A software developer writing at lzon.ca published a personal retrospective on March 28, 2026, tracing their experience with AI tools from ChatGPT‘s public debut in November 2022 through a two-month paid evaluation of Claude Code. The post, titled “The first 40 months of the AI era,” runs approximately nine minutes and covers hands-on experiments across creative writing, code generation, and autonomous project development. Author details were not available at time of publication.
- ChatGPT distinguished itself immediately from earlier chatbots like Cleverbot, but produced output the author described as “very boring and overtly inoffensive” — a limitation they say persists today.
- An attempt to vibe-code an MTG card placeholder app using Claude ended with the author replacing nearly all AI-generated code with their own across successive iterations.
- A Claude Pro subscription purchased roughly two months before publication led to a positive assessment of Claude Code as a new natural-language input modality alongside keyboard and mouse.
- The author distinguishes between AI’s limited utility for novel, iterative coding work and its “unambiguously good, useful” performance for well-defined tasks via Claude Code.
What Happened
The developer published “The first 40 months of the AI era” at lzon.ca on March 28, 2026, marking nearly 40 months since OpenAI launched ChatGPT in late November 2022. The post documents direct experience across three distinct phases: early experimentation with ChatGPT for content and code generation, a specific vibe-coding project that failed to reach its goals through AI prompting alone, and a recent two-month trial of Claude Code following a paid Claude Pro subscription. Author details were not available at time of publication.
Why It Matters
The post spans the full public arc of accessible large language models, from ChatGPT’s initial release through the current generation of agentic coding assistants. The author’s account captures a divide that has become common in developer discussions: broad skepticism about AI’s reliability for complex or novel code, alongside specific confidence in more constrained use cases.
The author first heard about ChatGPT’s code-generation capabilities from a segment on the Linus Tech Tips WAN Show shortly after the model’s November 2022 launch, where host Luke noted the model could be prompted to produce fully functional programs. That observation prompted their own testing. Their conclusion, reached over 40 months and across multiple AI tools, is that the technology’s practical value is real but unevenly distributed across task types.
Technical Details
Early ChatGPT experiments progressed from text generation — poems, Dungeons & Dragons character backgrounds, full fantasy world-building with kingdoms and lore — to code. The author found ChatGPT produced correct “hello world” programs without difficulty and could generate snippets for “common and well understood use cases,” reliably enough to replace Stack Overflow searches for routine problems. The consistent limitation was stylistic: outputs were described as “very boring and overtly inoffensive,” a quality the author characterizes as inherent to the technology and still present today.
The MTG card placeholder project is the post’s most detailed case study. The author specified requirements that included fetching card metadata from an external API, generating a QR code for each card, and laying out the combined data into a correctly formatted, printable page. The first output from Claude was described as “very impressive, and mostly worked.” Subsequent prompting to refine the result stalled, however, and the author made no meaningful progress through additional prompts alone. Across each iteration they replaced AI-generated code with their own. The final project “hardly used any AI generated code at all,” prompting the author to question whether the AI had saved any net time compared to writing the project from scratch.
Claude Code, accessed via a Claude Pro subscription purchased approximately two months before publication, produced a sharply different result. The author describes consistent success with natural-language commands given sufficient specificity: “So long as I was careful to clarify my intent, I was now able to tell my computer what I wanted it to do and it would consistently do what I asked.” They characterize this as “a brand new form of input and control of my computer along side my keyboard, mouse, and even command line terminal.”
Who’s Affected
The post is directed primarily at independent developers evaluating whether AI coding subscriptions deliver practical value. The author’s mixed record — reliable performance on well-defined tasks, failure on an iterative project requiring nuanced adjustments — is most relevant to developers using AI for personal or small-scale projects rather than large-team workflows. Their conclusion that Claude Code is “unambiguously good, useful, and just amazing” for natural-language computer control is qualified by stated doubts about AI’s broader coding utility, particularly for work that requires iterative refinement through follow-up prompts.
What’s Next
The author states they “would be very pleased to see this technology become fully commoditized,” framing current subscription pricing as a barrier to wider adoption. The post does not specify planned follow-up experiments or a technical roadmap. Its primary limitation as evidence is that it documents one developer’s experience across a small number of project types, without controlled comparisons, reproducible benchmarks, or multiple independent observations to support generalizations.
