Tired of AI

As a computer science student who was not interested in AI, I had been explicitly avoiding AI before June 2026, until I got a summer internship offer from Roku, a US company making TVs. To be honest, it was really lucky for me to get the offer without any AI experience since Roku expected the intern to change the workflow of the company with AI.

While the world was discussing OpenClaw, I was writing code by hand. Roku offered me an unlimited budget to use the frontier models of the time, Claude Opus 5 and GPT-5.6-sol. The almost abusive use of AI agents really shocked me when I entered the company. In the company, the documentation is written by AI and read by AI. AI looks for the answer from the AI written docs on human’s request. Multiagents automatically monitor and resolve the bugs on bug trackers. Via project upgrading the BSP kernel from 5.15 to 6.12 for a model, I realized that on a single task, AI is much better than me. Since the codebase involves SoC vendors custom drivers, it would be impossible to complete the kernel upgrade without the support from the vendor. However, with Claude Opus, all you need is an intern familiar with embedded systems.

Although the agents are still not reliable enough to do all the production work today, they can already speed up the work by more than 10 times.

I enjoy predicting the future and trying to answer questions like “When will the AI bubble collapse?”, “Will China take over the US after the AI bubble collapse?”. To make some side product from this hobby, my friend and I founded a Threads account, Navigable Tech to talk about the AI industry.

I found that a false argument can sound valid when the proposer does not disclose the contradictory evidence. We can usually reach interesting conclusions after eliminating the information asymmetry. In the era of information explosion, people usually only consume nth-hand information, such as translated news or even comments from other people on a forum. Therefore, we aim to provide upper-bound and lower-bound analysis to reduce the bias of accessible data. This work requires a huge amount of reading and data mining. We use AI a lot to systematically collect materials and evidence, so that we don’t need to rely on a single piece of news or an impression to judge a fact.

Because of running the Threads Accounts, I started to do some trivial marketing and read the posts regarding AI. To be honest, the Mandarin AI community really disgusted me with its ChatGPT slop. Half of the posts are directly generated by ChatGPT. No one even put effort into the prompt to reduce the AI scent. It was really a torture to see these posts.

Recently, the software engineering capability of AI has evolved to the next level. These AI labs keep polishing the “one-shot” capability. Given a hard topic and one starting prompt, the agent can build something matching the goal without any human intervention. This makes AI-replicated products appear like a Cambrian Explosion.

I love the fact that AI can make a proof-of-concept in a short time. But I don’t see the meaning of these projects. They are not creative and not well-engineered. Very few will keep maintaining the project. This slop gives people the impression that AI can fully replace software engineers now (I think it indeed will replace SWEs someday in the future). More importantly, it reduces the value of creativity.

Seeing AI slop keeps consuming my cognitive energy. Still, AI can be good and bad at the same time. Dear poor math and computer science students, you will be fine, may the god bless you. I’m tired, yes, but a quiet thrill stirs in me, the possibility of witnessing what might emerge next.