Frontier AI Paper Briefings

The paper that started the GPT paradigm

The staged release that changed AI safety discourse

The math behind 'bigger is better'

The model that made the world pay attention

Connecting vision and language at scale

Scale applied to speech recognition

State-of-the-art performance, unprecedented secrecy

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.

Safety evaluation for multimodal AI

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

The omnimodal model

The model that thinks before it speaks

Safety evaluation of reasoning models

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

Reasoning models get tools

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

OpenAI goes open-weight for the first time since GPT-2

The convergence of scale and reasoning

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

Expert-level performance across professional tasks

Native computer use meets frontier reasoning

Standardized RL benchmarks and environments

When language models learned to see and create

Teaching GPT to write code

Photorealistic text-to-image generation

The product that changed everything

OpenAI becomes a platform company

Text-to-video enters the frontier

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

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.

The product blitz

OpenAI enters the agent era

Extended reasoning meets web research

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.

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

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%.

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

The RL algorithm that would power RLHF

The prototype for RLHF on language models

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

The paper that made ChatGPT possible

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

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

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

Process supervision for reasoning

OpenAI's most ambitious safety bet

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

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

OpenAI's risk evaluation framework

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.

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

The safety exodus

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

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

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.

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

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.

Founding of the nonprofit AI research lab

The mission document that defined OpenAI's values

The CEO's roadmap to AGI

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

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

The governance crisis that shook AI

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.

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.

The for-profit transition

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