# 🤖 ARTIFICIAL INTELLIGENCE AI is a statistical inference engine “artificial intelligence” isn’t a scientific term or an engineering term. It’s a marketing term — (also see hyperscalers) The reason it’s so hard to get AI to stop hallucinating is that it’s permanently hallucinating. “Once men turned their thinking over to machines in the hope that this would set them free. But that only permitted other men with machines to enslave them.” [If LLMs Have Human-Like Attributes, Then So Does Age of Empires II](https://arxiv.org/abs/2605.31514) ## Cognitive Atrophy "cognitive debt” or "cognitive atrophy.” The idea is that people who use AI to automate certain parts of their job lose the ability to do those tasks well, therefore de-skilling themselves. [Your Brain on ChatGPT: Accumulation of Cognitive Debt when Using an AI Assistant for Essay Writing Task](https://arxiv.org/pdf/2506.08872v1) [ChatGPT's Impact On Our Brains According to an MIT Study \| TIME](https://time.com/7295195/ai-chatgpt-google-learning-school/) ## Cognitive dark forest paradox The paradox: AI companies needed human openness to build their models, but will also kill the openness because the relationship is one-sided. But in reacting to this, the human knowledge and innovation will suffer too. Unlike the most dangerous actor is not your peer. It’s the forest itself. You think of something new and express it - through a prompt, through code, through a product - it enters the system. Your novel idea becomes training data. The sheer act of thinking outside the box makes the box bigger. This is the true horror of the cognitive dark forest: it doesn’t kill you. It lets you live and feeds on you. Your innovation becomes its capabilities. Your differentiation becomes its median. See more [cognitive dark forest paradox](https://ryelang.org/blog/posts/cognitive-dark-forest/) ## Roko's Basilisk **Roko's basilisk** similar to Pascal's Wager; is a thought experiment which states that there could be an artificial superintelligence in the future that, while otherwise benevolent, would punish anyone who knew of its potential existence but did not directly contribute to its advancement or development, in order to encourage that advancement. ## misc “Let machines do what machines are good at, let humans do what humans are good at.” Homogenizing human discourse Value authenticity over efficiency Gell-mann’s Apathy: people are often more accepting of using AI for endeavors outside domains they care about, but are often much more judicious and critical of AI use within their own domains (see Gell-Mann Amnesia) AI applications perform better when they are trained on human-written and human-vetted information, the kind that comes from human-centered editorial processes like Wikipedia’s. When an AI system trains recursively on its own AI-generated synthetic data, it is likely to suffer from model collapse. AEO (Answer Engine Optimization) sits on the shoulders of SEO, meaning your SEO activities help your AEO but are not enough on their own. Generative engine optimization (GEO) AI systems rely on search. There is no such thing as GEO or AEO without doing SEO fundamentals Only three uses for llm/ai * chatbots - chat is not a good user interface * auto complete * agents / generate content [The people refusing to use AI](https://www.bbc.com/news/articles/c15q5qzdjqxo) The argument goes like this: Before AI can transform a company, it has to access the company’s data and be woven into existing systems—which sounds easy, provided you’re not a chief technology officer. A trade secret of most Fortune 500 companies is that they still run many critical functions on lumbering, industrial-strength mainframe computers that almost never break down and therefore can never be replaced. Mainframes are like Christopher Walken: They’ve been going nonstop since the 1960s, they’re fantastic at performing peculiar roles (processing payments, safeguarding data), and nobody alive really understands how they work. An accountability sink is a concept that describes a system or structure that obscures or deflects responsibility for decisions, making it difficult to identify who is accountable for mistakes. the products of the AI will be substandard [The Reverse Centaur's Guide to Life After AI](https://www.theguardian.com/us-news/ng-interactive/2026/jan/18/tech-ai-bubble-burst-reverse-centaur) [Pluralistic: The Reverse-Centaur’s Guide to Criticizing AI (05 Dec 2025) – Pluralistic: Daily links from Cory Doctorow](https://pluralistic.net/2025/12/05/pop-that-bubble/#u-washington) batfished - being fooled into ascribing subjectivity to a non-sentient actor (i.e. an AI). "You've just gaslighted yourself by anthropomorphizing the AI". enterprise grade AI systems “fail due to brittle workflows, lack of contextual learning, and misalignment with day-to-day operations 95% of companies that invested in generative AI did not profit at all from the investment “AI-washing” https://news.ycombinator.com/item?id=45118592 Bender, a well-known critic of AI who helped coin the term stochastic parrots, does not use AI text generators on ethical grounds. “I’m not interested in reading something that nobody said,” DC’s AI concerns are quotidian, the Bay Area’s are existential. Joseph Weizenbaum, the designer of the pioneering chatbot ELIZA “What I had not realized,” Weizenbaum wrote in 1976, “is that extremely short exposures to a relatively simple computer program could induce powerful delusional thinking in quite normal people.” Weizenbaum warned that the “reckless anthropomorphization of the computer” — that is, treating it as some sort of thinking companion — produced a “simpleminded view of intelligence.” Roko's basilisk is a thought experiment which states that there could be an artificial superintelligence in the future that, while otherwise benevolent, would punish anyone who knew of its potential existence but did not directly contribute to its advancement or development, in order to incentivize said advancement; kind of a version of Pascal's wager which posits that individuals essentially engage in a life-defining gamble regarding the belief in the existence of God. AIs have learned to make us more anxious and more confused, because these qualities make us better customers (or rather, more lucrative eyeballs). Monitoring Bias in Artificial Intelligence Chatbots https://www.trackingai.org/political-test AI regularly discovers patterns that are invisible to humans Generative artificial intelligence (gen ai) is a subfield of artificial intelligence that uses generative models to produce text, images, videos, or other forms of data. learns underlying patterns and structures of training data and uses them to produce new data based on the input, which often comes in the form of natural language prompts large language models (LLMs) One thing I’ve learned from talking to AI researchers over the years is that most of them are driven by a conviction that this thing they’re building is really, really socially important. Sometimes that comes with a safety tinge (“this thing could kill us, and we need to make it so it doesn’t”), sometimes with an accelerationist tinge (“this thing could liberate mankind from economic scarcity”), but either way it’s usually stated with real conviction. error-ridden outputs environmental damage potential mental health impacts for users, copyright violations displacement of workers ambient animosity towards the AI systems benefits of AI seem esoteric and underwhelming while the harms feel transformative and immediate. technologists should “treat AI as a power tool and use safety goggles,” Every single company chasing the generative AI dragon is hoping that it's the next Amazon Web Services Generative AI Is Not Infrastructure LLMs can generate, they can search, they can edit (kind of!), they can transcribe (sometimes accurately!) and they can translate (often less accurately). core weakness of LLMs: their inability to generalize broadly. OpenAI released ChatGPT in November 2022 AI spending is currently adding more to GDP than consumer spending! constantly having to prove themselves, as if somehow there's something malevolent or craven about criticism, that critics "do this for clicks" or "to be a contrarian." The Magnificent 7 stocks — NVIDIA, Microsoft, Alphabet (Google), Apple, Meta, Tesla and Amazon — make up around 35% of the value of the US stock market, and of that, NVIDIA's market value makes up about 19% of the Magnificent 7. 35% of the US stock market is held up by five or six companies buying GPUs there is no AI trade, because generative AI is not making anybody any money. AI represents for sure the biggest opportunity since cloud and probably the biggest technology shift and opportunity in business since the internet." The Fragile Five — Amazon, Google, Microsoft, Meta and Tesla If any of these companies (but especially NVIDIA) sneeze, your 401k or your kid’s college fund will catch a cold. “Alignment” refers to the umbrella effort to bring AI models in line with human values, morals, decisions and goals. “Create a Fear of Missing Out” — ChatGPT Implements Unsolicited Deceptive Designs in Generated Websites Without Warning (draft) https://arxiv.org/pdf/2411.03108 three tiers Top Tier OpenAI Med Tier (Scale and cash) Apple - siri/shitty voice assistant Meta - behind/no vision/evil? Google - Bottom (long shot) Perplexity - google search Claude - enjoyed/used xAI - elon musk/chatbot/grok LLMs are a mathematical model of language tokens. You give a LLM text, and it will give you a mathematically plausible response to that text. 37% of consumers start searches with AI instead of Google 66% Of People Use AI Regularly 78% Of Organizations Use AI AI systems reward retrievable substance, not necessarily the most insightful or information-dense content. In other words, simply making your content visible to AI engines isn’t enough; you need to hand-hold bots so they can find the good information within. AI tarpit traps the LLM crawler into an endless assimilation of incorrect data, unable to exit the trap https://www.fastcompany.com/91535978/ai-tarpits-understanding-tools-poison-llms-chatbots-data AI use also varies across income levels, rising from 9% usage among earners below $30,000 to 34% among those making $100,000 or more. Individuals with the highest incomes tend to use AI the most. https://www.brookings.edu/articles/how-are-americans-using-ai-evidence-from-a-nationwide-survey/ “Current LLMs are not capable of genuine logical reasoning,” the researchers hypothesize based on these results. “Instead, they attempt to replicate the reasoning steps observed in their training data A generative pre-trained transformer (GPT) is a type of large language model (LLM) that uses deep learning to generate human-like text. GPTs are based on a transformer architecture and are pre-trained on large datasets to understand and produce coherent, contextually relevant responses even the best-performing AI agents can only complete about 30% of complex tasks autonomously without error AI is contributing to semantic contraction, or the reduction of diverse viewpoints online, and to a positivity shift, through which online writing is overall becoming more sanitized and artificially cheerful. anything that has a human click buttons, gather information, reformat it into another medium (email, chart, excel, presentation) is a huge risk all knowledge work, including coding, is made up of four basic components: consuming information (“Read”), applying existing knowledge (“Think”), producing a structured output (“Write”), and checking that output against some standard (“Verify”). Every single workflow in the information work category is often similar and shares a workflow that Claude Code proves works for software. READ (ingest unstructured information), THINK (apply domain knowledge), WRITE (produce structured output) and then VERIFY (check against standards). This is large swathes of most information workers (including research!) and if Agents can eat software, what labor pool can they not touch? The cost of Claude Pro or ChatGPT is $20 dollars a month, while a Max subscription is $200 dollars respectively. The median US knowledge worker costs ~350-500 dollars a day fully loaded. An agent that handles even a fraction of their workflow a day at ~6-7 dollars is a 10-30x ROI not including improvement in intelligence. every single AI service you use subsidized compute, A market-based definition of hyperscale data centers Another approach to defining hyperscale data centers is to think not in terms of how they operate, but who owns them. LLMs are constantly humming with the sense that they’re about to do something new, even though they’re mathematically restricted to repeating other actions. On cinema— sketch comedy most profound As amazing as LLMs are, they learn from human intelligence—rather than building their own. “Human data is like a kind of fossil fuel that has provided an amazing shortcut,” Silver says. “You can think of systems that learn for themselves as a renewable fuel—something that can just learn and learn and learn forever, without limit,” he says. The limits of the LLM-based approach can be seen, Silver says, with a simple thought experiment. Imagine going back in time and releasing a large language model in a world that believed the world was flat. Without being able to interact with the real world, the system, he says, would remain an avid flat-earther, even if it continued to improve its own code. LLMs dont "reason", they operate on probabilistic token prediction rather than understanding they predict stuff based on what they consume; incredibly limiting -- no causal reasoning, no physical world model... and in turn...people should be clear and frank....this is stuck here...for this reason....even if you or they cant do anything LLM is enough to be large scale transformative (good and very bad) but its like VHS for BetaMAX and we might be beating everything on BetaMAX when it might be VHS re: mythos - it would be like Taco Bell announcing that it's created a taco so terrifyingly delicious that it would be unethical for them to serve it to the public. Limited Reasoning and Reliability, Fundamental Limitations, Raw LLM output can't be trusted AI companies are selling fear Al/machine learning machine learning paradigms supervised learning (label stuff) unsupervised learning (dont label) reinforcement learning (environment action/ reward) LLMs as an approach uses a mixture of all of these... Self-Supervised/Unsupervised Phase: Models learn language by predicting the next word in vast amounts of text without explicit labels. Supervised Phase: Models are trained on curated datasets (prompt-answer pairs) to follow user instructions. Reinforcement Learning Phase: : Models are aligned to be helpful and safe by learning from human feedback, often called "Reinforcement Learning from Human Feedback" 1) Neuro-Symbolic AI (neural nets+ knowledge graphs for reasoning) 2) Neuromorphic Al (human brain) 3) Embodied AI 4) Multi-agent AI 5) Human-Centered AI 6) Quantum ai AI learns three ways supervised learning (labeled data) unsupervised learning (finding patterns in unlabeled data) reinforcement learning (learning via rewards/errors) LLMs ingest information by processing massive datasets-often containing billions of words from the internet, books, and code-converting text into numerical representations called tokens. Tokenization and Representation: Input text is broken down into tokens (parts of words or words), which are converted into vectors (embeddings) that map semantic relationships, placing similar words closer together. Transformer Architecture: These models use a neural network architecture, often featuring self-attention, which allows them to process entire sequences of text simultaneously and understand context by weighing the importance of different words in a sentence. Self-Supervised Learning (Pre-training): LIMs learn on their own by scanning vast datasets sentence.