Artificial Super Intelligence - What will ASI be like? When? How will it change things?
In a Nutshell
AGI is expected by 2026-2027, rapidly evolving to ASI with PhD-level capabilities trickling down cheaply via clones and optimizations, saturating most cognitive needs and enabling humanoid robots like Tesla Optimus for expert manual labor. By 2050, transhumanism via gene editing, regenerative organs, and Neuralink-like implants will allow extreme body modifications within thermodynamic limits. AI integrates transparently into daily life (phones, cars, games) while universities should experiment with test-time compute and distillation; parents' social apps need AI moderation/fact-checking plus human connection to handle tribal debates.
These notes were generated by AI and may contain inaccuracies.
Blog posts from Microsoft, Google, OpenAI, Anthropic indicate CEOs and big researchers converge on AGI as smarter than all humans in all domains, superseding Nobel Prize winners. Expected in 2026 or 2027, 12 to 24 months away. Semantics between AGI and ASI, but raises question of AI's ultimate potential.
Question: How far can we change our bodies by 2050 via gene editing, nanomachines?
Biology is plastic; life forms range from fungi, prokaryotes, eukaryotes, vertebrates, invertebrates, spider-like pods. Possible forms: four brains in a jar, 20-ton ocean-going creatures. Thermodynamic limits exist, e.g., maximum animal size due to energy dissipation—Godzilla impossible due to overheating, core temperature hundreds of degrees; upright stance unfeasible. Elephants have columnar legs, can't jump; largest land animal like Argentinosaurus many times larger than humans.
By 2050 (25 years): Stem cell research followed since 2008-2009; predicted takeoff around 2020, accurate. Regenerative medicine market now $25-42 billion globally, from zero in 2010. First genetically engineered pig kidney transplanted into human. Innovation unlocks more: monoclonal antibodies, mRNA vaccines expand medicine classes quickly. Genetically engineered organs could become commonplace.
Implants like Neuralink augment. Nanomachines already exist in biology (nanometer/angstrom scale machinery in cells); biotech more efficient than stainless steel/titanium. Gold nanoparticles possible, but brain/liver/kidney chemical processes hard to replace.
Life 3.0 Framework (Max Tegmark)
Life 2.0 (humans): Software upgrades via learning (writing, math, languages, calculus) during lifetime, but biology changes slowly. Life 3.0 (machines): Entire stack changeable—swap hard drives, CPUs, OS. Computers design chips via ML algorithms; AI bootstraps itself.
Biological plasticity possible long-term, matching ASI level, though interconnected systems lack plug-and-play parts. Eventually like Ghost in the Shell body swaps or Cyberpunk arm swaps, not USB-easy but interchangeable/upgradable. No major physics laws block it.
OpenAI announcing PhD-level agent for $20,000/month. Lag from frontier models (OpenAI) to others was 2 years, now 4-6 months. Knockoffs like Perplexity, DeepSeek R1 reproduce functionality in days/weeks, 10-100x cheaper or free, good enough. o1 Pro, OpenAI deep research still superior.
OpenAI strategy: Release exclusive shiny models (o1 preview/Pro), then efficient distilled versions (o3). Tick-tock model like CPUs: new architecture (expensive/slow), then optimize (faster/cheaper).
Expect clones in days/weeks, broader distribution in months, feature parity in a year. Pattern compresses: 2026 days/weeks; 2027 days. Saturation at PhD+ level shifts competition to speed/efficiency/UX. Most (90+%) cognitive needs met by end of year/next year with Claude 4, GPT-4.5/5 (smart but slow/expensive now). Cycle: smarter (slower/expensive) → same intelligence cheaper/faster.
Arms race possible as people enhance businesses with AI, but at cognitive saturation, intelligence not differentiating; compete on service, location, red tape/permits.
People largely ignorant despite regen sciences/chatbots, except occasional use.
Usability key: Early AI (GPT-2, Google's Universal Sentence Encoder) required Python, training data knowledge—not useful. GPT-3 chatbots suboptimal form factor. ChatGPT 1000x more useful due to packaging (engine in train/jet vs. table).
First-order use: chatbots. Like steam engines initially replacing water wheels before trains/boats. AI to become transparent: baked into phones (generative models deployed), cars, Amazon integrating Claude into Alexa, Apple Siri (2-3 years). Video games with dynamic text/speech (Skyrim mods). Before long, buy real-life functioning R2-D2 at Best Buy.
R2-D2 or C3PO will signal the age of AI and robotics to the public, but it started five years ago. Usability depends on form factor and integration. In video games like Skyrim, AI integration is already happening with mods. By 2027, form factors will make AI more noticeable. Primary form factor: humanoid robots like Figure, Tesla Optimus, and Chinese ones (name forgotten). These robots will do manual labor and be PhD-level geniuses: best mechanic, lab helper, doctor's office assistant more competent than humans. This creates an 'aha' moment with autonomous portable form factors. Current chatbots are stuck in phones or browsers; in robots with hands, it's different.
For someone in early 20s not making much money, position yourself where passions lie. Older generations took practical jobs for 15-17 years before pursuing passions. Depends on circumstances, standard of living, opportunities. Keep true passions in mind. Speaker got into AI as a serious hobby while doing other work with overlap. No such thing as wasted experience: boring corporate meetings taught live streaming and interview skills. Humans have transfer learning. Skills from hated jobs aren't useless (e.g., Powershell scripting, VMware virtualization still remembered). Adopt abundance mindset: new skills are valuable, they may pop up later (e.g., wife's marketing video editing useful for YouTube). Not concrete, but an attitude.
Ethan at University of Florida, accepted to Perplexity's campus strategist program. Mission: maximize impact via bottom-up (clubs, demos, incentives) shifting to top-down institutional partnership.
Break into stakeholder buckets: students, faculty/teachers, admin (deans), families/parents, politicians/boards/alumni (public institution like UF has five). Each has different incentives/beliefs. Parents often clueless or non-factor, but allies if AI helps disabilities (e.g., reading issues; redo school prepared). Students push tech but fear plagiarism expulsion; some professors can't get students to use AI. Professors polar: AI maximalists (e.g., use ChatGPT for history paper, then critique errors). Collect stories to speak to everyone. Academia high contention; meet people where they are. Admin fear bias: needs thousands of studies; human bias ignored, AI bias sacrilegious.
Advice: make buckets, talk to people (Zoom calls), gather stories to advocate.
UF has Nvidia supercomputer donated by co-founder Chris Maly. Hooking Perplexity to it: test-time compute experiments (more think time = smarter models, computationally expensive). Distillation experiments (big model trains smaller). Advocate university program teaching frontier tasks like distillation, test-time compute, RL. Huge academic/corporate value: few people know this; gold mine with cluster.
Robots from factories as experts, connected via Starlink/5G irrelevant of factory location. Thought experiment: AGI form factors. Data centers have more compute (hundreds of thousands GPUs). Robots as peripherals/remote puppets (like iRobot). Tension: security (air-gapped, no hijacking 6ft/200lb robot) vs. cheap cloud intelligence (rent GPUs cheaper than portable). Engineering/economic optimization needed. Long-run: ASI chip in 20 years runs locally on robot, no cloud (faster if photonic/thermodynamic/quantum computing; 10W ASI possible but unproven).
Cloud connectivity critical for super intelligent robots first 5-10 years.
Not everyone needs warp drive designer now, but autonomous ASI might make itself compelling/useful. Anton corck (IMF?) theory: human cognition stays ahead of tech. Advanced AI makes users smarter (better writing, thinking; e.g., advanced math/science). Raises floor for heavy users. Cognitive abundance: AI may not need human help for warp drive, but can be useful. Fully autonomous AI/robots run farms, power plants, solar panels (grunt work). Users less excited by chatbots if too smart/autonomous; just ask hard things or 'make it work' (AI explains if wanted, e.g., code, health research = PhD in neuroendocrine systems).
Chatbots too smart for regular people? Historical: Google now billions queries/day. Interactive good info valuable (e.g., Claude explained dog resource guarding as fear, not dominance). Generational: kids grow up with it (speaker with AOL IM). Society acclimates over time. Speaker 48, started internet at 17-18; tech enthusiasm rarer in age group (e.g., Ars Technica comments like luddites). Silent majority smarter (use polls for signal).
Not many people speaker's age deep into tech lately.
David experiences audio issues during the live stream, gets muted temporarily, and resolves it by re-entering the stream and being invited as co-host.
A participant shares they are working at a startup building a social network for parents to solve issues, with high focus on AI and a human component that some hesitate to get from chatbots alone.
- Is this a good idea?
- Best ways to engage parents without relying too heavily on AI?
As humanity advances, social connection becomes critical.
Caution: Parent and family support groups online get extremely tribal because:
- People are emotional about raising children.
- Strong cultural values around children and parenthood.
- Many parents resist feedback due to pride or ego.
- Views on treating children vary across cultures.
Hardest challenge: Balancing different views and values.
Low-hanging fruit for AI:
- Moderation: Automatically intervene if users get heated, explain, ask to calm down.
- Fact-checking: Combat misinformation, e.g., vaccines.
- Historical data shows vaccines eradicated diseases, saved millions of lives.
- Not all vaccines are perfect, but counter claims they are bunk or more dangerous than helpful.
Expect issues with information literacy and moderation, e.g., varied reactions to vaccines.
For AI-centric experience:
- Track personal life and child development using long-form memory.
- Example tools: Claude projects, ChatGPT projects for context like medical cases, books, projects.
App could remember child's age, upload report cards, notes, logs, diaries to:
- Identify next needs.
- Analyze behavior causes.
Caveats:
- Many AI platforms' terms prohibit high-risk areas like child care.
- Need powerful disclaimers, safeguards.
Top of the hour; host ends to take dogs to dog park. Thanks viewers for super intelligence live stream, hopes they gained value, back next week.
Keep David Shapiro in your library
Save the videos and channels worth coming back to, and find them again in one place.
