Uploaded December 2025 | Updated September 2026, 2 weeks ago
Learn more about Udacity's refreshed GenAI Nanodegree here 👉 udacity.com/course/generative-ai--nd608
Register for a free live webinar with GenAI experts here 👉 https://riverside.fm/webinar/registration/eyJzbHVnIjoicGF0cmljay1kb25vdmFucy1zdHVkaW8teG54RngiLCJldmVudElkIjoiNjkyMGI5MDgxYWM5MDg5MDg1ODBmZmU0IiwicHJvamVjdElkIjoiNjkyMGI5MDhkN2Y3MjgyYmJlMDQwYTcyIn0=
TRANSCRIPT
1. AI equals LLM
When people think about AI today, ChatGPT usually comes to mind. But here's the thing: Large Language Models are just one slice of a much bigger pie.
AI actually includes things like computer vision systems analyzing medical images and monitoring traffic. It includes the recommendation algorithms behind your streaming services. It covers reinforcement learning agents optimizing supply chains, and robotics systems navigating warehouses.
LLMs are impressive, no question. They've made huge strides in understanding and generating language. But they're not the whole story. Most people don't know that most of the AI running in production right now consists of smaller, specialized models built to do one thing really well, rather than trying to do everything. So LLMs are AI, but AI is not only LLMs.
2. Generative AI solutions are turn-key solutions
The pitch sounds great: just plug in some AI and watch the magic happen. But anyone who's actually implemented these systems knows better.
Getting gen AI to work reliably takes serious prompt engineering. You iterate, test, refine, and repeat. Models behave differently between versions and providers, so you're constantly adapting. You need evaluation frameworks to catch problems and make sure outputs meet your standards.
What looks great in a demo can fall apart the moment you try to scale. You need testing, iteration, you need to gain expertise in how models behave, and continuous feedback loops. You shouldn't be afraid of this work, it is not rocket science, but it takes resilience, clarity of intent and rigor.
Underestimating this engineering work is probably the main reason why many AI deployments fail.
3. You don’t need software engineers anymore
"AI can code, so why hire developers?" It's a tempting thought, especially when tools like Cursor and Claude let non-technical people build working prototypes and personal apps, or help aspiring developers learn coding much faster than before.
I am personally very excited by this democratization. More people can turn ideas into reality, people who didn’t think they could code are now enjoying building software and web apps. Professional engineers are also more productive and enjoy their work more, as measured by multiple sources.
However, there's a massive gap between a working prototype and production-ready software. Professional engineering means understanding architecture that scales, security practices that protect users, and design patterns that keep systems maintainable over time. It means knowing the trade-offs between different approaches for your specific context.
4. Introducing AI is about introducing AI tools
Here's a striking stat from McKinsey: 88% of organizations use AI, but only 39% report any efficiency impact whatsoever, and just 6% see more than a 5% gain in efficiency. Most organizations are stuck in what you might call "pilot purgatory." They run experiments that never scale. Sound familiar? It's the same pattern we saw with Machine Learning and Data Science before LLMs, where initiatives rarely made it to production. The organizations actually seeing results do things differently. They combine strategic vision with education and grassroots experimentation. They give teams closest to the work space to learn and explore problems they deeply understand. Successful AI adoption isn't just about deploying tools. It requires organizational change and fostering a culture of experimentation and a growth mindset. It also requires the entire organization to be able to move faster, including legal and compliance.
5. Generative AI understands things like humans do
AI produces impressive outputs. But recent research from IBM and MIT shows it's processing information very differently than we do.
AI systems can generate great results without actually developing world models—those coherent understandings of how things work and relate to each other that we humans develop growing up. Humans develop understanding through physical interaction with the world. AI lacks this embodied grounding that connects symbols to meaning.
These systems are excellent at recognizing patterns in data. But they operate through statistical associations, not human-level understanding of reality.
This matters because when AI fails, it fails in ways no human would. Missing obvious context. Generating plausible-sounding nonsense. Breaking down when confronted with slight variations of familiar problems.
Learn more about Udacity's refreshed GenAI Nanodegree here 👉 udacity.com/course/generative-ai--nd608
Register for a free live webinar with GenAI experts here 👉 https://riverside.fm/webinar/registration/eyJzbHVnIjoicGF0cmljay1kb25vdmFucy1zdHVkaW8teG54RngiLCJldmVudElkIjoiNjkyMGI5MDgxYWM5MDg5MDg1ODBmZmU0IiwicHJvamVjdElkIjoiNjkyMGI5MDhkN2Y3MjgyYmJlMDQwYTcyIn0=
TRANSCRIPT
1. AI equals LLM
When people think about AI today, ChatGPT usually comes to mind. But here's the thing: Large Language Models are just one slice of a much bigger pie.
AI actually includes things like computer vision systems analyzing medical images and monitoring traffic. It includes the recommendation algorithms behind your streaming services. It covers reinforcement learning agents optimizing supply chains, and robotics systems navigating warehouses.
LLMs are impressive, no question. They've made huge strides in understanding and generating language. But they're not the whole story. Most people don't know that most of the AI running in production right now consists of smaller, specialized models built to do one thing really well, rather than trying to do everything. So LLMs are AI, but AI is not only LLMs.
2. Generative AI solutions are turn-key solutions
The pitch sounds great: just plug in some AI and watch the magic happen. But anyone who's actually implemented these systems knows better.
Getting gen AI to work reliably takes serious prompt engineering. You iterate, test, refine, and repeat. Models behave differently between versions and providers, so you're constantly adapting. You need evaluation frameworks to catch problems and make sure outputs meet your standards.
What looks great in a demo can fall apart the moment you try to scale. You need testing, iteration, you need to gain expertise in how models behave, and continuous feedback loops. You shouldn't be afraid of this work, it is not rocket science, but it takes resilience, clarity of intent and rigor.
Underestimating this engineering work is probably the main reason why many AI deployments fail.
3. You don’t need software engineers anymore
"AI can code, so why hire developers?" It's a tempting thought, especially when tools like Cursor and Claude let non-technical people build working prototypes and personal apps, or help aspiring developers learn coding much faster than before.
I am personally very excited by this democratization. More people can turn ideas into reality, people who didn’t think they could code are now enjoying building software and web apps. Professional engineers are also more productive and enjoy their work more, as measured by multiple sources.
However, there's a massive gap between a working prototype and production-ready software. Professional engineering means understanding architecture that scales, security practices that protect users, and design patterns that keep systems maintainable over time. It means knowing the trade-offs between different approaches for your specific context.
4. Introducing AI is about introducing AI tools
Here's a striking stat from McKinsey: 88% of organizations use AI, but only 39% report any efficiency impact whatsoever, and just 6% see more than a 5% gain in efficiency. Most organizations are stuck in what you might call "pilot purgatory." They run experiments that never scale. Sound familiar? It's the same pattern we saw with Machine Learning and Data Science before LLMs, where initiatives rarely made it to production. The organizations actually seeing results do things differently. They combine strategic vision with education and grassroots experimentation. They give teams closest to the work space to learn and explore problems they deeply understand. Successful AI adoption isn't just about deploying tools. It requires organizational change and fostering a culture of experimentation and a growth mindset. It also requires the entire organization to be able to move faster, including legal and compliance.
5. Generative AI understands things like humans do
AI produces impressive outputs. But recent research from IBM and MIT shows it's processing information very differently than we do.
AI systems can generate great results without actually developing world models—those coherent understandings of how things work and relate to each other that we humans develop growing up. Humans develop understanding through physical interaction with the world. AI lacks this embodied grounding that connects symbols to meaning.
These systems are excellent at recognizing patterns in data. But they operate through statistical associations, not human-level understanding of reality.
This matters because when AI fails, it fails in ways no human would. Missing obvious context. Generating plausible-sounding nonsense. Breaking down when confronted with slight variations of familiar problems.
