Speed of the Valley Pokes the Pentagon Awake

Day 2 at the AI Conference

Washington DC 9 May 2024

When the ‘Silicon Valley Way’ broke Primetime

There was a watershed moment at the conference yesterday where move fast and break things finally broke primetime.¹ The historian in me imagined it was a bit like being a fly on the wall of the Compiègne Wagon. The ‘Silicon Valley Way’ of doing defense appeared to have reached a tipping point. The creature created by WWII and Cold War - The Prime - went from predator to prey. The comfortable, predictable, sclerotic, process-obsessed, pin-stripe ways of doing business - where The Prime is treated like a rare protected species - appeared to finally collapse before our very eyes under the weight of its own self indulgence. The eager, casual shirt and jeans, mission-focused crowd, could no longer politely conceal their impatience. On a panel filled with symbolism, Google, Microsoft, and Rhombus, all lined up against Lockheed Martin - who defensively bleated at the open ‘we have an AI and it updated Aegis in the Red Sea to fight the Houthis doing in 10 hours what used to take 3 months’. At one point Dr Anshu Roy of Rhombus blurted out “Why cant we be made a prime, what’s the big deal”… at which point Dr Steven Walker of Lockheed Martin chortled to himself as if to say “you cant possibly understand what goes into becoming a Prime, and you will never get there”. It was an electric moment.

Rhombus versus Lockheed

The changing of the guard was personified in that moment. A tall lean immigrant, who developed an incredible AI that predicted the second Russian invasion of Ukraine a full month before it happened, was sat next to a well fed old white guy in a suit whose top shelf product costs more than $1.7Tn, is ten years late, 80% over budget, and is fully-mission capable only 29% of the time.²

The F-35 is not an isolated case. On a recent visit to Japan and South Korea, who both produce “high-quality ships, including Aegis destroyers, for a fraction of the cost that we do,” the Secretary of the Navy was stunned our allies were using “digitization and real-time monitoring of shipbuilding progress… down to individual pieces of stock materials.” He marveled that they “could tell us — to the day — when ships would actually be delivered.”

The crisis in American shipyards matters when we are on the verge of a war in East Asia. China’s navy is expanding by roughly the equivalent of the entire French navy every four years. BG Mark Clingan USMC - Deputy Commanding General of Marine Corps Combat Development Command in Quantico observed at a conference recently

One Chinese ship yard produces more ships in a year than all of ours combined". US "procurement is optimized for the 1990s and post cold war era. Our production may not be sufficient for the future conflict.

The slide below tells the story

Upon SECNAV’s return to the US he fired a broadside midships right into the rotten oak paneled boardrooms of The Primes - accusing them of profiteering.

You can’t be asking for the American taxpayer to make greater public investments while you continue to goose your stock prices through stock buybacks, deferring promised capital investments and other accounting maneuvers that—to some—seem to prioritize stock prices” instead of “making the needed, fundamental investments in the industrial base at a time when our nation needs us to be all ahead flank.

Clearly the revolving door will be jammed if Secretary Del Toro wants to enter the China shops of The Primes when he retires.

Dr Eric Schmidt pointed out at every opportunity that the old ways of doing business are failing America and its national security priorities. He recounted a series of anecdotes arising from his journeys in national security over the past decade where ridiculous inefficiency was absurdly hard to fix. Like the uniformed personnel who had to memorize numbers on one set of screens and manually enter them into another set of screens. When he asked how this could be fixed he was told a multi-year multi million dollar acquisition program would be needed. He lamented “even I knew how to write the code to fix this in a few minutes”.

Generals & Geeks Speed Dating

Day 2 of the AI conference was far more dynamic than the more lackluster Day 1. Perhaps this was a function of the big-set pieces on Day 1, vice the multiple simultaneous small stage events running every 15 to 30 mins that dominated day 2. It was ‘get to the point or die’. If an item did not meet expectations, people just got up and found something else that did. It was essentially speed dating between Generals and Geeks.

Admiral Sir Tony Radakin UK CDS, gave a stirring assessment of the war in Europe, reminding the audience that Putin has failed all of his strategic goals and his recent tactical gains are minor and at an unacceptably high cost. He impressed upon the audience that not only has NATO grown in size and capability, member states have also substantially increased spending in light of the threat to the continent by an increasingly desperate Russia whose economy is being “bent out of shape” by maintaining the war effort.

Always refreshingly outspoken SECAF Kendall did not disappoint. He prefers the term “Automation” to AI. This is not just semantics, it goes to the fact that humans exercise control and that systems are taking on more and more automatic tasks - like autopilot which has existed for years. His point being that development is more incremental than some might make out. Having said that, he then said most of space is automated and that when it comes to satellite management “we have a cell phone app for that”. War games hang on whether a satellite is taken out. He reminded the audience in the real world 1000s of satellites will be targets and we have to get on top of orbital warfare.

He spoke about his recent flight in an AI driven F-16, noting that it is not quite ready yet but it is only a matter of time. This is relevant to the USAFs plans for CCAs (collaborative combat aircraft) to augment the low numbers of manned aircraft we will be able to field against a numerically stronger China. Of course they will have CCA’s too. There was a conference-wide assumption that we are so far ahead in AI, that China having CCA’s will be insufficient to gain advantage. MIL always gets a little anxious with sweeping assumptions on big questions. On a different panel, Eric Schmidt made the point that certain types of chips are banned for export to China. His confidence that this would hold them back seemed to fly in the face of all the other US defense IP they have stolen… why should chips be any different? Manufacturing techniques perhaps? They are very capable, if they cant make them, they will steal them. Again, this is a big hinge to depend on.

AI & the IC

The last two panels MIL attended were the most AI centric and interesting of the whole conference. Both heavily featured current or former CIA personnel. The first was the State of AI in the IC and the second the Future of Open Source Intelligence. In the first, Dean Souleles, former Chief Technology Advisor (CTA) at the Office of the Director of National Intelligence (ODNI), Dr John Beieler, Assistant Director of National Intelligence (ADNI) for Science & Technology, were joined by an NVIDIA rep and a researcher from the Alan Turning Institute in the UK who seemed remarkably well versed in the IC.

There were a few really good take aways from this panel. The first was something that struck MIL as a future money maker. Pure data - data unblemished by intermixing with AI generated data. It was likened to low-background steel. The point being that AI needs massive data sets upon which to train. The phenomenons of hallucination and model collapse are created and exacerbated by data that has AI data within it. It has been estimated that it only takes 5 generations for total model collapse. Accordingly, data purity is going to be an incredible challenge - like keeping GM crop seeds from blowing into non-GM fields - and a likely big money spinner. The problem is data is time relevant. So there will be trade offs between purity and when the data was collected. Many use cases will need current data and will be adversely impacted by impurities.

Appropriately enough there was a lot of talk about AI bias. However, the response to this was quite simple in a sense: tradecraft and triangulation. No IC analyst runs with one datapoint and that does not change with AI inputs. They are subject to all the same checks and balances and QA processes as any other intelligence product. The issue of what different groups mean by bias also came up. There is a difference between the scientific and political uses of the term. Evidentially much hilarity ensued when CIA officers met with civil rights activists to discuss bias. From the scientific side, there is technical bias in everything - which sample set is selected, which AI used, and so on. All involve bias. Whereas the political is concerned about the biased purposes to which biased data its used. An important and not at all subtle distinction for example if you are selected from a facial recognition positive at an airport when your race has not yet been properly tested and coded in the AI.

Lots of talk about acceleration past Moore’s Law! RAGS - Retrieval-Augmented Generation came up as a solution to hallucinations. “It’s a technique that enhances the accuracy and reliability of generative AI models by incorporating facts fetched from external sources . An AI explains that “the concept is akin to a judge in a courtroom who, while having a broad understanding of the law, may need to refer to specific cases or precedents to make an informed decision.”

Finally, discussion ranged over the issue of IC contracting for AI. The problem is multi-fold. First, the flood of new companies. One govt official talked about their job requiring them to check every 90 days for new AIs that might help with a task. Second, most AIs are very narrow-band designed for very specific needs. Third, the licensing model. It is standard in the software industry but it does not fit the IC or DOD that have tens of thousands of users. So finding a way to make a whole lot of fast changing companies offering little AI work for a whole lot of people can be a big challenge. The panelists long for bigger, wider aperture, ‘do-it-all’ AIs to go some way to solve the problem. This sounded to MIL like a future AI Prime model as the integrator of AIs and provider of a ‘one-stop-shop’ solution. So maybe Lockheed will get the last laugh after all?

The biggest secret in the IC is that, by volume, secrets rarely matter to the IC

The Future of Open Source Intelligence was for MIL the most interesting panel. It started off with the usual fun statistics. For example, every single day, approximately 28.08 billion photos are being stored online. YouTube hosts over 720,000 hours of videos daily, which is equivalent to about 4.3 PB of data.³ But that does not capture the true scope of the challenge. This does: 90% of the world’s data has been produced in the past 2 years. Only an infinitesimal fraction of that is classified data. If one byte of data equals a grain of rice a zetabyte would fill the Pacific Ocean. In 2023, the world created around 120 zettabytes (ZB). As the foregoing suggests, the IC is in a high volume business.

The biggest secret in the IC is that, by volume, secrets rarely matter to the IC⁴. AI is the only way to sift and sort the tsunami of data being produced every second of every day. In fact, AI thrives on data at scale. Additionally, OSINT can act as an integrator across the other INTs, almost all of which either have open sources of their own or are reliant on other open source data. Finally, OSINT is essential for analytical tradecraft. Even exquisite data needs to be triangulated.

OSINT’s relative standing inside the IC will grow in direct proportion to the importance of AI to the IC mission. Exponential will displace the exquisite. Consequently, it will not be long before it finds itself in the drivers seat - a position once held by “the crown jewels”. This reality was reflected in the comments on the panel where it was pointed out that the DNI is personally invested in OSINT as the vehicle for transformation of the IC. How could it not be when you have to navigate zettabytes. Nothing else matters.

Some generic but juicy examples of AI augmented work were provided by active CIA staff on the panel.

1. A crisis last week required 2 analysts to come in for at least 10 hours. They used AI on OSINT and resolved it in 1 hour. It sounded like it took them longer to get to their desks and log in (that car park hike!) than it did to get the answer.

2. It is possible to circle an area inside Russia and find out what is being said inside that zone “instantly”.

3. OSINT AI can produce finished summaries drawing on reams of material from all sorts of open databases, all of it fully referenced, in a matter of seconds. The referencing function was a key element that they had to get right.

Compared to other INTs OSINT has a huge effect for negligible investments. The client does not care where the IC gets its analysis from so long as it is accurate, high quality and timely.

MIL was very pleased to see one of the panelists, Kristin Wood, get into China’s cognitive warfare plans. This is one of those terrible labels for an age-old skill - in this case propaganda and psyops. However, the tool kit has radically changed. The Cambridge Analytica story ripped the lid off mass psychometric manipulation of populations through social media. Extend that capability to surveillance, deep fakes, deception programs and all magnified and amplified by AI - and you get the gist of cognitive warfare.

Unmentioned by the panel, it is by far the greatest threat to the democracies because it uses our freedoms (of speech and protest) and our tech (social media platforms) against us - both of which are not available in China. Cognitive warfare opens the door to convincing another society (or elements therein) to adopt your narrative. This is how Putin turned the Republican Party into a pro-Putin anti-democracy party. Given enough time for cognitive warfare to take hold, it becomes possible to win without fighting.

Kristin Wood noted how we have no defenses against this threat. All the advantages lie with the autocracies. Add in the power of AI to create millions of bots, all with individual faces, voices, videos, backstories, email addresses, and all the things one checks for in social media for user legitimacy. This drives the costs of cognitive warfare through the floor and detection and reaction to fakes next to impossible to action.

Conclusion

General-Geek speed dating was a fantastic idea and was so well received that the organizers have already announced another one for next year. While the speed and style, the way, of Silicon Valley has become the obvious choice for getting things done, it does not mean it will become dominant overnight. As Dr Schmidt noted, it will likely take a war, and MIL would add, losing a war, as a forcing function to end the dominance of Primeworld and for the USG to fully embrace the way of the valley. However, this week the Pentagon was put on notice. Actions speak louder than words and AI and the valley are delivering serious outcomes that cannot be ignored.

1. Day 2 of the AI in National Security conference was a vast improvement on day 1. Please note there are so many panels and demos going on simultaneously there is no way to get the full scope of the program from one participant.

2. MIL can write this because he too is an immigrant and a well fed old white guy.

3. And so on. These numbers all go up as higher and higher quality images are used by more and more users.

4. ™ Military Innovation Lab.