AI Upends Physics: GPT-5.2, Seed 2.0 & Tools

Morning Brief: AI Rewrites Scientific Consensus — Plus the Week's Top AI Tools and Workflows

Good morning, {{ first_name | AI enthusiasts }}. OpenAI's GPT-5.2 just challenged a long-accepted answer in particle physics, proposed a corrected solution, and autonomously wrote the formal proof in 12 hours. The skeptics' debate over "can AI actually think?" continues, but the conversation is shifting: it's less about whether AI can contribute to science and more about how quickly it can upend our assumptions.

Top stories in today's AI rundown

• GPT-5.2 makes a theoretical physics discovery
• The Rundown Roundtable: practical AI use cases from our team
• How to launch an outbound calling agent in 15 minutes
• ByteDance's Seed 2.0: competitive models and new controversy
• New AI tools, community workflows, and industry moves

OPENAI: GPT-5.2 makes a theoretical physics discovery

OpenAI published a research preprint showing a specialized version of GPT-5.2 independently discovered a mathematical formula in particle physics, found that the accepted answer was incorrect, and produced a formal proof within 12 hours.

What happened

• GPT-5.2 addressed a problem in particle physics previously considered solved.
• The model identified an error in the accepted solution and proposed a corrected result.
• It wrote a formal, verifiable proof autonomously. Physicists at Harvard, Cambridge, and Princeton verified the result.
• OpenAI credited Kevin Weil as a co-author; Harvard physicist Andrew Strominger said the AI "chose a path no human would have tried."

AI is no longer just an assistant in labs — it is beginning to probe and revise long-held scientific claims.

Why it matters

This milestone tightens the gap between AI-assisted research and original scientific contribution. Skeptics will still debate whether AI truly generates novel ideas, but independent, peer-verified results make that argument harder to sustain.

SPONSORED: Unlock agentic AI potential (AWS Marketplace)

AWS Marketplace published an eBook about using agentic AI to streamline operations. It includes real-world examples, deployment frameworks, and guidance on procuring agents through partners.

THE RUNDOWN ROUNDTABLE: Our AI use cases

Each week we poll The Rundown staff on how they use AI. Here are two practical examples:

Rishi, Growth

Rishi connected Claude Code to Apify's API to scrape high-performing content from Instagram and TikTok for ad creative inspiration. He also integrated ElevenLabs for automatic transcription, enabling analysis of hooks, pacing, and language used by top creators. Rishi then built a scriptwriting system that:

• Turns rough hooks into ad scripts using proven copywriting principles.
• Grades output on a 12-point rubric and rewrites until it scores 90%.
• Incorporates human edits to refine the agent's understanding of his style.

Nate, University Educator

Nate uses Claude Artifacts to create custom webpages inside chat. He uploads long articles, spreadsheets, or surveys and asks Claude to turn them into interactive pages that highlight key insights. The front-end design makes the results easy to share with teams.

AI TRAINING: Launch an outbound calling agent in 15 minutes

Build an AI agent that makes real sales calls in about 15 minutes. This practical guide shows the end-to-end setup.

1. Create a new AI agent in ElevenLabs, choose a voice, and add system instructions with business and call objectives.
2. Sign up for Twilio (free + $15 in credits) and buy a phone number (~$1.20). Copy Account SID and Auth Token.
3. Connect Twilio to ElevenLabs by pasting credentials into ElevenLabs' Phone Numbers settings.
4. Create an outbound "batch" in ElevenLabs, download the CSV template, add leads, and upload to start calling.

Pro tip: Enable "Transfer to Number" so the agent can patch hot leads through to your phone.

PRESENTED BY CONCIERGE: Your brand's AI answer engine

Concierge builds Perplexity-style answer engines trained on your brand content. B2B companies use it to answer technical buyer questions, control conversation visibility, and build trust before demos.

BYTEDANCE: Seed 2.0 and the global pressure

ByteDance released Seed 2.0, a family of models that match or outperform GPT-5.2 and Gemini 3 Pro on many benchmarks at a fraction of the cost.

Key points

• Seed 2.0 Pro reportedly surpasses GPT-5.2 and Gemini 3 Pro on math, reasoning, and vision tests at about $0.47/M input tokens.
• The models demonstrate real-world agentic abilities, including long multi-step workflows (96-step CAD demos).
• ByteDance also faced controversy: its Seedance 2.0 video model raised copyright concerns in Hollywood.
• Seed 2.0 is available on ByteDance's Doubao app in "Expert Mode" and via API; wider consumer access remains limited outside China.

Why this matters: ByteDance's advances increase competitive pressure on Western labs and highlight geopolitical and IP tensions in the AI race.

Everything else in AI today

• Peter Steinberger (OpenClaw) is joining OpenAI to help "drive the next generation of personal agents."
• The Pentagon may reconsider a $200M deal with Anthropic over restrictions on military use of Claude.
• Spotify's CEO said top developers haven't written code this year as the company shifts to AI-led development.
• Alpha School reported students scoring in the 99th percentile after a 2-hour AI-first academic model test.
• Simile raised $100M to build AI simulations of human behavior and decision-making agents.

Community workflow: A reader's practical AI win

Reader Peter S. from Hollis, NH shared how he used Copilot to streamline an insurance claim after a basement water pipe burst. He uploaded photos and descriptions, and Copilot returned estimated replacement costs and purchase links. What would have taken days became a few hours.

How do you use AI? Tell us here.

See you soon,

Rowan, Joey, Zach, Shubham, and Jennifer — the humans behind The Rundown

Related reading

• AI and scientific discovery: examples and implications
• Agentic AI deployment tips for product teams
• How to evaluate model benchmarks and costs

Note: For readers interested in gaming news, we'll include AI-assisted game design updates next week — from procedural level generation to NPC behavior models that speed development.