Paris-based startup Twin Labs wants to automate repetitive tasks like onboarding new employees to all your internal services, reordering items when they’re out of stock, downloading financial reports across several SaaS products, reaching out to prospects, and more.
Twin begins with a science-fiction premise. Foundation models—LLMs—developed technological capabilities. Twin Labs co-founder and CEO Hugo Mercier told me they wondered whether they might replicate ourselves by teaching an AI agent to do our jobs.
Twin Labs’ approach to enhancing internal procedures is particularly intriguing. The business uses multimodal models with vision, including GPT-4 with Vision (GPT-4V), to mimic people.
Twin Labs attempted classic LLMs to construct autonomous agents before utilizing multimodal models. We’ve tried several items, including academic articles and open-source GitHub repositories. Conclusion: LLMs are untrustworthy. Mercier said this implies LLMs make bad judgments. Ultimately, the mission isn’t done.”
He says GPT-4V has been trained on several software interfaces and code bases, opening up new options. “When you show an interface, it understands the feature behind the button,” Mercier explained.
Twin Labs doesn’t use APIs or multi-step procedures like Zapier. Twin Labs is more browser-like. The gadget automatically loads websites, clicks buttons, and enters text.
If you’re recruiting someone, you may need to enter their details into your payroll system, invite them to Slack, establish a Google Workspace account, and invite them to the healthcare insurance provider.
Every time a new employee joins, companies review a broad list of duties. These activities are simple, but they must be done correctly, in sequence, and with certain checkboxes checked. It will be crucial to train Twin Labs’ AI helper utilizing screen capture and natural language descriptions.
Currently, the company is working toward this goal. Hugo Mercier and Joao Justi, the co-founders, spent six months constructing a product prototype. Betaworks, Motier Ventures, Factorial (Matthew Hartman & Clem Delangue), and many angel investors, including Florian Douetteau (Dataiku), Thomas Wolf (Hugging Face), Charles Gorintin (Alan), Mehdi Ghissassi (DeepMind), Romain Huet, Irwan Bello, Romuald Elie, Yan-David Erlich, Olivier Pomel, Rodolphe Saadé, and Thi
Twin Labs’ autonomous agent system faces several hurdles. In the AI area, API and infrastructure costs are falling quickly, yet accomplishing a job is expensive. To ensure product functionality, Twin Labs will supply a library of pre-trained jobs. The business plans to let customers create assignments on its platform thereafter.
Twin Labs’ novel technique to communicate with AI models is different from the chatbot interface most consumers associate with AI solutions. Mercier added, “We really wanted to get down to the nitty-gritty of what people do on a day-to-day basis and how we can take over some of the things that are actually a bit of a hassle for them.”
