At Rise of AI 2024 in Berlin, Irina Gontcharova, Director of Azure Infrastructure at Microsoft Deutschland, argues that AI does not solve human problems by itself but unleashes human creativity, much like the invention of the collapsible paint tube in 1841 freed painters to leave the studio and enabled impressionism. She traces how fast AI adoption is moving: ChatGPT reached mass adoption in about two months, far faster than mobile phones or the internet. She lays out what she calls the AI stack organizations need to get ready: infrastructure such as GPUs, a working data platform, application tooling built on top of models, and security and governance running through all of it. She stresses that Microsoft builds Copilot, not autopilot, keeping a human in the loop and in control. Using a live demo with an AI-generated balloon image, she shows how a model reasons about physics questions and where it still hedges or hallucinates. She closes by saying that readiness for AI requires imagination, which cannot come from a machine, combined with a responsible approach.
Key takeaways
- Gontcharova compares AI to the 1841 invention of the collapsible paint tube: a tool that unleashes human creativity rather than solving problems by itself.
- ChatGPT reached mass adoption in about two months, far faster than the internet or the mobile phone.
- She structures AI readiness as a stack: infrastructure, data platform, application tooling, and security, with policy and education running alongside it.
- Microsoft builds Copilot rather than autopilot so a human stays in control of AI-assisted work such as coding.
- A live demo shows a model reasoning about a balloon image with hedged, probabilistic answers, and that the same question against different data sources produces different results.
Detailed notes
A paint tube for the age of AI
Gontcharova opens with an exercise: naming what was significant about the year 1841. Search engines and copilots surface a US presidential inauguration or an Antarctic expedition, but she points instead to the invention of the collapsible metal paint tube. Before it, painters carried pigment in pig bladders that could not be resealed, so they worked in the studio from memory and sketches. The tube let them paint outdoors directly from life, and she credits it with unleashing the creativity behind impressionism and painters such as Monet, Renoir and Pissarro. Her point: AI, like the paint tube, does not solve human problems on its own. It is a tool that unleashes human creativity, and the imagination and judgment still belong to the person holding it.
Adoption is moving faster than any prior technology
She contrasts adoption speeds: it took the mobile phone about 16 years to reach 100 million users, the internet took less time, and ChatGPT reached comparable adoption in about two months. She frames this as a democratization of technology, comparable to how smartphone cameras let anyone create content without being a trained painter. She cites Microsoft investing 2.3 billion euros in AI in Germany and a large skilling initiative aimed at people in the country, though she is visibly unsure of the exact figures while speaking. She also cites an internal Microsoft study finding that AI technology investment pays back within roughly 14 months, which she frames as a deliberate choice companies in tight financial situations have to make.
The AI stack organizations need to get ready
Gontcharova breaks AI readiness into layers: infrastructure (GPUs and InfiniBand networking, from vendors such as AMD or Nvidia); a data platform, where she says few organizations manage to connect the people who build applications with the people who hold the data; AI platforms and toolkits for building and extending applications without being limited by scarce development resources; and security, which she says is non-negotiable because the same generative tools that create value can be used by bad actors to run scams. She notes GitHub Copilot can write code but is designed to work with a developer, not replace one, and that Microsoft deliberately calls its products copilots rather than autopilots for that reason. She adds that government policy and guardrails, plus education from K-12 through universities, have to be part of the same conversation, since most people preparing for this shift will not have studied computer science.
Demonstrating reasoning with a balloon image
She walks through a live example of asking Copilot to reason about a DALL-E generated image of balloons on strings: what would happen if the strings were cut, what gas is inside the balloons, and whether hydrogen would work at a barbecue. The model correctly infers the balloons would float away due to buoyancy, judges the gas is unlikely to be air and more likely helium or hydrogen, and warns against using hydrogen near a barbecue because of its flammability. She uses this to show the model reasoning with uncertainty ('unlikely', 'I'm not sure but') rather than stating facts outright, while noting that hallucination remains a real risk in these systems.
Different data produces a different answer
She recounts asking Copilot to analyze her own communications and say which superhero she resembles, getting 'Iron Man'. When a colleague asked the same question of their own data, they got the same answer, which she found unconvincing. Feeding the model a different dataset, a fandom wiki, changed her result to 'the Oracle from The Matrix'. Her conclusion: the answer a model gives depends entirely on the data connected to it, so organizations have to work out deliberately what data their AI applications should draw on rather than expecting a standard model alone to produce useful results.
Closing message
Gontcharova closes by saying no single company or country can do this alone and that partnership is required across the ecosystem. She summarizes what readiness for AI requires: imagination, which cannot come from a machine, and a responsible approach, framing these as the two things that let people 'paint the future' the way the paint tube let Monet, Renoir and Pissarro paint theirs.
Quotes
It's not the AI that solves the problem, it's us, human, holding this paint tube.
Irina Gontcharova
You need imagination, a lot of it, because that's nothing that can come from machine, and you need a responsible approach.
Irina Gontcharova
Questions answered in this talk
How fast did ChatGPT reach mass adoption compared to earlier technologies?
According to Irina Gontcharova at Rise of AI 2024, ChatGPT reached comparable adoption levels in about two months, compared with roughly 16 years for the mobile phone to reach 100 million users.
What layers does an organization need to get ready for AI, according to Microsoft?
Irina Gontcharova outlines infrastructure such as GPUs and networking, a working data platform, AI application tooling, and security and governance, on top of which organizations need imagination and a responsible approach.
Why does Microsoft call its AI products 'copilot' instead of 'autopilot'?
Irina Gontcharova explains that Microsoft deliberately keeps a human in control of tools like GitHub Copilot, using them to assist a developer rather than replace one.