Amazon, NVIDIA, Meta, and Oracle engineers explained how AI is changing their work—and where humans remain indispensable

Otabek Bobomurodov, City — Prague, Software Engineer at Oracle, LinkedIn

AI has significantly changed the way we work. On one hand, many tasks have become more automated because we now actively use AI agents to perform them. At the same time, responsibility for the final result still lies with the human.

I delegate a task to an AI agent, then check the result and literally go through every line of code. If I see that something needs to be replaced or done differently, I challenge the AI's decision and search for the correct solution together with it.

On the other hand, the problem is that there are fewer opportunities for independent problem-solving. AI can already handle even architectural decisions and operate at the level of a senior engineer. As a result, engineers rarely encounter the complex tasks that used to help them grow.

After a few months of working in this mode, I realized that I need to challenge myself more. Otherwise, work gradually turns into routine: delegate the task to AI, check the result, create a PR, and discuss it with colleagues.

Another change is the growth of expectations from management. If previously a task might have been given five days, now due to AI capabilities, the deadline can be one or two days. For some, this is additional pressure, but it is already a new reality.

In large technology companies, such as Oracle, a special CM ticket must be created for every change in production. Previously, preparing it could take several hours, but now AI can do this in a few minutes and even pass the necessary initial checks. For me, this is one of the main advantages, as such routine work is automated.

I also actively use AI for everyday tasks: writing professional emails, editing messages, preparing descriptions for PRs. Sometimes I use it to better understand the project architecture and gain additional knowledge about its structure and operating principles.

Despite the development of AI, we still need human knowledge and experience. The human remains responsible for the final result, and AI is a tool that helps speed up the process.

AI does not always provide the correct result. Sometimes it confidently suggests an incorrect solution. Moreover, senior engineers and architects still make higher-quality architectural decisions because they better understand the specific project, its principles, goals, and stages.

Therefore, at this moment, I do not think that AI can completely replace developers. Rather, it makes development faster and easier, but final decisions still remain with people.

Mukhriddin Ibragimov, City — Seoul, Senior Software Engineer at Instagram, LinkedIn

Over the past year, AI has become a practical assistant in my work at NVIDIA. I develop tools for 2D and 3D labeling and visualization used in NVIDIA DRIVE development. These tools are related to graphics, spatial data, complex user interactions, and large codebases covering frontend, backend, and infrastructure.

AI proved especially useful when transitioning to a new area of the project for me. It helps to quickly understand unfamiliar code, trace data and rendering flows, investigate visual and production issues, and connect different parts of the system. Currently, I use AI to study code, prepare draft implementations, write tests, analyze logs, documentation, and compare technical approaches.

At the same time, I still rely on my own experience in computer graphics to evaluate the correctness of geometry, coordinate transformations, and rendering performance, as well as the usability of 2D and 3D interactions for specialists involved in data labeling.

AI accelerates learning and development, but engineering decisions, visualization quality, result verification, security, and final responsibility still remain with me.

Daulat Almagambetov, City — London, Software Engineer at Meta, Instagram

Over the past year, I was able to accelerate my engineering goals to the next level. I found bottlenecks in projects where a lot of money was spent on GPUs and ultimately saved the company $1.2 million. I launched several experiments targeting different user segments, which allowed for rapid iteration of new solutions in the "move fast, break things" approach. Project tracking, deadlines, and context management have been greatly simplified using the Second Brain concept. It synergizes well with AI agents like Claude and Codex: the necessary context can be loaded directly into the model.

Although all this sounds great, AI does a lot: it writes code much faster and sometimes even better. At Meta, with the advent of AI, engineering excellence and the codebase have improved significantly, and most of the monotonous work has been automated. But there is one catch: this has had virtually no impact on business metrics. You seem to have become slightly faster everywhere, and by the principle of atomic habits, you should have reached a new level, but it didn't happen.

The human role and intuition are still in demand, and models still struggle to translate business requirements into working solutions. I think this is the sweet spot where humans remain particularly important. Conceptually, models are becoming smarter thanks to fine-tuning, but business is an environment where everything changes quickly: new requirements arrive and the rules of the game change. Predicting them in advance and fine-tuning for them becomes practically an impossible task.

Yashar Ahmadov, City — Luxembourg, Senior Data Scientist at Amazon, LinkedIn

Over the past year, AI has become a quite capable assistant in programming for me, but not much more than that. I use it for writing boilerplate code, refactoring, creating tests, debugging, documentation, as well as for quickly studying unfamiliar libraries or approaches. Tasks that used to take a whole day now take just a few minutes. AI greatly helps to move from a blank page to a working first version.

However, it still lacks autonomy. If given a clearly formulated task, it handles it well. But if given a complex multi-stage problem with many interrelated constraints, it can lose the thread, make confident but incorrect assumptions, or create something that looks correct but breaks in non-standard cases. Therefore, a human still needs to correctly formulate the problem, break it down into parts, and evaluate whether the result is truly correct.

This is where domain-specific knowledge remains indispensable. I develop software for modeling and optimizing supply chains, and no model can determine whether a warehouse operation simulation corresponds to real operational processes or whether the optimization result is reasonable and not just mathematically correct. Such evaluation comes from years of working with real systems, and AI does not replace it.

I would also approach "vibe coding" with caution. It is excellent for prototypes, demos, and validating an idea in a day. But this approach does not scale. Without thoughtful architecture, testing, and a human who understands exactly what the code does, you can end up with a system that no one can properly maintain or trust.

Therefore, I believe that AI significantly increases the productivity of experienced specialists, but makes expertise even more valuable, not less.