Staying up to date with AI

Every week it feels like there's a new advancement in the AI sector. There's new models, agent frameworks, benchmarks, and startups claiming that they've changed the industry forever. If you spend enough time on LinkedIn you'd convince yourself that software engineering is becoming obsolete every Thursday.

For developers who don't work at a company heavily invested in AI, keeping up can feel impossible. Unlike most areas of software engineering, where major changes happen every few months or years, AI moves at a pace that makes even experienced engineers feel behind.

When I started my journey, I tried my best to follow everything. New model releases, summaries of research papers, benchmark results, youtube videos, and even conference talks. Eventually I realised the amount of AI content being generated was growing faster than I could learn it.

The AI Firehose Problem

One of the biggest challenges with AI is that there is way too much information. A backend framework might receive a major update once every year, whereas AI companies seem to release something noteworthy every few weeks. By the time you've finished reading about one model, another company has announced a competitor with a larger context window, lower pricing, or a benchmark score that's slightly higher than last weeks.

This creates a strange feeling that you're constantly falling behind. There's always another paper to read, another model to try, or another video explaining why everything has changed again. The reality is that most people aren't keeping up with everything. Even the people who work in AI have to be selective about what they pay attention to. The hardest part is deciding which news and results actually matters.

Most AI news doesn't matter

This might sound bad to any of the recruiters reading this but most AI news has very little impact on how I build software. I don't really care that a startup raised another hundred million dollars, or that a model improved by 4% on a benchmark I've never used. I will say that most of these announcements are interesting, but only a few of them actually change how I work.

The updates that matter are usually practical. This includes significant improvements in model capability, large reductions in cost, new techniques for evaluation, or infrastructure that makes building AI systems easier. These are the changes that eventually influence real products. Everything else is often just noise disguised as news.

How I keep up to date

My approach today is much more simpler than it used to be. Rather than trying to consume everything I spend a small amount of time each week reading AI news from a few trusted sources. If something sounds genuinely interesting, I'll save it and revisit it later.

More importantly, I spend the majority of my learning time building things rather than focusing on theory. Over the last few months I've learnt far more from building AI features than I have from reading announcements. Adding RAG search to this blog taught me about chunking, embeddings, retrieval quality, and prompt design. Adding evals taught me how difficult it is to consistently determine whether an AI system is actually producing useful answers.

None of these lessons came from rage-baiting headlines. They came from trying to solve real problems and then figuring out that the solutions were rarely as straightforward as they looked in a demo video.

What I actually follow

As time has gone on, I've found myself paying less attention to individual model releases and more attention to the techniques behind them. New models are announced so frequently that it's difficult to remember which one was considered state of the art six months ago. What tends to stick around are the ideas that influence how systems are actually built.

Topics such as retrieval, evaluation, prompt design, infrastructure, and deployment continue to appear regardless of which company currently has the best model. These areas are usually where I spend most of my learning time because they stay relevant long after the excitement around a particular announcement has disappeared.

What I ignore

I rarely spend time reading startup funding announcements, or endless benchmark comparisons. These discussions generate a huge amount of attention but rarely make me a better software engineer. The AI industry is good at creating excitement. Every week there seems to be a new announcement that promises to change everything. Most of the time though, the fundamentals stay the same.

My take

The biggest lesson I've learnt is that keeping up with AI doesn't mean reading every announcement. If you tried to consume every single AI related reading, learning AI would quickly become a full-time job.

I've found it more useful to focus on a small number of important developments and spend the rest of my time building. The headlines change every week, but the fundamentals and lessons learnt from building something yourself tend to stick around much longer.