Notes on building in AI in 2024
Learnings on building and scaling our speech products
As we navigate the AGI race, the AI landscape continues to evolve at a breathtaking pace. Yet beneath the headlines and hype, there are deeper currents shaping how we build and scale AI companies. Let me share some observations from the trenches.
The Quality Conundrum
In today's AI ecosystem, there's an almost frenzied obsession with latency optimization and orchestration. Companies are racing to shave milliseconds off response times, building increasingly complex systems to handle requests. But in this rush for speed, something crucial is often lost - quality.
This tension between speed and quality becomes particularly evident in real-world applications. While many solutions focus on rapid outputs, the nuances of quality - be it in speech, code, or content generation - often get lost in translation. Each incremental improvement in quality requires exponentially more computational effort and data curation. When working with enterprises, we've discovered that "good enough" for casual use becomes insufficient for professional-grade output.
Our journey at Dubverse through 100,000+ GPU hours of training revealed something crucial: quality in speech synthesis isn't linear. Each incremental improvement in naturalness requires exponentially more computational effort and data curation. When working with production houses, we discovered that acceptable quality for casual content becomes insufficient for broadcast-grade output. The models need to capture not just pronunciation, but the emotional resonance that makes speech compelling.
This pattern extends beyond just speech synthesis - it's the reality for most AI applications today. At the end of the day, it's not about fancy optimizations or complex systems. It's about building something that just works, really well.
Quality isn't just a metric, it's what keeps founders up at night.
The Enterprise Edge
The true value in AI isn't in flashy demos or consumer applications - it's in the enterprise trenches where real problems meet scalable solutions. While generative AI has pushed capabilities from 80% to 85% in most use cases, the real opportunity lies in that critical last mile of customization and integration.
The numbers tell a compelling story. Enterprise AI spending surged to $13.8 billion in 2024, marking a 6x increase from 2023's $2.3 billion. But what's more interesting is where this money is going - 60% comes from innovation budgets, while 40% is now sourced from permanent operational budgets. This shift signals a fundamental change: AI is moving from experimental to essential.
At Dubverse, this reality became clear as we worked with enterprises. While our consumer growth to 2M users was exciting, the real transformation happened when we started working with production houses. Their needs weren't just about basic voice generation - they required broadcast-grade quality, consistent output across languages, and deep integration with existing workflows. This journey taught us something crucial: enterprise value isn't just about technology; it's about solving specific, complex problems that directly impact business outcomes.
The inefficiencies being tackled are massive. Organizations have identified an average of 10 potential use cases for AI transformation, with 24% prioritized for near-term implementation. These aren't just cost-saving measures - they're fundamental reimaginings of how work gets done. From automated documentation in healthcare to intelligent content localization in media, AI is filling gaps that traditional solutions couldn't touch.
But here's what makes the enterprise opportunity truly compelling: it's not about replacing humans or automating everything. It's about finding those critical points where AI can remove friction, enhance capabilities, and enable humans to focus on higher-value work.
The Human Element
The market has spoken decisively on one key point: tools that enhance human capabilities rather than replace them entirely are winning the race. This isn't just idealistic thinking - it's backed by usage patterns and revenue numbers. Look at the success of development tools like Cursor versus fully autonomous solutions like Devin - humans want to remain in the loop, directing and refining AI outputs rather than being sidelined by them. In healthcare, ambient scribing solutions like Abridge work alongside doctors, turning natural conversations into structured documentation rather than forcing physicians to adapt to rigid interfaces.
What's emerging is a new paradigm of human-AI collaboration, where the technology adapts to human workflows rather than forcing humans to adapt to technological constraints. The best AI implementations remember this simple truth: the best tool isn't the most powerful or the most autonomous - it's the one that a human can use most effectively to achieve their goals(Turns out humans actually like being useful. Surprising, I know. Maybe we should keep them around? 😉). It's about understanding the natural rhythms of human work and building tools that flow seamlessly into these patterns.
The Infrastructure Gap
The concept of reverse saliency - where certain components lag behind in technological development - is particularly evident in AI today. While foundation models advance rapidly, critical infrastructure components are struggling to keep pace. This creates immediate opportunities for founders building in AI:
Data Processing Infrastructure: Tools for efficient data ingestion, cleaning, and preparation at scale
Deployment and Monitoring: Solutions for model deployment, performance monitoring, and drift detection
Integration Layer: Tools that help enterprises integrate AI into existing workflows and systems
Resource Optimization: Solutions for managing compute costs and optimizing model inference
Compliance and Governance: Tools for model auditing, bias detection, and regulatory compliance
For founders looking to build in AI, these gaps represent significant opportunities. The winners won't necessarily be those building the next large language model, but those solving these critical infrastructure challenges.
Global Perspective
The AI landscape varies dramatically across different markets. While everyone's excited about AI, the reality is that market readiness differs significantly. Building in mature markets often provides clearer paths to revenue and adoption. As the saying goes, "Earning a dollar is easier than a rupee" - not because of currency conversion, but because of market maturity and adoption readiness.
Data: The Eternal King
Facebook AI Chief Yann LeCun introduced his now-famous “cake analogy” at NIPS 2016: “If intelligence is a cake, the bulk of the cake is unsupervised learning, the icing on the cake is supervised learning, and the cherry on the cake is reinforcement learning (RL).” The real revolution in AI isn't just about model architectures or compute - it's about how we structure, process, and feed data into these systems. From what we've seen in production, successful AI implementations aren't built on massive labeled datasets anymore, but rather on carefully curated, contextually-rich data that enables models to learn patterns autonomously.
The architecture of modern AI systems reflects this reality. Vector databases for efficient similarity search, sophisticated RAG implementations for context retrieval (which jumped from 31% to 51% adoption in enterprises), and specialized ETL pipelines for handling unstructured data - these aren't just infrastructure components, they're the foundation of reliable AI systems. In a landscape where everyone has access to similar foundation models, your data infrastructure and curation strategy become your true competitive moat.
The Road Ahead
Despite the challenges, there's never been a better time to build in AI - if you're in it for the long haul. The landscape is rapidly evolving, presenting opportunities in infrastructure, vertical applications, and enterprise solutions. As YannLeCun often emphasizes, building serious AI capabilities requires more than just technical talent - it needs a robust ecosystem of research, infrastructure, and industry collaboration.
The key to success isn't just about building impressive technology - it's about delivering real value. Whether you're solving infrastructure challenges, building vertical applications, or focusing on enterprise integration, the fundamentals remain the same: focus on real problems, invest in quality data, and build for the long term. The companies that will thrive are those that can move beyond the hype and focus on building sustainable, valuable solutions that make a real difference in how people and businesses work.
