Key Moments
AI Dev 26 x SF | Barun Singh & Kennith Jackson; The Hidden Cost of AI Velocity and AI Agents
DeepLearning.AIWant to know something specific about what's covered?
We've already dissected every moment. Ask and we will deliver (with timestamps).
Key Moments
AI coding assistants offer productivity boosts but may increase technical debt and erode developers' core skills. "Vibe coding," while fast for prototypes, risks creating unmaintainable systems.
Key Insights
AI agents can accelerate specific parts of the development chain, but organizational processes and bureaucratic structures need to evolve to realize overall value stream productivity gains.
The speed of AI model progress is unpredictable, but newer models are showing improved ability to solve bugs and produce less debt than older ones, as evidenced by an experiment where Opus and Codeex 54 solved a complex front-end issue faster than Sonnet.
Experienced engineers are crucial for assessing risk in AI-driven development, as the outcome of AI interventions can be non-deterministic and require human oversight.
The concept of an "AI Engineer" is likely a short-lived term, similar to early personal computer specialists; future engineers will inherently use AI as a standard tool, akin to how accountants now use computers.
The future of AI in development may lean towards "supervised agents," where AI acts as a collaborative colleague, a model that balances productivity with oversight for production systems.
To stay sharp, developers should prioritize learning and growth over immediate productivity gains, as chasing learning fosters adaptability in a rapidly changing technological landscape.
The illusion of immediate productivity gains
While AI coding assistants like GitHub Copilot and others are widely adopted (with nearly everyone in the audience reporting daily use), the promise of widespread productivity gains is not yet fully realized. Barun Singh explains that AI excels in specific parts of the development process, creating pockets of speed. However, bottlenecks persist in areas like product definition, QA, code review, and go-to-market strategies. Organizations have not yet adapted their bureaucratic structures to fully leverage these AI-driven increases in speed. Furthermore, current productivity metrics often focus on short-term outputs, such as the number of pull requests or features shipped. Singh predicts that the speed gained now will likely accrue significant technical debt, with a reckoning expected in the next 12-24 months as teams are forced to refactor code that was rushed, possibly leading to lower long-term productivity. The core issue lies in organizations not yet evolving their processes and structures to match the AI's speed.
AI's race against technical debt
The rapid advancement of AI models raises the question of whether future iterations will automatically solve the technical debt accumulated by current AI-generated code. While definitive predictions are impossible due to the unpredictable growth curve of AI models, there's evidence that newer models are performing better. In an experiment, Opus and Codeex 54 resolved a complex front-end bug within 10-13 minutes, whereas Sonnet struggled and gave up after 30 minutes. However, even advanced models can get stuck in "thinking loops" or "rationalizing" without reaching a solution, especially in large, complex codebases with years of accrued debt. There's a risk that future models might not uniformly solve all these issues. This leads to a critical point: developers must engage in risk assessment, as relying solely on future AI capabilities to fix present-day code "slop" is a dangerous gamble. The human engineer's judgment remains paramount in evaluating and mitigating these risks.
The overblown narrative of autonomous AI agents
The idea that engineers will soon become mere orchestrators of autonomous AI agents, simply writing a spec and receiving working software, is considered "hugely overblown." The analogy of cooking a meal or building a house illustrates why: human creativity and problem-solving involve iterative refinement and adaptation. You can't perfectly spec out a complex system like software without experiencing it. If a spec could be defined with absolute certainty, writing the code would be trivial. Trying to spec out an entire complex system upfront and then having AI build it without intermediate human feedback is inefficient, costly, and exacerbates cognitive load. It's akin to building a house and then deciding to tear down walls for design changes. While this spec-driven approach might work for internal tools or demos where iteration is rapid and stakes are low, it's ill-suited for foundational production software that requires continuous development and building upon.
Reframing the 'AI Engineer' role
The term 'AI Engineer' is seen as a transient label. Barun Singh likens it to the early days of personal computers when specialized roles like 'accountant who uses a computer' existed. Today, using computers is a baseline skill. Similarly, AI is becoming an embedded aspect of all engineering. While there's a niche for building and tuning AI systems (like RAG pipelines or fine-tuning models), this demand has decreased as these capabilities are commoditized by SaaS offerings and agentic engineering takes center stage. Andela has identified three archetypes: prototypers (rapid iteration), builders (production systems), and scalers (enterprise deployment). However, Singh suggests these are fuzzy distinctions, and a good engineer often performs aspects of all three. The emphasis should remain on core engineering skills, with AI being a tool, not the defining characteristic of a role that will likely disappear as a distinct title.
Skills that endure in the AI era
When hiring, Barun Singh prioritizes three enduring qualitative skills: 1. **Understanding and modeling the world:** The ability to conceptualize product requirements, architect systems, and grasp complex code structures. 2. **Execution with AI:** Competence in writing code independently, coupled with the skill to use AI tools productively for faster and more effective development. Both aspects must be assessed. 3. **Effective communication:** Articulating work through clear commits and PR descriptions, enabling others to understand and build upon the code with minimal debt. This skill signifies an expert's ability to explain complex topics at various levels of abstraction, which is crucial for self-critique, growth, and elevating team performance.
The spectrum of 'Vibe Coding' and AI interaction
The current hype cycle includes 'vibe coding,' which ranges from non-technical users leveraging AI to build software (akin to a modern website builder) to engineers interacting with AI assistants like Claude or Copilot via human language. While 'vibe coding' can be highly effective for rapid prototyping, internal tools, or exploring product-market fit in low-stakes environments, it's not recommended for foundational production systems due to potential technical debt. Singh proposes a spectrum: 'Artisanal coding' (fully hands-on, minimal AI), 'Supervised agents' (AI as a collaborative colleague, reviewing code), and 'Trusted agents' (broader team autonomy). For production systems, 'supervised agents' currently offers the best balance of productivity and oversight.
Prioritizing learning over output
To remain sharp and avoid skill degradation, developers should focus on continuous learning rather than chasing short-term productivity metrics. Immediate productivity gains can often lead to long-term costs, like technical debt. The key is to approach work with an open mind, acknowledging the exponential pace of technological change. This means constantly questioning existing skills and embracing new learning opportunities. Singh suggests framing personal development around questions of living meaningfully, contributing value, evolving as a human, and continuous learning. The pursuit of learning itself is intrinsically rewarding and also happens to drive productivity. The advice is to chase learning, and productivity will follow, especially when approaching change with awareness and openness.
Bridging the gap between executives and developers
Bridging the gap between executive pressure for immediate AI-driven productivity and developers' nuanced understanding requires empathy and clear communication. Executives are driven by a fast-changing market, competitive pressures, and fear of missing out, often fueled by market hype and online claims. Developers, in turn, need to be aware that their own long-held expertise might blind them to the present's rapid changes. Patient, consistent communication is key. Developers should aim to understand the executive perspective while also educating them on the nuances of AI implementation, like the potential for technical debt and the need for careful risk assessment. Ultimately, trusting the process, focusing on execution, and maintaining an open mind are crucial for navigating the hype cycle and aligning with the eventual reality of AI's impact, which will undoubtedly be transformative despite potentially exaggerated short-term claims.
Mentioned in This Episode
●Software & Apps
●Companies
●Organizations
●Concepts
●People Referenced
Common Questions
While AI assistants are improving speed in specific parts of the development chain, overall productivity gains are not yet fully realized. Bottlenecks in other areas, process evolution, and the potential for increased technical debt need to be considered.
Topics
Mentioned in this video
Kenneth Jackson previously worked here for over 15 years, building its professional services and AI solutions business.
An AI coding assistant mentioned as being used by most attendees daily.
An AI model that was part of an experiment to solve bugs in a codebase, taking 10 minutes to find a solution.
An AI development tool or platform mentioned in the context of 'vibe coding', similar to a website builder.
Mentioned as an AI tool used in 'vibe coding' where the engineer communicates with the machine about technical decisions at a higher level of abstraction.
An AI coding assistant mentioned as being used by most attendees daily.
An AI model that was part of an experiment to solve bugs in a codebase, but gave up after 30 minutes of confusion.
An AI talent marketplace that powers AI transformation for global enterprises through its talent network, with a global ecosystem of over 150,000 technologists across 135 countries.
Barun Singh's team won a hackathon at OpenAI a couple weeks prior to the recording, using 'vibe coding' for 4 hours.
Mentioned as a benchmark CRM system against which 'vibe coded' systems are compared, and also as a system that competitors claim to be able to replace quickly with AI agents.
Mentioned as a platform where claims about rapid AI development and productivity gains are often posted, and where the term 'AI engineer' was the fastest growing job description in 2025.
Chief Product and Technology Officer (CPTO) at Andela, with two graduate degrees in AI from MIT and over 25 years of software development experience. He is described as a builder at heart and an expert in the field.
SVP of AI solutions and operations at Andela, previously spent over 15 years with Google Cloud building its professional services and AI solutions business.
More from DeepLearningAI
View all 94 summaries
29 minAI Dev 26 x SF | Paul Everitt: The Shift to Agentic Engineering
26 minAI Dev 26 x SF | Brandon Waselnuk: Building the Context Engine AI Agents Need
27 minAI Dev 26 x SF | Diamond Bishop: The Next 100 Agents. Building the Agent Native Office
32 minAI Dev 26 x SF | Jerry Liu: My Agent Can't Read a PDF?
Ask anything from this episode.
Save it, chat with it, and connect it to Claude or ChatGPT. Get cited answers from the actual content — and build your own knowledge base of every podcast and video you care about.
Get Started Free