Transcription

• Speaker #0

  Welcome to Artifact's podcast on generative AI. In this deep dive, we're focusing on the latest developments for professionals like you already immersed in the Gen AI world, from NVIDIA bringing the power of 200 billion parameter models to your desktop, to AI robots combating loneliness in Japan, to AI assistants potentially revolutionizing SAAS entirely. Get ready for a deep dive into the most recent advancements.

• Speaker #1

  Let's kick things off with the bustling landscape of Gen AI. Acquisitions funding rounds the sheer speed of development. It's all happening at a breakneck pace. Yeah,

• Speaker #0

  keep Perplexity for example. They just secured a massive funding round, $500 million to scale their operations. Think about this from handling 100 million queries monthly to 100 million queries daily.

• Speaker #1

  That hints at the massive opportunity in Gen AI search and it's not just them. Swift Ventures is using LLMs to evaluate AI companies. It's kind of meta moment, AI analyzing its own growth.

• Speaker #0

  I know it's wild, right? Yeah. But what really excites me are the advancements in models and their capabilities. Like Grok might be getting an unhinged mode. It aims to deliver bolder, edgier AI responses.

• Speaker #1

  So imagine a model that's not just providing information, but also expressing opinions, using humor, even challenging your assumptions. It's like we're seeing the emergence of personality in AI.

• Speaker #0

  That's definitely uncharted territory. Then we have DeepSeek V3. It's outperforming giants like GPT-4.0 and Claude Sonnet 3.5. But here's the kicker. It's 10 times cheaper to run.

• Speaker #1

  Yeah, the final training cost was only $5.5 million, compared to the tens or even hundreds of millions others are spending. And on top of that, DeepSeek v3 is fast and efficient. 60 tokens per second with multi-token prediction.

• Speaker #0

  Lower development costs could really shake things up, right? Opening doors for smaller players, potentially disrupting the dominance of big tech.

• Speaker #1

  Absolutely. And then there's Microsoft's Phi 4, a 14 billion parameter model that excels at math reasoning tasks. It even outperforms Gemini Pro 1.5, showcasing how specific models are carving out niches beyond general language.

• Speaker #0

  It's fascinating to see how specialized these models are becoming. But let's talk about real-world applications. The show "L'Amour est dans le Pré" is using Gen-AI-generated images. They're virtually undetectable.

• Speaker #1

  This subtle integration raises some interesting questions, like how do we ensure transparency when AI creations become indistinguishable from human ones? It's a real challenge.

• Speaker #0

  Definitely. And then there's BunkBank. They're now transcribing client conversations in emergencies.

• Speaker #1

  That's a practical use case. But what about the potential implications for privacy and data security? It's a sensitive area.

• Speaker #0

  It definitely is. It's like a balancing act between innovation and protecting sensitive information.

• Speaker #1

  Exactly. And on a lighter note, Elle Arabia just featured a 10-page editorial spread that was entirely AI-driven.

• Speaker #0

  That's incredible. It's exciting to see the creative potential of AI. But what does this mean for human artists and designers? It's a question we need to consider.

• Speaker #1

  No doubt. This technology is impacting every corner of our lives, from professional applications to consumer products. Take CES 2025, for example. It brought us Romi, the AI robot designed to combat loneliness.

• Speaker #0

  I saw that. It's pretty remarkable.

• Speaker #1

  It is. RoMe has over 150 facial expressions. It initiates conversations, greets you when you come home, and it's actually being deployed in Japan for about $570 up front, plus a monthly subscription.

• Speaker #0

  It's more than just a novelty then. This raises interesting questions about our relationship with technology. Is it just a gimmick? Or could AI companions genuinely address the growing issue of social isolation?

• Speaker #1

  It's a valid question. And it's not just robots. LG and Samsung both unveiled smart TVs powered by Microsoft Copilot.

• Speaker #0

  Oh, wow. So your TV goes from passive entertainment to actively managing information, helping you find what you need. It's like your TV could essentially become your new assistant.

• Speaker #1

  Precisely.

• Speaker #0

  It's a whole new way of interacting with their devices. Now let's get a little more technical. Let's discuss retrieval augmented generation or RAG. What are its limitations?

• Speaker #1

  Well, RAG relies on real-time retrieval of documents, which can cause delays and lead to suboptimal answers.

• Speaker #0

  So it's not the most efficient system. So what's the alternative?

• Speaker #1

  Cache augmented generation or CAG. Instead of real-time retrieval, CAG preloads knowledge into a key value cache.

• Speaker #0

  So it's accessing information much faster.

• Speaker #1

  Exactly. We're talking 94% faster in some cases. For example, query time goes from 9 seconds down to 0.8 seconds. And because it's not constantly retrieving data, the architecture is simpler and much more scalable.

• Speaker #0

  So it's a significant improvement over RAG in terms of speed and efficiency.

• Speaker #1

  Yes, definitely. And to make things even more interesting, NVIDIA just unveiled a $3,000 desktop AI computer aimed at home researchers.

• Speaker #0

  Oh, wow. So that puts serious AI power within reach of more people.

• Speaker #1

  Yeah, this machine merges an NVIDIA Blackwell GPU with a 20-core Grace CPU, boasts 128 gigabytes of unified RAM, and can even run models with 200 billion parameters.

• Speaker #0

  So that's approaching the scale of some of the leading models out there.

• Speaker #1

  Exactly, and it comes bundled with NVIDIA's suite of tools for creating AI applications.

• Speaker #0

  So this could be a game changer for independent researchers and smaller companies.

• Speaker #1

  Definitely democratizing AI development in a big way.

• Speaker #0

  Yeah, for sure. On a lighter note, Instagram is planning to replace its AR filters. with AI generated videos.

• Speaker #1

  Interesting. So AI-generated videos offer more dynamic and personalized experiences compared to static filters.

• Speaker #0

  What's the thinking behind that?

• Speaker #1

  It seems to be about boosting user engagement by offering more immersive and creative experiences.

• Speaker #0

  Makes sense. And on the topic of creativity, MidJourney has a new feature called Moodboard.

• Speaker #1

  Oh yeah, Moodboard automatically creates moodboards from images.

• Speaker #0

  So you upload, say, 4 to 30 images reflecting your desired style.

• Speaker #1

  Correct. Then Midjourney's AI analyzes those images to generate a cohesive mood board.

• Speaker #0

  That could really streamline the creative process for designers and artists. Absolutely. That's pretty cool. And Palettes has added new functionalities for remixing 3D pictures.

• Speaker #1

  Oh, really? What does that entail?

• Speaker #0

  You can curate a collection of images, upload them to Palettes, and then adjust the influence of the original sketch on the final output.

• Speaker #1

  Interesting. So you have more control over the balance between the original concept and the AI's creative interpretation.

• Speaker #0

  Exactly. Moving on to enterprise solutions, Google Cloud has launched AgentSpace.

• Speaker #1

  AgentSpace? Okay, so what's that all about?

• Speaker #0

  Well, it's designed to automate search within enterprises through custom agents.

• Speaker #1

  So instead of manually searching through tons of data?

• Speaker #0

  Exactly. You have these AI agents doing the heavy lifting for you.

• Speaker #1

  Gotcha. And are they customizable?

• Speaker #0

  Yeah, they're tailored to specific needs.

• Speaker #1

  And what kind of AI are they using?

• Speaker #0

  They're using Notebook LM, which can generate audio synthesis and analyze complex information.

• Speaker #1

  Wow, that's impressive. So it's not just about retrieving text-based data, but also understanding and processing audio.

• Speaker #0

  Right. And these expert AI agents can automate workflows across various functions, like managing JRE tickets or handling expense reports.

• Speaker #1

  So this sounds like a significant step toward more streamlined and efficient business processes.

• Speaker #0

  It really does. Now, the thoughts of the week?

• Speaker #1

  Okay, so the evolution of AI assistance from embedded help to cross-system agents. The journey from simple AI assistance to powerful cross-system agents reflects a profound shift in how businesses leverage generative AI. In 2023, the rise of embedded gen AI promised seamless integration into daily workflows, enabling users to adopt advanced capabilities without disrupting habits. Yet high cost per user and limited customization left many questioning the value proposition of these black box solutions. By 2024, the narrative had shifted toward professionalization and industrialization. Tools like Microsoft Copilot and Salesforce's AgentForce began to blur the lines between passive assistance and active execution, hinting at a future where agents could do more than just assist. They could act autonomously. However, these systems were still bound by the ecosystems in which they operated. 2025 may mark a pivotal year. The emergence of industrialized agents capable of automating processes across multiple systems. This evolution presents a critical challenge. Most workflows span several platforms and embedding agents in isolated tools inherently limits their potential. Companies face two key options. They could open their platforms to allow broader control of external tools, enabling their agents to orchestrate processes across systems. Alternatively, they might allow their agents to operate outside their own ecosystem, enabling clients to call on these agents externally. A bold shift that would require new business models but could expand the agent's utility dramatically. This trajectory also forces a reconsideration of SaaS business models. As automation takes, center stage interfaces lose their primacy and the true value shifts to the intelligence and data that underpins these agents, data that ultimately belongs to the client. How SaaS platforms navigate this tension between control and openness will define the next era of innovation.

• Speaker #0

  Yeah, it's a fascinating question. Huge implications for SaaS business models as we move toward these More autonomous cross-system agents.

• Speaker #1

  Right. If these agents can operate across platforms, then the value shifts from the interface to the data itself.

• Speaker #0

  Exactly. Control might move from the software provider to the owner of the data.

• Speaker #1

  It's a fascinating dilemma as we move further into this age of intelligent agents.

• Speaker #0

  It really is. Thanks for breaking all this down. It's been an insightful deep dive.

• Speaker #1

  My pleasure. It's an exciting time to be following all of these developments.

• Speaker #0

  Absolutely. We'd like to thank the GenAI team for once again automatically generating us. And we'll see you in two weeks.