Thoughts on AI in 2025
Every corner of the internet is dominated by content on and by AI these days. I was dubious about writing this piece, but figured a few bits were worth outlining my thoughts on current AI systems. Being on the periphery of the tech ecosystem, I don’t subscribe to the doom or the hype of AI but I have felt a tangible impact this year.
I believe the introduction of agents this year was a turning point, systems that can take actions in an environment to complete a task. This has given non-technical folks like myself the tools needed to build prototypes and simple applications.
OpenAI's new benchmark for economically valuable tasks, GDPval, demonstrates that reasoning models perform really well on software development tasks. These agents have also found a fan in Jensen Huang.
While professional coders use these to boost productivity, I have often seen these agents creating messy codebases that lead to downstream headaches. There is growing consensus that these tools are far from perfect. They lack the intelligence needed to finish the tasks end-to-end, and rely heavily on prompting structure and explicit instructions. I believe there is a long way to go till we see these become truly autonomous, a view that was validated recently by Andrei Karpathy.
Hardware is also having a tremendous year, especially in the consumer space. There were some major launches like Friend that went rather sideways. I am not entirely sure of the value of these devices, and personally do not want another piece of metal overtaking my life. I would rather my phone have these additional capabilities but we have yet to see anything meaningful from Apple.
They seem to be taking a more deliberate and patient approach. While some people think Apple is asleep at the wheel, I believe they are letting other folks burn the capital to see what clicks with consumers. They are taking this time to build the infrastructure and distribution, a much needed edge against the newer players.
An area that I find really interesting is workflow automation. Using applications like Gumloop, n8n, and Microsoft Automate, you can stitch together a series of steps that can run in the background to finish a task. For example, I built a simple flow that took raw interview notes, ran it through OpenAI’s API to summarize, sent the output to an Excel table, and then published to PowerBI for collective analysis. This took me a few hours to build, but cut down 2hrs of manual work everyday.
I believe real productivity gains will be found here, in automating mundane tasks. We are already seeing a proliferation of startups tackling this in various industries like insurance, supply chain, and immigration that are seeing strong adoption.
The weak link remains the reliance of LLMs that often tend to be “black-boxes”. The performance remains unstable with a lack of understanding about how varying inputs affect the output. The work being done to advance interpretability, which is the understanding of how AI systems operate, is the next frontier I am paying attention to. Until we can reliably trace reasoning, AI will remain a tool we use cautiously rather than one we trust.
Despite the advancements in model capabilities, safety remains fragile. For open source models, it is possible to identify and remove safety direction through a simple operation at a minimal cost with no loss in performance. Strengthening the guardrails especially in open-source models will not only promote safety but also reduce over interference by regulators.
AI’s most meaningful progress in 2025 isn’t in general intelligence or flashy hardware, but in invisible automation where humans quietly offload repetitive tasks to agents and flows. I believe the tools will continue to mature, and somewhere between automation and autonomy, we’ll find the balance that makes AI truly useful.