The founder's look at what really speeds up an engineering job, what increases it, and what changes in the approach to hiring.

Teams with a high level of success have a 10-hour review of pull requests with a maximum of 91 hundred percent, according to Faros AI, based on data from over 10,000 customers from 1,255 teams. The writing of the code has become shvid. The result you are actually paying for is not the result: a working product that represents the very thing you asked for.

I run a company where engineering is a key part of the business, and I know what it's like to have a powerful team. Two years ago, one feature took up a whole day in the developer: time in the editor, little debugging, a lot of the same type of boilerplate code. Today, the same writer can spend an hour trying to clearly describe what he himself needs to do, launch three AI agents behind this description, and spend the rest of the day reading and revising what he has generated.. Prote, there are no fewer robots. This is a different job, and how you hire and hire retail workers, but how you hire and pay for them..

There is no need to be careful about what is being overestimated, and there is no need to believe that it’s all up to you. This is a test of the power that other founders should less often put: if it’s effective to make my team perform well, and if I’m just deceiving myself?

What has changed

The last two decades of the most expensive part of software development involved the transformation of business ideas into working code. Now the cheapest part. Today's AI models generate the first version of the code for a while. Those that previously took a whole working day are now more like a year of preparation and a quick first draft.

Sean Grove, an OpenAI engineer, argues that the code itself is worth 10-20 times the value that the engineer creates.. Reshta 80-90 hundred square meters, for his reasons, are what he calls structured communication: more precisely the formulation of what you need to do clearly, so that it can be done correctly. It’s not necessary to wait for exact numbers, so that you can learn more directly.

In practice, this means that your engineers now spend more than an hour writing a clear specification and revising what they generated the AI \u200b\u200bagent, and less than an hour writing the code by hand.. These commands go further and describe the specification itself as a rule of truth, and code generation is designated as those that no one manually edits. Most teams move between these two extremes. Everyone has the same thing: real work and real passion are now focused on the planned and transformation.

If it effectively speeds up the process of development, and if not.

This is the food that has the greatest impact on your budget and terms, and the honest truth is this: it’s all about what will happen and who will do it.

On simple, new projects, for example, the creation of an MVP, the evidence is unequivocally positive. In a controlled investigation, the developers created a simple HTTP server from scratch, and those who used the AI \u200b\u200bassistant completed the task remarkably quickly. Research shows that in such projects, the number of pull request measurements has increased by approximately 26 hundreds.. Also, if you create something new, it is difficult, most likely, to really speed up the progress.

On the great, already existing systems that support the advances of engineering, the picture is the same. METR's investigation tracked open-source code vendors who were working with powerful, large repositories involving millions of rows of code, and found that they were working on 19. This gap between perceived liquidity and real liquidity shows what a founder should know before he trusts the report about the status of the project. To be fair: METR later looked at the investigation through the selection process, so it is better to perceive it as one of the significant signals, and not as a residual verdict.

Underneath the result lies the problem of trust, about how to know, first of all, to rely on the result, generation. . The hour, which is written down, turns around in the sight of an important and expensive audit..

A practical rule, which is how I repent, if it seems that the project has succeeded in making a lot of money: ask that the whole thing is small, a self-sufficient part of the new work, and a change in the great system, which may already have business. In the first edition of the statement about the liquidity, you can believe it. Have another varto spit, which you yourself have verified and whoever has done it, first of all believe it.

Why do we need to respect, hire and evaluate retailers?.

If you are fluent in typing and know the syntax, remember to type less, it turns out that it is easier to filter candidates after them. The value of many new skills has grown disproportionately, but they are not reduced to a standard resume-filter.

Architectural vision: what people want to know about how the system can be controlled, and not just implement warehouses?

The accuracy of the formula is possible: how can you transform the often-used idea into the formula “we want it to work like X” into something specific that falls under the numerical metrics for verification?

Contextual engineering: if you want to give the AI \u200b\u200bsystem the required context and interchange, so that it sees the same result the first time or another time, and not just formally corresponds to the words of the query?

Orchestration: how can you launch and coordinate a bunch of AI-streams of work on different parts of the project overnight, without losing control over what is going on with them

AI-researcher, he still needs to enter the code to mark the wrong answer, because it’s impossible to evaluate those you can’t read. This means that the mere speed of writing code is no longer a competitive advantage, apparently, and the hiring process will need to follow new rules.

Who stands for mercy?.

The axis does not change, as long as the tools are not painted: the responsibility for the result is the responsibility of the one who confirms the change. You can't pull it up to fullness if it breaks down in production. A specific person can explain the news about what was published.

Whose has direct budgetary implications. If your team generates more code, lower before, review and revision - at no cost. A plan that doesn’t lie, that you wrote the first draft, but doesn’t lie, that you want to reliably check, but a plan that hasn’t been completed. This plan is unexpected, unplanned, and inevitably manifests itself at the most important moment: in production, in front of the client.. Also, clients, insure this.

What does this mean for your hiring plan?.

A number of simple positions in companies with a high rate of advancement are dwindling, while the premium for newbies in architecture and work with them is growing. If your hiring plan is still based on the fact that you will need a large number of employees for routine implementation, you should reconsider this idea.

Once it was reported that compilers would not allow business analysts to write powerful software without engineers. Fourth-generation movies and CASE-tools are equipped with plugins, generated directly from the specifications, without accessories. DevOps learn about the role of the database administrator. Nothing happened as predicted. Such an application lowered the threshold for entering the development of software, which increased the use of software that would be developed, and this, in its turn, increased the need for more qualified engineers who would be created. Compilers didn’t stop working with developers, they started writing code by hand in assembly language and raised the level on which they all work. The database administrator was not familiar with DevOps, this role was simply transformed to suit new capabilities.

A practical tip for your hiring plan is this: do not accept changes through AI as a reason for changing investments in engineering personnel. Take this as a reason to change the personnel you invest in. Fewer people hired to write routine code, more input for beginners in security and architectural design.

Three speeches that I made in another quarter.

Stop hiring engineers for their knowledge of syntax and clarity of logical thinking.. Precisely respect the architecture of the mind and the precision of the formulation is possible. Ask candidates to explain how to verify the result of the generation, and not just how to reject it.

Allow an hour to check the skin assessment, which transmits the code, generates III, as the planned income statement, and not those on which you intend to protect.

Do not misunderstand changes in the sphere as “We are replacing distributors” or “We are saving us an hour and pennies”. Similar innovations in the past raised the bar for engineers, and did not reduce the need for them. Plan your hiring and training accordingly.

The salesmen you hire are specialists in architecture, security, descriptions and inspections.. . These are other criteria for hiring, there are three reasons for this, as well as other points of expenses. Purchase price for additional tools, if cheaper? read on the HiTech website. Expert.