Ai

Code review, however, emerged as the most significant challenge. Reviewing the generated code line by line across all changes took me approximately 20 minutes. Unlike pairing with a human developer—where iterative discussions occur at a manageable pace—working with an AI system capable of generating entire modules within seconds creates a bottleneck on the human side, especially when attempting line-by-line scrutiny. This isn’t a limitation of the AI itself, but rather a reflection of human review capacity.

Seven years from GPT-1 to the plateau. How many more until we stop trying to build intelligence and start trying to understand what we’ve already built? That’s the real work now - not training the next model, but figuring out what to do with the ones we have. Turns out the singularity looks less like transcendence and more like integration work. Endless, necessary integration work.

I think this is it. The last 5 years and especially 3 have been amazing to watch. From GPT-2 to today AI has only gotten better at helping me code basic things. But it still tries to run non-sensical commands and just recently added a “Utilties-2” folder to my project because I already had one?

Some of us lean on AI coding to push side projects faster into the delivery pipeline. These are not core product features but experiments and MVP-style initiatives. For bringing that kind of work to its first version, the speed-up is real. … output quality gets worse the more context you add. The model starts pulling in irrelevant details from earlier prompts, and accuracy drops. … AI can get you 70% of the way, but the last 30% is the hard part. The assistant scaffolds a feature, but production readiness means edge cases, architecture fixes, tests, and cleanup

“We need to stop talking about AI as a magic fix and instead focus on the specifics: where are the biggest points of friction for developers, how can AI help alleviate that friction, and specifically how should developers use AI tools to overcome that friction and move faster?” Laura Tacho, CTO at DX told LeadDev earlier this year.

Not sure if I captured this before. but writing code was never the bottleneck in software development. And no engineer is able to drive a single feature to completion. Stakeholders are always involved.

NVIDIA’s earnings are, effectively, the US stock market’s confidence, and everything rides on five companies — and if we’re honest, really four companies — buying GPUs for generative AI services or to train generative AI models. Worse still, these services, while losing these companies massive amounts of money, don’t produce much revenue, meaning that the AI trade is not driving any real, meaningful revenue growth.

We’re three years in, and generative AI’s highest-grossing companies — outside OpenAI ($10 billion annualized as of early June) and Anthropic ($4 billion annualized as of July), and both lose billions a year after revenue — have three major problems:

The careful, coded corporate language executives once used in describing staff cuts is giving way to blunt boasts about ever-shrinking workforces. Gone are the days when trimming head count signaled retrenchment or trouble. Bosses are showing off to Wall Street that they are embracing artificial intelligence and serious about becoming lean. After all, it is no easy feat to cut head count for 20 consecutive quarters, an accomplishment Wells Fargo’s chief executive officer touted this month. The bank is using attrition “as our friend,” Charlie Scharf said on the bank’s quarterly earnings call as he told investors that its head count had fallen every quarter over the past five years—by a total of 23% over the period.

AI assistants optimize for making tests pass and errors disappear. Without clear direction, they’ll take the path of least resistance. Common shortcuts to watch for:

TypeScript any types appearing when proper typing gets complex Tests getting commented out or skipped when they’re hard to fix Quick fixes that address symptoms rather than root causes

Great write up on a using claude code cli. And this was in July before 3.5 sonnet! I found this list of ‘gotchas’ kinda funny because they’re the exactly the kind of thing a senior engineer would catch immediately but a novice wielding AI might not. Fare not, “Refactoring and Debugging AI vibe apps” coming to a O’Reily publication near you. (To borrow a joke from his post)

So from an executive perspective, lighting comically large piles of money on fire trying to teach graphics cards how to read is, surprisingly, the logical play. The rest, well, that’s all just creative marketing. It’s very difficult to show up to a quarterly shareholder meeting and tell your investors you just vaporized another $10 billion for absolutely no return-on-investment. At least, that is, without them questioning if you’ve completely lost your mind. Which is where leaning into the hype plays into it.

AI is increasingly automating many coding tasks, accelerating software development. As models and tools improve, we see the automation of more complex coding tasks under developers’ orchestration (like the ones we interviewed). This is already reality and no longer a future trend.

I think the reality is no one ever cared how nice my lines of code were. Anyone outside of software just wants the damn thing to work. SO using AI is just another tool. Adopt it.

I care about what gets merged into the codebase. I don’t care how the code got in your IDE.

I want you to care.

I want people to care about quality, I want them to care about consistency, I want them to care about the long-term effects of their work. LLMs are engineering marvels, and I have the utmost respect for the people who’ve created them. But we still need to build software, not productionize prototypes.