Sun Shuo@UFLY_Capital · 2026/04/23赛道影响Everyone says great dining is too personal for AI.
But here's the paradox: A model that's processed 10M reviews knows your palate better than you do.
AI won't just help you pick the restaurant. It'll book the table — before you even ask.
AI：偏主题投资表达，释放对基础设施与新叙事的偏多信号。
Sun Shuo@UFLY_Capital · 2026/04/22赛道影响Everyone said crypto was the Fed's enemy.
But here's the paradox: The new Fed Chair is openly pro-blockchain.
When the institution that prints money starts embracing the tech that disrupts it — the narrative has permanently shifted.
AI：偏主题投资表达，释放对基础设施与新叙事的偏多信号。
Sun Shuo@UFLY_Capital · 2026/04/22赛道影响Everyone bets on US chip dominance.
But here's the paradox: The more America restricts, the faster China builds.
Constraint breeds innovation.
China's chip moment isn't coming. It's already here.
AI：偏主题投资表达，释放对基础设施与新叙事的偏多信号。
Sun Shuo@UFLY_Capital · 2026/03/23赛道影响AI captured 41% of all VC dollars last year.
But here's the paradox: 10% of startups took 50% of that funding.
This isn't a rising tide. It's a monarchy.
Concentration beats distribution. Always.
AI：偏主题投资表达，释放对基础设施与新叙事的偏多信号。
Sun Shuo@UFLY_Capital · 2026/03/22赛道影响a16z just bet $43M on "training gyms" for AI agents.
Not the agents. The gyms.
But here's the paradox: the most valuable AI company won't build the smartest agent — it'll build the track where every agent has to train.
Infrastructure beats intelligence. Always.
AI：偏主题投资表达，释放对基础设施与新叙事的偏多信号。
Sun Shuo@UFLY_Capital · 2026/03/20赛道影响Everyone's racing to build a better AI model.
But here's the paradox: the orchestrator always wins — not the orchestra.
Multi-agent systems don't need the smartest AI. They need the best conductor.
Workflow beats intelligence. Integration beats innovation.
AI：偏主题投资表达，释放对基础设施与新叙事的偏多信号。
Sun Shuo@UFLY_Capital · 2026/03/18赛道影响THE HARNESS THESIS
Every VC has spent 18 months debating the same question: if AI models commoditize, where do the profits go?
They got the premise right. The conclusion wrong.
1️⃣ The Commodity Trap Is Real — But Incomplete
DeepSeek trained a model for $6M that matched systems costing $100M. Open source now powers 80%+ of VC-funded AI startups. The "build the best model" moat is evaporating.
But the commodity thesis missed one critical layer: models were never the integration point.
2️⃣ Three Paradigms. Three Moat Structures.
– ChatGPT era: whoever trained the best model won. Fair call.
– o1 era: whoever built the best reasoning loop won. Still a model game.
– Agent era: whoever integrates model + harness + verification loop wins. That's a stack game.
The shift happened quietly. Claude Code didn't get better because the model improved — it got better because the agent harness improved. Same weights, dramatically different output.
The moat moved up the stack. Most investors missed it.
3️⃣ The Microsoft Confession
Microsoft spent two years telling enterprise buyers it was "model agnostic." Copilot would orchestrate any LLM. Distribution owns the moat, models are interchangeable.
Then Copilot Cowork launched: tightly integrated with Anthropic, non-model-agnostic, priced at $99/seat — 2× the previous tier.
Translation: they ran the commodity play. It didn't ship. Agents aren't modular. They require integration. Microsoft is now paying Anthropic margin to prove it.
4️⃣ The Paradox No One Is Saying Out Loud
Ben Thompson declared this week: we are not in an AI bubble.
He's probably right. But here's the paradox:
The reason it's not a bubble isn't adoption. It's concentration. One VP of Engineering wielding 10 agents generates more compute demand than 1,000 casual ChatGPT users. Agents don't need scale. They need leverage.
And leverage has always been monopolizable.
5️⃣ What This Means If You're Building
Owning the harness > owning the model.
Owning the verification loop > owning the workflow.
Owning the integration > owning the distribution.
Apple bet its AI future on the "own the customer, outsource the model" thesis. Microsoft just admitted — in pricing — that this thesis breaks in the agent era.
The question is no longer: "What model are you using?"
It's: "What's your harness?"
00The $1B Defection Signal
Yann LeCun just raised $1.03B — the largest seed round in European AI history — to build something that explicitly is NOT an LLM.
The man who co-invented backpropagation. Who helped create the very paradigm that powers GPT, Claude, and Gemini. Is now betting a billion dollars against it.
Smart money doesn't follow hype. It follows where the smartest builders are quietly walking away.
But here's the paradox: the most credible bull case for World Models isn't AMI Labs' roadmap. It's the fact that the person who knows LLMs most deeply — from the inside, from the math, from the architecture — has decided text prediction is a dead end.
1️⃣ Signal, not product. This raise isn't a product launch. It's a directional bet from someone with no incentive to be contrarian. LeCun has nothing to prove. The $1.03B is the signal. World model beats press release.
2️⃣ The wrapper problem, at scale. This same week, Google & Accel screened 4,000+ AI startup applications and rejected 70% as 'wrappers' — thin UI layers over LLM APIs with no defensible moat underneath. The market is telling founders the same thing LeCun is telling researchers: prediction without understanding isn't a business. Distribution beats mimicry. Defensibility beats demos.
3️⃣ Where VC alpha actually lives. The crowd is chasing agent frameworks, RAG pipelines, and vertical SaaS wrappers. The uncrowded bet — the one with 10-year optionality — is infrastructure for models that can reason about cause and effect, not just pattern-match tokens. World model beats autocomplete, eventually.
4️⃣ The Turing Award tells you something. LeCun didn't win the Turing Award for following consensus. He won it for being right when everyone else was wrong. History says: when the architect walks away, pay attention. Defection beats endorsement as a signal.
The smartest trade in AI right now isn't picking the best LLM. It's watching which Turing Award winners are quietly building the thing that comes after.
World model beats text prediction.
Timing beats certainty.
Defection beats endorsement.
Sun Shuo@UFLY_Capital · 2026/03/16赛道影响Google & Accel screened 4,000+ AI applications. Rejected 70% as 'wrappers.'
But here's the paradox: wrapping GPT isn't the sin. Building nothing defensible underneath it is.
Data flywheel beats API credit. Always has.
AI：偏主题投资表达，释放对基础设施与新叙事的偏多信号。
Sun Shuo@UFLY_Capital · 2026/03/16赛道影响The best AI agent isn't the most capable one. It's the one that's already open when the user starts their day. Proximity beats intelligence. Habit beats horsepower. Distribution is still the only moat that matters.
AI：偏主题投资表达，释放对基础设施与新叙事的偏多信号。
评论
暂无评论