They can’t, and don’t. The secret is that experienced PhDs (mostly) dominate the high end of “AI” hiring, but don’t have much title or departmental differentiation from people who are “only” specialized software engineers in top tech companies. But this isn’t very well known, large companies want to recruit as much talent as they can regardless of role, and startups want to compete
on paper - therefore, you have the following effects:
1. Startups give whatever sexy title they want to the people they can afford, which makes titles fairly useless, because they almost never can afford the talent commanding “sky-high” salaries.
2. Within large tech companies like Google and Facebook, it’s hard to immediately tell which of the many data sciencey, machine learning-y titles correspond to the truly stratospheric salaries versus the engineers that work with those roles. For some teams, like DeepMind, Google Brain or FAIR, it’s easier to tell. But for others it’s a mixed bag.
For comparison, see the fashionable term of art “quant” in the financial industry, which has similarly devolved into marketing and a bimodal distribution. As a rule of thumb, you generally can’t trust AI titles or salaries at startups unless those startups are really known for their talent; further, you can safely assume that, at companies capable of paying for top talent, the very impressive salaries belong to titles which seem the most exotic and out of reach for general engineers in the job description.
Fortunately startups don’t really need to compete for top talent as a genuine technological differentiator, they just need to engage in signaling, so this is mostly a non-problem. Startups almost never have problems usefully improved by the cutting edge of machine learning research, and can instead use off the shelf tools and existing software to accomplish the same things. Frankly, it’s exceptionally rare for a startup to even have the massive data, pipeline and munging infrastructure requisite for actual research.