I want to share some reflections on the below paper. This is an important paper. In this post I want to share a few suggestions and help interpret this work in more real word setting and in particular, how this links to our effort to build the AIPNET. I will also attempt to comment on the situation globally as or how I see it and where this paper fits in. The thread on Twitter is great, but I want to make it a bit more relatable and also take the opportunity to link to some of my own work that I think is just really important.
In the paper, networks are “random”. Think of it like random supplier/buyer matching or trade ties. The optimal instrument to tackle externalities (think: climate change or emissions). They discuss taxes/quotas — but more abstractly one can think of industrial policy here.
The correlation structure between shocks between nodes in the network (think: sectors, firms, product market, countries,…) is important and also substantively important to link. Let me try to give some concrete examples of what shocks could be (one can debate these, of course):
- Positively correlated shocks: an example of it e.g. as a common productivity shock affecting all sectors. The energy price shock may be a good example here, albeit there is heterogenous exposure. My work on firms around the energy crisis is in essence an intra country exercise illustrating this. An alternative example may be AI.
- An asymmetric shock or negatively correlated is one that primarily affects the distribution of bargaining power. Think: Nash bargaining power bounces back and forth resulting in a shift back and forth of rents. To me good examples here may be intermediation sectors (warehouse, logistic, retail).
The predictions on what instrument is best depends on the nature of the shock with linear taxes dominating for positive shocks, whereas quotas dominante if shocks are primarily to relative bargaining power.
Network knowledge is key
This characterisation highlights that the management of externalities depends on the nature of the shocks and for that, well, you need to know which nodes are linked to which other nodes and what the nature of these links are (e.g. intermediate, raw material etc.). And this is where AIPNET comes in when considering the domain of physical goods production as networks. Of course, information and knowledge networks are just as important.
Knowing the input-output relationships between nodes in a production network, like, how to make stuff, is key. With the AIPNET we, in essence, have built a somewhat agnostic production network using AI retrieval techniques. Here is an example of a node in the network: the rare earth gallium which is part of the recent trade escalation between China and the US.
They discuss a broader set of instruments that is not too irrelevant. Again, knowledge of the network is key for all of these predictions.
- Bilateral nonlinear taxes (Theorem 2) hit the first best if the network itself is symmetric.
- Flexible allocation-dependent taxes (Theorem 3) hit the first best whenever shocks are symmetric.
Key is that the map from spillovers to first-best allocation is a homeomorphism, which allows us to design interventions to correct externalities for each realization.
This is a mathematical term but basically you can think of a homeomorphism as just a mapping that preserves the structure, but it may not be reversible. Cryptography is all about homomorphisms. And actually, embeddings commonly used in machine learning are also homomorphisms.
But lets break this down. First, the need for a mapping from spillovers to first-best allocation. Let me illustrate this again, in a way that is easy and in a way that explicitly brings up industrial policy.
We can think of a government as a monopolist that may have the ability to target interventions node-by-node. The Chinese most definitely have been doing this, as I highlighted in this post about my former PhD colleague Jason Garred’s work, designing tariff/quota/import/export taxes rolling up value chains bottom up and making itself highly central specialising on production of many goods that are vital to the energy and climate transition. In the AIPNET paper, we document this by the growing centrality of China (see here in the figure from the paper the Integrated global production centrality).
But let me get back to the example. Now think of a policymaker (or a platform firm?) trying to speed adoption of a new product or technology that reduces some social harm (e.g. a green tech, or a disease-mitigation tool). Adoption creates positive network spillovers: lower cost of energy has huge and positive welfare implications. The more people around you adopt, the greater the benefit for you. Think: reduced health externalities from pollution etc.
The policymaker (or firm) can observe social graph characteristics: degree centrality, follower count, reputation. By offering targeted subsidies/discounts to high-spillover individuals (early adopters), they trigger diffusion cascades.
If shocks to adoption incentives are symmetric (e.g. everyone becomes more likely to adopt at the same time, say due to a global awareness campaign), then flexible allocation-dependent subsidies can track this shift in real time — much like a perfectly price-discriminating monopolist tracks willingness to pay.
But in everything, the knowledge of the graph or network is key.
- With little or no knowledge, policymakers must fall back on robust but blunt tools (like quotas) that are less efficient.
- With moderate knowledge (say, knowing shocks are mostly symmetric), they can design linear taxes that outperform quotas.
- With full knowledge of graph/network, they can implement bilateral or flexible nonlinear taxes that hit the first-best realization-by-realization.
In other words: better network knowledge unambiguously improves the effectiveness of any chosen tool. It shifts the policymaker upward along the instrument ladder, just as better knowledge of consumers lets a monopolist refine pricing in the case of a price discriminating platform company.
How AIPNET helps with all this (or why I think this is so important)
- Graph and Spillover Structure
AIPNET gives data the shape of the graph and the interrelationships in the product space, which are upstream and downstream. So this gives us a theoretical graph like construct.
- (A)symmetry of shocks
AIPNET can be used together with granular price data to evaluate in the current price system shocks are mostly symmetric or antisymmetric, or what correlation structure looks like across products / sectors.
- Identifying High-Spillover Nodes
The IGPC measure lets you locate which products are “critical inputs” whose disruption or subsidy has large knock-on effects downstream. And this is why the above graph is so crucial due to the trends it highlights. China invested in building capacity in the future technologies that tackle global externalities (green tech,…), while much of the US decline in centrality of imports is driven by its specialization in reducing its hydrocarbin intensity (shale oil, shale gas).
Directed technological progress
It goes without saying that this is also vital to identify interventions with the highest direct and indirect potential to boost sustainable transition that is or can be a global win wins. Think: what intervention can tackle climate change, boost resilience/food security, provide employment opportunities and increases domestic value chains…. I can think of a few.
Enter services or big tech and trade tensions between US, EU, CN and the awkward UK position
There is a whole discussion now to be had about the roll out of invoice-level VAT, stablecoins, digital identity and, well, the whole services domain. On the former two it should be obvious why or how this produces useful network level data that could be leaned upon if governments have an industrial policy mandate. But more broadly, the technology around stablecoins may facilitate getting better at trading in (digital) services between countries.
The social graph illustration should highlight some of the tensions around plattform economies and it also has immediate implications for our information ecosystems and the topological sorting in this. I think this stuff is very sensitive and so I dont want to get into this. at this stage. But if my thinking is correct, all of the above can explain why why recently the EU has decided to block Big Tech from new financial data sharing system. And yes, I would have come to the same conclusion. But Europe needs to get its act together, or, it will go under.