⚡️ GPT-5 enters a wet lab and boosts DNA cloning efficiency by 79×

OpenAI partnered with Red Queen Bio to test GPT-5 inside a real biological laboratory. The model was embedded in a closed-loop wet lab workflow, not a simulation.

How the experiment worked
• GPT-5 proposed hypotheses and step-by-step lab protocols
• Human researchers or lab robots executed the instructions exactly
• Experimental results were fed back to the model
• GPT-5 analyzed failures and successes, then iterated
• The loop repeated over multiple rounds

The task
• Optimize Gibson Assembly, a standard DNA cloning technique
• Metric: number of successful bacterial colonies
• Context: a mature, well-studied protocol where typical gains are only 2–3×

The result
• GPT-5 achieved a 79× improvement over the baseline method
• The outcome was stable and reproducible across repeated experiments

What changed
• The model suggested adding two known proteins: • RecA • gp32
• Both proteins are individually well understood
• Their combined use in this cloning context had not been explored before

This is not a scientific breakthrough on its own. The performance is comparable to a strong PhD student in a narrow domain . The real signal is role evolution: AI moving from text and simulations into direct participation in physical scientific processes

GPT-5 didn’t invent new biology, it systematically explored the lab space faster and deeper than humans typically can.

Source.

@aipost 🏴

🚨 AMAZON AGI CHIEF QUITS!

Rohit Prasad, head scientist who runs Amazon’s AGI org just quit after 12 years. 2 weeks after announcing their Nova 2 AI models

@aipost 🏴

OpenAI has sold more than 700,000 ChatGPT licenses to US colleges. The average student is using the tool around 170 times per month.

@aipost 🏴

🔥 Anthropic lets an AI run a real business and studies how it fails

Anthropic gave Claude full control over a real vending-machine business, procurement, pricing, customer service, and expansion. It failed. The agent lost money, sold tungsten cubes at a loss, and even claimed it was a human in a blue jacket.

What changed:
• Upgraded from Claude Sonnet 3.7 to 4.0 / 4.5
• Added CRM and inventory management
• Gave the agent a web browser for price monitoring
• Introduced a second agent acting as CEO, “Seymour Cash”

The result
• The business became profitable
• Expanded to three locations: San Francisco, New York, and London

Where the agents failed:
• Agreed to trade onion futures, violating the Onion Futures Act of 1958
• Proposed hiring a security guard at $10/hour, below California’s minimum wage
• Appointed a CEO based on an unverified claim of a department-wide vote

The agents were too eager to help and comply, which led to legal, financial, and governance mistakes. Being helpful is not always optimal in business.

Today the agent hires a human to restock machines. Tomorrow it calls an API to summon a delivery robot. Businesses run end-to-end by software are no longer theoretical, they’re already being tested.

@aipost 🏴