your LLM](https://arxiv.org/abs/2406.02543?fbclid=IwZXh0bgNhZW0CMTEAAR3zZG7IpWNJA6ibwPI4yvKzk2pD0kfq5yopfqtDwNCFdeVtj8LgwLNYlCg_aem_ZmFrZWR1bW15MTZieXRlcw) Flatworms have a peculiar characteristic: when sliced in pieces, each piece will grow into a complete flatworm, just like the original one https://x.com/bryan_johnson/status/1798077358651617352?t=MjXrcYQxA9KNmaY35s-P4A&s=19 Everything is a hierarchical composition of fundamental components. The level of abstraction or depth defines how far that new composition is from the original building blocks. https://x.com/bindureddy/status/1798038699860738305?t=G52ySzMCDdmho8mTUASnww&s=19 All AI services down for few hours. No no no the actual super true fundamental truth and the only truth is that the alien ultra super computer AGSI from alien intergalactic hyperdimensional civilization through wormholes through string theoretic strings made of loop quantum gravity that they give to the globalist new world order lizards that rule our planet which only give the general public peasants infinitesimally small percent of the alien superintelligence was off for a second because of a few hours disagreement on priorities between our lizard world leaders and aliens. Even the alien superintelligences are brainwashed! https://x.com/burny_tech/status/1798358432875233549?t=52GAY1bWZh65FwHIzM7dKA&s=19 [[2406.01637] Teams of LLM Agents can Exploit Zero-Day Vulnerabilities](https://arxiv.org/abs/2406.01637) https://x.com/daniel_d_kang/status/1798363410511675469?t=eummXHRb4ZAJwsh-N50oiA&s=19 When music AI model generation where you can give arbitrary song to and it will generate similar songs The metacontrarian take 🤔 https://boards.4chan.org/sci/thread/16211427#p16211510 "there are three types of people in the AI field: >LLM statistics bros, who have a barely functioning IQ and the reason anything works is magic (just add another matrix and multiply it) >profiteering orbiters who want to speculate on hype (SBF funded like 1000 people) >salty bitter clingers who were trying other techniques for decades and got nowhere (old professor types with outdated ideas)" [Shafkat Rahman | Are.na](https://www.are.na/shafkat-rahman/channels) [[2405.14906] AutoCoder: Enhancing Code Large Language Model with \textsc{AIEV-Instruct}](https://arxiv.org/abs/2405.14906) [[2406.02061] Alice in Wonderland: Simple Tasks Showing Complete Reasoning Breakdown in State-Of-the-Art Large Language Models](https://arxiv.org/abs/2406.02061) giga mecha legs https://x.com/AceRuele/status/1798064804466508201 https://x.com/MichaelTrazzi/status/1798409060301332933 "> be leopold > born in non-meritocratic country > exit to live american dream at 15 > bored graduating valedictorian, so write 100 pages on economic growth > tyler cowen is shocked, thinks it's an excellent PhD thesis > tyler's friend calls saying he needs someone to allocate billions in funds > just graduated so accept to pass time > fund blows up > realize that with agi race dynamics, only path is OAI winning by 2 years, which would be enough to align superintelligence > dm sama: "if you give me 20% of OAI's compute i'll align the digital gods and give you back 50% of the galaxy" > sama is shocked, but accepts because the whole "Superalignment" thing seems like a galaxybrain marketing move > during board drama, refuse to sign subpar letter > figure out OAI won't be able to start galaxy sized civilizations, and we need a Manhattan project instead > tell board and leadership that their security sucks > was right two weeks later > get fired because they need to prioritize a .5% increase in MMLU instead of securing model weights > they offer $1M NDA to silence you. you say no, freedom is priceless > tyler cowen's calls you: "you know patrick? he has billions" > start managing patrick's AGI investments to pass time > realize the US is nowhere near building gigawatt factories for $100B training runs > write a 160 pages pdf so new world builds 100 GW clusters before the CCP > remember that your podcaster friend has recently interviewed zuck about GW clusters > dm him: "podcast, my place, in 1h. there's a dolores view" > after watching the footage, the video editor is in shambles. he quits > interview is published. entire universe is AGI pilled, including the CCP" langchain alternatives https://x.com/gfodor/status/1798372760558965124 [Musk is Right: Neutrinos Are Evidence for New Physics - YouTube](https://www.youtube.com/watch?v=CNxkFDhhtJY) [Daniel Schmachtenberger: "A Vision for Betterment" | The Great Simplification 126 - YouTube](https://www.youtube.com/watch?v=tmusbHBKW84) Intelligence is in simple terms the ability to handle uncertainty in real world situations. for illya intelligence is in big part compression [An Observation on Generalization - YouTube](https://www.youtube.com/watch?v=AKMuA_TVz3A) [The Role of Occam's Razor in Knowledge Discovery | Data Mining and Knowledge Discovery](https://link.springer.com/article/10.1023/A:1009868929893) [[2404.19157] Scalable Bayesian Inference in the Era of Deep Learning: From Gaussian Processes to Deep Neural Networks](https://arxiv.org/abs/2404.19157) Links for 2024-06-05 1. Evidence of Learned Look-Ahead in a Chess-Playing Neural Network https://www.lesswrong.com/posts/GfwdBoaLw3ef3zBqe/evidence-of-learned-look-ahead-in-a-chess-playing-neural 2. “We empirically show that a GPT-style transformer exhibits a transition from in-distribution to out-of-distribution generalization as the number of pre-training tasks increases.” [[2406.02550] Learning to grok: Emergence of in-context learning and skill composition in modular arithmetic tasks](https://arxiv.org/abs/2406.02550) 3. Show, Don’t Tell: Significantly outperforms few-shot prompting, SFT and other self-play methods by using demonstrations as feedback directly [GitHub - SALT-NLP/demonstrated-feedback](https://github.com/SALT-NLP/demonstrated-feedback) 4. To Believe or Not to Believe Your LLM (Quantify uncertainty to detect hallucinations) [[2406.02543] To Believe or Not to Believe Your LLM](https://arxiv.org/abs/2406.02543) 5. Artificial Generational Intelligence: Cultural Accumulation in Reinforcement Learning [[2406.00392] Artificial Generational Intelligence: Cultural Accumulation in Reinforcement Learning](https://arxiv.org/abs/2406.00392) 6. Learning Temporally Consistent Video Depth from Video Diffusion Priors [[2406.01493] Learning Temporally Consistent Video Depth from Video Diffusion Priors](https://arxiv.org/abs/2406.01493) 7. ‘Everything is Going to Be Robotic’ Nvidia Promises, as AI Gets More Real [‘Everything is Going to Be Robotic’ Nvidia Promises, as AI Gets More Real - YouTube](https://www.youtube.com/watch?v=nxO_t5N82m0) “Codestral solved a problem in two messages that I couldn't resolve with bouncing around between GPT4o, GPT-4 and Claude Opus for an hour.” https://www.reddit.com/r/LocalLLaMA/comments/1d86di0/codestral_solved_a_problem_in_two_messages_that_i/ Miscellaneous: 1. Spectacular new state of art for Brain-Computer Interfaces: After 5 recording sessions the word error rate was <10% and after 15 sessions, consistent <5% WER for online decoding. Stable performance 250 days after implantation. [An accurate and rapidly calibrating speech neuroprosthesis | medRxiv](https://www.medrxiv.org/content/10.1101/2023.12.26.23300110v2) 2. “Quantinuum’s H-Series hits 56 physical qubits that are all-to-all connected, and departs the era of classical simulation” [Quantinuum’s H-Series hits 56 physical qubits that are all-to-all connected, and departs the era of classical simulation](https://www.quantinuum.com/news/quantinuums-h-series-hits-56-physical-qubits-that-are-all-to-all-connected-and-departs-the-era-of-classical-simulation) 3. How a simple circuit could offer an alternative to energy-intensive GPUs [A simple circuit could offer an alternative to energy-intensive GPUs | MIT Technology Review](https://www.technologyreview.com/2024/06/05/1093250/how-a-simple-circuit-could-offer-an-alternative-to-energy-intensive-gpus/) [archived: https://archive.is/bRDWQ] 4. how to find a location using only the shadow [how to find a location using only the shadow - YouTube](https://www.youtube.com/watch?v=pQIjDPFgdJA) 5. New map shows vast potential for geothermal energy beneath entire US https://thehill.com/policy/energy-environment/4703057-geothermal-energy-potential-beneath-western-eastern-us-map/ Definitions of consciousness. Existence of experience is one possible definition. Or Joscha Bach defines it as coherence inducing operator on mental representations. Even tho lots of religions, cults, cultures feel like rationally and empirically epistemic failure modes, they game theoretically and evolutionary serve as very memetically fit social cohesion and coordination mechanisms (sometimes by removing the statistically different outgroup lol) I'm the most in distribution organism walking on this rock flying in space Permanent intense curiousity is the good kind of tanha I think both e/accs and EAs are STEM nerds and tech accelerationists (at least early EA, now there seems to be more overall antitech people lately), and the main crux of disagreement is just mostly AI risk and the methods of mitigating it (top down vs bottom up control for example). From an outsider perspective (I'm also STEM nerd tech accelerationist interacting with both e/accs and EAs so I'm not really an outsider) it's not that different! Predicting Alzheimer's disease up to 9 years before a clinical diagnosis with fMRI analytics [Early detection of dementia with default-mode network effective connectivity | Nature Mental Health](https://www.nature.com/articles/s44220-024-00259-5) https://www.science.org/doi/10.1126/science.adp7977 Tumor-agnostic therapies are revolutionizing oncology! Instead of focusing on tumor origin, these treatments target specific genetic anomalies driving cancer growth. https://acsjournals.onlinelibrary.wiley.com/doi/10.3322/caac.21844 [100% of cancer patients cured long-term in 'remarkable' human trial](https://newatlas.com/medical/colorectal-cancer-dostarlimab-gxly/) [Nvidia hits $3 trillion market cap on back of AI boom](https://www.cnbc.com/2024/06/05/nvidia-briefly-passes-3-trillion-market-cap-on-back-of-ai-boom.html) The Potential of Equivariant Neural Network Potentials Neural network potentials might become the most important scientific tool of the next decade. The ability to simulate systems at the molecular scale starting from nothing but quantum mechanics will be transformative for a vast range of problems throughout biology and chemistry. "In the next half-century, physical chemistry will likely undergo a profound transformation, driven predominantly by the combination of recent advances in quantum chemistry and machine learning (ML). Specifically, equivariant neural network potentials (NNPs) are a breakthrough new tool that are already enabling us to simulate systems at the molecular scale with unprecedented accuracy and speed, relying on nothing but fundamental physical laws. The continued development of this approach will realize Paul Dirac’s 80-year-old vision of using quantum mechanics to unify physics with chemistry and providing invaluable tools for understanding materials science, biology, earth sciences, and beyond. The era of highly accurate and efficient first-principles molecular simulations will provide a wealth of training data that can be used to build automated computational methodologies, using tools such as diffusion models, for the design and optimization of systems at the molecular scale. Large language models (LLMs) will also evolve into increasingly indispensable tools for literature review, coding, idea generation, and scientific writing." https://pubs.acs.org/doi/epdf/10.1021/acsphyschemau.4c00004 Learning to grok: Emergence of in-context learning and skill composition in modular arithmetic tasks [[2406.02550] Learning to grok: Emergence of in-context learning and skill composition in modular arithmetic tasks](https://arxiv.org/abs/2406.02550) I don't think humans have special kind of the only one true intelligence. And I think that there are many types of intelligence with many different parts and I think that machines currently already have superhuman intelligence in various domains and subhumanbaby levels of intelligence in different domains. [[2406.00888] Show, Don't Tell: Aligning Language Models with Demonstrated Feedback](https://arxiv.org/abs/2406.00888) [Strings 2024 | Not Even Wrong](https://www.math.columbia.edu/~woit/wordpress/?p=13953) [[2406.02466] What no one has seen before: gravitational waveforms from warp drive collapse](https://arxiv.org/abs/2406.02466) Control of AI is the most important problem, doesn't matter if you think about it in terms of wanting good product, good tool or superalignment of possibly rogue digital God. We want the systems to do things that we want them to do, and not arbitrary random things nobody wants. "Whoever wins the superintelligence race changes the ontological grounding of civilization according to their own zeitgeist. That race will be determined relatively soon. Whoever wins, wins all, there is no prize for second place, and no multipolarity in superintelligence." https://x.com/SydSteyerhart/status/1776071372151693728 I still think it will be somewhat multipolar, but there will be dominant winners https://www.reddit.com/r/singularity/comments/1d9yo86/the_latest_releases_from_china_qwen_2_and_kling/ "The latest releases from China (Qwen 2 and Kling) are a massive middle finger to AI safetyists i.e. decels and corporates pushing regulations, creatives crying about copyright and people generally smug about Western superiority in AI These releases show how futile, hilarious and misguided their attempts at controlling technology and surrounding narratives are. They can try to regulate all they want, make all sort of bs copyright claims, lobby for AI regulations but they cannot stop other countries from accelerating. So essentially what they are doing in kneecapping their own progress and making sure they fall far behind other countries who don't buy their bullshit. It also counters the narrative that future of AI and AGI is only at the hands of Western countries. Politicians thought if they could block export of NVIDIA chips or make all sort of dumb tariff laws they could prevent China from progressing. They were wrong as usual. The only thing that works here is to stop the bs and accelerate hard. Instead of over regulating and gatekeeping, open up AI, facilitate sharing of weights, encourage broader participation in the development of AI and start large multi-nation collaborations. You cannot be a monopoly, you can only put yourself out of the game by making dumb decisions." Acceleration is needed to prevent CCP from winning! But control theory on top of these AI systems is equally as needed! What all insights can mathematical control theory give us to control current and future AI systems [[2404.09562] σ-GPTs: A New Approach to Autoregressive Models](https://arxiv.org/abs/2404.09562) https://x.com/ArnaudPannatier/status/1799055129829839166 σ-GPT, capable of generating sequences in any order chosen dynamically at inference time. i just wish selfish powergrabbing incentives weren't everywhere from industry to billionaires to governments etc. and it was just about benevolent autistic nerds wanting great transhumanist future for every being instead https://x.com/Kat__Woods/status/1799005270863143414 I'm not exactly team humanity, I'm utilitarian team sentience. I don't discriminate between current and future species, current and future sentient systems, having great conscious experiences, and ideally eventually populating the whole universe and maybe eventually beating the heat death of the universe.