LinkedIn Co-Founder Opens Up on the Reality of AI Job Loss | EP #194
Reid Hoffman joins the Moonshots panel to unpack Erik Brynjolfsson's data on falling entry-level employment in AI-exposed fields, arguing job transformation (not permanent mass unemployment) is coming but on a jarringly compressed timeline. The group ranges across AI consciousness (Mustafa Suleyman's caution paper), the Hinton-vs-LeCun fight over whether superintelligence leaves any jobs at all, and a wave of infrastructure and funding news (OpenAI's India/Greece expansions, the Oracle-OpenAI $60B/year deal, Anthropic's $13B Series F, Broadcom/Nvidia ASIC competition). It closes with robotics and transportation stories - Optimus, autonomous surgical robots, and Zoox's robotaxi launch - and Hoffman's framing that computational/prompting literacy is becoming the core skill of the AI era.
Entry-level job loss and the transformation thesisEconomy▶ 3:32
Reid Hoffman responds to Erik Brynjolfsson's paper showing entry-level employment down in AI-exposed fields, arguing it's real job transformation happening on a much faster timeline than history's precedents (industrial revolution), disproportionately hitting new graduates.
Computational thinking as the new literacyAI▶ 8:11
Hoffman describes how heavy AI use changes cognition itself - using AI to write the prompt that will get the best answer, and expects individual contributors to be replaced by people who deploy suites of agents.
Discussion of Mustafa Suleyman's paper warning against prematurely ascribing consciousness to AI; Hoffman and Alex Wissner-Gross debate how humans over- and under-ascribe consciousness to animals, cars, and AI, and whether new categories of legal personhood are coming.
Interspecies communication and animal AIBiotech▶ 21:02
Salim Ismail reveals an XPRIZE in development with Palmer Luckey for AI-driven interspecies communication; group discusses Sarama (dogs), the Earth Species Project (whales, corvids, primates), and Daniela Rus's team capturing unprecedented humpback whale birth audio.
AI in education and declining test scoresAI▶ 23:52
US reading/math proficiency at historic lows is discussed alongside claims that AI tutoring lets kids learn 5-10x faster than classroom instruction, with the group arguing the credentialing/testing system itself will be rebuilt around AI-driven assessment.
AGI/ASI: Hinton vs. LeCun on whether any jobs surviveAI▶ 29:18
A Geoffrey Hinton clip warns superintelligent AI will eliminate nearly all jobs; Hoffman disagrees, sketching a Star Trek-like abundance future where humans still find purpose through competition, culture, and craft, while Alex Wissner-Gross argues for a multipolar (not singleton) superintelligence outcome.
AI regulatory sandboxes and governanceGeopolitics▶ 44:11
Reaction to Ted Cruz's proposed AI regulatory sandbox bill (HIPAA/FDA waivers); the panel discusses US tort-law fragmentation across 50 states, geographic regulatory arbitrage (India, Africa) for health AI startups, and Hoffman's push for safe-harbor rules enabling a 24/7 AI medical assistant.
OpenAI's planned 1-gigawatt India data center (part of the $500B Stargate project) and its new partnership with Greece are framed as both user-acquisition plays and a scramble for scale compute and energy access; Hoffman notes Microsoft's board-level view of the international AI landscape.
AI chip wars and the economics of fabsCompute▶ 57:26
OpenAI's new 3nm chip production run with Broadcom is analyzed as part of a broader shift toward inference-specific ASICs challenging Nvidia's general-purpose GPUs, with Alex Wissner-Gross citing 'Moore's second law' (fab costs doubling every four years) as the dominant constraint.
Larry Ellison overtakes Elon Musk as the world's wealthiest person on the back of Oracle's $60B/year, five-year OpenAI compute deal (4.5 gigawatts); Anthropic closes a $13B Series F at a $138B valuation with revenue run-rate jumping from $1B to $5B in eight months and shifts toward AWS Trainium 2 chips.
Prediction markets and patent explosionEconomy▶ 1:12:02
Polymarket's US launch prompts discussion of prediction markets as a truth signal and research tool for startups; separately, US computing patents have exploded since ChatGPT's release as AI accelerates both application drafting and the underlying discoveries.
AI agent capability scaling and benchmarksCompute▶ 1:19:14
Replit CEO Amjad Masad's claim that agent task-completion time is 10xing per generation (2 min to 20 min to 200 min) sparks debate over whether AI progress is exponential or 'hyper-exponential' (with Alex Wissner-Gross citing a possible late-2027/2028 capability asymptote) and whether multi-agent architectures are fundamentally different from scaled single models.
Robotics and autonomous transportationRobotics▶ 1:26:41
Coverage of Tesla's Optimus scaling plans, skepticism from iRobot founder Rodney Brooks about home robots by 2035, non-humanoid industrial robots for infrastructure inspection, a Johns Hopkins fully autonomous surgical robot, and Zoox's robotaxi launch in Las Vegas.
Predictions made
openReid Hoffman: The entry-level job of two years from now will look very different from today's entry-level job.
“The entry-level job of 2 years from now will be very different than the entry-level job today.”
Your call:
openReid Hoffman: Within a small number of years, essentially all educational and professional assessment will be conducted by AI, up to PhD-oral-defense-level evaluation.
“Within a small number of years all assessment will essentially be done by AI... the equivalent of being able to do PhD oral level defense.”
Your call:
openAlex Wissner-Gross: If current AI progress fits a hyper-exponential rather than simple exponential curve, there is an effective vertical asymptote in AI capability in late 2027 or early 2028.
“There's almost an effective vertical asymptote in late 2027 or early 2028.”
Your call:
openAlex Wissner-Gross: Multi-agent AI approaches will be absorbed into standard compute-scaling laws, producing transformative discoveries and near-magical AI capability within 2-3 years regardless of underlying architecture.
“We just find ourselves on a hyper exponential and all of this turns into transformative discoveries and almost magical AI on the time scale of 2 to 3 years.”
Your call:
openRodney Brooks (clip, via Dave Blundin): Most households will not have a robot in the home, due to supply-chain constraints rather than lack of technology.
“He said no... the technology will exist but the supply chain won't be there.”
Your call:
openElon Musk (referenced): Tesla will scale Optimus humanoid robot production to 1 million units per year within five years at a roughly $20,000 manufacturing cost, en route to 10 billion Optimus robots by 2040.
“He could imagine doubling the human lifespan in the next 5 to 10 years on the back of AI.”
Your call:
Numbers that matter
Entry-level employment down 16% in AI-exposed fieldsFrom Erik Brynjolfsson's paper discussed at the top of the episode.
India entry-level software jobs down 20-25%Salim Ismail's on-the-ground signal from a recent trip to India.
150 million GitHub users as of May 2025Cited by Peter Diamandis as evidence of programmer growth despite AI coding tools.
Number of programmers worldwide up 50% since 2022; salaries up 24% over 5 yearsUsed to argue AI is expanding, not shrinking, software engineering demand.
35% of 12th graders at/above reading proficiency (down from 40% in 1992); 22% proficient in math; 31% in scienceUS student achievement scores cited as historic lows.
Child learning with AI is 5-10x faster than in a classroom (updated from an earlier 2-6x figure)Statistic Peter Diamandis has cited, revised upward per a Stanford AI conference contact.
OpenAI's planned India data center: 1 gigawatt, ~22% of India's total data center capacity by 2030Part of OpenAI's $500 billion Stargate project.
OpenAI-Oracle deal: ~$60 billion/year for 5 years, 4.5 gigawatts of compute capacity ('two Hoover Dams')The deal that helped push Larry Ellison past Elon Musk as the wealthiest person in the world.
Anthropic Series F: $13 billion raised at a $138 billion valuation; revenue run-rate rose from $1B (January) to $5B (August); 300,000+ businesses with enterprise accountsCited as evidence of the fastest growth curve in tech history.
Amazon invested $4 billion into Anthropic and is building 1.3 gigawatts of AI training data-center capacity on Trainium 2 chipsPart of the broader Anthropic-AWS chip relationship discussion.
6,000 more US computing-related patents granted in 2024 than in 2023Attributed partly to AI accelerating both patent-application drafting and underlying discovery.
Replit coding agent task-completion time 10x'd across three generations: agent 1 (2 min), agent 2 (20 min), agent 3 (200 min)Cited by Amjad Masad and discussed as either evidence of hyper-exponential AI progress or a flawed benchmark.
Optimus generation 3 hand/forearm has 26 actuators; target manufacturing cost $20,000/unit; automotive sales estimated at 74-80% of Tesla's market valuation betFrom Elon Musk's public comments as relayed on the show.
Robot inspection/maintenance market estimated at $6.7 billion today, growing ~13%/year to $12.5 billion by 2030Cited from a tweet about non-humanoid industrial robots used in wind turbines, pipelines, and power lines.
Johns Hopkins autonomous surgical robot achieved 100% accuracy performing a gallbladder removal without human controlTrained via imitation learning on videos of human surgeons; contrasted with the human-operated da Vinci system.
Worth digging into
🕳️ Erik Brynjolfsson's entry-level job loss paper
The 16% entry-level employment decline in AI-exposed fields is the empirical anchor for the whole episode's jobs debate, but the panel only discusses charts secondhand.
🕳️ Mustafa Suleyman's 'seemingly conscious AI' paper
Both Hoffman and the panel treat this as an important warning shot but only summarize it from memory.
🕳️ The Universe 25 rodent-utopia experiment
Diamandis uses this 1960s NYU/John Calhoun experiment as a cautionary analogy for a post-scarcity AI society, but the historical experiment's scientific validity and relevance to human societies is contested.
🕳️ Johns Hopkins autonomous surgical robot
A robot achieving 100% accuracy on an unassisted gallbladder removal, trained via imitation learning, is a concrete near-term proof point for the panel's claim that AI will produce the world's best surgeons.
🕳️ Earth Species Project and Sarama - AI animal communication
Two different real efforts (Earth Species Project for whales/corvids/primates, Sarama for dogs) plus Daniela Rus's unreleased humpback whale birth audio point to a fast-moving but under-covered application of AI.
🕳️ Replit's Meter benchmark dispute
Amjad Masad's claim that agent task-completion time is 10xing per generation, and Alex Wissner-Gross's counter that the exponential-vs-hyperexponential fit could point to a 2027/2028 capability asymptote, is a live, checkable disagreement about AI's growth curve.