We can’t afford to ignore the warning signs about AGI

Three defensive tactics Israel should put at the core of its AGI strategy

Reading the results of the IDF’s assessment as to why it didn’t predict and prevent the terrible attacks of October 7, 2023, led me to wonder what intelligence signals we are currently receiving about likely future threats we are unprepared for. One has stood out more than any other: a rare consensus has emerged across technology and government that we have two or three years before artificial general intelligence (AGI) or a substantially equivalent set of algorithms change everything.

AGI is different from the Generative AI (GenAI) most of us got to know through interfaces such as ChatGPT in its broad and autonomous application: While GenAI can do incredible things in a narrow set of categories, AGI can accomplish any set of cognitive tasks – any task that can be accomplished digitally – at a speed and level of excellence far surpassing any number of humans. While this holds tremendous promise for our ability to solve problems facing humanity in the long run, Ben Buchanan, former special adviser for artificial intelligence in the Biden White House, explained to Ezra Klein why the US and China are so worried about the short term consequences of the other getting to AGI first: its national security implications.

In short, AGI’s ability to analyze data and write code at current processor speeds could upend human-written digital systems securing our digital infrastructure. As a result, the party who develops AGI first will have a tremendous advantage over the one that doesn’t. Trapped into a classic prisoner’s dilemma, the winner will have the incentive to first use AGI to cripple their adversary’s ability to advance to their level, leaving the fate of their rival’s digital systems in their hands. Most at risk are the legacy systems managing our critical infrastructure: health care systems, legal systems, financial exchanges, systems whose corruption pose a clear and present danger.

While the first party to develop AGI may turn out to be benevolent, and seek no more than to usher in a new age of abundance and stability, we cannot afford to be caught unprepared. Even assuming that we can protect against traditional hacks – infiltration and manipulation of digital control systems – AGI will be the most powerful tool of persuasion ever. Imagine an autonomous system setting up its own network of agents able to completely replicate any person’s digital signature, deepfake any person’s digital presence, able to manipulate communications simultaneously on a scale previously impossible. Social engineering on a scale previously unthinkable, all to benefit the objectives of those who control it.

Israel cannot win the race to AGI. We do not have the compute capacity and will not be able to build it up over the next 2 or 3 years while developing the algorithms that will midwife it to existence. Even if we could buy all the chips we need, the risk that one of the other countries in the race – China, for example, who has shown its anti-Israel bias during the war, or Saudi Arabia who invested $100B to buy its way to the forefront, will probably get there first. This means that those who control the models will most likely be either agnostic to our interests or opposed to them.

And even if America’s Digital Empires succeed in winning the race, it is evident that the models they are releasing are either non-Zionist or anti-Zionist. Ed Rettig demonstrated in his test of the GenAI model Perplexity that large-language models trained on an internet full of anti-Zionist materials will adopt that bias in their output. While there is a significant difference between Generative AI (GenAI) and AGI, the fact remains that AGI will be trained on the materials available to it – which are not in our favor. To say the least.

So we need to act now if we are to be prepared to defend ourselves if needed. But how? I’ve found it useful to apply a pandemic prevention model when thinking about the introduction of AGI, given the potential for disruption and destruction. From the moment AGI is introduced and until we are able to return to a reasonable approximation of life as usual, how can we minimize damage to ourselves, our loved ones, our communities?

If COVID19 taught us anything, it is that we can immediately do three things: Isolate, Inoculate, and (re) Introduce normal life (gradually).

1. Isolate: We need to ensure our critical infrastructure has air-gapped backup and, if possible, an analogue mode to ensure fail-safe, infiltration-proof operations that can be switched to once AGI is released. Until we know our cybersecurity can withstand an AGI-developed attack, we cannot take the risk of leaving our infrastructure vulnerable. We need to flatten the curve. Since time is short, we should take a Bottom-Up Maslow approach, as Aharon Horwitz called it: start with the base of Maslow’s pyramid, secure there, and then move up.
2. Inoculate: While Israel’s worldclass cybersecurity industry is already developing protections against AGI-driven attacks, the thing about unimaginable situations is that they cannot be easily predicted. That is why isolation is critical to help us develop the inoculation to create a new class of AGI resistant infrastructure. And as we learned in the pandemic, having a vaccine doesn’t help if you can’t efficiently distribute it. We can start planning now how to organize our efforts to roll out AGI-resilient tools to the underlying digital systems we depend on for everyday life.
3. Introduce: While we may never be fully safe from the threat of digital manipulation, once we can trust our systems to fail-safe and trust our digital immune systems to protect us once they identify a threat, we can begin the process of introducing AGI into our lives. Organizations such as Mosaic in Israel and the Global Governance Institute in Brussels are exploring how we will need to adapt our workforce and our democracy to an age of AGI. There will be other ways AGI will affect our lives, our societies, that need to be thought through and socialized. The more scenarios we prepare for today, the more society-wide wargames we run, the faster we will be able to adapt our society to the new normal AGI will create.

These, in addition to creating a second-strike capability to threaten a foreign adversary’s digital infrastructure to deter attacks (or neutralize attack vectors as I described in my 2013 novel, Do No Evil), are the least Israel can do to defend our citizens from an adversarial AGI. With an estimated two years until AGI, as Rabbi Tarfon once taught, “the time is short, and the work is plentiful.” We need to act quickly and best we do so in collaboration with other smaller nations facing similar threats who are capable of helping us achieve our goals: South Korea, Australia, Taiwan, and Singapore.