--- # consumer electronics device for ai ![[ai device.png]] **summary**: - the post questions whether users truly own and control the ai they use, as cloud-based ai is aligned with companies, not users - ai model costs are decreasing while quality increases, suggesting consumer-grade laptops will soon run powerful models locally - running numerous ai agents (10,000+) on personal laptops isn't feasible due to hardware limitations and shared access needs - cloud-based alternatives would cost $10-15k monthly now, potentially dropping to $100-300 monthly in two years - the author proposes a dedicated consumer ai computing device ($100/month lease) that could serve multiple users and sell unused capacity - ai is becoming a utility like electricity - decentralized, widely available and locally accessible - economics will eventually favor local ai infrastructure at community levels rather than dependence on a few global companies **post**: [[consumer electronics device for ai]] **date**: 20 july 2025 --- # expect ai magic to continue **summary**: - as assistants emerged from the base models through few-shot prompting for ai-2-user style responses to instruction tuning to rlhf, now reasoning capabilities are emerging from base via chain of thought prompting & traces to cot instruct and rlvr - the magic of ai is similar to traditional ideas of magic - saying words to make things happen without physical intervention - in-context learning (icl) is a groundbreaking capability that emerged unexpectedly from transformer models, allowing customized behaviors at a fraction of the cost of training or fine-tuning - these emergent phenomena represent "ai magic" that will continue with new capabilities beyond reasoning (eg, tool & computer use, vision, multimodality, ... n) - i suggest experimenting with icl to generate traces, fine-tuning models with reinforcement learning for specialized capabilities, and creating unique model collections **post**: [[expect ai magic to continue]] **date**: 9 may 2025 --- # google can do whatever the fuck they want with your licenses **summary**: - google is arbitrarily increasing prices, forcing gemini features on users without their consent - changing email, sso, or cloud providers is such a massive undertaking that most won't do it over a 5% price increase - big tech companies have walled off the internet and profit from it, making the author desire significant change **post**: [[google can do whatever the fuck they want with your licenses]] **date**: 1 may 2025 --- # palantir, tesla, morgan stanley are evil and there is no way around it **summary**: - palantir is about mass surveillance and population control, not freedom and protection - tesla has betrayed its mission with elon musk becoming a psychopathic oligarch who cares only about personal wealth - morgan stanley is responsible for the 2008/9 crisis, creates toxic work environments, and financializes things like housing and education that shouldn't be commodified - employees at these companies are morally complicit and should be held accountable - our collective tolerance of these organizations reveals a disturbing ethical landscape in 2025 **post**: [[palantir, tesla, morgan stanley are evil and there is no way around it]] **date**: 22 apr 2025 --- # reverse engineering commit history to train reasoning models on process rewards **summary**: - current language models are trained on static code repositories but miss the evolution process (requirements, implementations, changes) - commit history contains valuable information about how code evolves over time - proposal: reverse engineer commit history to generate prompts that would produce each commit's code changes - create a dataset of repository states interleaved with these generated prompts - use this to train reasoning models with reinforcement learning using "process rewards" - verify each step by comparing model-generated changes against actual commits **post**: [[reverse engineering commit history to train reasoning models on process rewards]] **date**: 21 apr 2025 --- # non-destructive revolution **summary**: - unlike communist revolutions that destroy existing systems, we can radically change systems without ruining them - system analysis gives us tools to understand complex systems, their components, and how rules impact outcomes - organization is a human strength that helped us dominate the planet, and most problems stem from poor organization "don't blame the player blame the game" - when the right ideas take hold, believers can gather and demand change **post**: [[non-destructive revolution]] **date**: 18 apr 2025 --- # ai adoption gap **summary**: - despite widespread ai awareness, actual adoption rates remain surprisingly low - workshops with 300 participants across 100 german companies revealed many professionals haven't used tools like chatgpt at all - while reportedly 13.5% of german enterprises have some ai knowledge, true power users likely represent only about 1% - the slow penetration rate is unexpected given the technology's groundbreaking nature and general awareness **post**: [[ai adoption gap]] **date**: 17 apr 2025 --- # ai agents transforming enterprises ![[cover.png]] **summary**: - ai has evolved from its classical definition as computer science simulation of human intelligence to include behavior-based definitions and advanced cognitive capabilities - modern ai systems can be categorized as: 1. assistants (dialogue-oriented user support) 2. reasoning models (logical thinking and problem-solving) 3. agents (autonomous systems that can act independently with tools) - organizations are implementing ai for: - developing company-specific knowledge bases - process automation for productivity gains - specific use cases include: - automated order processing (reducing manual data entry) - patent analysis (reducing analysis time from 4 hours to 15 minutes) - resume evaluation (providing detailed assessment beyond 6-second human scans) - major pitfalls of ai include: - technological challenges: hallucinations, context recall problems, reliability issues - economic/social concerns: job displacement, market concentration, intellectual property issues - the ai market is increasingly concentrated among a few companies with frontier-level models, raising concerns about power distribution **post**: [[keynote - ai agents transforming enterprises]] **date**: 15 apr 2025 --- # the difference between a dead man and a billionaire ![[Pasted image 20250415160641.png]] **summary**: - aaron swartz faced 35 years in prison and committed suicide at 26 for sharing scientific papers, while zuckerberg's company used pirated books for ai training with minimal consequences - the author questions the fairness of intellectual property systems that punish individuals like swartz while allowing tech giants to exploit similar practices - the piece criticizes figures like musk and dorsey who advocate abolishing IP rights, suggesting this would primarily benefit big tech companies **post**: [[the difference between a dead man and a billionaire]] **date**: 15 apr 2025 --- # alignment through structure ![[Pasted image 20250502125940.png]] **summary**: - ai alignment cannot be achieved through goodwill, reward functions, governance, or technical methods alone - humans and organizations will inevitably exploit any possible misuse of ai for power and wealth - key structural requirements for ai alignment: 1. distributed ownership of data, models, and compute 2. fair distribution of resources, capabilities, and usage 3. decentralized control, governance, and power - tech giants are currently shaping ai to serve their interests rather than society's - we have until 2030 to establish the structural foundations for aligned ai **post**: [[alignment through structure]] **date**: 13 apr 2025 --- # teaching ai to navigate screens **summary**: - ai agents struggle with screen navigation because they operate at the wrong level of abstraction - focusing on low-level clicking rather than interface patterns - the solution is creating an "imprint" system with two apis: "lm-in" (for language models to request actions) and "ui-out" (pre-learned interface navigation) - screen recording tools could capture human workflows, allowing ai to generate scripts that automate repetitive tasks - training on tutorial videos and simulated environments would enable more reliable, verifiable screen navigation - this approach would make ai navigation faster, more precise, and personalized to users' actual workflows **post**: [[teaching ai to navigate screens]] **date**: 12 apr 2025 --- # ai make ai **summary**: - agent frameworks are finally showing promise after years of failures, with tools like bolt.new and cursor building functional apps - recent research (meta agent search and ai scientist) demonstrates ai can create and improve other ai systems autonomously - we're approaching a point where ai creating ai will surpass human capabilities, potentially leading to a "kembrian explosion" of ai agents that we can create in a matter of minutes **post**: [[ai make ai]] **date**: 31 mar 2025 --- # os-complete navigation environment **summary**: - navigation/screen navigation is about language models looking at screens, understanding elements, and clicking to complete tasks - current issues: imprecise, slow due to continuous screenshots, vision processing, parsing, and command generation - proposed solution: "imprint api" - javascript-based api that translates lm commands into actual clicks - imprint api is generated by a screen reader and vision model pre-trained on interfaces, creating latent representations where similar functions across different uis map to similar endpoints - "portables" (html simulations of interfaces) can create training environments without using actual websites - an "os-complete navigation environment" would contain portables, screen reader models, and interaction traces to train language models - such environments will be essential for training models to effectively use computer systems **post**: [[os-complete navigation environment]] **date**: 10 mar 2025 --- # cafe still life with diffusion models **summary**: - reasoning models like sonnet37r or gemini20 flash fast thinking are great at extracting detailed descriptions of images - metaprompting allows for generation of consistent images quite similar to the original **post**: [[cafe still life with diffusion models]] **date**: 10 mar 2025 ![[_media/IMG_20250309_153048.jpg]] --- # cost of intelligence will drop 100.000x within 24 months **summary**: - cost of intelligence dropped over 10x within 18 months, as measured by the price of openai gpt api - in the next 18-24 months, it can drop by 100x due to model optimization, and another 100x due to hardware. those are orthogonal and can be multiplied - what will happen to the demand for ai? **post**: [[cost of intelligence]] **date**: 8 mar 2025 --- --start--