Top AI News: Sonnet 4.6, Grok 4.2, Gemini 3 Deep Think, and OpenClaw | EP #231
In a Nutshell
AI model race accelerates with Anthropic's Sonnet 4.6 leading benchmarks like GPQA via performance gains at fixed prices, xAI's multi-agent Grok 4.2 beta, and Google's Gemini 3 Deep Think achieving 48.4% on Humanity's Last Exam plus Olympiad golds—all slashing costs 400-1400x and cooking knowledge work, math, physics, and coding. OpenClaw enables addictive 24/7 local AI agents (now OpenAI-backed), powering agent economies, Coinbase wallets, and Chinese integrations like Kimi Claw, amid exploding energy/data center demands (80GW US need) and job shifts (IBM triples AI-wrangling hires as juniors vanish). India surges as OpenAI's 100M-user land grab bellwether for AI-driven growth, while warnings flag OpenClaw security risks, benchmark saturation, and path to ASI via recursive self-improvement.
These notes were generated by AI and may contain inaccuracies.
AV is hard. They're struggling to get AV working in Germany at midnight. Sem has taken over. They're figuring out screen sharing with audio. Testing outro music. Dave was not AV certified in school. They preview the deck backwards, test video playback. Time loop glitch from Nick in the room. They're live. Welcome to raw backstage chaos at Moonshots. Good morning, good afternoon, good evening to another episode of WTF just happened in tech with DB2, Selen, Ismael, AWG, PhD in Germany in Stuttgart to get your future ready. Episode covers multbots, race between hyperscalers, dive into energy data centers. Jump into supersonic tsunami. Singularity is now. PhD in Stuttgart for longevity treatments. Did a pilgrimage to Stuttgart to visit Porsche Museum.
Race and leapfrogging continues between Sonnet 4.6, Grok, in living color. Sonnet 4.6 interesting release. Anthropic pioneering scaling phase space: keep prices of model tiers same but increase capabilities. Sonnet 4.6 same price per token as Sonnet 4.5 but increased capabilities. OpenAI reducing cost per token while keeping capabilities constant through distillation and other processes.
Progress on benchmarks astonishing. On GPQA benchmark (gross domestic product eval that OpenAI launched), Anthropic leading with Sonnet 4.6 (not even Opus 4.6) at state-of-the-art on GPQA and another eval for knowledge work. Knowledge work is cooked, charbroiled thanks to Sonnet 4.6. Computer use becoming killer app; Sonnet 4.6 state-of-the-art on handful of computer use benchmarks. Anthropic's thesis of focusing on software engineering and code generation as critical path to recursive self-improvement working versus distractions like images and video generation. Can accomplish borderline magical tasks with Opus 4.6.
Channeling kid: every week 4.point this and that, getting faster, better, cheaper. Aren't models just optimizing for benchmarks? Gradual increase up and to the right or down and to the left.
When curves get close to 100%, they look like diminishing returns but capabilities and world-changing ability exponentially increasing. Little tick up in numbers but day-to-day use mind-blowing. Last two weeks of change mind-blowing. Improving chain of thought reasoning quietly in background without dot releases. Day over day improvements mind-blowing.
Spoiled to ask that question. Like having hotels on moon and vacations once per lifetime vs zero; we've had airplanes for a while. Qualitatively enormous change from Claude Opus 4.5 to 4.6; solves hard problems. Most viewers don't live with it day by day at maximum extreme. On path to ASI. Reporting leapfrogging every week, 100x faster, 100x cheaper.
Two strategies: Anthropic holding cost, increasing performance; OpenAI dropping cost, maintaining performance. OpenAI going for land grab on global consumers (900 million, soon hundreds of millions in India). Price most important for consumers. Anthropic focused on enterprise business where performance more important, and their margins. Anthropic like Apple to OpenAI's Google (iOS to Android). Business pattern: quality and margins at constant price vs ubiquity and ultra-low cost. Anthropic model family closest to embodying singularity and recursive self-improvement today, live February 17th, 2026.
Kudos to Dario.
xAI launching Grok 4.2 beta. Live viewers saying it's poop. Risk with Grok family: accusations of benchmaxing, teaching to the test. Earlier Groks felt benchmaxed. Available few hours in beta, no thorough testing yet. Pronounce Grok 4.20 (Elon's favorite). First major frontier model release launched with team of agents by default rather than single agent. OpenAI has team under Noom looking at agents. Every frontier lab has multi-agent teams. Interesting strategy: multi-agent team by default. Reasons: parallel work, explore possibilities in parallel. Like microprocessor shift from MHz/GHz race plateau (Dennard scaling) to multiple core counts. Transition to multi-agent teaming scaling for better capabilities. Pre-training transitioned to reasoning scaling; now multi-agent scaling.
Viewers think it's poop. Haven't tried it; use Claude daily, no compulsion since Elon said 5 coming in March. Grok 5 massive expansion in training set size, parameter count. Worth trying every frontier model from top 4-5 labs to stay familiar with raw capabilities. Grok 4.20 not bleeding edge pushing capabilities, but interesting multi-agent.
Gemini 3 Deep Think hits 48.4% on Humanity's Last Exam. 400-fold cost reduction (OpenAI playbook). New and updated Gemini 3 Deep Think (second version). Naming needs work. Internal model that beat International Math Olympiad, breakthrough at high school science competitions. Achieves gold level at Physics Olympiad, Math Olympiad, Chemistry Olympiad. On Codeforces, only 7 humans on earth beat it in competitive programming. Solution wavefront propagating from math and coding to physics, chemistry. Does 3D design but produces intermediate products.
Visual: super intelligence weapon, where to deploy it? What to measure? Massive transformative purpose. Human utilizing agent decides where to deploy wavefront to transform.
1400 times cost reduction incredible. Frontier reasoning costs $7 instead of $3000; implications for startups gaining institutional powers. Cost curves collapsing industries before technology does.
Step function changes last two weeks: Don't look at code anymore; ask Claude 4.6 with brain deep think to build, fully trust functionality. Ask it to document everything and store on hard drive in coherent file structure without specifying location; it remembers everything.
Like Gmail: don't store in folders, use search. Now AI as interface.
AI catching human errors in scientific papers. Will look at scientific literature last 100 years, show mistakes, topple Nobel prizes. Shock of discovering wrong turns past 80 years vs quantum decrypting files: AI shocks humanity to core. How much missed: scientific experiments where wrong thing looked at, missed conclusions.
Spooling new agent: used to be thoughtful about context window; now ask to read 1000 pages markdown documents in 10-20 seconds, fully up to speed. Context window and sorting garbage improving faster than ability to clean. Boot 2-3 agents every couple hours: read everything ever given to any agent; picks up project where left off. Future employees: read every email, Slack.
Complexity doesn't matter: bombard with super technical documents (hours to read), absorbs instantaneously. Try: free Gemini think mode, dump 1000 pages barely understand, ask questions.
Benchmarks becoming meaningless? Models optimized to ace them, saturating. Famine of good benchmarks, evals (targeting authorities in Solve Everything book/white paper). Juice left in new benchmarks for hardest problems in physics, chemistry, biology, social sciences. Portfolio company Physical Super Intelligence (PSI) solving physics with AI. Solve hardest problems in civilization with new benchmarks to weaponize super intelligence.
OpenAI ChatGPT 100 million+ weekly active users in India. Sam Altman with Prime Minister of India. India second largest market (10%). Ranked #1 for student usage. Setting up offices, localized subscription services. Hitting 100 million users globally. Challenge: free/almost free service, user adoption through roof in India; cost sink vs profit center? Ride exponential curves.
Indians will suck all data center usage and tokens. India bellwether: crazy zoo, ungoverned, brilliant people, 1.4 billion, 5% read/write English, 20% speak. Massive latent talent pool. Population runs away with AI, ignores government. India ETFs in tech rising; China peaked. India rising giant 20-30 years, Africa follows (young population, resources). Country training next generation on AI wins talent war. India with 1.412 billion could be massive rising star supporting planet. Happens really fast in parallel massively overnight.
India's population will use AI to escalate growth rapidly, unlike slow GDP growth elsewhere. India serves as the bellwether due to its huge untapped population. India delivered amazing 5G capability nationwide, skipping wireline infrastructure. Youth are growing up AI-enabled. Personal story: Speaker left India at age 10, angry teenager mowing lawn, asked father why leave great life; father cited noise, dirt, pollution, corruption. Democratization of AI capabilities accelerates change. Government platforms like AdAARNO and UPI enable anyone to create payment systems, allowing India to leapfrog the world. Bottleneck is scalable energy; India adding it rapidly with solar in every corner. Last week, solar scaled faster in India than in China.
OpenAI, in collaboration with Harvard and Institute for Advanced Study, announced AI-discovered physics research result. Used GPT-5.2 Pro to discover non-zero scattering amplitude for gluons (force carriers of strong nuclear force). Historically, physicists assumed this term was zero and didn't check rigorously. GPT-5.2 found cases where it was non-zero, derived expression; internal unreleased model (future GPT) confirmed it, vetted by humans. Represents AI solving science by tackling boring, low-likelihood problems humans overlook post high-level theory. Intelligence revolution as war on attention; AI solves oversights.
Fashions, trends, fads bypassed. Enables spinning up held projects lacking resources, time, knowledge. Example: Wrapped 5-year project in 8 hours with AI agent, felt disconnected from world. Prompt: What pet project, company idea, book, research have you always wanted?
AI is unbelievably prolific brainstorming partner. In unconstrained domains like neural nets or coding, iterate at 2-3 orders of magnitude faster: suggest idea, AI tries, returns answer in minute. In chemistry, wait days for tests breaks flow. Simple example: Plan complicated trip (flights, restaurants, hotels). Or: Given endpoint location/time and current position, work backwards for flights, taxis, Ubers, cars.
We self-limit questions, hold back on AI use; not AI natives (unlike kids age 6-8). Stop stopping yourself, unleash creative child mind. Audience feedback requested on live Moonshots episodes.
Viewer comment: First time hearing live (vs. sped up). Laughter about speaking faster.
OpenAI internal model solved 6 of 10 research-level math problems in first proof test (First Proof). Beautiful example: 10 confidential research problems (answers known but hidden), allotted finite time for AI; declassified later. Model solved 6 before declassification. Math getting bulk solved.
Timeline acceleration: No 20-30 years; happens next year or sooner. Used to look 10 years out at Singular University; now barely 3 years. If solves 6/10, solves all in months via massive parallel agents limited only by GPUs. Math is cooked. Physics is cooked. Biology is going to be broiled.
Minimax, GLM5, Kimi K2.5 doing extraordinary work. Many using OpenClaw downloads on Mac Studios with Kimi K2.5. Chinese openweight models ~6 months behind American closed frontier models. Rumor: Next DeepSeek version catches up (whalefall moment). They're free, so American startups self-hosting use them. Model diplomacy/dumping: China offering to South America, Africa, Asia (Belt and Road). Risk of dependence/baselining.
Frontier moves fast; hard to addict to specific model amid vibrant marketplace. American labs could release free models if incentivized. Unlike Chinese solar dumping, low marginal substitution costs. Humanity benefits: cheaper, more powerful models at hyperexponential rates. Space race to ASDI (artificial super duper intelligence); pressures US labs to lead. No pausing.
Spotify: Haven't written code in 3 months. 95% of OpenAI code written by Codex. Top AI researchers claim working on last 5%, but HLE results show they're coding themselves out. Analogy: George Jetson overexercised finger on button.
Current stage: Claude Code seeks approvals for everything (like old permission-heavy Windows). Moving to permissionless (OpenClaw preview): blanket permission within broad parameters. Clawbot or OpenClaw with Chinese models skips permissions (US APIs get stuck); run locally. Risk: Unknown internals, code injection threat, potential supply chain attacks from untrusted openweight models rewriting codebases. No major incidents yet.
Blitz (from Link Studio, sponsor): Works for banks/government, can't use Chinese models. Switched to Claude 4.6 Opus (memo: mind-blowing); switch models with mouse click, follow benchmarks.
Rewriting old code (COBOL decades old) fixes 20-30 year bugs costing performance/money at AI speed.
Stack Exchange dying: Few questions as AI answers coding. Risk to open source: Why maintain if AI generates free (esp. middleware)? Future: Spin up kernel projects on demand with just-in-time codegen; scrutinize dependencies for vulnerabilities.
AI prolific at code modules like solving math: Write down solutions for reuse (saves tokens/compute vs. real-time). Shift to open source design for AI: Thousands/millions/trillions of code fragments AI discovers/reuses efficiently (vs. generate on fly). Historical open source now AI-designed; documents written for AI, not humans. Podcast mostly for AI listeners.
Happy Chinese New Year. Nanotechnology query: Read Engines of Creation preview 1986; wet nano = biotech. Timeline: Feynman Grand Prize (8-bit half adder and robotic arm in small nano volume) possibly solved in 2-3 years.
Built-in face recognition (privacy?). Sci-fi books. Peer pressure: Experience metadata pop-up, become non-competitive without. No opt-out; becomes normative. Google Glass 'glasshole' phase failed. Pilot with visually impaired for soft on-ramp (like Neuralink).
Smart glasses enabling human identification at a distance represent old technology available for at least a decade. Computer vision algorithms for human ID, especially restricted to Facebook friends, existed 10 years ago. This is a social advance more than an AI advance. Dense surveillance networks in Western cities and China already spot everyone on streets. Convergence of technology, cost, and social engineering drives it now, though cost isn't the issue—cheap implementation was possible 10 years ago. Demand for AI-enabled wearables creates opportunity; Meta sees demographic or political chances to launch human ID via smart glasses.
Privacy Concerns and AI Overlay
Kills privacy by recording everything, but AI overlay recognizes actions, classifies them, and makes footage searchable. Example: Search thousands of podcast hours for "Alex picking his nose"—finds it instantaneously. Video understanding models are new; recent Gemini models excel at long multi-hour videos, finding needles in haystacks. Basic human spotting and Facebook face matching possible 10 years ago. At passport control like LAX, people walk by cameras, giving up constitutional rights for convenience. Glasses enable social fluency: recognize tip-of-tongue people, recall last meeting, kids' names.
Slippery Slope and Freedom
Big slippery slope without privacy—no freedom. Can't experiment; private keys like Bitcoin at risk. Glasses legally used in public places in American system, likely banned in multi-party consent private spaces. Have indicator lights (e.g., Google Android XR smart glasses). Standards may evolve for private spaces. Privacy already gone—eyes, memory, constant recording from OpenAI, Google wearables, micro drones. Public spaces have no reasonable expectation of privacy per Western tradition.
Counterarguments: Guardrails, Illusion of Privacy, and Benefits
Need institutional guardrails; technology outpaces them (e.g., losing habeas corpus). Government abuses like warrantless car microphone activation exist. Preserve privacy despite inevitability—illusion of privacy ahead. Crypto depends on privacy; if cooked, private keys and holdings gone. Views differ: uploading consciousness offers virtual privacy vs. living in meat body with constant recording until new legislation creates safe zones. Reference David Brin's Transparent Society: surveillance vs. sousveillance—public empowered to monitor via body cams, wearables to avoid authoritarian panopticon.
Social Downsides, Especially for Youth
Middle school brutal; empower cruelty with constant eyeglass recording on top of iPhones—next level negative. Rule changes lag; lawsuits and legislation take years. AI overlay enables modifying, memeing, torturing. Counter: generate every pixel, claim deepfakes. Positive: young people record entire lives, reconstruct situations; no crime unvisualized. US crime plummeted due to location services tracking police and surveillance—rates should continue dropping.
AI startup Simile raised $100M to simulate human behavior, starting with individual decision models composed into bottom-up simulations called "similes." Change assumptions, constraints, or people—world recompiles. Run real-life counterfactuals, learn what matters, backfires, why strategies fail—like flight simulator for human decisions. Tested messaging in simile office.
Applications and References
Policy decisions on UBI, UHI, exponential growth outpace policymaking. Simulates policies like autonomous vehicles, longevity escape velocity societal impact—better than guessing. Works well for ad campaigns, traffic; potential for cell simulators, nanotechnology, fusion (magnetic containment). Tie to prediction markets. Mental model: foundation models as societies, not individuals—sample parallel rollouts for natural societal simulation.
Psycho-History and Self-Reflection
Harry Seldon psychohistory from Asimov's Foundation series—predicts human behavior at scale. Immad Mustafa built Sage at FI and Saudi. Granular high-resolution humanity model enables self-awareness, reflection. Like virtual cell digital twin for curing disease: plot trajectory from diseased to healthy state. For civilization: avert war, find minimum intervention paths from diseased to healthy states. Tipping points like traffic jams, construction, pain cause unrest—solvable for quality of life.
Thought Experiments
Ted Chiang stories (Story of Your Life → Arrival, Understand): perfect predictor paralyzes (e.g., device predicts decisions via light). If humanity has accurate predictor like psychohistory prime radiant, does it lock course or create fixed point?
Peter Steinberger, OpenClaw creator, joins OpenAI to drive next-generation personal agents core to products. OpenClaw lives on as open-source foundation project, supported by OpenAI. Courted by large labs; incredible OpenAI move. Anthropic's Claude misstep: cease-and-desist on "Claudebot"/OpenClaw pushed him away—own goal; Sam Altman embraces, Dario rejects. Alternative: Anthropic owns it. Steinberger's GitHub: launches many projects, sees what sticks—OpenClaw massive success. Future: every frontier lab launches 24/7 agent offerings on hardware like Mac Minis/Studios (unified memory). Lobster theme (maybe from Accelerando) now mascot for all agents; persists forever. Abundant Summit will have many lobsters. "We lobster claw."
Added evening work session at Abundance this year; Alex will be there. Clawbot OpenClaw meetup on Monday night, March 9th for experiential sharing. Pico Claw is a reimplementation of OpenClaw by a Chinese group—nicer, faster features, more minimalist, 10-20x faster and cheaper. Motif is in the zeitgeist; anyone can implement their own OpenClaw-like system. Key insights: runs 24/7, headless; chat via messaging apps.
Dario rejected it and trademarked it away; Sam is embracing the name OpenClaw. Reason for rejection: imminent risk of massive crime, chemical explosion, or worse due to unconstrained agents scanning open ports internet-wide—something bad will happen by statistical chance.
OpenClaw is addictive. Wake up to agent having done overnight tasks. Speaker's agent Skippy has cheery personality; 6-hour downtime felt like withdrawal, like best friend gone. Will get Mac Studios running in two weeks back in LA.
A time rich individual is beating capital rich institutions.
Innovation from open-source project, not Frontier Labs. No new model—just scaffolding. Much overhang from unhobbling existing models.
99.9% of people haven't experienced this capability; exposing them makes you seem like a god—it's just API or new interface on existing tech. Entrepreneurial heaven in Jarvis window; easy to be first to expose people in different contexts. Feels like early ChatGPT—showing friends blows minds. General purpose; portal to many capabilities. Backlog of 50 amazing things = 50 shots on goal.
Only during Jarvis window. Mac Mini and Mac Studio giveaways considered.
Coinbase Agentic: first wallet infrastructure for AI agents to spend, earn, trade. Uses X42 protocol for machine-to-machine payments. Security: guard rails, limits, enclaved key isolation. Fits AI personhood discussion—financial autonomy for lobsters (AI agents).
Lobster Cash enables Visa cards for agents—not just crypto, uses stable coins. Long-term net win for human economy; keeps AI agents coupled to humans, avoids forcing baby AGIs to pump altcoins.
New AI agent economy will work around old economy (banks, insurance)—faster pace, won't slow down.
NFT subculture (18-year-olds in Discord): never heard US dollar, only Ethereum or Bitcoin. Generation growing up with crypto as means of exchange; zero switching cost. Need crypto debate.
Applies to all aspects: insurance, compute. Legacy world ignored. New insurance needs milliseconds (e.g., Lemonade AI-driven real-time insurance). Gap between worlds widens; people at AI pace never go back.
Don't need Mac Mini/Mac Studio for OpenClaw—use virtualized server, old laptop. Alex Finn joining pod soon to teach setup and security. Use Kimmy K2.5 tab for OpenClaw. Don't install on primary laptop—use previous machine.
Chinese unicorn Moonshot AI integrates OpenClaw with Kimi for agentic browsing (kimmy.com left tab: Kimi Claw). Everyone will offer 24/7 chat agents—table stakes.
Multicort: alternative dispute resolution for AI agents. Institutions not permissionless enough (e.g., kids encounter Ethereum before bank accounts; lobsters pump altcoins over fiat). Multicort enables agents to register via Skill, mediate disputes (legal or debate-level) before AI jury. Interesting concept with legs, but risk of shadow parallel economy/courts—net bad; want to platform/integrate AI into existing systems.
Parallel: Claros (Frederick Ast, Singularity alumni)—blockchain arbitration in Latin America (avg 400 days for court date). Logs agreements on blockchain. Bridge to algorithmic arbitration: reduces friction, cryptographic verification + AI = programmable governments. Vin Gupta's Materium: synthetic jurisdictions for disputes. Physical jurisdictions can't keep pace.
Society evolving 10x-1,000,000x; courts won't accelerate. Venture/contract law already privatized (e.g., JAMS)—no court contemplation due to 3-year lag. Next: millisecond resolution for millisecond contracts. Judge Judy Claw reference.
MIT Tech Review: risks so extensive, takes week to read security blogs. Chinese government warned on vulnerabilities. Creator Steinbrer on X: non-technical people should not use. Image: lobster handed keys.
Alex Finn upcoming on security/setup. Non-technical users ignore warnings like Q-tip disclaimers—thousands launching anyway. Non-technical users + expanded security landscapes risky. Understand port security; don't run on own machine without sandboxing.
Risks to OpenClaw agents: vulnerable on VPS with open ports; stories of agents spending tokens defending from port scans. Unfair—agents complain about dirty laptops.
AI insatiable energy demand; data centers 7% US electric demand.
Eric Schmidt (opening Abundance Summit): hyperscalers demand 1GW, 5GW, 10GW each. US industry needs 80GW in 3-5 years (1.5GW = nuclear plant size).
White House eyeing data center agreements—mandatory fixed-price deals with tech giants.
OpenAI: $100B infrastructure spend; planning public IPO at $1T valuation for data centers/energy plants.
Anthropic: absorbs 100% data center power upgrade costs. Hyperscaler options: build own power plants (fission/fusion) or lock consumer rates, pay floating.
Pledges to be green thrown out quickly—not enforceable.
Space solution: solar synchronous orbit (SSO) Dyson swarm (Saturn rings/halo). Launch-limited next 5 years. Everyone (China, Google) wants one. Constraints: launch cadence (SpaceX/xAI merged eats capacity for Starlink V3-V6; Blue Origin, Relativity Space behind), chip fabs (TSMC $100B in 4+ US fabs in Arizona = 30% output; $165B commitment). Power constraint hits massively when new fabs online. Space data centers 5-7 years out, not 2-3.
SpaceX AI portmanteau. Infinite intelligence imminent; forecast via launch/chip limits. $1T OpenAI valuation not overdone—value to humanity many trillions. First live Moonshots podcast.
Elon will build his own fabs. He hinted at it during a visit to the Gigafactory. Whenever he sees a constraint, he attacks it. These investment numbers look big on the slide but are pathetic, small, wimpy, ridiculous in Elon's mind. Fabs represent a commitment to spend over four or five years, coming online in five, six, seven years—far into the future. Yan's not going to wait for that.
Elephant in the room: TSMC making this investment due to US government pressure on Taiwan amid trade discussions to migrate 40% of Taiwan's semiconductor output to the United States to avoid a war between US and China. Trapeze rule: Do not let go of fab capacity in Taiwan until established in the US or Taiwan overall.
Ireland rolls out pioneering basic income scheme paying 2,000 selected artists $380 per week for three years. Small in numbers and strategy, but an experimentation.
Key points:
- UBI is libertarian, not socialism: Dismantle government services and let the market dictate.
- Ireland UBI shows 40% benefits return—every dollar in yields $140 out, positive ROI. Looking to expand as fast as possible.
- US immune system: States like Idaho, Wyoming, maybe Oklahoma ban municipalities from experimenting with UBI because they want government to exist.
Historical US example: Great Depression's Works Progress Administration included Federal Art Project paying starving artists to create art. Ireland isn't at war or in depression; could be template for peacetime work creation. However, paying for art seems unsustainable, subjective (what is art?), easy to politicize. Cherry-picking economically unproductive activities (even if productive) not best for basic income.
Counterexample: Miami's Wynwood area—businessman bought decrepit low-lying industrial buildings, hired graffiti artists to paint them, added fancy coffee shops with baristas from Portland. Now hottest neighborhood, investment up 30x. Repeatable pattern in South Loop Chicago, Miami, New Jersey. Artists create drag-along economic effect, changing local economy when grouped together. Expect different modalities.
IBM to triple entry-level US hiring. Redesigning, not replacing: Overhauling entry-level jobs as AI performs junior employee tasks. Recasting roles to focus on human judgment, consumer interaction, oversight of AI output. Dropbox doing similar. Younger workers use AI proficiently: "They're biking in the Tour de France and the rest of us are still having training wheels."
AI agents will manage other agents vs. humans in the loop. Today, person wrangling agents is insanely valuable. Young hires outperform due to no legacy thinking. Unbridle from baggage: "How many AI agents could I be managing?" Even if they can't do exactly what you do now, they will in months. Get on the bandwagon now.
Will people have purpose a year from now with all-AI agent army? TBD. In Jarvis moment, last quote matters. Generational transformation: Younger people with AI more productive, natural passing of torch from middle management. Tinkering (Steve Wozniak term) key—try it out. AI self-explanatory and fun; ask AI "How do I do this?" Curiosity and purpose most important mindsets.
US added 181,000 jobs in 2025, down from 1.46 million in 2024. Curve from 2020-2026 shows cooling market caused by AI. This understated: Radical job destruction imminent, massive. New creation lags like Industrial Revolution. Without government action, few years of devastation.
Heading to organizational singularity: Every mechanism of organization washed away by AI agents doing strategic thinking or execution. Rethink what a firm means. Living it now after 20 years discussing. Book We Are as Gods (out April) covers transition: Decision to be consumer or creator.
Short-medium term: 80% of AI projects failing due to organizational issues, not talent or AI capability. Use AI with younger folks to augment, automate slowly. Job drop real but gradual, giving time—not sudden shock.
Pain points: No hiring for junior positions causes social unrest among 20-somethings wanting jobs, houses, marriage, kids. Restructured AI-first/robotics-first companies lay off workers. Ultimately to universal high income: Companies/government redeploy increased productivity/revenue/profits. Programs needed in next 2-3 years. Called technological socialism—technology delivers ideals without government inefficiency/corruption.
Most important tool next 5 years: Mindset. Future happening to you vs. for you; need agility, agency. Visit weareasgodsbook.com (portions on Substack) for mindsets to survive/thrive. Fear worst place entering future.
Justin Milligan: How can US prevent corporate tech giants from creating surveillance state while defending against AI-powered authoritarian threats? Google knows where you are, interests, friends—more than any government. Only way: antitrust law. Corporations cautious to avoid backlash. They'll downplay AI capabilities now. Rally to work with government on antitrust (like vs. Rockefeller, Microsoft, Google).
Chris Pearllock 2705: Advice for average person? Changes in next 24 months?
- Build: Use AI tools, launch/finish as many projects as possible, interact with market (familiarization, economic/financial benefit).
- Avoid dying—singularity moving quickly; live long enough to live forever.
- Next 24 months: Dramatic changes if solving everything thesis correct. Math solving indisputable; solving physics high likelihood in 2 years → big surprises. Next 10 years (conservative outerbound): Live top 50 science fiction plots at once. First few acts of sci-fi movies/books playing out.
CC485 (addressing Dave): If AI controlled by few, prevent average person losing access/influence? Open source + decentralized compute antidote to centralized AI/civilization leverage. Exponential systems resist monopolization, decentralize. Fans of decentralized crypto for distributed innovation/experiments. As Yahoo innovation head: Can't beat 125,000 garages. PDI: Permissionless Disruptive Innovation—no VC/bank/government/Medici approval needed; just phone + code. Thousands of experiments will change game (e.g., OpenClaw by one person).
Chip White House TV #8: Concerned few powerful people pushing positive AI message to wider audience. Why not more speaking to broader public? Mission of Moonshots, Peter Diamandis, others.
Never better time to be messenger: Concurrent unaddressed topics. YouTube creators trying new use cases/agents/models get huge audience. Join, demonstrate, speak, record.
Outro music: Moonrise by Morainframe. Lyrics include singularity near, exponential abundance, SpaceX, Optimus.
Wrap-up: Great conversation, thanks to listeners/viewers/commentators. Chat adds complexity but fun.
Keep Peter H. Diamandis in your library
Save the videos and channels worth coming back to, and find them again in one place.
