2025-02-25

The Man Who Invented Prompt Engineering on AI, AGI & Humanoids w/ Richard Socher & Salim Ismail

Richard Socher, inventor of word vectors and prompt engineering and founder/CEO of You.com, joins Peter Diamandis and Salim Ismail to dissect a frantic week in AI: Grok 3's benchmark claims, the meaning (or lack thereof) of AGI, and why open-source models keep closing the gap on closed ones. The panel covers AI accelerating scientific discovery (protein design, antibiotic-resistance research, quantum-computing breakthroughs from Microsoft and Google), Peter's personal longevity stack, and a sharp disagreement between Peter and Richard over whether the AI data-center buildout is excessive (Peter) or justified by Jevons Paradox (Richard). A long robotics segment weighs humanoid robots (Clone, Unitree, 1X's Neo Gamma, Figure AI's Helix) against Salim's case for non-human-shaped machines, followed by a crypto check-in on the Bybit hack and Bitcoin's volatility. Richard closes by explaining why a true agentic 'Jarvis' assistant still doesn't exist and describes You.com's growing base of 50,000+ custom agents.

▶ Watch on YouTube

Topics

Grok 3 launch and benchmarking AI ▶ 2:00

The panel reacts to xAI's Grok 3 launch, built on a massive GPU cluster assembled in 122 days; Richard says benchmarks look strong but notes test-time compute, not raw model size, is becoming the key intelligence dimension.

Defining AGI and the dimensions of intelligence AI ▶ 10:40

Richard argues AGI has no single clean definition — a pragmatic financial one is roughly 80% of digitized work automated — and breaks intelligence into separate dimensions (visual, language, math, social, knowledge, learning efficiency) rather than one IQ-like score.

Open source vs. closed AI models AI ▶ 14:57

Richard and Salim agree open-source AI is winning the way open-source web servers overtook proprietary ones, and predict pure foundation-model companies will end up looking like capital-heavy telcos that can't capture the value they create.

AI accelerating scientific discovery AI ▶ 21:31

Richard (writing a book on AI for science) and Peter discuss breakthroughs like replicating 10 years of antibiotic-resistance research in 48 hours and Insilico Medicine's robotic lab, predicting most future scientific credit will go to AI with a human attached for the Nobel Prize.

Quantum computing breakthroughs Compute ▶ 28:59

The panel reacts to Microsoft's Majorana 1 topological-qubit chip and Satya Nadella's claim of a 'truly meaningful' quantum computer within years, while flagging skepticism (Sabine Hossenfelder) about how fast it will actually scale.

Peter's longevity and health-tracking stack Longevity ▶ 32:16

Peter details his three-part health routine — annual Fountain Life full-body diagnostics, Viome gut-microbiome testing to personalize diet, and OneSkin's peptide skincare — framed around 'health is the new wealth.'

AI data-center buildout: overbuilt or not Energy ▶ 35:20

Peter argues the US is overbuilding AI data-center capacity because training costs keep dropping 10x per generation; Richard disagrees, citing Jevons Paradox — cheaper intelligence just gets used more everywhere, including desalination and other energy-hungry fixes.

OpenAI's talent exodus and Thinking Machines Lab AI ▶ 38:40

The panel discusses Mira Murati's $30B-valued Thinking Machines Lab (staffed with ex-OpenAI researchers like John Schulman) and debates whether such high seed-stage valuations can ever be justified by revenue.

Humanoid robots: Clone, Unitree, and Neo Gamma Robotics ▶ 46:01

Peter showcases Clone's hydraulic-muscle humanoid, Unitree's wheeled four-legged robots, and 1X's Neo Gamma home robot, while Salim pushes back on the humanoid form factor itself, arguing purpose-built machines (like Richard's Machina Labs investment) are more efficient.

Figure AI's Helix software and robot data loops Robotics ▶ 53:05

Peter (an investor in Figure) shows its new in-house Helix AI system, built after Figure split from OpenAI, and the panel discusses how physical-world data collection, echoing Tesla's Autopilot, will be the real bottleneck for humanoid robots.

Bitcoin, crypto security, and the Bybit hack Crypto/Web3 ▶ 56:42

The panel covers the record-breaking Bybit cold-wallet hack, Michael Saylor's latest $2B Bitcoin purchase for Strategy, and why gas fees and custody risk still make crypto far less user-friendly than credit cards.

Agentic AI on You.com and why a personal 'Jarvis' doesn't exist yet AI ▶ 1:05:02

Richard describes You.com's 50,000+ user-built agents automating marketing, journalism, and VC due-diligence workflows, then lists the real blockers to a full personal-assistant AI: consent/recording law, the Microsoft Recall backlash, trust deficits, and the ad-based internet economy's incentive to block agents.

Predictions made

open Richard Socher: Given a couple billion dollars in funding, a focused team could build a digital superintelligence within about 1.5 to 2 years.

· due: 2027 · ▶ watch

“if you were given a couple of billion dollars you'd be able to build a digital super intelligence... probably like year and a half to two years”

Your call:

open Dario Amodei (clip): AI will compress roughly a century's worth of biomedical research into the next 5 to 10 years.

· due: 2035 · ▶ watch

“we're going to see a century worth of biomedical research in the next 5 to 10 years”

Your call:

open Peter Diamandis: Compressed biomedical research this decade could plausibly double human lifespan within the next 10 years.

· due: 2035 · ▶ watch

“it's not unlikely it could double the human lifespan... within the next decade”

Your call:

open Satya Nadella (clip): Microsoft's quantum breakthrough will enable a truly meaningful quantum computer within years rather than decades.

· due: unspecified · ▶ watch

“we believe this breakthrough will allow us to create a truly meaningful quantum computer not in decades but in years”

Your call:

open Elon Musk / Brett Adcock (cited by Peter Diamandis): There will be as many as 10 billion humanoid/general-purpose robots in use by 2040.

· due: 2040 · ▶ watch

“the projection is as many as 10 billion robots by 2040”

Your call:

open Richard Socher: AI will eventually be able to reconcile quantum mechanics with relativity.

· due: unspecified · ▶ watch

“I think AI will be able to figure it out”

Your call:

open Peter Diamandis: Clone's hydraulic humanoid robot will be walking within the next few months.

· due: mid 2025 · ▶ watch

“they hope to have it walking in the next few months”

Your call:

Numbers that matter

Richard Socher has over 200,000 citations Cited as evidence of his standing as one of the top five most-cited AI researchers.
xAI raised $6 billion and built a massive GPU training cluster in 122 days Elon Musk's Grok 3 cluster buildout, discussed as a feat of combined hardware/software execution.
You.com offers access to 40+ AI models Richard describing You.com's federated model-routing approach.
Socher's 2018 protein language model produced proteins 40% different from naturally occurring ones, vs. 3% for Francis Arnold's Nobel-winning directed-evolution method Evidence that the model had learned the underlying 'grammar' of protein folding.
Peter uploads about 200GB of personal health data per year via Fountain Life Part of his longevity-tracking routine to catch disease at inception.
Mira Murati's Thinking Machines Lab launched at a $30 billion valuation Cited as an example of ballooning seed-stage AI valuations.
Richard's VC fund is already 5x TVPI after about four years Richard citing his own fund's early returns.
Richard invested in Hugging Face (his former students) at a $5 million valuation; it's now worth $4.5 billion Example of a horizontal AI-infrastructure investment that broke out.
Industry projection of up to 10 billion humanoid robots by 2040 Cited by Peter from conversations with Elon Musk and Figure AI's Brett Adcock.
Michael Saylor's Strategy (formerly MicroStrategy) bought another 20,000 Bitcoin for about $2 billion Latest in Saylor's ongoing Bitcoin accumulation strategy.
You.com users have built over 50,000 custom AI agents on the platform Cited as evidence of growing agentic-AI adoption among You.com's business customers.
You.com has universities with roughly 30,000 students live on the platform Richard describing education customers who are rethinking assignments because AI can already complete them.

Worth digging into

🕳️ Richard Socher's book on AI for science

Socher says he's writing a book (working title heard on the recording as something like 'The Ur Machine') on AI-driven scientific discovery, arguing most future breakthroughs will come from AI with a human attached for credit — worth tracking as it develops.

🕳️ Insilico Medicine's robotic drug-discovery lab

Alex Zhavoronkov's lab reportedly runs AI-designed experiments on robotic hardware 100x faster than humans, fusing a 'theoretical world' and a 'physical world' loop — a template for accelerating science beyond pure simulation.

🕳️ Microsoft's Majorana 1 topological qubit

Described as a 'total left field' approach that nobody else was pursuing, with the surprising claim that stability increases as qubit count grows — flagged by skeptics like Sabine Hossenfelder as unproven at scale.

🕳️ Why no agentic 'Jarvis' assistant exists yet

Richard lays out four concrete blockers — consent-to-record law, the Microsoft Recall privacy backlash, trust deficits at AI-first startups, and the ad-based internet's incentive to block agents — a clear roadmap for what has to change before true personal AI assistants arrive.