All this highlights the need to consider AI economic impacts holistically, and to create incentives and infrastructure to reproduce AI economic impact analyses regularly to track longer term impacts via improved capabilities and job redesign.

The results are consistent with their hypotheses: AI drives firm growth, occupations with high mean exposure to AI decline but this is often more than offset by complementarities and productivity growth. Big AI impacts with different signs lead to small aggregate changes.

They use LinkedIn job ads to measure firm-level AI adoption and occupation / task distribution, and embedding distance to ONET tasks to quantify exposure. They use instruments to reduce selection biases and mitigate mis-measurement (read the paper for full details)

In their model, AI displaces an occupation if its mean task exposure to AI is high, and augments it if its exposure dispersion is high (some tasks are exposed to AI and others untouched). They also consider impacts on labour demand though AI-powered firm growth.

2. AI is being used to identify venom proteins that have antimicrobial properties. The haystack is the space of proteins present in venoms (40M!).

Both studies validate their results in the lab, highlighting complementarities between AI & experiments.

This is a VC investment so it doesn’t make sense to frame as something that has to be repaid. If the investment is creates improvements in model performance that generate massive value in the future then it makes sense. That’s the expectation.

I guess those models are generating income now + you need to invest to go down the learning curve?

This paper is a timely reminder that tech impacts on jobs are dynamic and complex - so we need dynamic and complex models and data to track / anticipate / respond to them!
Read it here: www.sciencedirect.com/science/arti...

They also forecast skill mismatches based on the network structure of job impacts e.g. whether a job’s neighbours see higher demand (bad for employers) or lower demand (bad for workers). This could inform timely policies to accelerate deployment and reduce disruption.

They find that the net impact of rapid decarbonisation on jobs is positive but depends on the stage of the deployment cycle. Many jobs see growth during the “scale up phase” (build up of low-carbon infrasructure) but decline afterwards.

They forecast the impact of energy tech adoption scenarios on output and labour demand in exposed sectors, and the implications for employers and workers based on impacts in related occupations.

Some limitations: use of a top down coarse taxonomy; innovation isn't always link-creation; neglects hierarchical nature of innovation; patents offer a narrow view of societal expectations; expectations not evenly distributed...
Still really worth reading - check it here:
arxiv.org/abs/2412.08092

They analyse the distribution of suspense and surprise in technology codes within higher level technologies and characterise a technology lifecycle of surprise - suspense - popularity. This is what it looks like for AI. It would be interesting to drill down further!