The first regression model is in effect a collection of 3 sub-models, with each sub-model being applicable to a different type of degree holder.
Sub-model (1a): PhD degree holders
log(annual_salary) = B2 + B1*yrs_experience + e
Sub-model (1b): Masters degree holders
log(annual_salary) = B3 + B1*yrs_experience + e
Sub-model (1c): Bachelors degree holders
log(annual_salary) = B4 + B1*yrs_experience + e
The sub-models assume that each type of degree holder starts out with a different baseline log(annual_salary) but after that the log(annual_salary) increases at the same rate for each additional year of experience.
The coefficient B1 in each sub-model represents the amount by which the log(annual_salary) changes, on average, for each year of additional experience among people holding the same degree. (The change will most likely be an increase.)
The second model looks at the relationship between log(annual_salary) and yrs_experience, regardless of (or ignoring) the degree type.
The coefficient B1 in the second model represents the amount by which the log(annual_salary) changes, on average, for each year of additional experience regardless of (or ignoring) the type of degree.
As pointed out by a_statistician, the second model implies that if someone - regardless of their degree - has zero years of experience (i.e., no experience), then their predicted log(annual_salary) would be equal to 0, which is to say that their predicted annual_salary would be equal to 1 (the units of 1 would be the same as the units of the data). This implication may be either too strong or nonsensical altogeher, so I concur with you that the second model desperately needs an intercept.
There could be a third model considered, which allows the rate of change in the log(annual_salary) as a function of yrs_experience to be different for different degree holders.