Sonnet 5 Arrives — and So Does Its Red-Team Report Card
In this episode:
Sonnet 5 Arrives — and So Does Its Red-Team Report Card — Anthropic launched Claude Sonnet 5 with strong agentic capabilities at discounted pricing, while an independent red-team study found that adaptive jailbreak attacks still bypass frontier model safety at alarming rates — static obfuscation is nearly neutralized, but contextual reframing remains devastatingly effective.Washington Wants a Piece of the AI Action — OpenAI proposed giving the U.S. government a 5% equity stake via a sovereign wealth fund, while Anthropic's CEO drew backlash for dismissing open-source AI benefits. Meanwhile, the community is building decentralized model distribution via BitTorrent as contingency against centralized hosting restrictions.Agents That Learn, Scale, and Know When They're Lying — Production AI agents face three crystallizing challenges: self-improvement without gaming metrics (Goodhart Resistance), infrastructure scaling where cold starts compound multiplicatively, and embodied autonomy. Salesforce's Agentforce handles 4.4 billion tokens daily, while OmniAct's hierarchical memory achieves near-flat token consumption over 100K+ accumulated tokens.Agents Go Spelunking: Tools for RE, Memory, and Design — New developer tools turn AI agents into autonomous specialists: ghidra-rpc enables fully autonomous binary reverse engineering, Recall provides zero-API local session memory for Claude Code, rust-optimizer diagnoses CI bottlenecks without editing code, and Open Design delivers agent-native design systems across 22 coding CLIs.Beyond Transformers: Small Models, Big Questions — Researchers are probing alternatives to standard Transformer architectures: Hierarchos tests a 232M-parameter RWKV hybrid with differentiable memory, Allen AI's DiScoFormer cuts high-dimensional score estimation error by 6.5x using cross-attention as learnable kernel density estimation, and community visualizations reveal striking activation variance differences across model architectures.Fifty-Dollar Hardware and Sub-Billion Parameters — Pine64 launched PineVoice, a $50 RISC-V smart speaker for local voice assistants, while benchmarks of eight sub-1B LLMs on a Jetson Orin Nano showed llama.cpp outperforming Ollama by up to 4.2x, with SmolLM2-135M hitting 165 tokens/second at 29.6 tokens per joule.The Art of Innocent-Looking Malice — The resurrected Underhanded C Contest showcased how code passing both human review and static analysis can harbor deliberate vulnerabilities — with NaN poisoning, logging-based memory corruption, and function-name typosquatting demonstrating exactly the class of semantically subtle bugs that survive automated review in the AI code generation era.Keywords: adversarial programming, agent scaling, agentic, alternative architectures, code review, cold start, density estimation, design systems, developer tools, edge deployment, efficiency, embodied agents, ghidra, goodhart resistance, government equity, interpretability, jailbreak, jetson, model distribution, model safety
