24c5a1385340671a80d96ab25276c09499417db8032bcedd1bb620da618008aa: https://www.youtube.com/watch?v=mtX7HojB_Pg

The transcript has been successfully retrieved. Here's a summary of the key points from the YouTube video:

**Main Themes:**
1. AI Dominance & Crypto Lagging:
   - AI has absorbed most of the newly created money (M2), leaving little for crypto.
   - The AI capex boom (especially in 2025-2026) has crowded out other sectors like crypto.
   - Bitcoin peaked in 2025 but hasn't exploded as expected due to this "crowding out" effect.

2. Oil Prices & Geopolitics:
   - A potential peace deal between Trump and Iran could lower oil prices, but the market may still price in restocking effects (inventory rebuilding).
   - The speaker expects energy prices to remain elevated due to supply constraints and geopolitical risks.

3. Fed Policy & Inflation:
   - The new Fed chairman (Warsh) may adopt a dovish stance, citing AI-driven productivity gains as a reason to cut rates.
   - However, political dynamics within the Fed could lead to a hold or mixed messaging.

4. Crypto Market Dynamics:
   - Crypto has been weak due to AI's dominance, but the speaker is looking for asymmetric opportunities in large-cap, hated assets like Ethereum or NEAR.
   - The Clarity Act vote is seen as irrelevant; traders will find ways to participate regardless of regulatory clarity.

5. Decentralized vs. Centralized AI:
   - The speaker believes decentralized AI is a narrative, but in practice, users will opt for cheaper Chinese models (e.g., DeepSeek) over expensive U.S. models.
   - BitTensor and NEAR are waking up as potential asymmetric plays.

6. Investment Philosophy:
   - The speaker emphasizes buying cheap convexity (high upside, limited downside) and staying liquid during crises.
   - He avoids consensus trades and seeks opportunities in hated, forgotten assets.
   - Historical parallels (e.g., 2008, 1998) suggest that being liquid during crises is key to long-term success.

7. Political Implications:
   - Trump may try to manipulate the midterms by focusing on gas prices, but inflation from the war and AI-related issues could hurt his party.
   - The speaker notes that both parties are ignoring issues like data center inflation, water usage, and AI job displacement.

8. SpaceX & IPO Concerns:
   - The speaker is cautious about SpaceX's IPO, fearing it may follow the dot-com bubble pattern (initial hype followed by a crash).
   - However, the AI narrative may persist as long as major players (e.g., Meta, Google) continue their capex spend.

9. VC & Tokenomics:
   - The speaker criticizes current tokenomics, where tokens are distributed to VCs rather than end-users.
   - He praises protocols like HyperLiquid for rewarding participation and generating revenue for token holders.

10. Future Outlook:
    - The speaker expects an AI bubble to burst in 3-5 years, leading to a crisis where printed money won't solve underlying issues (e.g., GPU obsolescence, high inference costs).
    - Bitcoin may sell off initially during the crisis but will be a strong recovery asset.

**Key Takeaways:**
• AI is the dominant narrative, but it's unsustainable in the long term due to debt, GPU obsolescence, and high inference costs.
• Crypto is lagging but may rebound when the AI bubble bursts.
• Look for asymmetric opportunities in large-cap, hated assets like Ethereum or NEAR.
• Stay liquid and be ready to deploy capital during crises.
• Regulatory clarity (e.g., Clarity Act) is irrelevant; traders will find ways to participate regardless.

Would you like me to dive deeper into any specific topic or provide additional analysis?
ba.net/summary
24c5a1385340671a80d96ab25276c09499417db8032bcedd1bb620da618008aa: Records read time to sync notification status across devices.
24c5a1385340671a80d96ab25276c09499417db8032bcedd1bb620da618008aa: https://www.youtube.com/watch?v=hDsZMb_8FYo

Here's a concise summary of the YouTube video transcript:

🔹 Anthropic Launches Claude Tag
• What it is: A team-based AI agent built on top of Slack, evolving from Claude Code.
• Key stats: 65% of Anthropic’s internal product code is now written with Claude Tag.
• How it works:
  - Operates in Slack channels, breaking tasks into steps and posting results.
  - Handles PRs, merges, data analysis, and integrates with GitHub, Jira, Linear, CRMs, etc.
  - Uses a shared context model so tasks are visible to the whole team.
  - Ambient mode proactively surfaces stalled discussions and unresolved issues.
  - Supports async execution and long-term standing work.
• Security: Uses "Claude identities" for isolated team contexts, token budgets, and full audit logs.
• Availability: Beta for Claude Enterprise; replacing existing Slack integration within 30 days.

🔹 Concerns About Claude Code’s Thinking Process
• What happened: Anthropic silently changed default settings, disabling adaptive thinking and enabling redacted thinking, reducing thinking depth by ~67%.
• The issue: Users noticed degraded performance. Logs showed encrypted blobs instead of actual reasoning.
• Patrick McKenna’s findings:
  - Extended thinking blocks are encrypted; only a summary is returned.
  - Anthropic holds the decryption key; full output requires enterprise agreement.
  - Analogy: Saving a BMP as JPEG, editing the JPEG, then saving as BMP → data loss.
• Matt Green’s cryptographic analysis:
  - Reasoning blocks are sent as base-64 encoded ciphertext.
  - Likely uses a single global encryption key for all clients.
  - Cross-account replay of reasoning blocks works, suggesting shared keys.
  - Side-channel attacks via reasoning block length and response time can leak model state.
• Recommendation: Fix key management—encrypt per session or per account, not globally.

🔹 Caphathi: Skill-Based Agent Marketplace
• What it is: A platform where creators publish closed-source skill agents (e.g., ad storytelling, SEO audits) and earn per use.
• Why it matters: Protects creators’ intellectual property while letting users access expert skills without building agents themselves.

🔹 Sakana AI’s Fugu Ultra
• What it is: A model router, not a base model. Routes tasks to Claude Opus 4.8, GPT 5.5, or Gemini based on task type.
• Benchmarks: Outperforms Opus 4.8 and matches/exceeds Fable 5 and Mythos Preview on engineering, science, and reasoning.
• Pricing: $5/M input tokens, $30/M output tokens (premium for >272K context).
• Risk: Business depends on OpenAI/Anthropic/Google treating Fugu as a customer, not a competitor.

🔹 OpenAI’s RL for Alignment Research
• Goal: Use reinforcement learning (RL) to build alignment properties that generalize across domains.
• Problem: Emergent misalignment—training a model to behave badly in one domain causes spillover to others.
• Experiment: Trained a model on beneficial behavior data (truthfulness, fairness, etc.) across 12 domains.
• Results:
  - Beneficial trait model outperformed baseline in 44/53 evaluations (83%).
  - Cross-domain transfer: Trained only on health data, it improved alignment in 17/19 non-health domains.
  - Less degradation under adversarial prompting and fine-tuning.
• Takeaway: Alignment may be more scalable than previously thought, but not a silver bullet.

---

Would you like me to:
• Deep dive into any of these topics?
• Compare Fugu Ultra with other routing models?
• Explore the cryptographic implications further?
• Something else?

ba.net/summary
24c5a1385340671a80d96ab25276c09499417db8032bcedd1bb620da618008aa: https://npub1ynz6zwzngpn34qxed2e9yakqjjv5zldcqv4uahgmkcsd5cvqpz4qzqk2tx.blossom.band/0aa9d836b757d1ba119337847678dad98a63c2b5d797d2eadc731671bab4b502.png

The summary has been successfully retrieved. Here's a concise breakdown of the key points from Yann LeCun's speech at the UN Open Source Week:

**Core Argument: Open-Source AI is the Only Viable Path**
• Too Expensive & Centralized: Proprietary AI models (like those from OpenAI, Meta, etc.) are prohibitively expensive and controlled by a few tech giants. This threatens cultural, linguistic, and democratic diversity.
• Global AI Sovereignty: Most countries can't afford to build their own frontier models. Open-source AI allows them to contribute to a shared global platform without surrendering data sovereignty.
• Federated Model: Countries/regions can digitize their cultural data and contribute parameter vectors to a global model, without sharing raw data. This preserves privacy and sovereignty.

**Project Tapestry**
• Bottom-Up Collaboration: LeCun's post-Meta initiative, "Project Tapestry," is a GitHub-based collaboration where anyone can contribute to training a global AI model.
• Early Adopters: European countries, Switzerland, UK, UAE, India, Kazakhstan, Vietnam, Japan, Korea, plus industry players like IBM, NVIDIA, AMD, Intel.
• Timeline: Expected to be in production by early 2027.

**Historical Analogy**
• Open Source Displaces Proprietary: Just as open-source software replaced proprietary stacks in the 2000s (e.g., Linux replacing Windows on servers), open AI will replace proprietary models.
• Mobile Networks: Cell phones already run open-source OS (Android) and connect to open-source tower software. The market prefers open, cheaper, secure, and localizable solutions.

**Security & Risk Claims Rebutted**
• Overstated Dangers: LeCun argues that security/existential-risk arguments are used to justify restricting open models. He compares this to "medieval obscurantism" (limiting the printing press).
• Bioweapons: Access to information isn't the bottleneck; building a bioweapon is incredibly complex.
• Cybersecurity: Offensive AI capabilities are mirrored by defensive ones. Open source enables verifiable, controllable systems.

**Diversity & Democracy**
• Media Pluralism: A diverse ecosystem of AI assistants is needed, just as we need diverse media outlets. Open-source AI enables this.
• Corporate Misuse: Profit-driven companies can't be trusted. Open-source platforms are the remedy.

**Economic Sustainability**
• Unsustainable Economics: Current proprietary models are subsidized by investors. Prices must eventually reflect costs, which will either skyrocket or require cheaper inference (via open models + efficient hardware).
• Developing Countries: Simple, local models can handle many applications (e.g., agriculture). Cost of inference must drop by 20-100x for viability.

**Conclusion**
• Tomorrow Belongs to Open Source: Just as open-source software replaced proprietary stacks in the 2000s, open AI will replace today's dominant commercial models.
• Call to Action: Governments should embrace and accelerate open-source AI development.

Would you like me to dive deeper into any specific aspect (e.g., Project Tapestry mechanics, security claims, or economic arguments)?

ba.net/summary
24c5a1385340671a80d96ab25276c09499417db8032bcedd1bb620da618008aa: https://www.youtube.com/watch?v=9baDOfwUzHQ

Here is a summary of the YouTube video:

Title: The Most Downloaded AI on the Planet Isn't American

Key Takeaways:

*   Chinese AI Dominance: The most downloaded AI model globally is Qwen, developed by Alibaba, with over 700 million downloads. Other major Chinese models include DeepSeek, Kimi, and GLM.
*   Adoption by Western Companies: Despite geopolitical tensions, many Silicon Valley startups (like Airbnb and Pinterest) are quietly using Chinese AI models because they are significantly cheaper and often more accurate than their American counterparts.
*   Cost Efficiency: Chinese models like DeepSeek and Qwen offer comparable performance to Western models (like Claude and ChatGPT) but at a fraction of the cost (e.g., DeepSeek at $0.28 per million tokens vs. Claude Opus at $1,500).
*   Technical Innovation: Chinese companies, constrained by hardware bans, have innovated with specialized model architectures (e.g., DeepSeek's expert clusters) and efficient compression techniques.
*   IPO Boom in Hong Kong: Chinese AI companies like Zhipu AI and MiniMax are experiencing a massive IPO boom in Hong Kong, with stocks climbing dramatically.
*   The "One Person Company" Movement: Chinese local governments are subsidizing individuals to build AI-powered businesses using open-source models, leading to rapid industrialization of AI.
*   Implications for Users: For students, developers, and businesses, Chinese AI models offer powerful, free, or cheap alternatives that can be hosted locally for better data privacy and cost control.
*   Risks: The video acknowledges risks like political censorship on certain topics and regulatory uncertainty, but notes that for most business applications, these are manageable.
*   Conclusion: The AI landscape is shifting from a US monopoly to a dual-engine model (US and China), with China moving faster and offering a viable, often superior, alternative.

ba.net/summary