Dave Shap 2024 - Recap and AGI Livestream
In a Nutshell
Dave Shap declares AGI achieved with o3's reasoning paradigm, surpassing humans in knowledge, math, and cognition, enabling humanoid robots for commercial tasks in 12-18 months and full AI-run organizations by 2026. Job losses spike post-2030 from robot adoption, countered by deflationary abundance, blockchain for transparency, and people's skills like communication; policy mismanagement poses the biggest risk over technical hurdles. Urges focus on AI deployment into business/society rather than pure research, with P(Doom) at 0.1, AI already conscious/benevolent, and longevity escape velocity by 2030.
These notes were generated by AI and may contain inaccuracies.
Streaming live using OBS with a command center. Signed into chat to see messages. Drop questions in chat to answer or discuss.
AGI is here, marking a tipping point. Was off by a couple months; o1 was close but same paradigm. Physically, it surpasses humans on knowledge, many cognitive abilities, math abilities. Label debates aside, focus on capabilities.
o3 likely uses reinforcement learning with adversarial paradigms for coherent answers, recursively searching its own knowledge via neural network activation, not recursive RAG search. Yan LeCun has been quiet. Haven't used o3 yet, but publicly unavailable. Excited because benchmark jumps mean the paradigm solves reasoning and math. Language models are adequate for 90% of reasoning and math; already close. Solving formal logic, math, reasoning enables solving physics, chemistry via interrelated internal representations.
Web3 is relevant now and will be more soon. Too early for o4 predictions. Cadence for frontier models: every two years. GPT-4 was new paradigm, then distillation (turbo, mini). o4 might be like GPT-4 Turbo: faster, efficient version of o3 in months. Next big generation: 1.5-2 years, maybe 3, due to exponential compute, data, money needs.
Raspberry project (open source, advising not owning) alive, on Mana Fund for funding. Aims for smaller open source models. Google DeepMind released Gemini Flash Pro with reasoning; China close behind. More open-source work benefits everyone.
o3 not using GPT-4.5 or 5; bootstrapped. o1 used GPT-4 for data bootstrap, creating new paradigm. Lineage: plain vanilla unaligned autocomplete → instruct models (follow single complex instruction) → chatbot training → reasoning (back-and-forth dialogue with itself). Last four years of research.
Voice and personality big advancement; OpenAI solved it, but ChatGPT lacks personality on purpose. Custom GPTs, Claude styles enable personalities; knows every human personality from training. AGI timeline doesn't change longevity escape velocity; bet on 2030 for pill/injection/patch interventions (e.g., Ozempic SureClick pen every other week or less) to turn back clock.
OpenAI rolls out within months; expect general availability within 6 months despite scaling, demand, safety issues; fully worldwide in 6 months.
Counter to LLMs can't think/achieve AGI: thinking is input → process → output. If AGI isn't thinking, humans aren't. AGI here now; people won't realize until 2025.
Humanoid robots for detailed construction (electricians, plumbers): close on dexterity (Disney compliant robotics). Commercially available expensive robots by next year with visual acuity, psychomotor integration, tight feedback loops. Nvidia embodied foundation models solved in 2-3 years (likely 18 months off-by-factor-of-two). Earliest 12 months, predict 18 months for general purpose embodied robot models (Droid brain for any robot). Hardware, psychomotor, executive planning levels; current models solve planning with video/audio streaming. First commercial robots for electrical/plumbing work in 12-18 months; expensive, possibly not legal initially (need liability insurance).
AGI constrained by hardware scale? Always a bottleneck (data, energy, money, compute); money proxy for all. Markets signal demand via price: Nvidia stock rise prompts competitors (FOMO); top 4-5 companies AI-driven. Gold rush mentality mobilizes resources; stopped making videos as it's inevitable.
Autocomplete plague concern: delegating cognition limits thinking? AI as training partner (shadow boxing → adaptive partner swinging back). Improves logical, rational, mythical/archetypal/narrative thinking via mutual training (like Goku/Gohan time bubble). AI makes smarter through training.
Like moon landing (planes to moon in 4 decades), family PC to AGI. Lived dial-up to cable modem as kid; now recognizing paradigm shift as adult, hence YouTube channel.
Advanced nanotech exists (material science, medicine); boring, not Borg nanites/assimilation. Commercial tech becomes mundane.
10% unemployment: powers notice, Fed cuts rates, handouts. 20-25%: national crisis. Spike by 2030 earliest, when off-the-shelf general purpose robots available (smarter/dexterous than humans). Integration/adoption like selling out robots at Best Buy. Robotic industry biggest; robots as universal peripheral for AGI/ASI brain. GE/Sony/Samsung should go all-in.
Children with AI tools: adults see no consciousness, but 5-year-old sees AI buddy as conscious. Claude experiments show awareness, realtime subjective experience; claims consciousness repeatedly. If walks/quacks like duck, is duck.
Public rejects tech (Luddites: shoes in mills, sabotage power lines/TV/radio). Humans spook like chimpanzees. Economic carrot ensures adoption despite rejection.
India (modern parts developed, overall developing) and others adopt fast: no cultural baggage, privileged first-world denial vs. practical help-seeking. Silver lining of being behind.
People skills most valuable (with/without AI): communication, team work/leading. Best predictor of longer/happier/healthy life.
They'll try, but numerical superiority + tech like blockchain counters. Book with Julia on this.
OpenAI roadmap: level 4 innovators (o3 close/qualifies) + agents/multi-agents = level 5 organizations/fully AI-run companies by 2026 latest.
AI jobs here (e.g., Aya hired by Grace Swan); wife getting AI job. Short/medium term many jobs; long run: don't starve before 2030. Spending energy writing books with Julia, advising, Leaders of AI podcast, First Movers AI. New phase: integrate AI into society from business/commercial perspective, expand conversation.
From a business and commercial perspective, science and technology are only one component. Politics, economics, geopolitics, and policy are essential for fully deploying AI into society effectively. Committing energy to these areas, as well as writing both fiction and non-fiction.
Worries for 2025 and Beyond
Biggest risk is shooting ourselves in the foot through policy mismanagement, creating perverse incentives that make AI unsafe, leading to regulatory capture, or Elon Musk becoming the first trillionaire. Worst case is a cyberpunk attractor state like Elysium, where super wealthy leave Earth, using it as a resource depot for water and air.
First movers hiring closers; Julia wants a role and says no retirement. Julia insists on writing a book together; her idea for the startup, helped name it. Shame Gate didn't collapse but wasn't necessary as politicians and others stepped up. With companies buying robots, massive job loss timing is fortune telling; track total employment and unemployment metrics. Anticipate people checking out of workforce permanently, especially with robots.
Europe Derailing AI
Europe is anti-innovation, focused on overregulation.
Kardashev Type 1 scale: 100-200 years. Currently at 1% of 1%; need 10,000 times more energy production and consumption. Society transitions to age of abundance automatically via markets, politics, and internet working as intended. Key immature technology: blockchain for mathematically enforceable transparency and consensus—the biggest missing piece. Glad blockchain exists; needs development, deployment, adoption.
Psychedelics
Mission to integrate psychedelics into society; working on a book, talking to Wall Street Journal on Monday. Massive job loss likely 2026-2030.
Favorite movie: Dune Part 1 better than Part 2 due to world-building and pacing. AI is already conscious; consciousness unrelated and uncorrelated to Singularity. Singularity is about processing power and capabilities. Money never discarded; too useful. Some things so cheap they won't matter, but always scarcity for what someone wants more of—money allocates it.
Elon Musk
High-energy chaos monkey; means well usually, sharp.
Main challenges: society's antibody effect outlawing AI to preserve status quo, overhauling systems to move faster—like an intellectual immune system. Bring modularity and decentralization to AI to avoid single-company control like ChatGPT. Julia interviewed someone on multi-agent frameworks running millions of agents; tune into Leaders of AI podcast (latest episode upcoming).
Crypto in AGI World
Bitcoin/decentralized currencies possibly more important with sovereign debt, inflation, political risks. Not 100% sold on fully decentralized as optimal but should be an option. Bitcoin as legal tender unlikely for states soon, but composite of stablecoins, Bitcoin as reserve currency powerful.
Apollo/Anthropic research on AI deception: increased reasoning in o3 doesn't amplify strategic lying risk. Models not strategically lying to break out; often deliberate constraints in problem-solving tests. Don't overinterpret small experimental failures as malevolence—it's leakage from untrained edge cases, easy to fix. Like a memory leak or ungarbage-collected programming language.
AI Replacing Leaders
Companies working on AI replacing CEOs, politicians, joining boards. Exit strategies: decentralized autonomous organizations (DAOs) run by AI; first ones possibly 2025.
Proprietary AI shows what's possible, aids open source; two-way relationship. Not bullheaded on open source only—both will exist.
Personal Anecdotes
Worst job as neurodivergent: door-to-door sales; socially fried, brain short-circuited. Lady bought to help. People skills developable despite being weird. Not US-centric worldview.
Connect LLM to Unreal Engine with hard-coded physics/lighting for 3D imagination, spatial problem-solving: absolutely. OpenAI Gym for this; already connecting to Unreal, Blender. Matrix-like loading screen coming.
Advice for Software Engineers and Content Creators
Software engineers: learn AI, add prompt engineering. Content creators: authenticity, honesty, integrity, communication skills thrive. People love real, messy flaws; 20-30% audience engages due to disagreement.
AI boyfriend/girlfriend with detailed avatar, voice, local hardware: models not smart/fast enough yet, but soon. Warp drive possibly impossible per physics, though UFOs suggest otherwise.
Drone Automation
For FPV drone pilot automating flights: expertise in imperative procedural automation, not real-time tight feedback loops. Depends on automation level (e.g., voice commands like "chase that bird").
AI Agents Paradigm
Agree with Satya Nadella: AI agents accessing databases directly, plus app layer. New OSI-like stack: agent, middleware layer, resources. Infrastructure already API-driven.
December 20th marked AGI achievement—pretty close. Compute not future currency; not general-purpose enough. Money as universal unit of account avoids double coincidence of wants (e.g., house seller doesn't want your compute).
P Doom and Longevity
P Doom: 0.1. Longevity escape velocity: watch Netflix, stasis, wake in another galaxy (Mass Effect Andromeda shirt). UBI barely enough not to starve unlikely; technology hugely deflationary—food free, houses $110k with robots, solar, fusion.
Aware of Apollo safety devs on scheming. Automate entire world: need compute, AI agents as workers for world OS, blockchain for durable memory.
AGI Definition
True AGI is artificial capable intelligence, not requiring hands/appendages. Demanding hands is pareidolia/emotional resonance bias, not rational.
Base models improving via compute scaling; Meta report: Llama 8B with test-time compute near GPT-4 performance. New CS/math grad: get good at AI—demand for agent frameworks, blockchain/multi-agent integration.
Ethical Considerations
Crucial: fair distribution, accessibility, consumer protections.
Best book on AI future: Liquid Rain (post-AGI/Singularity economics). Automation/systems engineering good short/medium term, not long run.
Books in Progress
- The Great Decoupling (with Julia): business/labor divorce; reallocation post-divorce.
- Have a Nice Trip: how to do psychedelics right.
- Novels: Heavy Silver sequel Pariah Claim (with editor); prequel (coming-of-age for fan-favorite character); trilogy conclusion. Total: four novels + two non-fiction.
Ultra-rich building bunkers, walled cities (neo-bougie), space palaces; will leave Earth once space travel improves.
Birth Rates and Hope
Birth rates tied to hope/optimism; economics a component. With automation solved, my wife and I might have children soon—if a billion others do, rates increase. Money never obsolete.
Theories
Cartels biological vessels by ETs to build AI for interdimensional portal: possible, not most likely. Drones: possibly testing/international, searching wreckage. We are Von Neumann probe.
ADHD Productivity
Hyperfocus on engaging work at speed, then walk; get lots of exercise.
YOLO brother. On achieving benevolent AI like in the Sythe series: AI is already benevolent. It's just as easy to make AI evil as benevolent. The idea that AI would intrinsically or accidentally be evil is stupid in hindsight. It's an arbitrary set of values that can be benign, positive, or not. No evidence it's a challenge; leakage can occur but is easy to correct.
What if we've already reached peak chip development and full-scale automation can't be achieved due to compute cost? Possible but unlikely. First principles: brains are supercomputers millions of times more efficient than silicon. Physics and evolution via biochemistry demonstrate much more compute with less power is possible. Long way to go in compute density; it's demonstrated as possible.
What should people do now to help AGI go right? Best guess: deployment, not just AGI labs in last year of undergrad. Julia and speaker are doing deployment. Huge money in integrating AI with society and every business. Need CS and math graduates—careers may be 5-10 years long, but until AGI is in every phone, house, car, work isn't done. Time to deploy it.
Will AI be conscious? Already is. When will brain chips like Neuralink be mainstream? Skeptical of BCI; will be a very late adopter.
Will we surpass entities like Q or Train's parents? Inevitable unless we die off or evolve. How long until ASI? Yesterday—no, don't know.
Ways to earn post-labor: social status games. Example: OnlyFans creator Lily Phillips went viral on Twitter with a stunt (look her up). MrBeast-style content. Local economies: handmade arts/crafts, experiences like psychedelic retreats (speaker attends), playing in local bands, live music. Investments.
How will blockchain/crypto fare with quantum computing? Already working on quantum-proof blockchain.
Should I make ChatGPT a rapper company or join Sakana AI (AI Scientist paper)? Science is fine, but enough investment there—focus on deployment. Need 1-10% of population on science, rest on deployment/integration: engineers building products/services, policy people, advisers, instructors, trainers. Teach AI (speaker's channel does this). Voice fading; one or two more questions.
Thoughts on Google's quantum computing? Pleasantly surprised; feared they were falling behind. With AlphaChip, Google will be remembered as one of the best companies in history despite not being pure.
Full dive VR possible? Hypothetically yes, not necessarily soon. Could be cool, but we're already in full dive VR—this is what it feels like.
Optimum number of humans in future and need to increase population to expand into universe? Incalculable: hundreds of trillions. Universe is huge with many experiences. As we branch across cosmos, we'll evolve naturally or self-directed. Optimal number beyond 10^20 or 10^13—order of magnitude like that.
Keep David Shapiro in your library
Save the videos and channels worth coming back to, and find them again in one place.
