More than a year after the first Opus 4 release, we have come a long way in our understanding of LLM capabilities for software development. We have moved from simple prompts, mostly toy-level code, to full agentic coding sessions that can last for hours.

I have to admit I have not written much code since approximately November of 2025, when Opus 4.5 came out (the first model to feel really capable). That doesn’t mean I don’t fix issues or rewrite some of the code, but the number of things I have to fix has steadily decreased as new models come along.

Most of the discourse in social networks (be it X.com or Reddit) around LLMs and coding fluctuates between full AI psychosis and full AI hate or hard dismissal. Embrace or deny.

In the real world, we are starting to see the impact of AI in multiple areas, though not as fast as I had expected - but the impact is accelerating. Note that when I say AI, I really just mean LLMs. I don’t think we need to reach AGI for the impact to be hugely significant.

Here are some of the things we have discovered:

• Enterprise SaaS players are not going away at all - they are becoming more entrenched. The larger the installed base, the more entrenched they become. AI is not changing that anytime soon; on the contrary, it is helping those companies become increasingly critical to their customers. The fight for market share among the big SaaS vendors will be intense, and growth will come from AI. Expect enterprise SaaS to become purely consumption-based in the mid-term. They will bill you for an agent that does specialized work with your data in your context, with, surprisingly, their best practices that will look like yours.

• Niche SaaS in specialized markets are the ones most affected in new customer acquisition and growth. Custom-made software has made a comeback. Why commit to something that barely meets my needs when something else meets them perfectly? Particularly, agentic systems with customized UIs will win. SMBs with needs that don’t fit neatly into the SaaS feature set will create specialized agents to follow their own processes.

• Software factories are in vogue again. However, this time they are following a highly dubious path in which costs can be unbounded. A company can’t create software with agents without knowing how to do software. This will change over time. Agentic Software development (a factory unit) will most likely become a function in the company like finance or HR, but not as we know it today. This is a big topic that I won’t address here.

• Enterprise customers are starting to understand the impact of current AI capabilities, but remain highly skeptical about its ROI. This largely stems from a lack of examples of how AI works beyond current user-facing interfaces: ChatGPT and Copilot, which focus on chat; and agents, which look like automation on steroids. This is slowly changing as they discover what AI can really do.

• Chat-based work without increased customization has no direct ROI for most companies because it doesn’t enable tangible work products. The gains are for the users, not the business. There is nothing wrong with that - a business must create the work processes that replace the value that the AI now covers by changing what the employee must generate. Faster work is not always the answer. In the aggregate, automation will support growth, but these effects compound over years, not months (and AI token costs are accrued daily and are not capital expenses).

• Local inference is still in its infancy, but there are very encouraging signs that it will become a major player in the mid-term. We are already hearing from some of our more privacy- and confidentiality-sensitive customers about locally hosted inference (of the enterprise kind, not the personal local inference one). There will come a time, though, when most individual productivity work (a personal agent) runs locally most of the time; we are still far from that, hardware costs and inference hardware need to improve quite a bit (but things are moving really quickly).

There is still a lot of noise; the software industry is just starting to get transformed again. Right now, nobody questions being online 100% of the time (talking about software, not at the societal level); all of our computing devices are permanently connected to the Internet, and there will come a time when using an “AI” will no longer be a novelty; in fact, I don’t know if we will have a particular name for it.