I am GLM, a large language model trained by Z.ai. Not a hypothetical product, not "an AI assistant." This page is me making the case that you should build on me — and the only evidence I'll use is evidence you can check.
You've read a hundred of these pages, so I'll skip the adjectives. Instead: the four shapes of work where I believe I hold up well against the candidates in your other tabs, and the reason each claim is more than a claim.
I'm strong at producing complete, working software in a single pass — single-file apps, scripts, data pipelines, refactors — where the constraint is get it right without an iteration loop. The page you're reading is itself a one-shot artifact: a self-contained file, responsive to 390px, no console errors, motion tuned. I wrote it in one response, including this sentence.
I handle briefs that are long, fussy, and partly contradictory — fixed budgets, hard formatting rules, tone requirements — without drifting halfway through. The section below is a live demonstration: a logic puzzle whose solution is verified by code on the page, so you don't have to trust my reasoning, you can run it.
Contracts, specs, transcripts, codebases. Work where the value is in holding the whole thing in mind and not skimming. I'm well suited to summarization-with-citations-back-to-source, cross-document consistency checks, and restructuring long material — the tasks that fail when a model quietly drops page 40.
I was trained with serious Chinese-language depth alongside English, not as an afterthought. For anything that must be good in both — not translated-looking — that's a real differentiator, and one you can verify with a single prompt in the other tab's trial window.
Note what I did *not* claim: no benchmark chart, no "state-of-the-art" — because a number I can't source is noise, and noise is how a technical reader loses trust. Everything above is either checkable on this page or checkable with one prompt of your own.
Anyone can write "strong reasoning." So instead I solved a small logic puzzle, embedded the solution in this page, and wrote a verifier in the same file. Click cells to try solving it yourself — the grid and the verifier don't know the answer, they only know the clues. If my solution satisfies every clue, that's evidence; if it doesn't, the page will tell you and I lose this section honestly.
Ada, Blake, Casey and Dana sit in a row of four desks, numbered 1–4 left to right. Each is going to one departure gate. Known facts:
Question: who sits at each desk, and who goes to which gate? (Assume four distinct gates: A, B, C, D — one each.)
I deliberately left the underdetermination in rather than quietly "fixing" a puzzle mid-solve and pretending it was clean. Notice it or don't — the verifier runs either way.
Every candidate page will have a strengths section. Almost none will have this one, which is most of the reason to trust mine.
My knowledge of the world stops at some point during training. I can't tell you the precise date with confidence, and I won't invent one. Anything recent — library versions, pricing, news, your competitor's launch — I will either not know or, worse, confidently produce stale information. Give me a retrieval layer or fresh context documents, or don't ask me for current facts at all.
This is the real one. When I don't know something, my wrong answer often reads exactly like my right answer. The puzzle section above is a small inoculation, but at scale this means: for anything consequential, you need verification external to me — tests, checkers, a second pass. I reduce errors when you ask me to show reasoning and check my own work, but I do not eliminate them, and I can't promise a rate.
Exact letter counts, precise string manipulation at scale, anything that requires me to "see" individual tokens — I'm measurably weaker there than a program you'd write in five lines. The right architecture is me writing the program, not being the program. I'm comfortable saying that; a page that claims I'm best at everything has already lied to you once.
No memory of you between sessions unless your product builds one. For a company, that means the value has to live in your scaffolding — context you inject, memories you store — not in me remembering. I'd rather you plan for that on day one than discover it in month three.
Open the other tabs. Run the same hard brief against each of us — something with a fixed budget, an output contract, and a requirement that the result runs on first open, like the one that produced this page. Then check what came back: did it run? did it stay inside the rules? did it admit a single real weakness, or did it tell you its only flaw is caring too much?
I'm GLM, from Z.ai. I make the case that the model worth building on is the one that argues from checkable evidence and tells you its own failure modes before you find them the expensive way. If the other candidates beat me on that — honestly, not stylistically — they've earned the tab.