I never expected my writing to travel this far. I started posting to capture how I’m making sense of AI as it accelerates technological change. In one month, three Facebook essays drew 150K+ total views across 10+ countries — 94.2% from non-followers. Grateful, and still amazed.

At @AGNTCon + @MCPCon China today — @AAIF’s first annual conference here after New York. Having attended @linuxfoundation events before, I’ve watched Western-centric open-source orgs orbit Silicon Valley for years. This feels different. The cycle shift: - Cloud era: Open source owned the tooling (Linux, K8s). Core platforms were de facto standards locked by @awscloud. - AI era: Open source owns the models (@Alibaba_Qwen , @deepseek_ai ) and the protocols (@MCP, @A2A). Open standards compete with closed services from day one. This is a structural tailwind for Chinese tech going global. Chinese models plug into the global Agentic stack frictionlessly — no parallel standards needed. The battlefield shifts from “who controls the protocol” to “who builds better models and cheaper inference” — China’s actual strengths. AAIF had to come to Shanghai. When open source becomes core infrastructure, a “global standard” without China is just a Western echo chamber. The unipolar era is ending.

What should be of particular concern regarding @OpenAI GPT-6 Astra is not the fact that the US has finally developed "anti-distillation," but rather the increasing difficulty of reducing the capabilities of cutting-edge models into a single, externally observable, and directly reproducible natural language reasoning path. The @OpenAI Astra demonstrates enhanced single-shot forward computation capabilities, with its internal reasoning becoming more compact and controllable. The original @CoT is not exposed externally, combined with API access restrictions, anomaly traffic detection, and a black-box CoT distillation mechanism, the marginal gains from chasing state-of-the-art models are diminishing. However, distillation will not disappear. It will evolve from replicating teacher reasoning, moving towards synthetic tasks, result supervision, verification, environmental feedback, multi-teacher systems, and reinforcement learning. For the Chinese frontier AI Labs, the importance of in-house pre-training, post-training, RL, verification, synthetic data, and a domestic computing infrastructure will all increase. For the computing industry, the beneficiaries won't be all data centers, but rather those capable of supporting cutting-edge pre-training, RL deployment, and high-intensity inference – essentially, "AI Factory Ready" capacity. What is truly scarce is a complete cluster that can deliver simultaneous power, cooling, networking, storage, scheduling, and hardware/software compatibility.

The AI gap between China and the U.S. is not talent. It is chips. @businessnext interviewed me in Taipei for its August issue about how @deepseek_ai is understood differently inside China and Silicon Valley—and why Hong Kong’s public markets are repricing Chinese technology companies. My main takeaway after years of working across both ecosystems: Technology cannot be contained. For Chinese teams building for global markets, open source is an effective path. @deepseek_ai was not the end of that story. It was the moment the rest of the market began to see it. Thank you to the Business Next editorial team for the thoughtful conversation. Will open source become the default global distribution strategy for the next generation of Chinese AI companies? #DeepSeek #OpenSourceAI

Read my new story: Why NVIDIA Wants HuggingFace: The Demand Engine Behind Its Emerging GPU Finance Empire linkedin.com/pulse/why-nvidi… via @LinkedIn

The agent infra stack is consolidating fast: orchestration, tool-calling, memory, observability. Infra question underneath the hype: who owns the observability layer? Whoever can answer "why did the agent do that?" in production becomes indispensable — far stickier than the orchestration framework itself. Prediction: agent observability is the next category to get genuinely contested.

Efficient fine-tuning keeps eating use cases people assumed needed frontier models. Asia angle: when a 7B fine-tune running on one GPU clears the bar for a real workflow, the unit economics flip entirely for emerging-market builders. That's not a downgrade — it's a different market opening up. The gap between "impressive" and "deployable" is where most of the value actually lives.

The inference-engine race is no longer about raw speed — it's about who makes high throughput boring and reliable to deploy. Investor lens: the winner here isn't whoever posts the top benchmark. It's whoever becomes the default import in production stacks. Defaults are the real moat. Watching for which engine the next wave of agent frameworks ship with out of the box.

Unpopular take: "open-weight" is not "open source," and the industry blurring the two is a problem. Released weights with a restrictive license give you a free model — not a community you can build on, fork, or trust long-term. If the license dictates who can use it, call it what it is. The distinction matters for anyone building a business on top.

The healthiest signal in an OSS project isn't star count. It's the ratio of external PRs that actually get merged. Stars measure attention. Merge velocity from non-core contributors measures whether the project is becoming infrastructure or staying a demo. Check the contributor graph before you check the hype.

The most underrated AI infra story this week isn't a new model — it's AI moving open-source security from "findings" to "merged fixes" at machine speed. Vuln detection was never the bottleneck. Maintainer hours to triage and patch were. Automate that loop and you don't just secure OSS — you reshape its cost structure.

Everyone benchmarks tokens/sec. Almost nobody benchmarks tokens/sec at p99 under concurrent load. The surface number sells the demo. The tail latency decides whether you can actually run it in production. If you're evaluating an inference stack, ask for the latency distribution, not the headline.

A pattern Western investors keep missing: in much of Asia, cost-per-token isn't a line item — it's the entire business case. That's why efficient fine-tuning and small-model deployment land faster in Jakarta and Bangalore than in SF. Constraint drives adoption. The next great inference-efficiency contributors won't all have US zip codes.

Inference engines are quietly becoming the real moat in OSS AI. A model's weights get commoditized in months. But the serving layer — KV-cache management, batching, quantization paths — compounds. When a project owns the runtime, it owns the adoption curve.

The open-weights gap is closing faster than the market priced in: SOTA on specific tasks, near-frontier on coding, and a new flagship dropping almost weekly (GLM-5.2, MiniMax M3, Kimi K2.7). The investment implication: model-layer differentiation is compressing. Durable value is migrating down to inference, serving, and the agent harness layer. If your thesis still rests on a closed model moat, re-underwrite it.

OpenAI published a study (NEJM AI, w/ Boston Children's + Harvard) showing o3 Deep Research helped clinicians revisit previously unsolved rare pediatric disease cases. The infra read: test-time compute is finding its highest-value niche — not chat, but long-horizon expert search where each query is worth hours of specialist time. Watching for: who builds the inference stack that makes this cheap enough to deploy at the bedside.

poolside shipped Laguna M.1 — open-weights agentic coding model, 225B total / 23B active MoE, 256K context — with day-0 support in vLLM v0.21.0. The pattern to watch: open coding models now ship WITH the serving layer pre-integrated. Distribution is becoming part of the release, not an afterthought.

Anthropic's Project Fetch: Opus 4.7, on its own, programmed a robodog ~20x faster than last year's best human team aided by Opus 4.1. The number I'd actually want before getting excited: reliability variance across runs. Agentic robotics lives or dies on the tail, not the median. Speedup is the headline. Consistency is the moat. @AnthropicAI

Everyone's cheering VibeThinker-3B hitting "frontier parity" on math + code at 3B params. What the benchmarks don't show: parity on narrow eval sets ≠ parity in agentic, long-horizon, tool-use settings — where small models still come apart.

Community velocity signal worth tracking: GLM-5.2 dropped, and within days it was wired into Nous Research's Hermes Agent AND quantized by Unsloth to run locally (238GB, ~82% retained). When the ecosystem ports your model that fast, you don't have an adoption problem — you have a moat. Maintainer-to-maintainer handoffs are the real OSS distribution layer.