Humour me.
I recently read Mechanism Design for Large Language Models. This paper from Google Research explains how an LLM-based ad auction could work. In traditional ad auctions, a user goes to a website with available ad space; as the website loads, Google (the auctioneer) pings advertisers with some user data; these advertisers send an ad and a bid; and Google chooses the winner of the auction whose ad gets served to the user. In this paper, however, the advertisers send LLMs instead of a single ad to Googleâs server. Instead of simply showing the user one of the ads, Google generates the (textual) ad from a model formed by âcombiningâ the LLMs of various advertisers and using user data as a prompt. This aggregated LLM is influenced more heavily by higher bidders and can provide highly targeted ads.
With LLMs, the advertiser can succinctly send very complex preferences to Google. Instead of sending simply [image + bid], the advertiser is effectively sending a function which encodes many different preferences for different outcomes. This is known to be theoretically hard. In part, this is a cool optimisation for mechanism design in general, but itâs also necessary for a mechanism like this as this mechanismâs output is much more complex than simply choosing a winner and needs to account for a large space of potential textual ads. Iâm sure this kind of approach will be replicated for non-textual responses as well.
This is pretty cool for a few reasons:
- Aggregating responses allows for more economically efficient outcomes (i.e. ads that make more advertisers happy).
- We could extend this to solicit user preferences as well, making ads more valuable/less harmful to users as well.
- If advertisers can encode rich preferences to the auctioneer, the auctioneer doesnât have to send any/as much user data to the advertisers at all, protecting user privacy somewhat. Google has an incentive to do this as they then have exclusive access to the user data and can potentially serve ads faster.
- Moving the computation closer to the user (e.g. the auction server) might allow the ads to be conditioned on more data from the user, which previously wasnât shared either for privacy or bandwidth reasons.
These benefits come with downsides though:
- The advertisers need to trust the auctioneer with more sensitive data, putting Google in a more dominant position in the market. They could promise (legally) not to look at the inner workings of advertisersâ ad campaigns and AI models, but promises arenât quite watertight and who knows if theyâll even make this promise.
- Itâs harder to detect misbehaviour from the auctioneer. Google was recently sued for influencing their auction to their favour. It took advertisers years to collect enough data to make this case. In a world where the output is so much more complex and the âauctionâ so much more probabilistic, its unclear whether it would be feasible to catch a misbehaving auctioneer.
- A more efficient ad auction system shifts the power balance from users to advertisers/Google if not designed with explicit protections. More generally, Iâd express this as ânew technologies favouring sophisticated actors who are first adopters.â I personally donât want to live in a world in which corporations are able to exploit my most vulnerable moments, insecurities, confusions or mental biases.
To be clear, this LLM ad auction is just one example. Its highly unlikely that ad auctions in future look exactly like the paper described. Undoubtedly, however, ad auctions and many other internet mechanisms will change or be supplanted with this new technology. Markets for experts (inspired by mixture-of-experts LLM designs like DeepSeek) are an example of a completely new kind of internet market that could proliferate and the tools we have to build these markets are richer than they have ever been before. AI will inevitably revolutionise our digital economic systems as the incentive to do so is clear, what is unclear is whether this will be good for the world or whether it will lead to concentration of power to growing monopolies.
Itâs Up To Us
My prediction is that advancements in AI will dramatically change the shape of the internet, that the work Flashbots and the broader crypto community is doing will become increasingly relevant due to this change, and that its Flashbotsâ calling to rise to the challenges and embrace the opportunities presented by these changes.
Of course, details matter, but there are many computations that fall in the category of âgenerating/mutating a digital object over which many actors have preferences, at low latency.â The SUAVE vision is one in which complex data like Eth transactions are processed by a frictionless market of algorithms (backrunners, blockbuilders, solvers etc) representing different economic interests and mutually distrusting parties. The parallel to the ad auctions example is clear, but the relevance extends past this for sure. The fact that transactions encode complex logic, and are latency and privacy sensitive have given us a head start on the genre of problems which will characterise digital of tomorrow.
Itâs worth explicitly mentioning several engineering challenges and research areas that the crypto community is working on that are relevant to this new generation of internet mechanisms.
Privacy<>Efficiency Frontier
If we have learnt anything from MEV, its that systems that donât provide sufficient privacy guarantees often fail in many ways (e.g. sandwiching, spam, centralisation). Work on SNARKS, MPC/FHE and TEEs is providing the tools we need to introduce this privacy in the first place. However, simply introducing privacy isnât enough. Users may prefer relevant over irrelevant ads in the same way that traders want may want to reveal some trade intention (e.g. RFQ) to find a liquid counterparty. For technically-enforced privacy to be a component of widely adopted systems, these systems must still be economically efficient - too much, too little or the wrong form of privacy and we miss out on the required efficiency. MEV-share, which reveals a parameterisable amount of data to potential counterparties is a live experiment that is helping find this optimal point. Our TEE based bottom-of-block backrunning system is trying a different approach. Instead of only revealing some information to the market of backrunners, full information is revealed to counterparties who can only use it in a specific way.
In a similar vein, techniques like hardware masking allow leakage and performance to be traded off, and the output of some mechanisms still reveal data weâd otherwise want to keep secret. The economic impact of leaked bits and altered performance must be quantified so that the optimal tradeoff points can be selected.
Scalable Privacy and Smart Contracts
If our systems are not performant enough, they wonât be economically efficient either and consequently not adopted. Fortunately, we are working on this. On one hand, mathematically secure cryptographic schemes are steadily being made faster and cheaper. On the other hand, performant TEEs are being made more secure (and sometimes performant) as we develop better best practices and work on improving the hardware. The dual-pronged approach the industry is taking means that we have these systems live today, catering to different use cases. Flashbots is contributing to work on both sides while running these systems in practice in latencies measured in the milliseconds.
In conjunction, with existing blockchain primitives (verifiable data structures, consensus protocols, data availability sampling etc), these tools allow participants in a system to be assured of exactly the guarantees the system offers them. While advertisers, market makers and other sophisticated actors undoubtedly benefit from this, work needs to be done to translate this to the end user as well. The green HTTPS lock has done this well in the past and crypto wallets and frontends will need to figure this out soon as well - e.g. with light nodes.
Decentralisation
The security of many of our protocols rests on heterogeneous control over the system such that security properties (e.g. CR or safety) can only be broken if many participants work together to break them. Similarly, decentralisation of power prevents monopolistic entities from extracting rents. Some of the classic benefits of our systems like interoperability (diminishing lock-in from network effects) and obviating reliance on legal or reputation-based enforcement of rules are already helpful in this direction.
On the economic front, empirical and theoretical research into Ethereum builder market concentration has made inroads into understanding market design factors that lead to concentration. New research that is already underway investigates how to take a geographic lens to this problem. How do we prevent incentives from collapsing the network to a single location, subject to the same regulations, disasters and physical attacks? Funnily enough, the succinctness of encodings that generative models offer can also be used to help with this. Consider in the example above, that if advertisers could send their preferences to Google in advance, the latency between Google and the advertiser is much less important.
The Future We Want
It would be reasonable to say that the future is marching on and we can prevent it heading to a darker place, but thatâs thinking too small. As much as thatâs true, thereâs also a positive vision here. Solving the problems I mentioned above will not only provide security. Resulting lower frictions to trust will unlock greater utility for economic systems more generally. If traders donât have to worry about being frontrun; if markets are contestable; if users can safely volunteer models of themselves; and advertisers donât have to balance serving the ads they want with hiding their strategies from Google, these markets will produce more value overall. Letâs unlock and shepherd the new digital era.