Searle (1980) introduced the terms to separate two claims. According to "weak" or "cautious" AI, the computer's value for studying the mind is that of a very powerful tool, one that lets us formulate and test hypotheses rigorously. According to "strong" AI, the appropriately programmed computer is a mind: it can literally be said to understand and to have other cognitive states, and its programs do not test psychological explanations, they are the explanations (1980, p. 417). Searle had no objection to weak AI; the Chinese Room targets the strong version.
His own answers: could a machine think? Yes: we are such machines. A digital computer? In a trivial sense, yes as well. What he denies is that anything thinks solely by virtue of running the right program, because formal symbol manipulation has no intentionality (1980, p. 422). No one supposes that a simulated fire will burn the neighborhood down: why would a simulation of understanding understand anything? And he charges strong AI with a residual dualism: if only the program matters, it could as well run on a machine or on a Hegelian world spirit (1980, p. 423) (see multiple realizability).
Over time the meaning drifted. Today "strong AI" often means AI able to do nearly everything a human can, close to AGI, and "weak AI" the narrow AI that already exists; M. Mitchell (2019) notes that this usage, which she associates with Kurzweil, is not Searle's (2019, ch. 3). The axes are worth keeping apart: generality (what a system can do) and mentality (whether it has a mind). For Searle, not even a general system would have a mind merely by running a program.
The argument targeted hand-written programs in the style of GOFAI; in 1990 Searle extended it to networks with his "Chinese gym" (Searle, 1990). Current models follow no rules anyone wrote, only learned weights. Does that change anything, given that they still run as programs?