Why AI makes engineering discipline more valuable

There's a pattern I keep seeing in organizations adopting AI for software development. A team gets access to a coding assistant. Productivity jumps for a few weeks. Then quality problems start surfacing: inconsistent patterns across the codebase, security practices that vary from file to file, architectural decisions that contradict each other. The CTO asks: what happened?

The organization treated AI as a faster typewriter instead of a new team member who needs onboarding, direction, and oversight.

Richard Socher's comparison of AI work with management gets to the heart of that problem. In his public post, he emphasizes clear delegation, requirements, and the development of trust. Those are skills engineering leaders already need. AI makes them more consequential as teams can produce more implementation than they can responsibly accept without a method for reviewing it.

From language research to enterprise work

Socher's perspective connects research, enterprise deployment and product leadership. He co-authored GloVe, a method for learning word vectors from word co-occurrence across a corpus. Those numerical representations capture useful relationships between words. His co-authored decaNLP paper in 2018 represented ten language tasks as question answering over a supplied context and trained a model across those tasks. It connected language research with the practical problem of telling one system which task we want it to perform.

When Salesforce introduced Einstein in 2016, it also announced a research group led by Socher as chief scientist, bringing advances in language, vision and deep learning to its product and engineering teams. At You.com, which he co-founded, he carried that work into AI products for knowledge workers. In the company's September 2024 account, he emphasized accuracy, trust and useful adoption by enterprise customers.

His co-authored work earned ACL Test-of-Time awards in 2023 for the sentiment-treebank paper and 2024 for GloVe. He was also named to TIME100 AI in 2023. That history gives his management argument weight: he has worked on both the capabilities of AI systems and the problem of making them useful inside organizations.

The gap between a demonstration and dependable work

Socher has described the difficulty directly: "It's easy to make a quick prototype with an LLM. It's difficult to make them accurate at scale." His last-mile comparison is about the engineering needed after an impressive demonstration.

In an August 2025 interview, he described companies assembling their own interfaces to language models: "like 70% accurate and you can cherry pick five examples and look, it's perfect. But then when you really use it, people don't have adoption." Reliability and knowing how to use the system both mattered. A demonstration can conceal the cases that make a tool frustrating or unsafe to depend on.

He had raised the code-quality problem earlier. In a January 2023 interview, he said: "For coding, the challenge will be the quality of the code. The opportunity there could be the accuracy and quality as validated by the top 50% and how will coders incorporate these tools into their work streams."

I see the same management problem when an organization moves from letting engineers experiment with a coding assistant to making AI a standard part of development. Better models expand what a team can attempt. The team still needs a way to maintain quality, consistency and accountability as more of its implementation is generated.

Learning to manage AI

Socher's management analogy is the clearest way I know to explain the shift.

"The future of work is all of us becoming managers of AI. Similar to moving from individual contributor to people manager: learning to delegate clearly, specify requirements, build trust. That's the skill we need now."

A manager must understand the system well enough to evaluate work they did not personally carry out.

An engineer directing a coding assistant faces those responsibilities too. Architectural judgment, quality standards and the ability to inspect another party's work become central. The ability to produce code remains useful; it also supports the judgment needed to catch an implementation that looks plausible but solves the wrong problem.

I wrote about Socher's argument in December 2025 because it described, in management terms, what I have been building as an engineering discipline. Synthesis engineering is the professional discipline of systematic human-AI collaboration for complex work. Synthesis coding applies that discipline to software development.

PrincipleWhat the team doesWhy it matters
Human architectural authorityHumans decide the technology choices, system boundaries and security model; AI implements within them.A succession of locally plausible changes must still form a coherent system.
Systematic quality standardsApply the same rigor to generated code as to human-written code.Faster implementation otherwise accelerates the accumulation of defects and technical debt.
Active system understandingEngineers understand the behavior and dependencies well enough to debug what they build.Someone has to diagnose the failure at 2 AM.
Iterative context buildingKeep requirements, decisions and constraints in maintained artifacts that the assistant can consult.Useful context should accumulate across the work rather than vanish with a conversation.
Transferable knowledgeLeave code, explanations and evidence that another engineer can use.Individual productivity must become work the team can maintain.

Clear delegation starts with architectural authority: define the structure and constraints before asking for implementation. Specifying requirements becomes a durable practice when those requirements live with the project and are updated as decisions change. Trust grows through review and verification. Maintaining an understanding of the whole system is the leader's continuing responsibility, even when someone or something else writes the code.

Choose the approach for the work

When I discuss AI-assisted development with engineering leaders, the comparison table that gets the most traction distinguishes three approaches:

ApproachHuman roleAI roleWhere I would use it
Vibe codingExplore an idea with limited inspection of the implementation.Generate and revise a prototype in response to feedback.Learning, experiments and throwaway prototypes.
Agentic codingDefine the goal and boundaries, then review the result.Carry out a bounded sequence of implementation tasks with greater autonomy.Well-defined work whose result can be checked.
Synthesis codingSet architecture and requirements, review evidence, and authorize consequential decisions.Implement, investigate and test within that structure.Production systems and complex codebases that teams must maintain.

The same developer can use all three in a day. Agentic execution can also sit inside synthesis coding: autonomy within a task leaves architectural and release decisions with the responsible people. The leadership skill is matching the approach to the work and ensuring that production systems receive the rigor they require.

Delegating a CSV export

I would apply that analogy to a team adding a CSV export to a customer dashboard. An AI assistant can investigate the existing implementation, propose a change, write code, and produce tests. The responsible engineer still has to decide which data a user should be able to export and what evidence would establish that the feature respects that decision.

What result are we delegating?

“Add CSV export” leaves several product decisions open. Does the export contain the current page or every matching record? Does it use the active filters? Which columns belong in the file? What should happen when the result is empty or too large to handle in one request?

For this example, suppose the product decision is to export the records matching the dashboard’s current filters, with a documented set of columns. A user must receive only records they are allowed to see. The engineer should settle the meaning of those requirements before asking for an implementation.

The assignment can then point to the existing authorization code and the dashboard’s query logic. It should identify the behavior the assistant must preserve, the files it may change, and the questions it must return for a decision. A reference to the current code is useful context; a specific statement about the required behavior gives the reviewer something to check.

If the existing query mixes filtering with pagination, the assistant may discover that reusing it would export only one page. That finding calls for a decision about the implementation. Silently changing the requirement to match the easiest available function would produce a different feature.

What evidence would make the result acceptable?

A demonstration with one account and a handful of rows establishes little about access controls. I would ask for evidence at the places where the feature could violate its requirements.

Use an authorized account with known records to check the selected columns and filters. Use a second account with different permissions to check that inaccessible records remain absent. Attempt the export without the required permission and check the server’s response. These cases should exercise the actual export path, including its query and authorization checks.

The output file also needs examination. A passing unit test for the serializer says nothing about whether the endpoint selected the right records. Open or parse a file produced by the endpoint, and compare it with a known fixture. Check empty results and text containing commas, quotes, and line breaks. If the product supports large exports, require evidence for the agreed size and failure behavior.

A review should explain how the evidence was produced. “Tests pass” is incomplete without the tested revision, the cases covered, and any checks that could not run. If the assistant could run only a local fixture, that limitation belongs beside the result. It cannot stand in for a check that requires another environment.

The engineer also needs to inspect the change itself. Tests can confirm an expected behavior while missing a second route that exposes the same data. For this feature, I would trace the request through authorization, record selection, and file generation before accepting it.

Which decisions remain with the responsible person?

In this example, the assistant can propose changes to implementation details. It cannot decide to weaken a permission check, expand the exported fields, or deploy the feature merely because those actions would complete its task.

Those boundaries should be part of the assignment. If the existing access model is ambiguous, the assistant should identify the ambiguity and the affected code. The responsible engineer can then resolve it with the appropriate product or security owner. Inventing a policy in generated code would leave the team with a decision nobody knowingly made.

Acceptance and release are separate decisions. A reviewed change may be ready to merge while a rollout still requires coordination with an owner of the affected service. The person delegating the work needs to know which decision they hold and which belongs elsewhere.

Build trust around observed work

The management analogy has limits. An assistant’s fluent explanation does not establish that it understood the requirement, ran the stated checks, or noticed a missing case. I would base permission to do more work on the results the team can inspect.

For the CSV export, that might mean beginning with investigation and a proposed diff, then allowing implementation once the requirements are settled. Repeated success on a well-understood task can justify less intervention within that task. A change to authorization or data exposure deserves fresh scrutiny even if earlier work went smoothly.

For a training exercise, ask engineers to review the same proposed export, find an unstated product decision, identify a missing permissions test, and explain what would make them accept the change. Review disagreements together. One engineer may accept a passing permissions test while another notices that it bypasses the actual export path. Resolving that disagreement gives the team a more precise review standard to use on its next change.

Develop engineering judgment

The scarce resource in AI-assisted development is judgment. An engineer needs to recognize a security assumption, an architectural inconsistency or an edge case the prompt never mentioned. Hiring and professional development should reward people who can explain how a system works and evaluate the consequences of changing it.

In a 2025 interview, Socher put the skill shift this way: "The ability to discern and evaluate contents and outputs is going to become more and more important than the initial creation of them."

Code review deserves investment as implementation accelerates. A tiered review framework can direct attention toward the risk of a change. An authorization change needs different scrutiny from a local presentation adjustment; a small diff can carry a large exposure. Automated tests, security analysis and performance checks should run where they address the change's risks, and the reviewer must understand what those checks actually establish.

This is a broader skill than finding mistakes in a diff. It includes choosing a sound design, recognizing useful work, asking for missing evidence and deciding when the team understands enough to accept responsibility for the result.

Train people for the work you expect

Organizations that hand engineers AI tools without a methodology get inconsistent results. I have seen structured training, clear principles and defined workflows turn adoption into better outcomes. Teaching someone to delegate well includes teaching them to recognize an incomplete instruction and to review the answer against the actual requirement.

Socher described role-specific training and improved adoption in the August 2025 interview.

September 2026 update: In a later account, published February 12, 2026, he said:

"Most individual contributors aren't managers and the adoption would be very low until we said, 'here's a training programme and a certification programme.' When people had to do it, adoption in older organizations really picked up. But that managing and delegation mindset doesn't come naturally to most people."

That is his experience of adoption; it supports the practical case for teaching the management skills alongside the tool.

A shared discipline that compounds

The terminology and methodology of synthesis engineering are released under CC0 1.0 Universal. Any organization can use them in hiring, training, internal standards and derivative materials without seeking my permission. I want them to become a shared industry discipline that people can adapt and improve.

The connection I draw across Socher's research and product work is that useful AI depends on the systems built around its capabilities. Learning better representations, expressing tasks clearly and delivering a dependable product are related parts of that work. Engineering leaders can develop the organizational counterpart: a team that knows how to direct AI and take responsibility for what it builds.

Socher also forecast: "Not being able to work with AI will soon be like not knowing how to use a computer or the internet." I expect the organizations that learn this discipline early to build an advantage that compounds. Context accumulates in useful artifacts. Quality standards prevent avoidable debt. Architectural authority keeps a growing system coherent. Knowledge transfers to people who were not present for the original conversation.

The methodology is being tested in production. The ambition is a team that can attempt more, build it well and keep improving after any one engineer or assistant has moved on. Directing AI well deserves a place alongside architecture, testing and operations in the professional practice of engineering.

Related reading: Socher’s research and engineering context, and the research connections in depth.