At BoS Europe 2025, Greg Baugues, took the stage with a talk that challenged a lot of assumptions about how we think AI works.
Using the quote “any sufficiently advanced technology is indistinguishable from magic”, Greg made a clever connection between what feels magical about AI and what’s actually going on under the hood in his talk “AI You’ve Been Doing It All Wrong”, available now in the BoS talks library.
He showed how understanding that difference helps us use AI in a smarter, more effective way. Get to know more about it:
1. How LLMs Work
Greg kicked off his talk with an experiment: he asked ChatGPT to pick a number between 1 and 50. With the audience watching, it kept choosing the number 27. That wasn’t a glitch, it was the point.
He explained that this behavior highlights something essential about LLMs: they’re basically incredibly advanced autocomplete systems trained on massive amounts of text (around 10 terabytes) to predict the next token, not the next ‘word’ in the way we normally think.
“The difference is, LLMs are not guessing the next word. They’re guessing the next token.“
Unlike traditional computer programs that generate truly random numbers, LLMs have “reasoning mode” now, that is similar to reasoning, following a chain of thought. His point was that, in the case of random numbers, they aren’t running a random number generator as you might assume.
He highlighted that magicians know that when humans are asked to pick a number between one and 50, they often choose a number ending in seven because it ‘feels random’. This distinction is crucial: LLMs are not just computers you chat with. They are fundamentally different.
2. The Impact Of AI Tools
Greg also discussed how LLMs are completely reshaping creative work, doing things in seconds that would’ve felt like straight-up sorcery just a couple of years ago.
He shared two great examples:
• Design and Image Generation: ChatGPT’s image tools have gotten way better recently. Lettering actually looks like lettering now and you can iterate on images while keeping the state of the previous version.
Want a heavy-metal squirrel? A Pixar-style robot? A logo that doesn’t look like it escaped from the uncanny valley? You can get there in a couple of clicks.
Greg makes the point that AI is about to give non-designers “superpowers”, letting anyone mock up ideas fast, try wild variations, and ship concepts without waiting for a full design cycle.
He also suggests that AI will make creative processes accessible to those without design ability, enabling them to rapidly iterate, experiment and ship ideas.
• Translation and Dubbing: LLMs are shockingly good at translation because they’ve been trained on basically every language on the internet. Pair that with tools like Eleven Labs for voice cloning and video dubbing, and suddenly any business can sound fluent in dozens of languages… Instantly.
Greg’s take? Reaching a global audience is about to become so easy and cheap that it’ll stop being a competitive advantage and start being table stakes.
Magic? No. But it sure feels like it.
3. AI Coding Agents and Vibe Coding
For software development, Greg asserted that “coding agents is this one place where I feel like we have unquestionably found product market fit with LLMs”.
He marked the release (still very recent) of ChatGPT (November 30, 2022) as the moment the way he coded completely changed.
“I’ve been writing code my whole life, and I’ve had multiple times over the last two years when, like, stuff like this happened, and I just had to close my laptop and go for a walk and let my brain reconfigure what’s happening there.“
One of his favorite ideas is Andrej Karpathy’s concept of “vibe coding”: basically, stop obsessing over the code and just… Ride the vibes. Describe what you want, embrace the exponential power and let the AI handle the rest.
While vibe coding offers immense speed (allowing everyone to bang out stuff while watching TV), Greg strongly cautioned that it is best for throwaway weekend projects. Senior engineers are right to be wary, as it can lead to security issues and unmaintainable code.
4. AI-Assisted Coding Best Practices
For software developers and businesses employing developers, he shared several best practices for AI-assisted coding:
4.1. Write Tests (by hand): Tests are essential, especially with AI-written code, as LLMs may pass easy tests but fail to hit crucial edge cases.
“You’re going to want to write tests and really understand what’s happening in those tests. I do think that if you are writing tests and you’re comfortable with your tests, and then those tests are passing, you can have a lot more confidence in the code these things are writing, and it makes it easier just to trust and use cursor rules.”
4.2. Use Cursor Rules (System Prompts): Use system prompts to define company styling and conventions, ensuring the LLMs follow specific guidelines.
4.3. Work Small. Small files, small changes: LLMs work better with smaller amounts of context, making languages/frameworks with smaller files and good separation of concerns preferable.
4.4. Commit Often and Revert: Since the cost of code is so much cheaper now, developers should be willing to revert large blocks of AI-generated code if a session is unproductive.
4.5. Train and Hire for AI Tools: Greg stressed that developers who refuse to use AI tools will likely have a hard time getting employed, comparing it to hiring a contractor who refuses to use power tools.
‘I’m not saying everyone becomes vibe coders, but I think that the developers who refuse to use these tools are going to have a very hard time getting employed.’
4.6 Onboarding and Apprenticeships: AI agents serve as an infinitely patient tutor who understands the code base, significantly helping with software development onboarding and relieving the burden on senior developers.
5. Model Context Protocol (MCP)
Greg also touched on the future of AI interfaces and introduced the Model Context Protocol (MCP): basically an interoperability layer that lets LLMs plug into real tools and APIs.
With MCP, an LLM can act as a natural-language wrapper around services like Stripe or Postgres, which Greg demoed live.
His advice for founders and product leaders was clear: start thinking about building an MCP server for your product, because customers will soon expect to interact with your service directly through tools like Claude or the next OpenAI client.
6. Conclusion: AI Is The Electrical Bike Of The Mind
Greg wrapped up his talk with a perfect analogy:
‘If the computer was the bicycle of the mind, AI is the electric bike of the mind.’
Just like an e-bike didn’t just make his family pedal faster, it opened up totally new possibilities, AI is doing the same for his work as a developer. It’s expanding what he can build, how quickly he can explore ideas, and the level of ambition he brings to every project.
He summed it up beautifully: ‘These last two years have been the most fun and the most exciting I’ve ever had coding.’
And it raises the question: What could AI unlock for you?
Watch Greg’s talk and share your thoughts: