The Fifth Elephant - Pune Edition

Other possible talk titles:

• What Do Engineers Do When the Machine Writes the Code?
• Learning to Work with Something That Almost Works
• Old disciplines for new machines
• Mise en place pour machines capricieuses

Mechanical engineers aren’t assemblers—they set up factories. Civil engineers aren’t construction workers—construction companies do the construction. Similarly, creating software should go beyond producing code—we should cultivate the sociotechnical system that produces the code. While building StoryMachine, our experimental open source tool that helps PMs and Engineers cut out “common sense” units of work, we have uncovered a lot more about what it means to build software with on-demand, stochastic and jagged machine intelligence.

In this session I’ll talk about how to refactor your engineering team to accommodate this new reality. I’ll make a case for:

• Tightening the collaborative loop between engineering and product—what this looks like in practice when you have AI involved (”Taste and Adjust”/”The value is in the conversation”)
• How Sutton’s bitter lesson calls for radical simplicity and mise en place thinking (don’t fight the chef, just set up their kitchen)
• Why scientific thinking and basic statistical literacy is table stakes in a world where we have to deal with stochastic outputs from AI (how to avoid data and benchmarks leading you astray)

All of this will be backed up by our experience building and evaluating StoryMachine—along with a decade’s worth of lessons from deploying complex software engineering projects at nilenso.

Takeaways

• The main takeaway will be our high-confidence recommendations for all the new skills that Software Engineers and Engineering Leaders need to learn in order to work in a future dominated by use of machine intelligence.
• Specifically, I will cover this triad: the essence of iterative refinement (Agile, OODA whatever you want to call it) and why it is more powerful today, how to design an environment that gets out of the way of AI agents and how to adapt your thinking to account for risk and uncertainty—and why this triad is actually nothing new at all.

References

Links that contain about some of the ideas I’ll bring up in the talk. (actual slides to come soon)

Taste and Adjust

Minimum Viable Benchmark

Artisanal shims for the bitter lesson age

The quality of AI-assisted software depends on unit of work management

Bio

Atharva, a member of nilenso, has been tinkering with LLMs to figure out how to build products that deliver on the hype in production.

His first exposure to serious software development was with the Git project, where he rewrote parts of the submodule functionality by emailing patches to the maintainer. Since then, he’s been fascinated with the sociotechnical dynamics of building software.

While at nilenso, he helped a hyperlocal delivery startup revamp their payout systems, solved data integration challenges at a non-profit building population-scale software and helped build voice AI agents in production.