Tencent open-sources 770B model as MIRI warns of extinction odds

Tencent just open-sourced a 770-billion-parameter model, and I keep coming back to the same question: why is a Chinese state-adjacent tech giant more comfortable publishing its weights than any major American lab? OpenAI, Anthropic, Google DeepMind — all sit on frontier models they treat as crown jewels. Tencent, meanwhile, is shipping this thing alongside a second, smaller open-source release aimed squarely at coding and research workflows. This isn't a one-off. It's a pattern across Chinese labs — DeepSeek, Alibaba's Qwen, and now Tencent at serious scale — and it's starting to look like a deliberate strategy rather than a side effect of catching up.

Why does this matter beyond the headline? Open weights change who gets to build on the frontier. A 770B model you can download and fine-tune is a different kind of asset than an API you rent by the token. It shifts leverage toward developers, startups, and governments who want sovereignty over their AI stack instead of dependency on a handful of US-based providers. I'd argue this is as much geopolitical positioning as it is technical generosity — China gets to shape the open ecosystem's defaults, the way Meta once tried to with Llama before pulling back toward more restrictive releases. If the best open models increasingly come from Chinese labs, the "safety by secrecy" argument that Western labs lean on starts to look less like caution and more like a business model.

Meanwhile, on the actual safety question, MIRI's CEO is out there putting a number on AI-driven extinction risk that's genuinely startling — high double digits, meaning more likely than not that AI wipes out humanity. I don't think that estimate is well-calibrated, and I say that as someone who takes alignment seriously; probabilities that precise about civilizational outcomes tend to say more about the forecaster's priors than about the world. But the gap between that view and the industry's default optimism is the real story. We're at a point where the people building frontier systems and the people most focused on their risks aren't just disagreeing on details — they're operating with completely different mental models of what's coming.

There's a smaller, more grounded thread worth noting too: a new study testing LLMs as uncertainty-aware optimizers in lab discovery, having models estimate their own confidence before proposing the next experimental step. It's a modest idea, but it's the kind of practical scaffolding that actually makes AI-assisted science more trustworthy — teaching models to know what they don't know, rather than asking us to just trust the output. That's the unglamorous work that rarely makes headlines but quietly determines whether any of this becomes reliable infrastructure or stays a demo. By the way, it's a nice contrast to the extinction-risk debate — one is arguing about whether AI ends civilization, the other is just trying to get a language model to admit uncertainty about a chemistry experiment. Maybe more of the field should sit closer to that second conversation.