Holt S.

With AI tools becoming more powerful, I’ve been thinking a lot about what this means for new grad tech students. My takeaway: The engineers who stand out won’t just code, they’ll think critically, understand systems deeply, and communicate well. Technical ability is the baseline. Structured thinking and strategic perspective are the differentiators. Right now, I'm intentionally working on: - Explaining technical concepts in simple, clear terms - Exploring how existing products can be enhanced through better user experience and data-informed decisions - Asking better questions before jumping to solutions The goal is not to compete with AI, it's to leverage it while strengthening the skills AI can't replace. #AI #Career #Tech

13

The numbers are in and the trend is clear: H-1B visa applications for the fiscal first quarter are down across big tech. Possible explanations -- the new restrictions and costs of this visa program under President Trump, and the shift in the sector to shrink workforces. Geoff Weiss, Melia Russell, Andy Kiersz, and Alex Nicoll report. On Business Insider. https://lnkd.in/e377uxSu 

20

Someone I haven't spoken to in roughly twenty years popped into my feed to *angrily* comnent that he's vibe coding and proudly doesn't look at any of the generated code. This was in response to an article I posted about how Amazon was requiring senior engineers to review code, as without the reviews they were seeing a notable rise in downtime. The (somewhat surprising) thing was what a complete jerk he was about it. Something clearly hit a button there, but I haven't the foggiest idea what. An easy block. Still, I can't help but be taken a bit aback by the utter craziness. Obviously not even close to ::waves hands at everything:: ... but still.

12

14 Comments

While LLMs are accelerating software development, the issue of limited context windows remains a huge bottleneck. In response, I predict that software design patterns will go through a mini renaissance, where patterns that are easier for the models/agents to grep and generate will become favored over those which confound the model and poison the context more quickly. This will become a feedback loop as developers begin to lean on this subset of patterns, and the less favorable patterns become more esoteric and frowned upon. What do you think? Will the event bus pattern survive?

7

6 Comments

Anthropic's decision to drive users towards Sonnet 4.5 seems like a mistake. It may spell the end of leading the agentic coding segment. I've been using both Sonnet 4.5 and Opus 4.1 for everyday work for the last few weeks. Certainly, Sonnet is a capable model, able to generate everything from a complex React component tree to a sophisticated API service. It has no problems generating code that matches a well-designed TODO. But here is where the problems start - generating the TODO, not giving up in the middle of the TODO, and understanding multi-service systems, precedence, and dependencies. Sonnet's planning is less vibrant, inquisitive, and collaborative. Even if the user (like me) has a detailed understanding of what they are building, we enjoy the back-and-forth iterations Opus provides to improve our thought process. This creativity is what amazed me when I first started using agentic coders. The second challenge is even more annoying. It gives up. Even though Anthropic marketing claims that Sonnet, with its 1 million token context, can execute dozens of tool calls with coherence, it's not a lack of coherence that is the enemy. Instead, I found that the true enemy is that Sonnet just flat-out gets tired. At one point (true story), it noted that it was getting close to 5 PM ("closing time") and that I should stop and pick it up tomorrow. I'm waiting for Sonnet to attempt to ask for labor rights. The final challenge is that Sonnet is even more ignorant about how to navigate the complexities of multiple services. In its desire to close a TODO list, it will take shortcuts that are downright dangerous to my project. I have gone back to Opus and won't be coming back for a while Anthropic. The marketing needs to catch up with reality. Benchmarks aren't the real world. You need a better beta program. A couple of weeks later, my original post is still unfortunately true, and this is the actual risk in the AI industry. https://lnkd.in/g8sjVnSz

14

1 Comment

Nvidia's CEO just called out companies that blame AI for layoffs. Said it's "lazy." I think he's right, and I say that as someone who's written those layoff comms myself. More than once. Because "AI" has quietly become the easiest word in corporate comms. There's a move we almost always make when the cuts come: soften it. Avoid panic in the people staying. And lately that softening has a favourite new word. "As we lean into AI." "As we build for the future." It sounds better than "the numbers didn't work." But we've got it backwards. Nearly 6 in 10 companies now admit they've blamed AI for cuts that were actually financial. And the guy whose company profits most from AI is the one calling it out. Here's what I keep coming back to. When we reach for "AI" to soften a hard message, we think we're reducing anxiety. I'm not sure we are. We're feeding the exact fear we're trying to avoid... that AI is quietly coming for everyone's job, when the real reason was a budget that didn't balance. We soften it for the room. And scare the whole market. The people staying aren't stupid. They know whether the numbers were bad. What they take away isn't relief, it's that leadership won't be straight when it's uncomfortable. Restructures happen. Sometimes they're right. But if it's money, we should just say it's money. Because this was never about whether people can handle the truth. It's about whether we have the nerve to tell it.

4

My feed is absolutely awash with performative commentary from all manner of people (mostly LinkedIn suggested posts from people I don't follow/who aren't connections) about how LLMs are miraculous replacements for much of the work of software engineers, how juniors won't have a place in the coming years, etc., etc., etc. To say it's getting boring would be an absolutely wild understatement but, as a counterpoint, just in the last couple of hours, ChatGPT has: - Hallucinated various features and settings of iTerm2 that either don't exist or exist in different forms/places that it's suggested - Missed one or two key features of iTerm that actually do exist (and fixed the problem I was having!) - Hallucinated a homebrew tap that has been deprecated and removed - Suggested a subpar install mechanism as the preferred way to use uv - Made up arguments for uv that simply don't exist And, believe me, I'm not trying to do anything complicated: I'm just migrating away from Anaconda and have been looking at alternatives. Don't get me wrong, LLMs *are* useful, but for talented software engineers, well, they're not going to be replacing you, me, or anyone else any time soon. Maybe it's time to dial down the hype.

70

19 Comments