Industry shifts going into 2025, both perceived and predicted
It’s been a few months!
This post was partially written last year so some links are old news now, but I wanted to both finish what I started and remind myself and readers that my intention with this blog is to provide a weekly learning forum for myself and the wider engineering community interested in such broad topics as AI/ML, productivity, engineering culture & talent, corporate & industry trends, etc. etc. It’ll be fun!
Neel Nanda has a weekly review self-reflection habit that I discovered via this Tweet and as I read through the prompts — “What am I currently procrastinating?” / “What’s weighing on my mind?” — I realized that I dropped writing this tech blog last year because life and work got busy and I felt I had to de-prioritize my extracurricular learning habit to stay afloat, but I never wanted to stop learning. I shifted to consuming content more passively (mostly via Twitter), but Neel’s tip to “make it the default” (re: self-reflection, but for me in this case active learning outside of my direct areas of practice and expertise) stuck with me.
Here’s my attempt to start a weekly Sunday habit to summarize everything I’ve read, learned, and generally been curious about in the last week in AI/ML & engineering. No length requirement!
Online privacy and deepfakes
Andrew Carr tweeted this paper yesterday (“The GAN is dead; long live the GAN! A Modern Baseline GAN”, Huang et al., 2025) and while most of the math cruised at a solid 10,000 feet over my head, I couldn’t help but be reminded of the AI-generated Will Smith eating spaghetti meme (and how remarkable a difference a year in model development can make!). The authors describe how “existing popular GAN [generative adversarial network] backbones like StyleGAN use many poorly-understood empirical tricks with little theory”, “derive a well-behaved regularized relativistic GAN loss that addresses issues of mode dropping and non-convergence that were previously tackled via a bag of ad-hoc tricks”, and propose a roadmap to go from StyleGAN2 to R3GAN. In short, they only keep the “raw network backbone” and "basic image generation capability” of StyleGAN2 (removing all other features), add their derived well-behaved loss function and regularization.
The result is a streamlined GAN baseline that is generally stabler and more explainable and maintainable (fewer ad hoc “tricks”). I like this paper’s approach because it references the same core engineering principles that distributed systems and databases wonks are intimately familiar with — scalability, stability, and abstracting and automating away parts of a brittle system.
I would have liked to see way more emphasis on online privacy and security implications. It’s briefly alluded to in paper disclosures (“it is not needed to point out that a generic algorithm for optimizing neural networks could enable people to train models that generate Deepfakes faster” and “The cost of training these large GANs is not prohibitive and is often done by hobbyists. As such, it is doubtful that these models will unlock any new capabilities for mis-use or dual-use.”) which is valid, but I’d be curious to find any study that tracks advances in generative image and video with money or worse lost en masse to scammers or other bad actors. Meta Video Seal and Audio Seal are efforts to detect and “watermark” such generated content.
This will become especially complicated with the rise of “AI engineers” and copilots, which necessitates another heading.
The rise of the “AI engineer” and copilots
GitHub Copilot (for code). Microsoft Copilot (for life on a PC). Windsurf (Codeium’s IDE). And of course, ChatGPT, Claude, Perplexity.
Talk to most software engineers in Silicon Valley and you probably won’t hear any actual fear of being completely automated away; there will always need to be a shepherd for the sheep, the thinking goes, and Big Tech Companies claiming to generate more and more of their code (thereby cutting hiring costs) are still generally brushed off as standard news cycle marketing strategy and not taking a long-term sustainable approach to build high-quality products. A common party joke around here is “yeah, Google does generate their code. Have you seen their new Pixels?”
But clearly there’s a trend beyond a fad that’s sticking here. Last year YC’s Lightcone podcast described a future of unicorn companies with fewer than 10 employees, aided by generated code. This is becoming a common early-stage startup pitch. Software engineers raised on the bread and butter of writing all their own code from scratch and all-nighters agonizing over Vim and C on tiny screens in computer labs may now play wise and shake our heads over the new generation of software engineers that go to ChatGPT for everything from homework help to plant identification, but the programming landscape is changing in a major way and we’ll just want to be as careful as ever about what we’re shipping to production.
To read:
Introduction to Computing Systems: From Bits & Gates to C/C++ and Beyond, Patt & Patel, found via Twitter
A Hitchhiker’s Guide to Scaling Law Estimation (Choshen et al., 2024), found via Twitter
Good reads from late 2024:
Google's NotebookLM is a great product that will die on the vine, Dave Friedman
Models of Life, Abhishaike Mahajan