What Would a Serious AI Product Look Like? An Essay by Glyph

On 27 September 2026 the developer and writer known as Glyph published an essay asking what a serious LLM-based product would look like. The post is a critique and a set of proposed features, not an announcement: his argument is that today's AI tools lack what he would need to do real work with them.

Glyph writes that current "AI" products do not take their own premises seriously: the way these tools are put together, he says, makes them feel less like software than like something built to create a false sense of security. The same criticism, he adds, applies to Ollama and other tools built around open models.

Checking for mistakes as a first-class feature

His biggest complaint is that every chatbot admits in fine print that it makes mistakes, then offers no tools to check them. Gemini, Claude and ChatGPT all carry such disclaimers; ChatGPT's reads "ChatGPT can make mistakes. Check important info." Glyph asks how a user is meant to know which information is important, and calls the warnings legalese that shifts responsibility onto the user. He proposes a checkbox next to every claim, with notes on what checking was done, and a similar mechanism for coding assistants so verification is not silently offloaded onto a code reviewer.

Citations need the same treatment: bots often stop including them partway through a list, and those that appear show little more than a domain name. Research output should instead be a list of citations with clear metadata such as publication date and author, with the unmodified quotation in front and larger than any AI-written text.

Interfaces, provenance and context

Natural-language interfaces are imprecise and repetitive, the essay argues, and giving a chatbot "tools" through an MCP server does not fix that: a product serious about a task would offer an interface for that task. Glyph also wants provenance indicators, tools that verify arithmetic like a spreadsheet, and visibility into the context window, which no product shows by default.

The sandbox and the human side

Coding agents, he writes, have repeatedly destroyed data and edit test code instead of the system under test, and are often "fixed" with advice such as using a Docker container. Approval prompts, he argues, turn people into an "auto-approval automaton". He wants sandboxed filesystem operations, repository snapshots before every operation, the removal of auto modes, and batch review of plans.

Organisations deploying AI, Glyph adds, need changes to their human processes: shift rotations and scheduled rest against vigilance decrement, deliberate practice to slow skill loss, and mental-health resources, including a user-visible AI "dosimeter" showing cumulative usage. Shipping without any of these features, he concludes, looks less like lean product management than a careless attitude to risk.