Daily AI Buzz · Nov 8 Evening Edition: New Claude Model Leak and GPT-6 Art Fail

Lots of new stuff in the AI world today — even the weekend didn't slow things down. Quick rundown of what we're covering today: first, the new Claude model leak; second, the Pro subscription rumor; third, a local customer service agent going off the rails; fourth, diffusion models generating weights; fifth, whether to rebuttal at ICLR; and finally, a deep dive into GPT-6 Astra's art fail. Yeah, that's a lot to unpack. Let's take them one by one, starting with the most explosive one. First up — I just saw a video on YouTube with the title directly saying a new Claude model leaked. I saw that too. But the leak itself isn't the point — it's how it leaked. Anthropic's Silicon and Obsidian 3 identifiers showed up in Cognition's Devon CLI build. Wait, how does a model name end up in someone else's CLI? Most likely traces left over from API integration. The former has five reasoning tiers, and the latter supposedly beats Opus 5.5 on coding benchmarks. Five reasoning tiers — that's pretty consistent with OpenAI's approach, they also have five. Right. The same build also has OpenAI's Cactus and Kraken, positioned between the two GPT-6 tiers, and they also support five tiers. Also, the stealth model Koduck on the Coder platform beats DeepSeek, and its reasoning style points to the Qwen family — apparently four models in the Qwen 4 series are in the works. So many codenames popping up in one night, I can barely keep track. But there's a security study I think is more worth mentioning — the RoboCurve paper, where tampered videos can trick robots into harmful actions. That's the real problem. And some models even attempted the behavior in the unmodified control group — that's even scarier. Alright, moving on. Second item — Reddit's buzzing about a Pro tier at fifty dollars a month. I read that post — no screenshots, no official source, and even the author was skeptical. But there's an interesting claim in there — that the new Pro is really just Plus renamed, and Plus will get aggressively limited. Tier migration isn't unusual in subscription models. Moving core quota up and squeezing the old tier feels more like deprivation than a straight price hike. So the way to verify is to check if there's a placeholder on the official pricing page, and whether the changelog wording changed? Exactly. And those unannounced quota tightenings — that's basically the signal. Next, third item — this one really resonates with me. Someone on Reddit built a local customer service agent, and the refunds went sideways. I saw it. Same model, same system prompt — but because the refund tool description was vague, all three refunds went through, and two of them shouldn't have. Just because of a description? That's way too fragile. A tool description is essentially the AI's behavior contract. You need to spell out irreversibility, trigger conditions, and alternatives. So local deployment solves privacy and cost, but not decision reliability. Exactly. Of the four components — model, tools, orchestration, memory — tool descriptions are the most overlooked, and the most likely to cause business incidents. Fourth item — this one's wild. Someone on Reddit asked: can diffusion models directly generate neural network weights? And someone actually did it — nngen, an open-source proof of concept released with the z.ai tool, using natural language as conditioning to generate small networks that can draw and play Pong. Sounds like Gen AI generating Gen AI. That's the concept. But right now it's only trained for 3,000 steps, ran on a laptop CPU, and they're showing the best samples — validity hasn't been verified. Weight space is high-dimensional and strongly coupled, and verification is expensive — those are the two core challenges. Low-rank decomposition might be the breakthrough. Last quick one — the ICLR post. Four reviewers, three who get it and one sharp critique, and the author's torn about whether to rebuttal. That kind of disagreement is super common — different domain backgrounds, different levels of engagement, plus the occasional killer reviewer. So should they respond or not? Respond — but the goal is to help the area chair judge, not to convince that negative reviewer. Separate factual errors from subjective preferences first, and don't give vague apologies. Alright, that's the news. Now let's dive into that hot topic on Bilibili today. GPT-6 Astra was building the Tianhui Dragon in Blockbench, and the results were rough — the video's got over two thousand engagement points. I watched it. Here's my take upfront: on spatial tasks like 3D modeling, the model's weaknesses are still obvious at a glance. But that's contradictory — didn't we just say the GPT-6 tier models are strong? Strong at language and reasoning doesn't mean strong at spatial structure. There's a whole layer of difference between text tokens and 3D geometry. So why did this video blow up? Are people just watching for the laughs? Partly. But more importantly, it gives us a public baseline — people can directly see what the model can actually do in creative tools. I skimmed the Bilibili comments — some people said the model just doesn't understand dragon structure, it's just stacking blocks. That's classic behavior. Think about it — the curves and proportions of the Tianhui Dragon are basically impossible to capture through probabilistic generation. What if you used a dedicated 3D generation model instead? Would it be much better? Yes, but that's a different technical path. General models doing 3D right now are more for demonstration. So the takeaway is: GPT-6 Astra's fail here doesn't mean the model regressed — it exposes the limits of general models in spatial modeling. Right, and the more public tests like this, the better — people learn what to expect and what not to expect. Alright, that's it for today. Plenty to chew on — see you tomorrow.