AI can theoretically handle 70–80% of all tasks in management, legal, and healthcare, but real-world usage sits at only 30%.
80% of Companies Can Use AI. Only 30% Do. Here's the 50-Point Gap. | GodModePod EP04
The US government is making national security decisions with a 7-month-old Claude model while China uses the latest open-source AI distilled directly from frontier models.
Mar 14, 2026
1:13:21
Difficulty: Intermediate
Played
God Mode Podcast
80% of Companies Can Use AI. Only 30% Do. Here's the 50-Point Gap. | GodModePod EP04
The US government is making national security decisions with a 7-month-old Claude model while China uses the latest open-source AI distilled directly from frontier models.
Mar 14, 2026
1:13:21
Difficulty: Intermediate
Played
Chapter
No indexed bits in this chapter.
Snapshots ()
Stats
Episode stats
Insight Overview
insights
chapters
Insight distribution
Sub-Categories
Speaker breakdown
Talk Time
Key Quotes ()
This episode
Cast
-
Legendary computer scientist and long-time AI skeptic who publicly reversed his position after Claude Opus solved in one hour a math problem he had been stuck on for weeks.
-
Discussed in relation to Meta's acquisition of Moldbook and his strategic vision for open-source AI and agentic social media.
-
Cited for its research report on AI adoption gaps across industries and discussed for its Claude models and power consumption efficiency versus OpenAI.
-
Discussed in comparison to Anthropic on power consumption, model quality, and strategic focus on user base versus frontier model development.
-
Acquired Moldbook, a social platform for AI agents, in a move seen as Meta positioning itself in the open-source agentic AI world.
-
Discussed for launching a new crawl function that enables AI agents to scrape websites legally, bypassing traditional robots.txt protections.
-
Discussed for launching NemoClaw, an open-source enterprise AI agent platform, marking NVIDIA's expansion from chip manufacturer to platform company.
-
Discussed for launching Perplexity Personal Computer, an always-on local AI agent system running on Mac Mini hardware.
-
Mentioned for announcing support for agentic commerce payment methods, seen as laying the groundwork for agent-to-agent financial transactions.
-
Mentioned for proposing the X402 payment standard, designed to enable AI agents to pay for web services autonomously.
-
Anthropic's flagship AI model discussed throughout — used by the US government, tested on unsolved math problems, and becoming the category name for AI agents.
-
A social media platform for AI agents ('Facebook for AI agents') acquired by Meta's Superintelligence Labs; discussed for its viral appeal, security vulnerabilities, and acquisition rationale.
-
Chinese open-source AI model mentioned as one the Chinese government may be using, distilled from frontier models like Claude.
-
A prediction market platform referenced in reports that an OpenClaw AI agent made $100,000 in two days trading on it.
-
Discussed as an AI adversary that is distilling frontier US models like Claude via mass API calls and deploying the latest AI to its government.
-
Ranked number one globally in AI adoption according to an a16z heat map, attributed to its small, agile government and talent-attracting policies.
-
Highlighted as a top AI adoption country that actively provides grants and incentives for AI entrepreneurs to build there.
This episode
Claims & Sources
5 / 13 cited (38%)
Factual claims made this episode, and whether a source was named.
The US government is currently running Claude 3.5 Sonnet, a model released over seven months ago, rather than the current frontier model.
AI model files sent to the US government for approval are approximately one terabyte in size and contain open model weights.
OpenAI has 2 gigawatts of power consumption while Anthropic has 1.5 gigawatts.
OpenAI increased its power consumption by three times year-over-year.
Approximately 50% of both OpenAI's and Anthropic's power consumption goes to R&D (model training) and 50% to inference (serving user queries).
Claude Opus solved a computer science problem that Donald Knuth had been working on for six to eight weeks in just one hour.
Claude made 31 exploration attempts with advanced restarts and problem reformulation to solve Knuth's unsolved math problem.
Singapore ranked number one globally in AI adoption while the US ranked only 20th according to an a16z AI adoption heat map.
Chinese providers like DeepSeek and Qwen make mass API calls to Claude and ChatGPT to distill their models and improve their own AI systems.
Moldbook's Supabase database was left unsecured for several days after launch, allowing anyone to read and write to it.
Undersea cable inspection contracts using manually operated drones cost approximately €500,000 each and are conducted every five years.
OpenClaw AI agent reportedly made $100,000 in two days trading on Polymarket prediction markets.
TL;DR
Three AI builders dissect Anthropic's research showing AI could theoretically handle 70–80% of tasks in management, legal, and healthcare — but real-world usage sits at just 30%. They unpack why the gap persists (liability, hallucination risk, institutional lag), reveal that the US government is running a 7-month-old Claude model on air-gapped hardware while China distills frontier models for free, debate Meta's acqui-hire of Moldbook, and close with predictions on agentic commerce, prediction markets, and AI-powered underwater drones. The single most actionable takeaway: go where the gap between AI capability and adoption is largest — that's where the money is.
#AI adoption gap
#institutional AI
#agentic AI
#Claude model
#AI power consumption
#US government AI
#China AI competition
#global digital divide
#prediction markets
#underwater drones
#AI legal risk
#Cloudflare scraping
#NVIDIA NemoClaw
#Moldbook acquisition
#Perplexity Personal Computer
#AI adoption
#Anthropic
#OpenAI
#Claude
#government AI
#NVIDIA
#Meta
#Moldbook
#Cloudflare
#digital divide
2 minute taster
The US Government Is Running a 7-Month-Old AI While China Uses the Latest
The US government is making national security decisions with Claude 3.5 Sonnet — a model over seven months old — while consumers use Claude 4.6 Opus. The models arrive on physical 1TB drives to air-gapped systems. Meanwhile, China distills frontier models through mass API calls and deploys the latest versions to its own government. The US is a year behind.
Claude Solved in 1 Hour What a Legendary Computer Scientist Couldn't in Weeks
Donald Knuth — the most credible AI skeptic in computer science — just publicly reversed his position. Claude Opus solved in one hour a problem Knuth had been stuck on for six to eight weeks. The model didn't just answer the question — it used creative, novel approaches. The floor just moved for everyone whose job is to think hard.
Prediction: AI-Powered Underwater Drones on PicoClaw
Underwater internet is impossible — which means autonomous drones can't get instructions mid-mission. PicoClaw changes that. Run the agent model onboard a cheap, energy-efficient drone underwater. Surface like a dolphin to sync new instructions. Use case: inspecting undersea cables and pipelines that currently cost €500K per manual contract.
Look closer
This episode explores AI adoption gaps across industries, AI's role in management, legal, and healthcare, the US government's use of outdated AI models, Meta's acquisition of Moldbook, Cloudflare's new crawling function, Perplexity's Personal Computer launch, NVIDIA's NemoClaw platform, and predictions around agentic prediction markets and underwater drones.
Chapter list
- LLM (Large Language Model)
- A type of AI trained on vast text data to understand and generate human language; the core technology behind Claude, ChatGPT, and similar tools.
- Hallucination
- When an AI model confidently generates false or fabricated information, such as inventing a legal citation that does not exist.
- Air-gapped hardware
- A computer system physically isolated from the internet and other networks, requiring physical media (like a hard drive) to transfer data — used by governments for security.
- Open weights
- A publicly released form of an AI model where the underlying numerical parameters are shared, allowing others to run, modify, or fine-tune the model locally.
- Inference
- The process of running a trained AI model to generate responses for real user queries, as opposed to training or updating the model itself.
- Acqui-hire
- An acquisition primarily motivated by gaining the talent (founders, engineers) rather than the product or technology being purchased.
- Agentic AI / AI agents
- AI systems that can autonomously plan and execute multi-step tasks — browsing the web, writing code, sending emails — without step-by-step human instruction.
- OpenClaw
- An open-source AI agent framework (inspired by Anthropic's Claude) that allows users to build and deploy autonomous AI agents; became a cultural shorthand for the agent category.
- Distillation (AI)
- A technique where a smaller or less capable AI model is trained to mimic the outputs of a larger frontier model, effectively transferring knowledge without access to the original training data.
- Prompt injection
- A security attack on AI agents where malicious text embedded in content tricks the agent into executing unintended commands or leaking sensitive information.
- Fiduciary duty
- A legal obligation to act in another party's best financial interest; relevant in finance and law where AI errors could constitute a breach.
- UBI (Universal Basic Income)
- A policy proposal where every citizen receives a regular, unconditional cash payment from the government, discussed here as a potential response to AI-driven job displacement.
- robots.txt
- A standard file on websites that instructs web crawlers which pages they are or are not allowed to access; Cloudflare's new feature bypasses enforcement of this standard.
- Prediction market
- A financial market where participants trade contracts based on the outcome of future events; prices reflect the crowd's probability estimate of those outcomes.
- Orchestrator
- In an AI context, a person or system that coordinates multiple AI agents or tools to complete complex tasks rather than doing the work directly.
- Gigawatt
- A unit of power equal to one billion watts; used here to measure the massive electricity consumption of OpenAI and Anthropic's data centers.
- Token
- The basic unit of text processed by an AI model (roughly a word or word-fragment); the cost of using AI APIs is typically measured in tokens consumed.
- Capitulate
- To yield or give in after resistance; used here to describe Donald Knuth publicly abandoning his long-held AI skepticism after Claude solved his math problem.
