Дороничев: ИИ — пузырь, который скоро ЛОПНЕТ. Какие перемены ждут мир?

High-level thesis

• Andrey Doronichev (ex-Google product lead who built YouTube Mobile) argues that AI today is in a speculative “bubble” comparable to past technology waves (electricity, early internet).
• The bubble is likely to deflate, but—similar to the internet-era collapse—it will leave durable infrastructure (GPUs, data centers, power, software stacks) and lasting change.
• Key practical message: everyone should work on “AI readiness” — both personal and business — because adoption will be fast and unevenly distributed.

Technical concepts and industry analysis

Infrastructure and economics

• Massive capital is flowing into data centers and GPUs; electricity and power provisioning are becoming limiting factors.
• Unlike early-internet “dark fiber” assets that sat idle, GPUs are actively consumed now, so investments are being used immediately.

Models and training trends

• The initial “pre-train on more data / bigger transformer” era is maturing.
• Next phase emphasizes:
  • Vertical / domain-specific post-training (fine-tuning).
  • Reinforcement loops and synthetic data.
• Specialized models (vertical language models) are expected to improve predictably in domains like math, biology, and law by training with domain experts and feedback loops.

Reliability and failure modes

• Hallucinations remain a core problem: models can invent facts when pressured to answer.
• Research uncertainty is high: creators often discover surprising capabilities and behavior that is not fully interpretable.

Mitigation strategies

• Fine-tuning, reward/reinforcement loops, synthetic data, and retrieval-augmented designs to improve domain accuracy.
• Ensemble / “cross-check” approaches: run multiple models (or instances) and have them review/verify each other to reduce hallucinations; tradeoff is increased compute/cost.
• Be model-agnostic: build stacks that can swap underlying LMs (GPT, Gemini, Grok, etc.) to avoid provider lock-in and leverage comparative strengths.

Product concepts, features and examples

1. Longji (health / longevity product)

   • Goal: help users reduce biological age and extend healthspan by combining AI with a human team.
   • Data sources: wearables (Oura, Apple Watch), continuous glucose monitors, lab/blood tests, photo food logs, PDF lab uploads.
   • Human + AI blend: nutritionists, endocrinologists and telemedicine support monitor data and give recommendations; AI agents aggregate and synthesize context.
   • Onboarding: define user-specific functional goals (e.g., surf at 70, lift grandchildren) to align interventions to meaningful outcomes, not just lifespan numbers.
   • UX/engagement challenges: motivation loops, avoiding user stress (biofeedback that “scolds”), retention (people sign up but don’t follow through).
   • Measurement limits: biological age is noisy; recommended focus is on healthspan and functional goals rather than raw lifespan metrics.

2. Doronichev’s personal AI health/fitness stack (practical how-to)

   • Trained a GPT-style model to emulate his real coach (personality, biomechanics) and used voice prompts during workouts.
   • Integrated context: uploaded training history, weights, blood tests, wearable metrics; nutritionist prompts and calorie tracking.
   • Built a prompt-driven “operating system” for health: persistent context (prompts, history) stored and managed (examples: Git/GitHub and Cursor used for context/prompt management).
   • Outcome: improved adherence and results (reduced body fat, visible abs) by combining personalized AI attention with human validation.
   • Guide takeaway: create personas/prompts that capture your coach’s domain knowledge; feed continuous device and lab data; iterate with human-in-the-loop checks.

3. How-to for building vertical AI agents (summary guidance)

   • Start with base models, benchmark on domain tasks, then apply targeted fine-tuning and reinforcement feedback from domain experts.
   • Use synthetic / domain-specific datasets when real data is scarce.
   • Implement factuality checks: ensembles, retrieval from authoritative sources, and human verification—especially in regulated domains (biotech, medicine).
   • Maintain provider-agnostic architecture to combine or replace foundational models easily.

4. Career / human-skill guidance (actionable recommendations)

   • Two educational priorities:
     1. Fundamental knowledge: math, physics, systems thinking to remain adaptable.
     2. Practical applied skills: entrepreneurship, hands-on trades, or operational execution.
   • Psychological and behavioral skills:
     • Responsibility and ownership: people who accept risk and legal/financial responsibility remain valuable because investors and organizations demand accountable humans.
     • Will, intention, and self-discipline: meditation and practices (e.g., mindfulness) reduce reactive behavior and strengthen consistent habits.
     • Corporeality: physical skills, live performance and embodied experiences (sports, in-person events) retain unique human value as digital content becomes easier to synthesize.

Risks, trade-offs and social points

• Bubble dynamics: many speculative companies will fail when hype subsides, but infrastructure and talent allocation will seed long-term capabilities.
• Responsibility/liability economics: firms and investors often prefer humans in advisory or accountable roles because legal liability remains human-centric; this preserves market value for accountable practitioners.
• Societal risks: over-reliance on AI may erode cognitive skills (e.g., writing structured essays, deep thinking). Education should emphasize systematic thinking and argument structure.
• Engagement problems in health products: user motivation, stress from constant feedback, and translating accurate measurement into actionable guidance are core product design challenges.

Practical takeaways (concise)

• Work on AI readiness now: develop personal skills (leadership, ownership, fundamental learning + applied skills) and company readiness (how to integrate AI agents, fine-tuning, vendor-agnostic stacks).
• Use LLMs as thought partners and assistants — don’t fully delegate reasoning. Combine human judgment with AI outputs and verify (especially using ensembles and retrieval).
• In regulated/high-stakes domains, build pipelines with domain-specific fine-tuning, human-in-the-loop verification, and multi-model cross-checking to reduce hallucinations.
• For health/fitness: integrate wearables and lab results with a persistent AI context plus human oversight; focus on healthspan/functional goals and strong onboarding to align motivation.

Practical message: prepare for rapid, uneven AI adoption—build readiness now at both the personal and organizational level.

Products / tools mentioned

• GPT (ChatGPT), Google Gemini Pro, Grok (xAI)
• Cursor (prompt/context tool), Git / GitHub (prompt/context storage)
• Wearables: Oura, Apple Watch
• Continuous glucose monitors
• Longji (early-access longevity/health AI + human team product)

Main speakers / sources

• Andrey Doronichev — Russian‑American technologist, ex‑Google (led YouTube Mobile/product work), serial entrepreneur; primary interviewee.
• Gleb (host/interviewer) — conducts the interview and presents product Longji.
• Other referenced people/sources: Dima Matskevich (prompt/context practices), Robert Sapolsky (on determinism/free will), various AI company founders and commentators (Sam Altman, Elon Musk referenced), and teams working on decentralizing compute (Lieberman brothers / Race project).