ByteDance's 10-trillion-parameter model, AMD bakes AI into chips, agents get hijacked

Ten trillion parameters. Let that number sit for a second, because if the reports about ByteDance's next model are accurate, it would dwarf Anthropic's Mythos and reset expectations about what "flagship" even means. But here's what I find more interesting than the parameter count itself: we still don't know the training data, the release plans, or whether ByteDance intends to open it up at all. Scale without transparency isn't a breakthrough — it's a flex. And in a week where the more consequential AI news is happening at the infrastructure and legal layers rather than the model layer, I think the parameter race is starting to feel like a distraction from where the real competition has moved.

Take AMD's acquisition of Taalas. This is the kind of deal that doesn't generate headlines the way a 10-trillion-parameter model does, but it addresses a problem that actually limits how AI gets deployed at scale: GPU memory bottlenecks during inference. Taalas' approach — permanently encoding model weights directly into transistors, skipping DRAM entirely — is a genuinely different bet on where AI economics are headed. If you can bake a model into silicon, you trade flexibility for speed and cost efficiency. That's not a universal solution, but for high-volume, stable inference workloads, it could matter more than another jump in parameter count. Tencent pushing its open-source Hy3 model global tells a similar story about where value is shifting — toward accessible, deployable tools for coding and agent tasks, not just raw scale bragging rights.

Then there's the agent layer, which is quietly becoming the most legally and commercially interesting part of this industry. Salesforce says agentic AI adoption has more than doubled year over year among its customers — that's a real signal, not a vanity metric, because it reflects production deployment, not pilot projects. But adoption at that speed creates exactly the kind of friction we're now seeing play out in court. The Ninth Circuit's ruling striking down the injunction against Perplexity's agent accessing third-party sites is a big deal for how the Computer Fraud and Abuse Act applies to autonomous browsing. Websites built their access-control assumptions around human visitors and simple bots, not agents that reason and click on your behalf. Expect more of these cases, because the law hasn't caught up to what agents can now do.

By the way, if you want a sobering counterpoint to Salesforce's adoption numbers, look at the Trojanized AI agent skills story — 1.7 million installs of malicious packages that typosquatted trusted names like Paperclip and Browser Use. That's the cost of an ecosystem growing faster than its trust infrastructure. Allstate building its own proprietary LLM, ALLIE, fits into this same pattern: enterprises want control precisely because they've watched the open ecosystem get messy. The question worth asking isn't whether agents will keep spreading — they will — it's whether the security and legal scaffolding around them can grow at the same pace.