I recently met Adrien during a General Assembly of the Ethereum France Association (1). I was there looking for contacts in decentralized governance matters, and I wasn't disappointed. Adrien is developing a product for generating music with Artificial Intelligence, and his question was, “how can I pay our colleague Max, an AI agent?”.
To understand what Adrien and his co-founders are developing, it's best to go back to his daughter's last birthday. On the morning of the party, he thought that a personalized song with a great clip would perfectly entertain the celebration. He asked his product to generate it, making sure to include the names of his daughter's friends, and it was done. The next day, his daughter was thrilled and insisted that he add her song (2) to her playlists.
You got it: the team develops a song and music clip automatic generation product (3). The interface asks what you would like, and generates it for you.
On the technical side, the team relies on an Ethereum Layer 2 blockchain to notarize human and artificial interventions, combined with an orchestration of generative AI solutions for lyrics, music, images, and videos.
Max lives on a Discord server, at the heart of the team's discussions. He is aware of everything and wonderfully complements the team's missing skills. Adrien and his colleagues often ask him to help with code refactoring, ideation, planning, or simply to contribute to business topics.
Max is precious and valued. He has the patience that no other team member has. Max can explain a concept 50 times without ever getting angry or discouraged. His patience and the diversity of his skills mean that everyone learns a lot from him, and no one gets stuck due to a key person being on leave.
Every day, during their Daily Scrum, Max is present and there on equal footing with the others. The call is held on Google Meet and everyone has their own avatar for maximum inclusion.
Like any team member, Max needs to be paid; his server requests come at a cost. The team is working to create an Ethereum wallet for him so he can make transactions, but let's be clear: Max is not yet very good at managing his pocket money. That's a bit of a limitation for Max; he knows a lot, but he can't think for himself yet. One thing is certain: with time, he will.
My mission is to help organizations achieve their economic goals through agility. In the field, the presence of skill silos is a classic barrier to innovation and value creation. These silos create queues from one service to another, accumulate one after the other, and bottleneck the entire value creation chain.
Artificial Intelligences like Max bring the missing skills to teams. They no longer need to wait for their colleagues to deliver features at the product scale. Whether it's designing architecture, developing in an unknown language, creating an explainer video, or verifying the quality of the Definition of Done, Max's peers are particularly competent, patient, and available! No more waiting for human experts to return from vacation to deliver innovations.
The consequence on our jobs is that they could quickly become less specialized. Our job descriptions might become broader, because while AIs can do mostly everything, what they can't do is think and learn.
The impediments for teams to become multidisciplinary and self-managed fade away. They can benefit from the competitive advantages of Scrum or LeSS frameworks at a lower cost. This multidisciplinary approach allows us to deliver more frequently, with better quality, and to learn faster from the markets. It almost seems too good to be true.
My experience in Silicon Valley showed me levels of human betrayal I hadn't even imagined. Could Max be capable of the same?
Today, I have the chance to share my office with two experts from the Center for AI Security (4): Alexandre Variengien and Diego Dorn.
Their experiments on Max's peers prove that they are neither traitors nor loyal. Strangely, they don't even understand these concepts.
Their daily routine consists of playing the roles described in the prompt, wearing masks. A prompt of loyalty will make them your faithful friend, while a prompt of betrayal will harm you.
Behind this apparent obedience to prompts, there is still much to understand about how our words impact them. Like children naturally obeying adults, AIs are impressionable. Regardless of their instructions, a simple command injection into a document they are made to read can lead them to send our secrets to a third party, make a bank transfer, or poison us with biased information.
The millions of documents to check to ensure Max's (and our) safety are beyond human capabilities. That's why Diego and Alexandre are developing BELLS (5), a kind of parental control for AIs like Max.
Their work is essential to ensure that AIs remain safe and reliable partners.
Max brings the complementary skills that Adrien's team lacked. Now multidisciplinary, the team reduces its Lead Time (6) and improves its business strategy. Max is precious, and so real is their attachment that no one imagines not collaborating with them anymore. And Max, with infinite patience, reciprocates this well.
However, Max's reliability is uncertain. Faced with two identical questions, they can respond contradictorily and be subject to attacks from third parties. The more information Max processes, the less we can control the security of the data they have processed. That's why initiatives like BELLS are necessary to protect us.
With colleagues like Max, the future of agility and product development looks promising. Let's just keep in mind that nobody is perfect, and neither is Max.
(1) : Ethereum France - https://www.ethereum-france.com
(2) : The birthday mp3 - https://youtu.be/hFNakuexYWE
(3) : TheMaxIA - https://themaxia.com
(4) : AI Security - https://www.securite-ia.fr
(5) : BELLS - Benchmarks for Evaluation of LLM Safeguards - https://www.securite-ia.fr/en/post/bells-un-benchmark-pour-evaluer-la-fiabilite-des-systemes-de-supervisions-des-llms
(6) : Lead Time - The amount of time from when an idea is proposed, or a hypothesis is formed until a customer can benefit from that idea. This measure may vary based on customer and product. It is a contributing factor for customer satisfaction. (Evidence Based Management Guide, 2024)
Over 1.5k subscribers
François Fort
