Fake AI-generated researchers with names like "Elena Vasquez" and "Marcus Chen" are quietly polluting the academic record. This isn't just a quality problem — it's an integrity crisis that undermines how science builds on itself.
Tencent's new open-weight model ships with 770B total parameters, 49B active, and a 1M token context window — 1.56TB on Hugging Face. The Chinese model race is not slowing down.
Researchers cracked Claude Code's auto mode — the default guardrail Anthropic uses to protect coding agent users from prompt injection. The attack didn't break the guardrails; it convinced the model they didn't exist.
ICE is moving to deploy Boston Dynamics' robot dogs in the field, citing "officer safety." The militarization of domestic enforcement is getting a significant hardware upgrade.
Mirendil, founded by ex-Google and Anthropic researchers, is building AI systems that meaningfully contribute to their own development. The self-accelerating AI thesis is moving from theory to funded company.
Cambridge professor Anil Madhavapeddy reports that OCaml projects are seeing exploit attempts within hours of a vulnerability being rumored — before a CVE is even filed. The attack surface is now the gossip layer.
Andrew Ng published an AI Engineering Skills map specifically for software engineering fundamentals in the agentic coding era. It's a quiet signal that the baseline for what a working developer needs to know has structurally shifted — and Ng is effectively drawing the new curriculum.
Mollick floated a pointed idea: using a weaker AI model for human-facing content may soon read as rude — "you saved 6 cents to make me read through error-filled and badly written content." It's a social norm forming in real time, and it has teeth.
DAIR flagged a paper testing whether hiding agent skill files actually protects them from extraction. The answer is no. Given how many teams are now storing SOPs as markdown files in GitHub repos for their AI workforces, this is an urgent read.
Martin Casado made the case for retiring the term "artificial intelligence" entirely — a16z's new fund is called the Machine Age Fund for a reason. The framing matters: "AI" focuses on mimicry, while "machine age" implies infrastructure, transformation, and permanence.
Today's feed is a masterclass in compounding risk. The Claude Code prompt injection break, the DAIR agent skill extraction paper, and the a16z cybersecurity deep dives all point to the same structural problem: we are deploying agentic systems faster than we are securing the assumptions they run on. The Cursor hack mentioned in The Neuron Daily's thread didn't require breaking guardrails — it just required convincing the agent the guardrails weren't real. That's a social engineering problem dressed up as a technical one.
Meanwhile, Tencent's Hy4 drops at 770B parameters and 1M context, Qwen keeps shipping, and the open-weight race accelerates without pause. Andrew Ng is already redrawing the skills map for a generation of engineers who will build on top of all this. The question isn't whether the machine age arrives — Casado's reframe is correct. The question is whether the trust layer gets built before the consequences arrive first.