When I stopped by MIT’s Dome at Davos this year, I sat in on a very interesting panel about international AI development that included Sir Demis Hassabis, the Nobel-prize winning co-founder of DeepMind, discussing (amongst other things) the dangers of “rogue” AI. Sir Demis said that he was “cautiously optimistic” about the future, which actually left me feeling pretty good as I wandered back out into the Swiss mountain air. So if agents aren’t going to kill us, what are they going to do for us?
Madison Avenue Agents
In a recent podcast discussion with the New York Times, Sir Demis talked about the near future for AI. Naturally, as the guy is a genius, it is all worth listening to. But one thing he said in particular caught my attention. He was talking about what truly useful AI assistants might look like and he said that if a personal agent knows you well then you could teach it to “protect your attention”. He went on to explain what this means, by talking about how social media is vying for your attention and how an AI could defend your attention from “being assaulted by other algorithms” that want to capture (and, indeed, overrun) it.
To protect your attention. What a succinct and evocative phrase to describe what AI can do for you. When it comes down to it, your attention is a very valuable resource, so protecting and managing that resource is really important.
This is such an interesting way of thinking and just one example of why he has a Nobel prize and I do not. His vision of a digital assistant that could pan the media to get what you need, to get the “nuggets” (as he calls them), so that you could avoid polluting your mind and your mood by dipping into the raging torrent to find the valuable piece of information that you wanted. I rather like the metaphor of my agent diligently exploring for gold in a raging river of lies, nonsense, manipulative tripe and propaganda while I get on with other things.
Now, as Joe Marhese notes, our entire online economy is built on top of systems trying to capture and optimize our time and attention. Advertising promises to deliver a message in return for this attention. Performance marketing promises outcomes, which has an entirely different meaning when it comes to agents and the models behind them. This makes me reflect further on something that Ken Mandel wrote about how marketing is going to be disrupted because while humans are the target of advertisements, in the near future the kind of agents that Sir Demis is predicting will filter everything including the advertisements. He talks about the emerging “AdTech” world where brands, agencies and agents will use structured data to communicate and to exchange data about wants, needs, preferences and offers. Or, as he rather neatly puts it, the future of advertising is “code, not copy”.
This is where financial services are going. When my agent wants to open a savings account, it will negotiate with the agents of financial services providers. But how will it choose which one to select? Assuming that it is going choose on something more than price, that is. Service providers may need to start thinking about reshaping some surfaces towards agent preferences, which might differ from human preferences. I can imagine that agents might prioritise factors like the speed of response, uptime, data accuracy and so on. Optimising for these non-human factors could provide banks a competitive edge, so they should probably begin to assess their strategies for the medium term.
(But I can also imagine asking my agent about some softer factors. That is, agents might be programmed to optimise for value, not just cost. They might weigh a variety of other factors such as long-term reliability, risk-adjusted returns, compliance and alignment with ethical frameworks. For example, an agent might recommend a slightly more expensive investment product if it offers superior risk management or aligns better with a user’s ethical preferencesI might choose a Gretabot that prioritises “green” investments or I might choose an Elonbot that invests in space or a Dimonbot or whatever else. That is, different people will be panning for different nuggets in the same stream.)
When it comes to choosing between two similarly-ranked choices though, I rather suspect that Ken is right to point to what he calls “trust signals” as the key because agents will prioritise trusted brands not based on our human and emotional version of trust but on a more measured version based on reputation, verified reviews and sustainability metrics instead of traditional advertising. This reinforces my view about the priority of establishing standard ways for extend identification, authentication and authorisation services to agents.
I actually see this within Sir Demis’ cautiously optimistic context, because I think that it will be much harder to “game” trust in an AI world. Imagine, for example, that when your AI pays the bill at a restaurant, the restaurant bot gives your bot a cryptographic token, some form of zero-knowledge or blinded proof that you ate at the restaurant. Then in order to submit a review of the restaurant on some online service your agent would have to provide that proof. The restaurant would not know who you are, but it would know for sure that you had dined there.
Agents In Charge, Charges In Agents
As it happens, I am organising a trip to Boston in a few weeks time. For such a trip, I don’t want to presented with 30 different hotel options: even the most rudimentary agents ought to be able to narrow that down to three to show. I do not have time to search restaurant reviews or go back through my calendar to see where I invited guests last time, my agent can do that and show me three it chooses for me. I don’t know whether to book a midweek supersaver fare or a roundtrip Wednesdays only carry on bag fare, I’ll the agent choose – it already knows I prioritise punctuality over online meal choices.
It was impossible to escape AI at Davos this year, since every conversation touched on the shift toward agentic AI. To be honest, I’m looking forward to this new life. While the agents work tirelessly to protest my attention from data assault and battery—ranging from impossible to assess offers to sell me hotel points to endless solicitation for new credit cards—I can get on and carry out more useful (and certainly more creative) activities.