The Signal — February 26, 2026

Three stories today, all circling the same question: who actually controls AI's next chapter?

Alibaba's Qwen 3.5 Matches Frontier Models — and It's Open Source

Alibaba's Qwen team released four new language models this week, and the benchmarks are hard to ignore. The flagship Qwen3.5-122B-A10B scores 72.2 on BFCL-V4 for tool use, compared to 55.5 for OpenAI's GPT-5 mini. The smaller Qwen3.5-35B-A3B handles a million-token context window on consumer GPUs with 32GB of VRAM.

Three of the four models ship under Apache 2.0. Download them, run them locally, build commercial products. No API bill, no usage restrictions. The models also retain accuracy under 4-bit quantization, meaning you can run frontier-grade capabilities on hardware you already own.

This matters because the gap between open and closed models was supposed to be the moat. If a team in Hangzhou can match Anthropic and OpenAI at a fraction of the compute cost, the pricing pressure on proprietary APIs gets real fast.

Sources: VentureBeat · The Decoder

Karpathy: Programming Is "Unrecognizable" Now

Andrej Karpathy posted that programming has fundamentally changed in the past two months. His example: an AI agent independently built a video analysis dashboard over a weekend. He typed the task in plain English, the agent worked for 30 minutes, solved problems on its own, and delivered a finished result. Three months ago, that would have been an entire weekend project.

What makes this notable is the reversal. As recently as October 2025, Karpathy called agentic AI hype exaggerated and the products far from ready. He changed his mind after Opus 4.5 and Codex 5.2 shipped in winter. He still says agents need "high-level direction, judgment, taste, oversight" but the nature of the work has shifted from writing code to managing the things that write code.

Bloomberg ran a separate piece today on how coding agents are fueling productivity anxiety across the industry. The timing isn't a coincidence.

Sources: The Decoder · Bloomberg

Perplexity Computer: 19 Models, One Orchestrator, $200/Month

Perplexity launched Perplexity Computer, a platform that routes tasks across 19 different AI models based on what each does best. Claude Opus 4.6 serves as the orchestrator. When you submit a complex query, the system breaks it apart and farms pieces to specialized models. Available to Max subscribers at $200/month with additional per-token billing.

The pitch is that no single model excels at everything, so let a smarter layer decide which model handles which subtask. It's a bet that orchestration will matter more than any individual model's capability. The approach also sidesteps vendor lock-in: if a better model appears from any provider, it slots into the system.

At $200/month, this is aimed at power users and small teams, not casual searchers. Whether multi-model routing delivers enough value over just using one good model remains the open question.

Sources: Heise · Implicator

On the Editor's Desk

Seven stories came through the pipeline today. We ran three. Here's what didn't make it and why.

OpenAI's London expansion got coverage everywhere, but it's a real estate announcement. They're opening their largest research hub outside the US. Good for London, but "company opens office" isn't a story unless something ships from it. We'll cover what comes out of that lab when it produces results.

OpenAI Codex + Figma integration is an incremental product update. Useful for designers, but it's a feature addition to an existing tool, not a shift in how anything works.

Trump telling tech companies to fund their own data center power sounds like policy but reads like a press conference quote without follow-through. No executive order, no legislation, no concrete mechanism. When it becomes actual policy, it becomes a story.

PNNL partnering with OpenAI on federal permitting is genuinely interesting — a national lab using AI to speed up environmental reviews — but too niche for a general audience without more detail on what it changes in practice.