AI Daily: Astra Breaks Math Barriers & Alibaba Unleashes Qwen3.8-Max

On August first, an OpenAI model named Astra solved ten previously open problems in mathematics. It did it for an estimated compute cost of two THOUSAND dollars. In episode 128, we talked about what actually moves the needle in AI. THIS is what it looks like. So what else is moving? A lot. First, Alibaba just unveiled Qwen3.8-Max. That's a two-point-four trillion parameter model aimed squarely at the biggest players like Anthropic and Moonshot. The real news? They plan to release the weights publicly next week, a major move in the open-source race. ByteDance is also shipping. Its new model, Seedance 2.5, is generating thirty-second, high-quality video clips from a mix of images, audio, and other videos. This is rolling out now on platforms like Jimeng AI. And the money is following.

Chinese venture capital firms are raising a new thirty-five-billion-dollar war chest for AI, signaling a massive renewal of confidence and a hedge against US dominance. Finally, for a reality check, Cloudflare just dropped its latest traffic data. No surprise, ChatGPT is still king, with Claude and Gemini trailing. But the report’s real finding is about reliability. Ask an AI for a list of top ten anything, and there's a less than one percent chance you'll get the same list twice. You’re fighting for visibility in a system that doesn’t even agree with itself. Okay, let's go back to OpenAI's Astra. Because this is more than just another benchmark score. This is different. For two thousand dollars in compute, Astra didn't just answer questions. It generated formal, machine-verifiable proofs in Lean for problems that have been open for years.

We're talking about the existence of non-sofic groups and new sphere-packing bounds. This is not rote memorization. This is original academic research, done by a machine, and published on GitHub for anyone to check. The quote that nails it comes from Build Fast with AI: "AI is progressing beyond test benchmarks to original, machine-verifiable academic research." And the cost... two thousand dollars. That is the revolution. It means this level of discovery is no longer the exclusive domain of labs with billion-dollar data centers. It collapses the cost of creating new knowledge. That changes EVERYTHING. Now, let's talk about the other big shift happening right now. It’s the battle over who gets to use these powerful tools. On one side, you have Alibaba. They are not just building a massive two-point-four trillion parameter model.

They are preparing to give it away, promising to make Qwen3.8-Max open-weight. This is a direct shot across the bow from Chinese AI, following the success of Moonshot's Kimi K3. They're competing on performance AND on access. On the OTHER side, you have Anthropic. They just published a new position paper, and it's a tightrope walk. They're saying open weights are beneficial for transparency and security… up to a point. For their most powerful, frontier models? Absolutely not. They’re keeping those locked down, citing the risk of misuse for weapons of mass destruction or autonomous cyberattacks. So you have three camps forming. Meta is 'open everything.' OpenAI is 'open nothing.' And now Anthropic is trying to invent a third way: 'open, but conditional.' Into this debate, Alibaba is about to drop a bomb, betting that radical openness will accelerate their ecosystem faster than their rivals.

So here's where we are. One part of the industry is achieving verifiable, superhuman reasoning for the price of a used car. The other part is fighting a philosophical war over whether to give the keys away. The models are getting exponentially smarter. The debate is no longer about if they can change the world. It’s about who gets to decide how.