The degree to which awe at AI occurs correlates closely with how much AI is used for coding

At the same time, a free, older Advanced Voice mode can stumble on dumb questions in short-form clips while the highest-tier paid Codex model can spend an hour restructuring code or uncovering vulnerabilities

TLDR: These groups are talking past each other.

Those capabilities have improved fastest because 1) the domains provide explicit, verifiable reward functions amenable to RL (e.g., unit tests pass/fail), and 2) they're highly valuable in B2B settings,

When given a terminal, these models can solve programming tasks that used to take days or weeks, restructure codebases coherently, or find and exploit system vulnerabilities.

This group experiences the strongest sense of "AI psychosis" because recent advances in those technical areas have been dramatic.

That yields a cohort that both pays for frontier agentic models and uses them professionally in programming, math, or research.

Part of that is the technical nature of reinforcement learning, which benefits from verifiable reward signals, and part is product focus: companies prioritize areas that generate more B2B value, so

Second, even paid access (e.g., ~$200/month) doesn't mean broad improvements are obvious. Most everyday prompts—search, general writing, basic advice—haven't seen the dramatic leaps found in highly technical

Those free or deprecated models don't reflect the capabilities of this year's state-of-the-art agentic models, especially OpenAI Codex and Claude Code

Many people tried the free tier of ChatGPT last year and let those experiences—model quirks, hallucinations, and viral short-form videos of an older Advanced Voice mode flubbing simple queries—shape their

First, recency and tier of use matter.

There appears to be a growing gap in understanding AI capability.

TLDR: Raw data from multiple sources is collected, compiled by an LLM into a .md wiki, then operated on by CLIs and the LLM for Q&A and incremental enhancement, all viewable in an editor. Manual editing

As the repo grows, synthetic data generation and finetuning could embed the wiki's knowledge in model weights rather than relying solely on context windows

Additional tools can process the data, for example a simple search engine over the wiki usable via a web UI or exposed to an LLM via CLI as a tool for larger queries Further explorations:

Linting: LLM health checks can find inconsistent data, impute missing information (with web searchers), and suggest connections or new article candidates to incrementally improve data integrity Extra tools:

Outputs can be rendered as markdown files, slide decks, or matplotlib images, then viewed in the editor. Outputs are often filed back into the wiki to enrich it for future queries, so explorations accumulate

Output:

Once the wiki is large enough (e.g., ~100 articles and ~400K words), an LLM agent can handle complex queries against it, researching answers. Rather than relying on a complex RAG setup, the LLM can