Key Concepts
AI-first company, anatomy of an AI team, generalist vs. specialist, upskilling/reskilling, hiring strategy, Ampere's wager, inner/outer loops, domain expertise, technology adoption, bottlenecks, MLOps platform, model training, model serving, business acumen.
Anatomy of an AI Team
The speaker, Dennis Linkov, emphasizes that building an effective AI team requires understanding the type of company:
- Technology Companies: (e.g., Big Tech, startups) Technology is the core value proposition. Challenges often involve a lack of domain knowledge and business misalignment. They typically buy data or expertise.
- Verticalized Solutions/Services Companies: (e.g., Palantir, Wisdocs) Everything either goes right or awfully poorly.
- Tech-Enabled Companies: (e.g., banks, retailers, SMBs) The core product isn't technology, but benefits from it. Challenges often involve technology limitations. They typically buy technology through service providers or turn-key solutions.
Linkov argues that technology isn't the primary limitation to success; it's how technology is used. He cites examples like the continued use of fax machines, contact-based payments, checks, and the slow adoption of electronic medical records to illustrate this point. The core question is: "Is technology the limitation of our success?"
He cautions against blindly hiring AI researchers without a clear need or scale. Transformation work is often necessary before advanced model work becomes valuable. However, for companies like OpenAI or Anthropic, top-tier AI researchers are essential.
Ampere's Wager: A thought experiment: Would you trade your existing team for five researchers from top AI labs, even with added incentives? This highlights the importance of domain knowledge and practical skills over pure research expertise in many contexts.
An AI team's responsibilities are broad, including defining use cases, integrating with products, measuring ROI, finding data, testing workflows, building interfaces, selling the product, and ensuring customer adoption. Success requires a comprehensive team, not just individual AI researchers.
It's crucial to identify the company's bottleneck (e.g., shipping features, acquiring/retaining users, monetization, scalability, reliability) and prioritize hiring accordingly. The key takeaway is to understand what kind of team you need.
Evolution of a Generalist
In 2021, Linkov built his first machine learning team at a conversational AI company, focusing on generalists supported by automation. The mandate was "We want AI," which translated into goals like serving hundreds of thousands of concurrent models, multi-domain support, low cost, and real-time training/serving.
The team built a custom MLOps platform and focused on fine-tuning encoder models and building RAG as a service. The team owned six microservices on tend.
Three key areas of focus were:
- Model Training: Prioritized general knowledge of model architectures, encoder fine-tuning, and basic data engineering using Hugging Face.
- Model Serving: Leveraged abstractions to minimize the need for deep Kubernetes expertise, focusing on understanding trade-offs.
- Business Acumen: Emphasized engineers' ability to interact with customers and understand their needs.
In 2024, building another team, Linkov noted the advancements in open-source tools and commercial models. This shifted priorities, with commercial APIs and prompt tuning becoming more important. The domain knowledge bar also increased due to the specific needs of medical record processing.
The key is to balance skills and budget, recognizing that skills can be distributed across multiple team members.
What if you already have a team? Focus on reskilling and upskilling. Think about inner and outer loops.
- Inner Loop: Daily activities essential for the team's success (e.g., model training, prompting, product requirements, model serving, domain expertise, building business cases).
- Outer Loop: Broader activities that differentiate the team (e.g., technical expertise, domain expertise).
A weak technical inner loop hinders execution, while a weak domain outer loop prevents product-market fit.
Generalists are valuable at the beginning of an AI strategy, helping to find product-market fit. Specialists become necessary later to push for incremental performance gains. Generalists are adaptable and often "good enough" for many tasks.
Upskilling, Reskilling, and Hiring
Three key areas for upskilling:
- Learn to Build: Move from static requirements to functional prototypes.
- Become a Domain Expert: Domain experts should write use cases and work directly with LLMs.
- Be Human-Facing: Engineers should participate in customer calls.
Linkov's team has weekly 30-minute learning sessions to stay updated.
When to Hire: Hire when you need to hold context and act on context. Humans are needed for accountability and expertise.
Who to Hire: Know your team composition and needs to set a budget. Avoid blindly following trends (e.g., dismissing junior engineers). Verify trends and think from first principles.
Ask relevant questions during the hiring process, avoiding irrelevant LeetCode-style questions.
Reiterating Ampere's Wager: Prioritize domain expertise, sales skills, and customer empathy over pure research talent, depending on the company's needs.
Conclusion
The key takeaways are:
- Define the team you need to win: Understand your company type, bottlenecks, and required skills.
- Cross-functional teams are evolving: The overlap between roles will increase as AI becomes more integrated.
- Continuous learning is essential: The AI landscape is rapidly changing, requiring constant upskilling.
AI summaries can miss context or contain errors. Check important details against the original video.
