Emad Mostaque on the Intelligent Internet and Universal Basic AI

Episode Summary

Executive Summary: Ahmad Mostak argues that AI is now both essential and dangerous: survival depends on building it correctly, transparently, and with human agency at the center. He outlines a new vision, the Intelligent Internet, to provide open AI infrastructure for healthcare, education, and government through layered nodes, local assistants, and a funding mechanism tied to beneficial compute.

Main Topics: AI existential risk and the case for 'P-doom' (Priority: 5/5): Mostak says AI could either help humanity immensely or create catastrophic failure modes through misuse, emergent behavior, and misuse by powerful actors; he estimates roughly 50% odds of doom over an indefinite horizon. Human agency as the central alignment goal (Priority: 5/5): He rejects purely utility-maximizing or truth-maximizing framings and instead argues AI should expand individual, community, and societal agency, with children's agency as a moral anchor. The Intelligent Internet for regulated-industry AI (Priority: 5/5): His new project aims to build open, transparent AI infrastructure for healthcare, education, government, and other 'stuff for living,' using common knowledge bases and locally adaptable models. Layered infrastructure: personal devices, distributed nodes, hypernodes (Priority: 4/5): Mostak proposes a three-tier stack: local AI on consumer devices, hospital/organization-level distributed nodes for tuning and services, and national or international hypernodes for common knowledge and large-scale model work. Open data, open weights, and transparency as safety tools (Priority: 4/5): He argues that decision-making AI must be based on transparent, gold-standard datasets and open models, because interpretability is still unsolved and hidden data can hide dangerous behaviors or biases. Crypto funding via proof of beneficial compute (Priority: 4/5): To finance a public-good AI infrastructure, Mostak proposes a token-based mechanism akin to Bitcoin, where participating helps fund compute for beneficial uses like cancer research and AI tutors. Regulation, speech risk, and military use (Priority: 4/5): He is skeptical of broad pauses or heavy-handed regulation, but strongly favors transparency requirements for regulated domains and special regulation for speech models because voice is highly persuasive and dangerous to misuse.

Key Arguments: AI is already powerful enough to transform most lives, but only if it is made accessible, aligned, and locally relevant. The biggest danger is not one single failure mode but a mix of human misuse, hidden objectives, unsafe deployment, and emergent behaviors from increasingly autonomous systems. Agency, not abstract 'truth-seeking,' should be the main design principle; AI should enhance human choice rather than replace people. Regulated sectors like healthcare and education should not rely on black-box proprietary systems; they need open, auditable, standardized infrastructure. Most current frontier models are architecturally similar; the decisive issue is objective function, training data, and governance rather than minor model differences. Speech models are uniquely dangerous because they can imitate trusted voices and directly manipulate behavior, especially children. The economics of AI will favor cheap, local inference and specialized systems over ever-larger central models for most daily tasks. A public-good funding mechanism can align incentives by tying token value to the growth of beneficial compute and real-world utility. The most practical safety intervention is not broad model prohibition but transparent data logging for AI used in regulated decision-making. Open datasets and model standards can make AI more trustworthy, more reusable, and more democratic across countries and cultures.

Data Points: Estimated P-doom: 50% - Mostak’s rough estimate of humanity’s survival odds over an indefinite AI timeline Humanity without AI help: Very bad scenario - He argues that failing to use AI to assist society is itself dangerous Stable Diffusion adoption: 300 million downloads - He cites Stability AI model downloads as evidence that open models can spread rapidly Stability AI research community size: 500,000 people - He says the company built large communities around research and development Research hiring at Stability: 20,000 applicants / 120 offers / 83% acceptance - Used to illustrate how attracting motivated researchers and giving freedom worked Stable Diffusion training compute: 250,000 A100 hours - Original model training cost estimate Llama 70B training compute: 20 million H100 hours - He cites Meta-scale training as a reference point Llama 70B training cost estimate: ~$40 million - Approximate spend implied for the large Llama run Medical cluster example: 1,000 H100s - He references the Chan Zuckerberg biomedical cluster as a national-scale benchmark Hospital fine-tuning scale: 128-256 H100s - He says that is enough for a bank or hospital to fine-tune specialized models National hypernode scale: Thousands of H100 equivalents - He believes this is enough for a country to maintain specialized AI infrastructure Inference energy for GPT-4-level local AI: ~5 watts - He argues a GPT-4-ish model can run on a Mac neural engine or similar low-power device H100 power draw: ~1,000 watts - Contrasted with local low-power inference Human reading speed / token speed: ~10 tokens per second - He notes that speed beyond human reading pace has diminishing practical value Electricity cost example: $0.20/kWh - Used to estimate the yearly cost of running a 5-watt AI continuously Continuous 5W AI energy cost: About a tenth of a cent an hour / a few dollars per year - Demonstrates how cheap local AI inference can be Gemini Flash price: 7 cents per million tokens - He cites this as evidence that intelligence is getting very cheap Model context windows: 2 million to 10 million tokens - He points to Google’s large-context models as a major capability shift People without smartphones: ~600 million - Used to argue that infrastructure gaps remain large Average global IQ: 90 - He frames this as an infrastructure problem rather than an intelligence problem Cancer impact: ~half of everyone listening will get cancer - Used to motivate universal health AI infrastructure

Pivotal Quotes: "If AI is built right and is aligned and works with us, then we're in a very good scenario." — Ahmad Mostak: Opening framing of AI as both existential risk and enormous opportunity "We call universal basic AI." — Ahmad Mostak: His term for free, baseline AI infrastructure for everyone in essential life domains "Who owns that AI that is your best friend, that is your kid's best friend?" — Ahmad Mostak: On the importance of transparency, ownership, and access in personal and family-facing AI

Implications: The interview points toward a future where AI is treated like public infrastructure: cheap, local, open, and domain-specific. For industry and governments, the message is that trust, transparency, and agency will matter more than raw model power.