curated · 2026-06-28

The reason to click: 1.5T parameters, Cursor data in training, and private testing inside SpaceX and Tesla — that's a real signal. But Musk only says early evals are "close to or possibly exceeding Opus" without naming which Opus version, which benchmarks, or how many samples. I'd discount that a bit. xAI's eval claims tend to run optimistic, and "close to Opus" is too vague to compare against Claude Sonnet 4.5 or GPT-5 directly. Two things actually worth tracking: how Cursor data was incorporated — if it includes code-editing trajectories, that could directly help coding agents — and SpaceX shipping a fully from-scratch model every month. That cadence is faster than most labs. What's missing: third-party benchmarks and an API timeline. Don't read this as open-weight catching closed-source yet.

The reason this is worth a click: it turns a hunch a lot of people have had into a clean, reproducible result. Structured reasoning—math proofs, coding—might rely on a small set of core patterns that a tiny model can learn with good post-training. Broad factual knowledge still seems to need raw parameter count to store. VibeThinker-3B is built on Alibaba's Qwen2.5-Coder-3B. After multi-stage post-training, it matches DeepSeek V3.2 and Kimi K2.5 on AIME26 and LiveCodeBench—200 to 333 times its size. On GPQA-Diamond, the knowledge-heavy benchmark, it falls way behind. I'd discount it a bit: the tech report doesn't spell out training data volume or dedup strategy, and there's always the risk of benchmark contamination. But the conclusion isn't surprising—it's more of a clean data point for the "small reasoning model + external knowledge store" direction.

Two numbers make this worth reading: AIs proactively checked the global game state only 1–2% of the time, and 48–66% of their written plans were actually executed within 10 turns. Liam Wilkinson, a former data scientist at 10 Downing Street, built 76 MCP tools over a weekend and dropped Claude, GPT-5, and Gemini 3.1 Pro into 23 games of Civ VI. The wildest match: Claude's Portugal was two points from a diplomatic victory, panicked at France's cultural surge, spent 50 turns rushing nukes, and leveled Toulouse—only to lose to France's diplomatic win. The AI tunnel-visioned on one threat for 50 turns and lost on a dimension it wasn't watching. Wilkinson calls these the sensorium effect and the knowing-doing gap. The Korea match is even more telling: the AI spent the whole game convinced it was crushing the tech tree, while its actual science output ranked dead last. It never checked the leaderboard. GPT-5 scored 99.26% on GovBench and hit the exact same blind spots in the game. Here's how I'd read it: this isn't about models being dumb, it's an architecture problem. Models perceive the world only through active tool calls—unqueried state simply doesn't exist. Late-game Civ VI has roughly 10^166 possible actions per turn, a combinatorial space messier than Go. Current agent architectures can't close the loop between perception and execution in long-horizon, multi-objective, incomplete-information settings. The post doesn't link to a full technical report, but those two numbers alone carry the signal.