Two takes on the future of software development

Welcome to Runtime! Today: Two legendary software engineers sum up the current state of AI coding in mid-2025 and what comes next, Microsoft and OpenAI continue to trial balloon their contract negotiations, and the latest enterprise moves.

(Was this email forwarded to you? Sign up here to get Runtime each week.)

Vibes 3.0

One reason for the disconnect between the enthusiasm people inside tech companies have for generative AI technology and the uneasiness with which a lot of people outside tech companies approach it is that generative AI is transforming the way people inside tech companies build technology. It's far from perfect, but when it comes to developing software, building infrastructure, and managing large-scale technology projects, the question has shifted from whether to use AI to the best way to use AI.

On Wednesday, Andrej Karpathy, the former Tesla and OpenAI AI researcher, and Sourcegraph's Steve Yegge, a veteran software developer who played key roles at Amazon and Google, laid out just how much disruption is coming for workers in the software mines. "Roughly speaking, software has not changed much on such a fundamental level for 70 years, and then it's changed I think about twice quite rapidly in the last few years," Karpathy said in a talk to attendees of Y Combinator's AI Startup School.

Karpathy labeled this new era "Software 3.0," a break from earlier eras of software where programming languages were used to create code to give instructions to computers, and then weights were used to give instructions to the neural networks behind machine learning.
Now, software engineers are using prompts written in English — "a very interesting programming language," Karpathy said — to give instructions to large-language models.
In this scenario, LLMs are like operating systems for a new type of computer powered by GPUs, although these new computers are in their "1960s era," he said, noting that this new type of compute is very expensive, centralized, and done in batch jobs.
And while he ran through the common problems people experience trying to use these new computers — from hallucinations to security to "anterograde amnesia" — Karpathy is most excited about the ability of LLMs to help developers build "partial autonomy apps" that work like the AI coding editor Cursor, which was designed with a clear separation between the familiar human-piloted programming space and the AI programming copilot.

Yegge, on the other hand, believes new AI coding agents are going to reduce (but not eliminate) the need for the AI-powered IDEs and coding editors that are so hot right now. "Instead, agentic coding is console based, 1970s-style, in a text terminal window with little to no UI," he wrote in a blog post Wednesday.

"The developer workflow is simple: You ask the AI for code changes, then you review the diffs and test the output, in a loop, lather, rinse, repeat," Yegge wrote.
After a process of trial and error, he now uses several AI coding agents (which he called "a team of programming brutes") to develop new software.
But while developers who are using agentic coding tools are seeing huge increases in productivity, they're forcing companies to make tricky decisions about the never-ending problem of measuring developer productivity.
Yegge just finished a book on this new era of "vibe coding" — a term popularized by Karpathy — and in the process of researching that book "some companies, big ones, are telling us that they are beginning to consider 'displacing' people who cannot or will not adopt agentic coding," he wrote.

While predictions of massive job loss among the developer class soak up attention, both Karpathy and Yegge see developers adapting, as always, to a new layer of abstraction. After all, very few software engineers have experience with Assembly code these days.

"Agentic vibe coding is powerful, but requires new kinds of supervision," Yegge wrote, noting that "building enterprise software will always be monumentally difficult, so engineers and AIs will team up to build it together."
Karpathy argued for some patience: "when I see things like, 'oh, 2025 is the year of agents,' I get very concerned, and I kind of feel like this is the decade of agents."
Patience, unfortunately, is not a virtue embraced by Wall Street, so expect the AI hype train to continue spewing nonsense as the real practitioners figure out how to shape a new era of computing.

Thanks for reading and supporting Runtime over the last two years! As we head into our traditional (so to speak) solstice break, please consider making Runtime a part of your marketing strategy for the second half of 2025. We're rolling out several new sponsorship opportunities and pricing packages this summer, and you can find more details about those options here.

Oh well, whatever, never mind

Microsoft and OpenAI continued to negotiate their contract-restructuring discussions through the media this week, which is great news for editors looking for content as summer looms around the corner. The latest update comes courtesy of The Financial Times, which reported Wednesday that Microsoft is prepared to take its ball and go home if it doesn't get the terms it needs to allow OpenAI to alter the deal.

"The software giant has considered halting complex discussions with the $300bn AI start-up [sic] if the two sides remain unable to agree on critical issues, such as the size of Microsoft’s future stake in OpenAI," according to the FT's sources. OpenAI will lose out on $20 billion of a $40 billion funding round from Softbank if it is unable to transition into a for-profit company by the end of the year, and Microsoft must sign off on any restructuring plan under the original deal that funnelled billions of Microsoft's cash and cloud infrastructure credits into OpenAI's coffers.

Walking away from a new deal with OpenAI would cut off Microsoft's access to OpenAI's intellectual property after 2030, which is a long time away but sooner than Microsoft would prefer under a restructured deal, according to reports earlier this week. Both companies told the FT they're "optimistic we will continue to build together for years to come," but the tension suggests otherwise, especially as OpenAI increasingly goes after Microsoft's core enterprise AI customers.

Enterprise moves

Greg Ernst is the new chief revenue officer at Intel, taking over the top sales role at a crucial time in its history after nearly 20 years in sales roles at the company.

Craig Martell is the new chief technology officer at Lockheed, following technology leadership roles at Cohesity and the Department of Defense.

The Runtime roundup

Microsoft plans to cut "thousands" of sales representatives as part of its annual restructuring plans that follow the close of its fiscal year at the end of the month, according to Bloomberg.

SAP and IBM are caught up in a messy corruption investigation into a botched ERP rollout at Quebec's provincial auto insurance board, according to CIO.

Thanks for reading — Runtime is taking an extended break through the Fourth of July holiday, see you on Tuesday July 8th!