[x.com](https://twitter.com/fchollet/status/1748780260295164114?t=ushdliEKfXa42Tb2pqdk5g&s=19) "Meanwhile what I mean is AI with general cognitive abilities, capable of picking up new skills with similar efficiency (or higher!) as humans, over a similar scope of problems (or greater!). It would be a tremendously useful tool in pretty much every domain, in particular science. Fluid intelligence is an information conversion ratio, therefore it has an upper bound: optimal efficiency. At some point in the distant future, our AI will get there (it will not take the form of a curve fitted to a large dataset, because that has very low efficiency). Reaching optimality will not confer the system will omnipotence. It simply means that it will no longer be bottlenecked by information processing. Instead, it will be bottlenecked by everything else -- starting with information collection." Oscillating between the industrial revolution and its consequences have been a disaster for the human race and technology if directed properly, not into dystopia with control in the hands of few, not into extinction, but into protopia, has the potential to create infinite sentient flourishing, growth, wellbeing, meaning, immortality, freedom through reverse engineering biology and the universe for advanced neurotech and physics and merging with machines and synthetic life forms, and building optimal physical and social technologies for collective flourishing and sustainability [Post AGI Education - What is the optimal educational experience and how can we get there? - YouTube](https://www.youtube.com/watch?v=N8PeXGhENNI) [#51 FRANCOIS CHOLLET - Intelligence and Generalisation - YouTube](https://www.youtube.com/watch?v=J0p_thJJnoo&t=1933s&pp=ygUQRnJhbmNvaXMgQ2hvbGxldA%3D%3D) [François Chollet: Keras, Deep Learning, and the Progress of AI | Lex Fridman Podcast #38 - YouTube](https://www.youtube.com/watch?v=Bo8MY4JpiXE&pp=ygUQRnJhbmNvaXMgQ2hvbGxldA%3D%3D) [François Chollet: Measures of Intelligence | Lex Fridman Podcast #120 - YouTube](https://www.youtube.com/watch?v=PUAdj3w3wO4&pp=ygUQRnJhbmNvaXMgQ2hvbGxldA%3D%3D) pions make most of mass [You Probably Don't Know Why You Really Have Mass - YouTube](https://www.youtube.com/watch?v=MyGIQ3RlKkU) superdoomer episode on mechanistic interpretability is doomed and we should halt AGI for decades i just shared [Uncut version! Roman Yampolskiy debates Roko Mijic: Basilisk, AGI Safety. Dan Faggella moderating. - YouTube](https://www.youtube.com/watch?v=FsEAnGllDCk) or mechainstic intepretability is too dual use and will be eplxoited by bad actors for cotnrol too much [Can (and should) we build godlike AI? | Robert Wright & Connor Leahy | Nonzero Clips - YouTube](https://www.youtube.com/watch?v=KggCQ5q6vfU) [Francois Chollet - On the Measure Of Intelligence - YouTube](https://www.youtube.com/watch?v=mEVnu-KZjq4&list=WL&index=2&pp=gAQBiAQB) [Uncut version! Roman Yampolskiy debates Roko Mijic: Basilisk, AGI Safety. Dan Faggella moderating. - YouTube](https://www.youtube.com/watch?v=FsEAnGllDCk) [x.com](https://twitter.com/burny_tech/status/1746681738691002804) Given the the research's design (způsob jak se učilo), benchmarks (jeden specifickej benchmark od Googlu), specialization, modalities (text) and other context, it seems it vastly outperform humans in that particular context, and in that context can technically save more human lives better than other skilled humans, measured by that particular (in certain ways limited) benchmark, and thats an an amazing technological achievement, and that performance will most likely accelerate and generalize out of its context even faster. But its still so limited, je to konverzační LLM, neřeší to co všechno je v nemocnici teď možný lidmama a jinýma technolgiema kde se ještě tolik neutomatizovalo, přímou human to human interakci, různý formy irl dialogů co ten benchmark nezahrnuje, různý testy na těle bez nebo s technolgiemi, léky, různý typy operací, a jiný věci co se ještě neautomatizovaly a co možná půjdou automatizovat víc těžce (některý doktory jsem viděl jak na ten paper odpovídali že by byli rádi kdyby jim LLMs zautomatizovaly domlouvání s pojšťovnama =D), co se pomalu automatizuje, podobně jako se tam využívají celkově počítače a nebo jak se začínají integrovat ML algoritmy např na rozeznávání rakovin ze skenů apod. Zrovna tohle bych byl rád kdyby google open sourcnul, podobně jako to jejich nový math AI co řeší geometrii na úrovni medalistů, aby to mohlo začít pomáhat, např aby ti co v Americe nebo v developed nations nemají vindru na healthcare mohli dostat alespoň něco, aby to pomohlo krizi nedostatku doktorů s alespoň něčím, nebo aby to někteří doktoři začalii využívat jako asistenta jako se používá Google (než se tyhle věci adaptují irl skrz tu byrokracii to bude asi realisticky roky), a by to se na to hodily různý benchmarky tuna jiných od jiných skupin než jen Google, vytvořily se další různý extensions, mutace apod., tohle si myslím by mohlo šíleně benefitovat hodně lidem. Jak se to začne víc zlepšovat a i specializovat, integrovat s více modalitama (víc než text), s robotikou, napojení na IT infrastrukturu uvnitř zdravotníctví, (alespoň tam kde se digitalizuje), napojení na internet apod. tak to má úžasný potenciál solidně transformovat zdravotníctví k levnějšímu a různých aspektech lepšímu stavu a různý části potenciálně i víc zdemokratizovat. Myslím že dost můžou přidat insighty z jiných projektů s LLMs s robotikou co např dělají autonomně chemii nebo materiální vědu. Myslím že dost můžou přidat insighty z jiných projektů s LLMs s robotikou co např dělají autonomně chemii nebo materiální vědu. Myslím že se může eventulně s velkým progressem celkově tato statistika dostat na menší počet smrtí strojema vs lidmama, ne jen v tomto niche benchmarku, jako např u selfdriving cars. [Intel's German fab will be most advanced in the world and make 1.5nm chips, CEO says | Tom's Hardware](https://www.tomshardware.com/tech-industry/manufacturing/intels-german-fab-will-be-most-advanced-in-the-world-and-make-15nm-chips-ceo-say) American Intel company wants a fab in Germany and the Taiwanese TSMC wants a fab in Arizona. Battling climate change with a mechanical tree that does the work of 1000 trees by the Arizona state university! [MechanicalTreeTM – Carbon collect](https://carboncollect.com/mechanical-tree/) Mechinterp backdoors [x.com](https://twitter.com/StephenLCasper/status/1748872347699081682?t=d2Hb1jKdRMny-tzY5lXiIg&s=19) [A generative model of memory construction and consolidation | Nature Human Behaviour](https://www.nature.com/articles/s41562-023-01799-z) https://medicalxpress.com/news/2024-01-generative-ai-human-memory.html https://phys.org/news/2024-01-machine-molecular-properties.html https://www.sciencedirect.com/science/article/pii/S2667318523000338 https://openai.com/research/practices-for-governing-agentic-ai-systems since pretty good AI models are small now, one could theoretically create an autonomous agent that feels more lively by running him as a program in an environment that observes thinks acts in a loop with hardcoded need for sleep, it could have vector database long term memory, context window short term memory,... or maybe after certain experiences do a very fast lots of finetuning to for example change emotion or personality or knowledge more than just storing it in context window and dabase, hardcode input of "im lonely" "im bored" etc. to stimulate various actions stored in the tool box (there could be a tool that could help create new tools or experimenting) this can be hardcoded maybe too, the basis for curiousity: RL agent doing actions and a world model predicting the conseauences of those actions. RL agent tried to generate actions such that the world model's error is maximized, leading to exploratory curiousity, and Active inference [Artificial Curiosity Since 1990](https://people.idsia.ch/~juergen/artificial-curiosity-since-1990.html) [Dynamic and selective engrams emerge with memory consolidation | Nature Neuroscience](https://www.nature.com/articles/s41593-023-01551-w) [Bloomberg - Are you a robot?](https://www.bloomberg.com/news/articles/2024-01-19/altman-seeks-to-raise-billions-for-network-of-ai-chip-factories) Once you figure out your brain, you become unstoppable. [David Goggins: How to Build Immense Inner Strength - YouTube](https://www.youtube.com/watch?v=nDLb8_wgX50) Visualizing RAG [x.com](https://twitter.com/hwchase17/status/1748730564130283914?t=8qLWfkdkLBi99MMHOsB2Ug&s=19) [[2401.07103] Leveraging Large Language Models for NLG Evaluation: Advances and Challenges](https://arxiv.org/abs/2401.07103) Imagine you could learn and realtime work with all of the quality knowledge on the internet billions times faster billion times more effectively with billion times more memory like the current AI models that are exponentially approaching that while humans are stuck in their evolutionary baseline of information processing capacity. Effective Neurotech Intelligence Augmentation is already starting to happen. Only in AI you're like "Oh this is from few months ago, that's like stone age in AI timelines" ML podcasts [x.com](https://twitter.com/chrisalbon/status/1749101112710565956) [x.com](https://twitter.com/ESYudkowsky/status/1749157823785931220) "My intuition is that the human brain is likely far from the optimality bound (20% of the way there?), but because getting closer to it has decreasing benefits, humans already possess most of the benefits intelligence confers (80%?). Hence modern technology and our collective ability to solve virtually any problem we put our minds to, by leveraging networked brains and externalized cognitive resources. Your neurons fire at 100Hz, send signals at 100 m/s along myelinated fibers, dissipate 20W, and can't multiply 6-digit numbers unaided. The ultimate thinking artifact permitted by the laws of physics is apparently FIVE TIMES more optimal than that, but has only 25% more benefit." How would we percieve if we upgraded our hardware [x.com](https://twitter.com/BasedBeffJezos/status/1749316697033691406?t=vU2A65ZcDDDehRPmWjSQIw&s=19) [Dan Hendrycks on Catastrophic AI Risks - YouTube](https://youtu.be/57y7DxWfOS0?si=MDDthoQyewT8aUfy) AI: Unpredictable, Unexplainable, Uncontrollable Roman yampolskyi [x.com](https://twitter.com/romanyam/status/1748790916586889385?t=WXJqBe2af1BCk6iRmzTqrw&s=19) AI GPT4 in harward course [x.com](https://twitter.com/emollick/status/1749239679788974247?t=SH_0llDKaOnt8o4sX5EFzA&s=19) AI agent Finding news and tweeting [x.com](https://twitter.com/DivGarg9/status/1749149133317996796?t=2kJOCBAHdTWg8ULUk5jrBQ&s=19) Longevity protocols [x.com](https://twitter.com/powerfultakes/status/1749090751693099188?t=hiasn5ux9hy6VwVsh_duAg&s=19) BCI/acc [x.com](https://twitter.com/trentmc0/status/1749095289909072063?t=sO5i5FzgxdHnhT-3Aif4Rw&s=19) There is no such thing as perfect moment to start. Start now! Amazon robotics acceleration [x.com](https://twitter.com/LinusEkenstam/status/1749216813416636791?t=C4FgnhO1S9L9N7yNywlwKA&s=19) Alignment in general is a game theoretic problem for arbitrary agents Lying being result of training data is the most common cause, and we dont have proper tools to properly remove it yet. Another cause could be instrumental convergence, it being useful emergent algorithm in internal dynamics for many objective functions regardless of training data. One of the hottest things in AI is currently getting the most quality data, and instrumental convergence is still a thing. Mechanistic interpretability is about finding ways to: even if it learns any of these circuits from training data or instrumentally emergently regardless of training data, they can be localized and removed. You can do higher of analysis of data using sentiment analysis to prevent those patterns you dont want Or you can create synthetic data, that's a big trend now too I think most humans agree on wanting machines to not ignore all our values. Like in the extreme if one isnt antinatalist we wouldnt want them to kill all humans or sentience in general. Certain prompt engineering and jailbreaks make them go completely unhinged and out of the distribution more than avarage humans We dont have a proper mathematical theory of instrumental patterns. I agree humanity isnt perfect. But i think mostly everyone agrees on at least the simple ground of various examples i sketched and alignment researchers work on before getting overly global and philosophical. Like for example your LLM or beyond being okay with killing and persuading everyone. Without the existing technical research's results from this, ChatGPT would be much more unhinged than it currently is, and we're slowly cracking more Certain agents use AI to make the world a better place as well. It is be exploited by bad agents or just terrible economic incentives too. But not everything is completely dark. I disagree with some people saying all technical results from AI alignment research and AI reverse engineering research (that we use in practice daily to control existing AIs not going completely unhinged and in all sorts of ways direct their training/inference/capabilities in any ways we want (like for example designing specific personalities in specific LLMs or making the LLM output as helpful, useful, direct responces, or making LLMs lie less, less machiavellianism), etc.) is not valid because some humans lie or because humanity isnt perfect on its own. This is a concrete technical problem, very concrete set of patterns that we are slowly mathematically identifying and localizing which makes it potentially fully controllable. More ideal might be if the AI was operating on more forced emergent or more hardcoded world model and halting it there. I like Connor as he also tries to be as pragmatic in engineering way as possible about this even if he also loves global and philosophical discussions. But i love Joscha too, but i feel like he takes many wanted good aspects of AI for granted instead as technical problems. [Joscha Bach and Connor Leahy [HQ VERSION] - YouTube](https://youtu.be/Z02Obj8j6FQ?si=PHLJhAy262dDc6Q_) Generative AI landscape [Generative AI in a Nutshell - how to survive and thrive in the age of AI - YouTube](https://www.youtube.com/watch?v=2IK3DFHRFfw) Democratizing Artificial Intelligence to Benefit Everyone: Shaping a Better Future in the Smart Technology Era [x.com](https://twitter.com/burny_tech/status/1749422918096822457) [The core message of this... - Zarathustra Amareis Goertzel](https://www.facebook.com/zarathustra.goertzel/posts/pfbid0Kznb1GoDtZF12XuZaxNZpK5sriHmtREx93wSJZw3m4MCikRhrHe3PykVEGerSwaKl) Against centralization against risk [x.com](https://twitter.com/BasedBeffJezos/status/1736221260827226360) [[2006.10739] Fourier Features Let Networks Learn High Frequency Functions in Low Dimensional Domains](https://arxiv.org/abs/2006.10739) "A jak by si rozlišil rozdíl mezi něčím, co má vědomí a co ne, proč / jestli by ho mohlo AI mít? ( nemusí bejt nutně, co říkají ostatní, spíš tvůj názor na to )" nikdo reálně netuší, což je trochu problém, já jsem agnostický nejdřív bychom museli mít verifikovaný kompletnější model vědomí v neurovědách, což nemáme osobně pocitově podle dosavadních pokusů o modely osciluju mezi pohledama třeba libovolný výpočty mají nějakou úroveň prožitku, třeba je potřeba sebereferenční struktura, třeba je potřeba konkrétní fyyzikální nebo biologický hardware, třeba nějaký integrující software, třeba je potřeba jenom nějaká fieldlike struktura, třeba je to nějaká hyperspecifická konkrétní struktura co je v mozku,... všichni od neurovědců po fyziky po filozofy po Ai researchers mi příjde všude jenom hádají, je strašně těžký to verifikovat a o tom se taky hádá, či to vůbec jde verifikovat, ale já myslím, že jo, přes například prožitkový mosty mezi systémama jako náš mozek a ostatníma tohle bude potřeba vyřešit, protože už teď upgradujeme mozky přes napojení na mašiny, [Welcome to the Cyborg Era: Brain Implants Transformed Lives This Year](https://singularityhub.com/2023/12/29/welcome-to-the-cyborg-era-brain-implants-transformed-lives-this-year/) a dřív nebo později se budeme pokoušet spojovat mozky s jinýma námi vytvořenýma systémama na silikonu nebo na biologickým substrátu už to přestává být scifi ale realita na druhou stranu, myslím že přes korelace nebo prožitkový mosty mezi systémama to půjde eventulně vyřešit aby tyhle pokusy nebyly jenom hádání co to reálně udělalo, když už se dějou Prožitkový mosty: (Experience bridges) Tyhle siamský dvojčata mají spojený mozek a můžou si navzájem posílat myšlenky. Pravděpodobně uvnitř pořád jsou dva individuálové. Něco takovýho, ale na druhé straně místo evolucí vytvořenýho člověka tam máš námi vytvořený organismus nebo stroj. [Conjoined Twins Who Share A Brain Hear Each Other's Thoughts](https://www.shared.com/conjoined-hogan-twins/) myslím že už dneska na týhle planetě žijou lidi co si přes výsledky z těchto a podobných relevantních výzkumů udělají nesmrtelnost a šílenou inteligenci [x.com](https://twitter.com/gsarti_/status/1749366992132292652) [🔍 Daily Picks in Interpretability & Analysis of LMs - a gsarti Collection](https://huggingface.co/collections/gsarti/daily-picks-in-interpretability-and-analysis-of-lms-65ae3339949c5675d25de2f9) Problém je v tom že v podstatě nerozumíme jak to uvnitř funguje takže to ani nemůžeme pořádně ovládat, protože AIs se škálují rychleji než stiháme identifikovat vnitřní strukturu https://www.lesswrong.com/posts/2roZtSr5TGmLjXMnT/ nebo jiný nástroje na kontrolu [Representation Engineering: A Top-Down Approach to AI Transparency](https://www.ai-transparency.org/) [GitHub - JShollaj/awesome-llm-interpretability: A curated list of Large Language Model (LLM) Interpretability resources.](https://github.com/JShollaj/awesome-llm-interpretability) Snad každej týden vidím nový automatizovaný jailbreaks co nějaký safety mechanismy rozkládají https://chats-lab.github.io/persuasive_jailbreaker/ [[2307.15043] Universal and Transferable Adversarial Attacks on Aligned Language Models](https://arxiv.org/abs/2307.15043) Prompt injection https://cdn.discordapp.com/attachments/631426993992499210/1198543946628997190/GBmegGhXEAA4xmY.jpg?ex=65bf4a08&is=65acd508&hm=89774c0a9014e74d10c8dc11afa77b0d224a0a14710058ee0fd26578359c50f2& https://cdn.discordapp.com/attachments/631426993992499210/1198543946893230140/chatgpt-car-dealerships-chatgpt-goes-awry-when-internet-gets-to-it.jpg?ex=65bf4a08&is=65acd508&hm=6e6dd04132ee38544f9974a812475081167bd89c2f8c2887ad26bad0820abf87& Ono je problém že často když do toho initial promptu 20x zminis "nedělej x", tak to stejně někdy má tendence udělat XD Hmm, to chce nějaký druhý checking LLM co discardne nerelevenat replies, to by mělo být o dost úspěšnější, třeba to tak dost companies dělají (další problém je vědět jak to přesně specifikovat aby se všechny edge cases covernuli) Problém často bývá že dost tehle multiagent triků zvětší cost a to se pak hůř škáluje Sleeper agents [Sleeper Agents: Training Deceptive LLMs that Persist Through Safety Training \ Anthropic](https://www.anthropic.com/news/sleeper-agents-training-deceptive-llms-that-persist-through-safety-training) Nejlepší tldr [x.com](https://twitter.com/karpathy/status/1745921205020799433) Taky je problém že když máš v trénovacích datech např lhaní, a to LLM se to naučí, tak zatím nemáme dobrej mechanismus jak to efektivně odstranit s dosavadníma metodama na gigantických modelech (nebo se podaří opak, že lze ještě víc, tzv. Waluigi effect) (taky jsou argumenty že lhaní může být dobrý emergentní instrumental algorithm ať to v trénovacích datech je nebo ne) Něco málo máme u malých modelech [[2310.06824] The Geometry of Truth: Emergent Linear Structure in Large Language Model Representations of True/False Datasets](https://arxiv.org/abs/2310.06824) Nebo honestly vektor [Representation Engineering: A Top-Down Approach to AI Transparency](https://www.ai-transparency.org/) A pak se dějí tyto věci: [[2311.07590] Large Language Models can Strategically Deceive their Users when Put Under Pressure](https://arxiv.org/abs/2311.07590) LLM strategically deceiving their users in a realistic situation without direct instructions or training for deception Ještě v tomto oboru dobrý progres nedávno bylo reverse engineering internal reprezentace chess boardu v ChessGPT https://fxtwitter.com/a_karvonen/status/1743666230127411389?t=3UDIFxYPmJljaVzK1B0F0w&s=19 https://fxtwitter.com/a_karvonen/status/1743666236180119706?t=3gyzwQ26mR3cf_f0qkPX9Q&s=19 [Chess-GPT’s Internal World Model | Adam Karvonen](https://adamkarvonen.github.io/machine_learning/2024/01/03/chess-world-models.html) [Two-thirds of Americans say AI could do their job | Fox Business](https://www.foxbusiness.com/technology/two-thirds-americans-say-ai-can-do-their-job) [Elon Musk says to expect roughly 1 billion humanoid robots in 2040s | Fox Business](https://www.foxbusiness.com/technology/elon-musk-says-expect-roughly-1-billion-humanoid-robots-in-2040s.amp) Trycyclic antidepressant meta-analysis - its weak [x.com](https://twitter.com/EikoFried/status/1749400762369581202) [Reddit - Dive into anything](https://www.reddit.com/r/machinelearningnews/comments/19ciynk/this_ai_paper_from_johns_hopkins_and_microsoft/) rag agents finetuning jobs education alphacodium law intelligence augmentation and beyond automateing mechanical tasks, automatic flexible tasks controllability open source autonomous unhinged AIs Will exist google leak What is not a construct of human perception? Is the process of seeing things as constructs of human perception just another construct of human perception? And is the process of seeing the process of seeing things as constructs of human perception as just another construct of human perception yet another construct of human perception? https://venturebeat.com/ai/is-openais-moonshot-to-integrate-democracy-into-ai-tech-more-than-pr-the-ai-beat/ [We need to research Full Autonomy RIGHT NOW - A call to action for OpenAI and others - Reduce X-Risk - YouTube](https://youtu.be/wKfgx8uvEqQ?si=x2R2HcuyM6WTd2Qh) selfcorrection selfdirection selfimprovement [[2310.10603] Exploring the Power of Graph Neural Networks in Solving Linear Optimization Problems](https://arxiv.org/abs/2310.10603) [The Mastermind Behind GPT-4 and the Future of AI | Ilya Sutskever - YouTube](https://www.youtube.com/watch?v=SjhIlw3Iffs&t=1305s) [Jim Fan: The next grand challenge for AI | TED Talk](https://www.ted.com/talks/jim_fan_the_next_grand_challenge_for_ai) One possible future i imagine is collapse of global top down control with somalialike warzone of powerful dictators in postapocalyptic environment walking in antiradiation suits with transhumanist augmentations using individual embodied AGIs as weapons or AGIs being autonomous entities on their own in this battle for power and control [Ben Goertzel VS Robin Hanson: The AI Disagreement - YouTube](https://www.youtube.com/watch?v=q6ZJYP5LJuA) [x.com](https://twitter.com/LakeBrenden/status/1749473226466402638?t=vGi7zu0kom6xiQZ0j2nMvg&s=19) how people learn compositional visual concepts through both Bayesian program induction and meta-learning [Bernardo Kastrup on Sean Carroll, Illusionism, & More! - YouTube](https://youtu.be/tTjtAWu_yOY?si=5mzVd9xNXf4f2RgD)