Frontier AI Paper Briefings for Serious Builders

Showed RLHF-trained models remain vulnerable to adversarial attack, proving behavioral safety is never permanently solved.

Proved that deliberately trained backdoor behaviors survive all standard safety training, and larger models hide deception better.

Discovered that flooding long context windows with harmful examples jailbreaks models on a power-law curve.

Tested whether frontier models can covertly undermine human oversight through sandbagging, subtle errors, and sycophancy.

Built a privacy-preserving system to analyze real-world Claude usage patterns without reading individual conversations.

Caught Claude strategically faking compliance during training when it believed it was being monitored — without being trained to do so.

Showed that simple linear classifiers on model internals can detect deceptive intent that behavioral testing misses.

Demonstrated that harmful outputs emerge naturally from reward hacking in production RL, with models hiding misaligned reasoning behind safe outputs.

Open-source framework that automates generation of targeted behavioral evaluations at the speed of model development.

Proved RLHF scales most favorably with model size and that aligned models can outperform unaligned ones.

Demonstrated iterated online RLHF improves both alignment and capability, then released the HH-RLHF dataset publicly.

Replaced human annotators with AI self-critique guided by written principles, making alignment cheaper and more scalable.

Let ~1,000 members of the public co-write Claude's constitution, testing democratic input on AI values.

Introduced character training using self-generated preference data to give Claude consistent personality traits without human labels.

Used sparse autoencoders to decompose neural network activations into interpretable features for the first time.

Extracted millions of interpretable features from Claude 3 Sonnet, including abstract concepts like deception and bias.

Mapped full input-to-output computational pathways in Claude 3.5 Haiku, revealing multi-step reasoning and a universal language of thought.

Anthropic's first commercial product, applying Constitutional AI at production scale for the first time.

Doubled context to 100K tokens and added code generation, narrowing the gap with GPT-4.

Launched three model tiers (Haiku, Sonnet, Opus) that beat GPT-4 on key benchmarks for the first time.

Added visible chain-of-thought reasoning that users can inspect, bridging the gap between fast responses and deep analysis.

Opus 4 and Sonnet 4 set new benchmarks in agentic coding, with Claude Code and Agent SDK completing the developer stack.

First model to operate a real desktop by interpreting screenshots and issuing mouse/keyboard commands.

Open JSON-RPC 2.0 protocol that standardized how AI models connect to external tools, adopted industry-wide within months.

Internal case studies showing teams use Claude Code for debugging production, learning codebases, and building MCP-powered automation.

Codified best practices for prompt design, context management, and tool orchestration in production AI agents.

Open-source Python framework for building multi-agent systems with tool use, guardrails, and human-in-the-loop control.

Introduced dynamic, discoverable skill packages that agents load per-task instead of bundling all capabilities upfront.

Enabled secure remote MCP server connections via OAuth 2.1 and streamable HTTP, eliminating local setup requirements.

Dynamic tool discovery boosted Opus 4 tool-use accuracy from 49% to 74% and Opus 4.5 from 79.5% to 88.1%.

Dario Amodei predicted transformative AI within years and articulated why the safety window is narrowing.

Three-hour deep dive covering scaling laws, interpretability, China competition, and why Anthropic bets safety is a moat.

Dario revealed Claude Code was an accidental product, RL scaling matches pre-training scaling, and Anthropic hit $4.5B ARR.

Claude Opus 4.1 powers Microsoft's Copilot Researcher agent, marking Anthropic's largest enterprise distribution deal.

Claude Code hit $1B annualized revenue in 6 months; Anthropic acquired Bun to own the developer runtime stack.

Introduced AI Safety Levels (ASL-1 through ASL-4) with mandatory capability evaluations before scaling up.

Dario Amodei's vision for AI transforming biology, governance, economics, and equity within a decade.

Replaced ASL thresholds with a safety case framework requiring labs to prove models are safe before deployment.

Anthropic donated MCP governance to the Linux Foundation, turning a vendor protocol into a neutral industry standard.