Gergely Orosz on AI in development: slow down to speed up (Category AI4SDLC)
Watch Keynote Gergely OroszSlow down to speed up: AI and software engineering” The Craft Conference was held in Budapest. The main idea of Gergely is similar to my theses with highload AI has made it easier to write code, but questions have arisen about the credibility of these AI changes. If the team generates pull requests faster than they can understand, validate, and associate them with the product result, then local acceleration easily turns into system slowdown.
The first part of the speech Gergely analyzes the current state of affairs in the industry Instagram/Meta history with their account takeover incident via Meta AI It wasn’t just a bug, he said, but a symptom of an organization where the focus on AI began to squeeze out safety, reliability and engineering culture. Next, he talks about the current state of affairs at Anthropic and OpenAI (examples of large AI labs), Cursor (Example of a company that makes tools for engineers)Google and Meta (Bigtech with a wide variety of internal tools)as well as Uber (An example of a company with full-fledged information for engineers around AI)
These stories are interesting, showing that individual output per developer has really grown, but team productivity and quality don’t have to grow with it. Gergely shares Linear and Cursor data: more PR, more lines of code, bigger changes, more acceptance without human review. From this you can see that the code really comes more, but then it reaches the stage of the review. And to make sure that doesn't become a bottleneck, we need architectural thinking engineering practices around testing, surveillance, reliability, safety - they used to be nice to have, and now must have. Only they allow you to somehow trust the volume of changes that have now become the norm with the help of AI.
Interestingly, Gergely actively praises Uber and talks about their internal infrastructure: MCP Gateway, Agent Builder, Minion, Code Inbox, risk profiles, AI code review. That is, AI is embedded not as a toy next to an IDE, but as part of a value delivery system: change routing, risk assessment, review streams, migrations, feedback loops.
Separately, Gergely actively scolds AI metrics that give improper motivation to employees. If an organization measures token usage or simply adoption, people will maximize AI usage, even when it worsens the outcome. This is an old disease of curve metrics multiplied by the rate of code generation. You can look very modern and produce more technical debt at the same time.
AI does not negate good engineering, it makes it more valuable. Naming, module boundaries, testability, understandable contracts, normal review, the ability to remove old code – all this becomes not an aesthetic, but an infrastructure of trust. When code is written by an agent, a person needs a way to understand what has changed, why it is safe, and what the consequences will be in three months.
Near the end of the speech, Gergely shares recommendations 1Slow down not in the sense of “do not use AI”, but in the sense of “do not generate more than you can check”. Speed must obey the bottleneck. (reviewability)Not the other way around. 2Start not with a tool, but with a business result and a weak point of the process. Where is time lost: discovery, onboarding, review, tests, migration, incident response, documentation, old code? AI is useful where built into a specific feedback loop. 3It is worth investing in verification systems. Evals, tests, static analysis, risk profiles, observability, staged rollout, code ownership is a way to scale AI without completely losing control. 4The engineer of the future needs product manager skills, domain expertise, architectural taste and the ability to build systems on top of LLM: RAG, agents, evals, workflow automation, internal platforms. The cheaper the code, the more valuable it is to know what code is worth writing. 5AI adoption cannot be delegated only to enthusiasts and counted under licenses. Managers will have to stay hands-on: understand the limitations of models, see real bottlenecks teams, protect quality and rebuild review. The middle layer of managers who only drive their hands here is not much needed.
#AI #AI4SDLC #Engineering #Software #Architecture #Management