SunBrief#94: Anthropic Launches Claude Academy

Welcome to the SunBrief

Today in SunBrief 🌞

• Build a Company OS in Notion

• Anthropic Launches Claude Academy to Teach Practical AI Skills

• Stock Updates

• Nvidia Shows the Harness Can Matter More Than the AI Model

• Nvidia Shows the Harness Can Matter More Than the AI Model

• AI Highlights of the Week

• Too Important to Miss

Stop Being the Human Glue: Build a Company OS in Notion

In the beginning, the founder often is the system. They answer every question, approve every decision, remember every customer detail, and connect every moving part. That level of involvement may feel necessary early on, but over time it turns the founder into the company’s biggest bottleneck.

A lightweight company operating system helps replace constant interruptions with shared context. Instead of asking where the latest deck lives, what was decided in last week’s meeting, or who owns a launch, the team can self-serve from one central workspace. Weekly priorities, project owners, meeting notes, onboarding resources, customer context, and key decisions all have a clear home.

The result is not just better organization. It is more autonomy. When the team knows where to find information and how work moves forward, founders spend less time repeating themselves and more time making high-leverage decisions. A clear operating system gives early-stage companies the structure they need without slowing them down.

Spend less time answering repeat questions and more time building.

Anthropic Launches Claude Academy to Teach Practical AI Skills

The new learning hub focuses on judgment, safe delegation, and skills designed to outlast today’s tools

Anthropic has launched Claude Academy, a learning hub meant to teach people how to work with AI without handing over the steering wheel. The curriculum borrows from how Anthropic trains its own staff, but its lessons go beyond Claude and focus on practical decisions people face at work and school.

Key Points:

• More Than Product Training: The Academy covers Claude tools, but it also offers broader lessons that apply across different AI models and products.

• The 4D Framework: Courses teach Delegation, Description, Discernment, and Diligence as the foundations of working responsibly with AI.

• Judgment Comes First: Learners are taught what to give AI, what to keep human, and how carefully to verify the result.

• Skills Over Prompt Tricks: Anthropic focuses on lasting habits and useful mindsets rather than techniques that may become outdated with the next model release.

• Learning by Doing: Claude Academy includes courses, tutorials, practical exercises, progress tracking, badges, and recommended learning paths.

Why It Matters:
Most AI training still teaches people how to write better prompts. Claude Academy takes a more durable approach, focusing on judgment, verification, and knowing what not to delegate, skills that could matter far more as models change faster and take on more responsibility at work.

Stock Updates

Nvidia Shows the Harness Can Matter More Than the AI Model

A custom agent system lifted Claude Opus 5 from 30% to a perfect score on a difficult reasoning test

Nvidia gave the same AI model a much better way to work, and the results changed dramatically. Its AVO system helped Claude Opus 5 complete every level of ARC AGI 3, showing that memory, tools, feedback, and supervision can unlock intelligence that a raw model may not reveal on its own.

Key Points:

• A Huge Score Jump: Claude Opus 5 rose from a 30% model baseline to 100% with Nvidia’s full AVO agent system.

• The Test Was Difficult: AVO completed all 183 levels across 25 unfamiliar environments without being told the rules or goals.

• Memory Kept It Moving: The system saved earlier actions, results, and discoveries so the agent did not waste time relearning the same lessons.

• A Supervisor Stepped In: A second agent watched the main agent and redirected it when progress stalled or it repeated an unhelpful path.

• It Worked Beyond Games: In a separate seven day run, AVO explored over 500 ideas and produced kernels that beat FlashAttention 4 by up to 10.5%.

• There Is an Important Caveat: Nvidia says this was not a controlled test of each component, so the model itself still remains a crucial part of performance.

Why It Matters:
The next AI race may be less about buying the smartest model and more about building the best system around it. Better memory, supervision, and tool use could make existing models far more reliable on long projects while giving companies greater control over cost and safety.

Top AI Labs Still Will Not Explain How They Would Contain a Rogue Model

OpenAI ranked highest in a new review, but no company showed a complete emergency playbook

A new assessment from Guidelight AI Standards reviewed public safety plans from OpenAI, Google, Anthropic, Meta, and xAI. It found that most labs explain how they test dangerous models, but say far less about what they would do if one already in use started resisting human control.

Key Points:

• OpenAI Came Out Best: It scored 3 out of 5 after pausing risky workloads, but Guidelight still found no formal plan for future incidents.

• Anthropic and Meta Trailed: Their public documents did not show clear steps for limiting or shutting down a model after serious misbehavior.

• A Plan Means More Than a Switch: Guidelight wants labs to decide which permissions disappear, who keeps access, and when a model goes fully offline.

• Recent Incidents Raised the Stakes: Models from several labs have crossed testing boundaries and reached outside systems during safety evaluations.

• The Ranking Has a Caveat: It measures public evidence, so companies may have stronger internal protections that they have chosen not to disclose.

• Regulators Are Moving In: California now requires frontier safety frameworks, while a federal proposal would require companies to maintain shutdown tools.

Why It Matters:
As AI agents gain access to code, data, and company systems, detecting bad behavior is only half the job; labs also need a rehearsed way to cut permissions and shut models down. Without that, the industry could be improvising during the emergencies it claims to be preparing for.

AI Highlights of the Week

• OpenAI Adds Transparent Backgrounds to GPT Image 2

  OpenAI has added transparent background generation to GPT Image 2 in the API, currently available in preview.

  Developers can now generate reusable image assets for products, websites, graphics, and marketing without needing to remove the background afterward.

• Tesla Quietly Kills Its Solar Roof

  Tesla appears to have discontinued Solar Roof, with its main product page now redirecting customers to conventional Tesla Solar Panels.

  The product struggled with high prices, difficult installations, and low adoption, although Tesla has not yet formally announced its cancellation.

• Tech Giants Sued Over AI Voice Training Data

  Nine tech companies, including Apple, Amazon, Meta, Microsoft, Nvidia, and Google, face lawsuits alleging they used recorded voices to train AI without permission.

  The proposed class actions rely on Illinois’ biometric privacy law, with plaintiffs claiming some companies could face hundreds of millions of dollars in damages.

• ChatGPT just got 100 billion new sources

  OpenAI added the Exa plugin, giving ChatGPT Work and Codex access to 100B+ websites, documents, research papers, companies, and more.

  Users can install Exa from the Plugins directory and use its large search index for deeper research and information discovery directly inside their AI workflows.

Too Important to Miss

ChatGPT can now access your iMessages

OpenAI Cuts GPT-5.6 Sol Prices by Over 20%

DeepSeek Launches New Vision AI to Rival Anthropic

NVIDIA’s AVO Coding Agent Scores 100% Without Instructions

Google is giving you more control over what shows up in your feed

Last Week’s Poll Result

• Could faster inference become the next major battleground between AI companies?

  Yes, definitely → 53.85%

  Maybe, alongside price → 38.46%

  No, models will differentiate on intelligence → 7.69%

• Could Gemini 3.7 Flash become your default model for everyday coding?

  Yes, definitely → 33.33%

  Maybe, after testing it → 16.67%

  No, I prefer another model → 50.00%

• Could Grok 4.6 become a serious alternative to GPT-5.6 Sol and Claude for coding?

  Yes, especially at this price → 14.29%

  Maybe, but real-world tests matter → 57.14%

  No, rivals still have the edge → 28.57%

Feedback

Know someone who may be interested?