On July 29, one month after leaving NVIDIA, Yangqing Jia — creator of Caffe, core contributor to PyTorch, former Alibaba Cloud VP, and founder of Lepton AI — announced his new company: Intent Lab.
The news alone carries weight: this is the AI researcher who sold Lepton AI to NVIDIA for $700 million just last year. But the interesting part is not “he founded another company.” It is what the company builds.
Intent Lab’s core product is called Fleet — not another coding assistant, but an “AI engineering team.” In Jia’s framing, it turns a user’s intent directly into production-grade software systems.
Three early results are public, and each deserves individual attention:
-
A 534% inference speedup for GLM-5.2: lifting Zhipu’s flagship model from 102 tokens/s to 647 tokens/s — claimed as the fastest inference speed yet recorded for that model.
-
A SQLite-compatible database from one instruction: from the single prompt “build a SQLite-compatible SQL database engine,” Fleet constructed the database from scratch and passed all 6 million acceptance tests. Running entirely on open-source models, it cost about $350.
-
AgentFS, a distributed filesystem for AI agents: designed for agent workloads, with directory state queries 626× faster than AWS EFS.
But the results matter less than how Fleet works — because that is where the underlying paradigm shift in AI coding shows up.
What Fleet Actually Did
Take the GLM-5.2 optimization first. Fleet attacked it from four angles:
-
Kernel level: kernel fusion plus agent-generated PTX/SASS code cut kernel launches, adding 24% inference performance.
-
Runtime level: H2D batching achieved zero-copy steady-state decoding, worth 16%.
-
Communication level: fusing collective communication, residual adds, and RMSNorm added 18%.
-
Speculative decoding: introducing DSpark to propose and verify multiple candidate tokens at once delivered roughly another 4× in inference speed.
Note where the credit goes: none of this was one engineer hand-tuning. Fleet analyzed the bottlenecks itself, proposed the fixes, validated them, discarded what failed verification, kept what passed, and iterated — with zero human intervention in the loop.
The second result is more striking. Starting from one instruction, Fleet dispatched work across distinct roles — decision-making, architecture design, coding, testing, review, quality assurance. The full run produced 87.2 million output tokens, of which coding accounted for only 38.6%; the remaining 61.4% went to design, decision, and verification processes. Real software engineering was never just “writing code.”
The third result shows the same engineering depth. While building AgentFS, Fleet also ran formal model checking (about 1.9 million reachable states) and roughly 300 integration tests. During verification it discovered and fixed a concurrency bug on its own — one that could have corrupted data.
The Frame: From AI Writing Code to AI Building Systems
Place the launch in a wider frame and the paradigm shift comes into focus. AI coding has moved through three stages:
Stage one: code completion (2023–2024). Copilot, CodeWhisperer — a smarter autocomplete.
Stage two: agent coding (2024–2025). Claude Code, Cursor, Devin — AI that understands context and writes code autonomously.
Stage three: system construction (2026–). What Fleet represents: AI stops being a programmer who writes code and becomes an engineering team that understands requirements, designs architecture, coordinates tasks, writes code, tests, verifies, deploys, and operates.
Fleet’s workflow runs six stages: understand → design → coordinate → build → verify → evolve.
That means the user supplies neither a PRD nor even a clean specification. Fleet works with the user to clarify goals and convert them into concrete constraints and acceptance criteria. When new information emerges mid-build, it updates architecture and code in step. After launch, it keeps watching how the system behaves and feeds operating data back into design.
A closed loop: intent in → system out → operational feedback → continuous improvement — dynamic, iterative, self-correcting, in fundamental contrast to the linear pipeline of requirements → design → code → test → deploy.
What It Means for the Industry
1. Software engineering is de-skilling at the bottom
When AI can build databases, tune inference engines, and create distributed filesystems from fuzzy intent — and handle testing and bug fixes on its own — the market value of “can write code” as a professional skill drops fast. The traditional senior-engineer role is due for redefinition.
2. The product surplus is expanding
Jia’s own trajectory is a signal: from Lepton AI (GPU cloud orchestration) to Intent Lab (autonomous AI engineering teams) is a completely different bet — less “stack GPUs,” more “let AI use compute to build productive systems.” It echoes a concept Boris Cherny, creator of Claude Code, articulated: the product surplus — capabilities models already have that products have not yet unleashed. Jia sees that gap and is betting it widens.
3. Cost collapse has structural consequences
Fleet rebuilt a database for $350 on open-source models — about $2,000 on Opus 4.8. Either figure is orders of magnitude below a traditional multi-person-month effort. When building software approaches the marginal cost of describing intent, the industry’s supply curve shifts right, and projects that never penciled out — custom ERPs, vertical CRMs, niche tools — suddenly become viable.
What Developers Should Take From It
1. High-level engineering design beats writing code. If Fleet’s path is right, “can code” keeps losing value while “can design systems” keeps gaining it — architecture trade-offs, system constraints, distributed consistency, performance bottlenecks. AI does not replace those easily.
2. Asking is itself a technique. Fleet starts from user intent but works actively to sharpen goals with the user. Turning a vague idea into clear constraints is becoming a meta-skill: good requirement descriptions equal good software systems.
3. Verification capability is the moat. Fleet’s most underrated part is its verification system. In a Kubernetes codebase checkout test, S3FS failed the check; AgentFS passed. It did not just “generate code” — it generated a verified, production-ready system. Whoever builds the fastest verification loop behind AI-generated code can bring AI into production more safely. Verification speed is displacing coding speed as the competitive dimension.
Two honest caveats: these are early results, and whether Fleet can handle messier real projects stably, at an acceptable long-run cost, still needs practice to prove.
But the choice itself is the loudest signal. When one of the industry’s most engineering-minded builders bets his next company on “AI constructing systems autonomously,” it is time to stop describing what is happening as “AI writing code.”
This is not AI writing code. This is AI building systems — a watershed in the history of programming.
