Uploaded November 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 Reasoning Models & Chain-of-Thought Breakthrough
The emergence of models like OpenAI's o1/o3 series and all the other commercial and open source reasoning models represents an important evolution in AI capabilities. These systems can plan, reflect, self-correct, and work over extended time horizons. Adding thinking tokens allow the models to spend more compute and time on a problem before giving an answer, increasing the overall capacity of the system to solve complicated problems.
Why this matters: Thinking models can now write much better software, solve some science and math problems, complete more complex and long agentic tasks, and much more. And all of this without a significant change in the underlying architecture of the neural network, as the thinking capability is added in the reinforcement learning, post-training stage.
#2 AI Agents & Vibe Coding
AI has shifted from answering questions to actually doing things. Salesforce's agents handle 380,000 support conversations with 84% resolution rates. Some enterprises report 50% efficiency gains. But the industry is waking up to reality—these systems are more brittle and harder to monetize than we thought.
Why it matters: AI now tackles repetitive tasks on its own and handles complex workflows with some oversight. Take "Vibe Coding"—coding agents can build sophisticated apps with minimal supervision, so less technical people can create prototypes and personal apps. You still need experts to bring things into production with proper code and security, but the barrier to entry has dropped significantly.
#3 Dramatic Efficiency Gains
AI tech hasn't changed dramatically in the last year. We're still in the scaling era where more parameters and compute generally make better models, although we are starting to see systems like Alpha Evolve that mix pure gradient-optimized systems with symbolic systems. But incremental improvements matter. Microsoft's Phi-3-mini hit a 60% MMLU score with just 3.8 billion parameters. A year ago, you'd need twice as many parameters for that performance. Two years ago you needed over 100 times more.
Why it matters: Smaller models mean capable AI runs on consumer hardware and high-end phones. The original ChatGPT's capabilities now work locally on a decent GPU. Some smaller and efficient models can even run on edge devices and robots, bringing us closer to AI that acts in the physical world, not just the digital one.
#4 Open-Weights Models Closing the Gap
The performance difference between open-weight and closed models shrank significantly in the last year. DeepSeek, Qwen, Alibaba, Mistral, Kimi, Meta, and many other players publish models that are genuine competitors, challenging US commercial player dominance and creating alternative AI development paths.
Why this matters: This prevents any single company or country from having a monopoly on advanced AI. It accelerates innovation through competition, provides alternatives for organizations concerned about vendor lock-in or with significant privacy concerns, and distributes AI capabilities globally.
#5 Video and Image Generation and Editing
2025 marked the leap from uncanny animations to stunning, audio-complete videos. Systems like Sora 2, Veo 3, and open source LTX2 now generate high-quality, controllable video content at scale, with capabilities for editing characters, objects, and styles. Meanwhile, image generation reached new heights with tools like Nano Banana and Qwen Image Edit enabling text-driven creation and modification.
Why this matters: Video and images are the most information-dense and engaging content format. AI-generated video and images impact entertainment, education, marketing, content creation, and — critically — provides synthetic data for training embodied AI and robotics systems.
#generativeai #artificialintelligence #udacity
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 Reasoning Models & Chain-of-Thought Breakthrough
The emergence of models like OpenAI's o1/o3 series and all the other commercial and open source reasoning models represents an important evolution in AI capabilities. These systems can plan, reflect, self-correct, and work over extended time horizons. Adding thinking tokens allow the models to spend more compute and time on a problem before giving an answer, increasing the overall capacity of the system to solve complicated problems.
Why this matters: Thinking models can now write much better software, solve some science and math problems, complete more complex and long agentic tasks, and much more. And all of this without a significant change in the underlying architecture of the neural network, as the thinking capability is added in the reinforcement learning, post-training stage.
#2 AI Agents & Vibe Coding
AI has shifted from answering questions to actually doing things. Salesforce's agents handle 380,000 support conversations with 84% resolution rates. Some enterprises report 50% efficiency gains. But the industry is waking up to reality—these systems are more brittle and harder to monetize than we thought.
Why it matters: AI now tackles repetitive tasks on its own and handles complex workflows with some oversight. Take "Vibe Coding"—coding agents can build sophisticated apps with minimal supervision, so less technical people can create prototypes and personal apps. You still need experts to bring things into production with proper code and security, but the barrier to entry has dropped significantly.
#3 Dramatic Efficiency Gains
AI tech hasn't changed dramatically in the last year. We're still in the scaling era where more parameters and compute generally make better models, although we are starting to see systems like Alpha Evolve that mix pure gradient-optimized systems with symbolic systems. But incremental improvements matter. Microsoft's Phi-3-mini hit a 60% MMLU score with just 3.8 billion parameters. A year ago, you'd need twice as many parameters for that performance. Two years ago you needed over 100 times more.
Why it matters: Smaller models mean capable AI runs on consumer hardware and high-end phones. The original ChatGPT's capabilities now work locally on a decent GPU. Some smaller and efficient models can even run on edge devices and robots, bringing us closer to AI that acts in the physical world, not just the digital one.
#4 Open-Weights Models Closing the Gap
The performance difference between open-weight and closed models shrank significantly in the last year. DeepSeek, Qwen, Alibaba, Mistral, Kimi, Meta, and many other players publish models that are genuine competitors, challenging US commercial player dominance and creating alternative AI development paths.
Why this matters: This prevents any single company or country from having a monopoly on advanced AI. It accelerates innovation through competition, provides alternatives for organizations concerned about vendor lock-in or with significant privacy concerns, and distributes AI capabilities globally.
#5 Video and Image Generation and Editing
2025 marked the leap from uncanny animations to stunning, audio-complete videos. Systems like Sora 2, Veo 3, and open source LTX2 now generate high-quality, controllable video content at scale, with capabilities for editing characters, objects, and styles. Meanwhile, image generation reached new heights with tools like Nano Banana and Qwen Image Edit enabling text-driven creation and modification.
Why this matters: Video and images are the most information-dense and engaging content format. AI-generated video and images impact entertainment, education, marketing, content creation, and — critically — provides synthetic data for training embodied AI and robotics systems.
#generativeai #artificialintelligence #udacity
