Software Gets Truly Soft with Generative AI
In their post on May 12 (What every CEO should know about generative AI), a gaggle of McKinsey specialists included a section on "Changing the work of software engineering". In it, they said (emphasis mine):
The biggest part of a software engineer’s job is writing code. It’s a labor-intensive process that requires extensive trial and error and research into private and public documentation. At this company, a shortage of skilled software engineers has led to a large backlog of requests for features and bug fixes.
To improve engineers’ productivity, the company is implementing an AI-based code-completion product that integrates with the software the engineers use to code. This allows engineers to write code descriptions in natural language, while the AI suggests several variants of code blocks that will satisfy the description. Engineers can select one of the AI’s proposals, make needed refinements, and click on it to insert the code.
Our research has shown that such tools can speed up a developer’s code generation by as much as 50 percent. It can also help in debugging, which may improve the quality of the developed product. But today, generative AI cannot replace skilled software engineers. In fact, more-experienced engineers appear to reap the greatest productivity benefits from the tools, with inexperienced developers seeing less impressive—and sometimes negative—results. A known risk is that the AI-generated code may contain vulnerabilities or other bugs, so software engineers must be involved to ensure the quality and security of the code (see the final section in this article for ways to mitigate risks).
"But today, generative AI cannot replace skilled software engineers".
But today.
I see this as a momentary state of affairs. I can clearly see a time - soon - when Generative AI does replace skilled software engineers. A properly trained model will include the ability to review code for vulnerabilities and bugs, assure quality and security, and test the code in virtual environments.
I see a series of bespoke models built by the internal Data Science team that automatically comb applications for data exhaust, build and monitor API's and pipelines, perform proper munging, create an internal data dictionary, maintain a data catalog, track data provenance, provide data governance, and suggest new data associations to answer new questions. These models will include a continuously updated cascade of guidelines and guardrails based on code tests, outcome monitoring, and human supervision.
As a result, 'coding' will become
"Run payroll based on this employee spreadsheet with current tax rules. Include commissions and bonuses."
Rather than being, "a labor-intensive process that requires extensive trial and error and research into private and public documentation," code will be generated, tested, cross tested, and certified. Data remains the single most important resource the company has, worthy of the highest degree of protection affordable.
领英推荐
Just as secretaries became personal assistants and then department managers, software developers will upskill to become subject matter experts who interface between the business and the system. Their job will be teaching business stakeholders the fine art of asking good questions and providing tweaks to the systems as problems arise and new guidelines are needed.
Software is no longer hard coded but truly soft; flexible, malleable, changing with the weather with proper explainability and versioning for look-back audits.
Machines will work the same way the payroll department did before the time of machines. They will crunch the numbers, stick to the cadence, handle exceptions, and bring confusing issues to the attention of those in charge. Those in charge will be subject matter experts who can make gut-check risk management decisions on behalf of the firm.
If your company makes and sells software packages, it's not too late to pivot. But it will be soon.
Full Sterne Ahead?001: Marketing Analytics Live Online interviewees
Full Sterne Ahead 002: How Generative AI Changes the Role of the Analyst - a conversation with Tom Davenport
CEO/Chairman at DMscore
1 年English majors will finally have a job, as prompt engineers!
Author, Managing Partner at Verity7, Speaker, Director Emeritus Digital Analytics Association. Author of "Army of Liars" published September 3rd, 2024.
1 年No doubt AI and generative bots will change life for knowledge workers everywhere. But my larger concern is what happens when the wrong people get hold of chatbots. I feel like the world needs training right away in how to handle the amount of disinformation likely to come out of AI and chatbots generally.
Senior Data Strategist @ ODOSCOPE | Modern, innovative and sustainable business growth by better using your own data
1 年here's another opinion, I tend to believe it makes a point: https://www.dhirubhai.net/posts/milanmilanovic_ai-artificialintelligence-technology-activity-7063078809486409728-QyVT ;-) Edit: Why does the image not show up? The linked quote says: "To replace programmers with AI, clients will need to accurately describe what they want. We're safe."
I agree with the great majority of Jim's comments, and feel that any technical obstacles to broad-scale use of gen AI-driven coding will disappear quickly. There may well be sociological and cultural barriers to acceptance of these approaches, however. Professional programmers/data scientists/data analysts/etc. will be likely to resist. IT organizations may prohibit such citizen-driven activity. Some highly decentralized system developers will undoubtedly do something stupid, and that will set back the entire movement. I think these changes are ultimately inevitable, however. For an example of slow adoption, take citizen data science based on automated machine learning. Takeup has been slow, and has been resisted by data scientists. It may be that gen AI-based development will be much harder to prevent, however. It is a consumer trend, not a tech trend.
Data, Products, Leadership
1 年Your scenario is plausible and a natural evolution of software engineering since we moved from machine code to assembly to higher-level languages like Fortran, Cobol, and Basic.?These days there are whole superstructures (libraries, frameworks, etc) already built, and software developers can be seen as the machinists that hook it all together.?Generative AI adds higher-level tools to their belt. Before generative AI transforms software or any “creative” digital industry, I think many important questions must be answered first.?Two at the top of my mind are 1. Copyrights (or maybe better said: Learning rights).?Who gets to decide what gets into the training in the first place - the model builder or the originator(s) of the "training" work? ?2. When the machine generates something that's incorrect — maybe criminally so — who’s at fault??Today the answer seems to be, “Oh well, we’ll let it know, and the model will evolve,” but there are no consequences for being wrong. Until provenance and liability have reasonable answers, GenAI will remain fringe work for commercial software development.