ai strategy

Bilingual

The Wizard With the Very Defensible Pond
2 months ago
- A wizard believed his small pond and proprietary data (frog counts, sentiment, outcomes) gave him a defensible advantage over competitors.
- A sorcerer demonstrated that a general-purpose AI ('goblin') could match or surpass the wizard's specialized pond intelligence using publicly available knowledge.
- The wizard's proprietary data included historical usage and outcomes, but was incomplete, inconsistent, and lacked a clear strategic purpose.
- The sorcerer highlighted that acting on proprietary data reveals strategy, allowing competitors to cheaply copy successful moves.
- The wizard's apprentice, who understood the pond's purpose and judgment, was hired by the sorcerer over the AI because AI lacks decision-making ability.
- The wizard ultimately replaced his apprentice with a goblin AI but realized he had never defined what 'optimizing' the pond meant.
- The key lesson is that data without a clear goal is insufficient; judgment and understanding of purpose are essential.
Wenfeng Liang: Four-Hour Investor Meeting Transcript
2 months ago
- DeepSeek prioritizes AGI development over product-driven profit maximization, viewing products as stepping stones.
- The company focuses on a single main thread—AGI—excluding areas like 3D/video generation and world models.
- AGI path involves stages: CoT, Agent, continuous learning, AI self-iteration, and finally embodied intelligence.
- DeepSeek adopts restrained commercialization, taking only reasonable profits and prioritizing cost efficiency.
- Open source is a strategic choice for team cohesion, societal benefit, and higher AGI success probability.
- Liang believes the gap in AI talent is minimal—domestic talent is sufficient, and resource efficiency matters more.
- Model competition hinges on cost, time, and user experience, with cost being paramount.
- DeepSeek has no ambition to become a super app like ByteDance or Tencent, focusing instead on technical breakthroughs.
- Team stability is core and protected through non-confrontational strategies and a vision-driven culture.
- The company's vision is practiced through actions, not slogans, emphasizing restraint as a path to AGI.
Chinese AI Open Weights Grow Safer as American AI Becomes Dead Weight
2 months ago
- Daniel Miessler published a post on July 18 claiming Chinese AI open-weight models are a 'CCP mousetrap' to undermine U.S. AI dominance and influence Taiwan.
- Moonshot's Kimi K3 model is seen as competitive with leading U.S. models like Opus 4.8, challenging previous assumptions about Chinese AI capabilities.
- Miessler's argument relies on predatory subsidy claims, but Moonshot's pricing for K3 aligns with Western rates, contradicting the subsidy narrative.
- Open weights make models permanently available once released, preventing any 'snap' or recall, which undermines the mousetrap analogy.
- Chinese AI growth is driven by industrial policy like the AI Plus directive, fostering open-source ecosystems and market-based competition, not covert plots.
- Claims linking AI developments to Taiwan's reunification or U.S. market selloffs lack evidence and ignore broader economic and political factors.
Kimi K3, Qwen 3.8, and Anthropic's (Potential) Unravelling
2 months ago
- Two SOTA foundation models, Kimi K3 and Qwen 3.8, were launched, challenging top-tier models and proving open models can reach the frontier.
- Foundation models are expensive due to research, compute, and electricity; post-training, inference costs (mainly compute and electricity) dominate.
- Companies vary in infrastructure ownership: leasing data centers (e.g., Anthropic), building data centers (e.g., Meta, Alibaba), or owning power plants and data centers (e.g., SpaceX), affecting cost structures and margins.
- Model-only providers face high risk as success depends on being the best or cheapest, requiring constant innovation or unique products to avoid commoditization.
- Anthropic’s focus on regulation and self-improvement, along with high costs, puts it at risk compared to OpenAI’s product moats and infrastructure investments, increasing unbundling vulnerability.
American AI is locked down and proprietary. It's losing
2 months ago
- US AI strategy relies on proprietary, locked-down models, while China adopts an open-weights approach.
- AI models lack strong moats; switching between them is easy via APIs, making enterprise services the real differentiator.
- China's open models exploit US export controls by turning compute disadvantages into distribution advantages and commoditizing the AI layer.
- The performance gap between US frontier models and open models is closing, with Chinese models claiming comparable capabilities at lower costs.
- US companies prioritize short-term profits and face misaligned incentives, risking economic fallout if AI spending declines due to ecosystem weaknesses.
- Open technologies foster permissionless innovation and ecosystem benefits, contrasting with locked-down practices that limit global adoption.
A Script for Mark Zuckerberg
2 months ago
- Meta CEO Mark Zuckerberg (in a fictional 2026 earnings call) reflects on company history, mistakes, and the strategic importance of AI investment.
- Key historical lessons: human curiosity drove Facebook's early growth; the feed increased engagement; mobile expanded access; Instagram evolved with technology.
- Mistakes include obsessing over becoming a platform (delaying mobile native experience), over-investing in Reality Labs (VR), and underestimating Facebook's shift to entertainment (a blindspot TikTok exploited).
- Biggest failing as CEO was undervaluing Meta's ad business; ads connect entrepreneurs to consumers uniquely, unlike search-based ad models.
- AI is critical because it addresses existential risks to digital businesses, enhances ad targeting and content recommendations, and expands monetizable inventory (e.g., every pixel).
- AI investment leverages Meta's strength in entertainment and human connection, not productivity tools; it may also enable new hardware for AI access beyond phones.
- Aggressive compute (GPU/data center) investment is necessary; Meta will rent out excess compute short-term to fund build-out and use hurdle rates to discipline decision-making.
- Meta's focus: using compute to enhance human connection, entertainment, and commerce—not competing in chatbots or productivity subscriptions—ensuring AI investment pays off via ads.
Al Vigier: Canada's AI strategy shouldn't include secret Palantir bills
3 months ago
- The Canadian government's 'AI for All' strategy aims to become a strategic anchor customer for domestic AI to boost adoption, targeting 60% of businesses using AI by 2034.
- Despite the strategy, the government currently buys AI secretly from foreign vendors like Palantir, with undisclosed contracts worth millions in defense and policing.
- The strategy focuses on equity investments, compute funding, certification, and health missions, but lacks clear procurement orders to buy openly from Canadian firms.
- Equity stakes in companies may deter private investment and create state-dependent firms, rather than fostering competitive market players.
- Starting the missions program in health, a slow and risk-averse sector, may lead to delays rather than proving effective public AI procurement.
- A true anchor customer should buy openly from Canadian firms, set quotas for domestic spending, ensure auditability, and publish receipts quarterly.
- The author, founder of a Canadian AI company, urges the government to transparently buy domestic AI, as it does for foreign products, to support sovereignty.

ai strategy

The Wizard With the Very Defensible Pond

Wenfeng Liang: Four-Hour Investor Meeting Transcript

Chinese AI Open Weights Grow Safer as American AI Becomes Dead Weight

Kimi K3, Qwen 3.8, and Anthropic's (Potential) Unravelling

American AI is locked down and proprietary. It's losing

A Script for Mark Zuckerberg

Al Vigier: Canada's AI strategy shouldn't include secret Palantir bills