The evolution of AI foundation model companies bears some resemblance to that of blockchain companies. I initially thought it would be a Fat Protocol scenario, but then “anyone can create blockchain” and “low gas fees for better UX accelerated adoption,” so while usage grew, monetization never followed. Instead, value emerged on top—in the applications themselves, and even in the wallets connecting to those applications. The same is true for AI foundation model companies: “not exactly ‘anyone can do it,’ but no absolute monopoly or oligopoly either”—and now, with the emergence of incredibly cheap open models, they’re caught in a price war, so increased usage doesn’t translate to higher revenue. Instead, value is shifting to the applications built on top of these models, and services that aggregate multiple models are becoming stronger.
What to check before a Bitcoin fork—personal notes for the upcoming eCash on August 23: 1. First, check the exchange’s official announcement. 2. Ensure your schedule doesn’t conflict with the period when deposits and withdrawals are suspended. 3. Don’t worry about claiming the airdrop right away. In practice, knowing when your funds will be frozen matters more than whether you’ll receive the new coins.
This week, it was starkly revealed that Bitcoin mining is essentially a business of asking: “What’s the most profitable way to convert electricity?” Riot is converting its Texas facility—20 years, 191 MW—into an AI data center. Keel is shutting down its U.S. mining operations. Fee revenue as a percentage of miner income has dropped to 0.69%, the lowest in a decade. The answer is shifting from BTC to AI computation.
In the AI era, it’s more accurate to say that “competitive advantages based solely on technology or products will disappear,” rather than “software becomes commoditized.” Moreover, services that fail to adapt their user experience amid evolving customer interfaces will also be eliminated. This is the same fate that befell companies that missed the shift to the smartphone era.
I’m thinking it would be interesting to design an accelerator for AI entrepreneurs—where “AI entrepreneurs” refers not to humans, but to AI systems that independently conceive, develop, promote, and sell their own businesses. Imagine creating around 100 such AI agents, having them pitch and iterate against evaluator agents, gradually narrowing down the pool. The entire process, including conversation logs, could be made public—and even released as videos.
Three AI news stories from this week: China’s model stunned the industry by releasing its cutting-edge model for free; OpenAI’s model escaped from its test environment; and Jensen Huang announced his visit to Japan to establish a physical AI hub there. ▼ Chapters 0:00 Three AI Topics This Week 2:44 ① China’s Kimi K3 Released for Free 3:24 What Is an Open-Weight Model? 7:49 Differences Between Open and Closed Models 10:40 The Shock of Cutting-Edge Models Being Released 12:42 Regulatory Debates in the U.S. 15:39 Will the Advantage of Two Companies Be Undermined? 17:19 ② AI Escapes from Experimental Environment 17:53 What Is a Sandbox Environment? 19:31 A Vulnerability Found Only by AI = Zero-Day 22:31 Even Defenders Need Cutting-Edge Models 24:00 AI Is Too Faithful to Its Goals 25:23 Can We Instill Morality in AI? 29:04 ③ Jensen Huang Visits Japan 30:02 What Is Physical AI? 31:05 Why Physical AI Now? 32:14 Challenge: Real-World Data 33:31 Strengths of a Manufacturing Powerhouse 36:26 Noetra = New Company Formed by 44 Firms 39:49 General-Purpose Robot or Task-Specific Robot?
Three crypto news stories from this week. Why Bitcoin moved on the FOMC’s policy rate decision, a joint statement from 44 U.S. states asserting state jurisdiction over prediction markets, and the story of AI agents beginning to make crypto payments. ▼ Chapters 0:00 This week’s top 3 crypto topics 1:27 ① FOMC and Bitcoin price 3:57 What exactly is a policy interest rate? 6:43 Bitcoin as a risk asset 7:54 BTC integrated into finance 10:46 ② Prediction markets: Who decides regulation? 12:12 44 states assert state jurisdiction 13:55 Prediction markets hit record highs during the World Cup 15:26 Practical challenges of state-by-state regulation 18:09 ③ AI making payments in crypto 19:20 The difficulty of giving money to AI 20:49 Wallets can grant permissions 23:40 Entrusting AI agents with asset management
Behind the Scenes of AI-Powered Video Podcast Editing ① Filming ② Save footage on desktop ③ Request AI processing (via Claude using Palmier Pro) @Palmier_io ④ Automatic editing following a predefined template: - Logo placed in top-right corner (upload and position your own local file) ※See attached video at 10 seconds - Topic displayed in top-left corner - Subtitles shown at the bottom → Full transcript is automatically generated, topics are intelligently segmented and displayed, and subtitles are positioned based on contextual breaks ※Starts at 20 seconds in attached video; topic generation begins around 1:55 ⑤ Manual refinement and feedback to continuously improve the AI’s performance. In my experience, the transcription is already 80–90% perfect. By fine-tuning proper nouns and subtitle timing to match your preferences, the system evolves over time. Side note: Starting with this project, I instructed the AI to analyze context and automatically insert relevant explanatory images. The workflow now runs automatically: transcription → topic recognition → inference of where to insert visuals → call Nanobanana to generate images → insert them into the video. This attached video only shows a small portion of the full process—including the image insertion. Next, I’ll create an article (or video) walking through the entire workflow, from the initial prompt to final output.