The Ghost Of Neutron Jack To The Rescue?

Today's guest post, by our friend David Janello, PhD, CFA, addresses that question.

Dr. Janello is the founder of SpreadHunter, and the author of The Nuclear Option: Trading To Win With Options Momentum Strategies.

Before we get to Dr. Janello's post, two quick programming notes. 

1. We've got two options trades teed up for later today, a bullish bet on a biotech (a previous winner for us), and a bearish bet against a retail name. If you want a heads up when we place them, you can subscribe to our trading Substack/occasional email list below. As usual we're enabling free subscribers to unlock this post. 
   
2. If you want to hedge market risk here, you can use the Portfolio Armor website or iPhone app to scan for the optimal ones given your risk tolerance and time frame. 

Now on to Dr. Janello's guest post. 

Authored by David Janello, PhD, CFA on Nuclear's Substack

Jack Welch Takes On AI's Biggest Challenge

There’s a very technical AI word for it called distillation, but you and I would call it theft…

This administration supports open source models, but what we do not support is IP theft. And if we see, especially that overseas models are stealing from our great companies, we have the ability to sanction them because of this theft.

–Scott Bessent, U.S. Treasury Secretary

Scott Bessent, US Treasury Secretary

Taking a cue from previous wins in banning the Best Selling Electric Vehicle In the World,, the Fastest Electric Vehicle In the World, and more importantly, numerous Gasoline Powered Motorcycles from the domestic US market, the US Government is now discussing imposing economic sanctions on Award Winning Open Source Chinese AI Models, because of presumed distillation attacks against major Artificial Intelligence firms.

What is a Distillation Attack?

Here is Anthropic, PBC (Public Benefit Corporation) outline of the distillation process:

These labs used a technique called “distillation,” which involves training a less capable model on the outputs of a stronger one. Distillation is a widely used and legitimate training method. For example, frontier AI labs routinely distill their own models to create smaller, cheaper versions for their customers. But distillation can also be used for illicit purposes: competitors can use it to acquire powerful capabilities from other labs in a fraction of the time, and at a fraction of the cost, that it would take to develop them independently.

As Anthropic — with great insight — points out, distillation is a unique crime. Unlike embezzlement, burglary, armed robbery and other, better known variations of theft, AI Distillation can in certain contexts even be “widely used and legitimate.” Distillation also stands apart from other forms of theft because the offending party is actually paying, at least in part, for the item that they are stealing.

What is the solution to the problem?

Here is what Anthropic recommends:

Anthropic has consistently supported export controls to help maintain America’s lead in AI.

Export Controls occupy an intriguing role in the history of technology. Younger readers might be surprised to discover that the ubiquitous Secure Sockets Layer – that is, every website beginning with the prefix ‘https’ – was once subject to extremely strict Export Controls. In fact, until 1996 Cryptographic Software was categorized as a ‘Munition’ under the International Traffic in Arms Regulations (ITAR) before being moved to the Department of Commerce. This fact was mentioned (almost always in hushed tones) whenever someone installed a new Cryptographic package.

Did you know that the encryption package we are about to install is categorized … as a Munition?

To keep vulnerable consumers from using dangerous, weapons grade encryption to protect their personal emails and web pages, in the late 1990s the US Government mandated weak 40-bit encryption on all non-domestic Web Browsers, and slightly strong encryption on domestic Web Browsers. Which worked in the short term to “protect” users. However, when the exact same Government abandoned the Export Restrictions a few years later, the now obsolete weak encryption keys were left sitting in legacy code bases. The dead code provided a perfect entry point for hackers to exploit. FREAK and Logjam were the most famous attacks using this Government Mandated Security Hole.

Reading Anthropic’s document, we can see clearly that they are in trouble with regards to Distillation Attacks and need help. Will Export Controls do the trick this time? Or perhaps the sanctions suggested by our Treasury Secretary?

One important question to ask is: what will the Government do about American firms that return the favor and (re-) distill the Chinese Models? Will they be prosecuted too? Or face sanctions?

The economics are certainly in favor of anyone distilling or hosting the open weight models.

Let’s imagine that you want to pay full freight and lease a 16-node H200/B200-class Amazon EC2 cluster to run an Open Source Chinese Model like Kimi3. This costs roughly $0.20 to $1.20 per million tokens at high utilization (greater than 75% capacity).

If you use the kimi.com API to access the results, the tokens cost a flat $0.60 to $3.00 per million input tokens and $2.50 to $15.00 per million output tokens.

So there is room for profit even compared to the ‘cheap’ Chinese Models.

Note well that the Amazon cluster costs thousands of dollars per hour to run, so be sure to have billions of tokens of solid customer demand lined up before signing up for your cluster. If demand slacks off, you can easily lose thousands of dollar per hour.

In contrast to the Amazon and Chinese API costs, here is the rate for Anthropic’s comparable current top-end flagship model: $5.00 per million input tokens and $25.00 per million output tokens

This pricing gap leaves a tractor trailer sized opportunity for sophisticated, well capitalized investors to exploit. The economics are even better for State Sponsored Actors and Sovereign Wealth Funds with large energy resources to help cover the expensive data center power requirement. To cite one example, the Iranian Republican Guards have used (and continue to use) highly subsidized energy prices to profitably mine Bitcoin and to at least partially avoid international banking sanctions at the same time. Artificial Intelligence Hosting offers potentially higher margins than crypto mining, as evidenced by the numerous firms pivoting their data centers from crypto mining to AI.

Since Export Controls probably won’t work to prevent Distillation, what should Anthropic and OpenAI do instead?

The obvious answer is to use a highly touted, ‘Too Dangerous To Release to the Public’ AI security tool like Claude Mythos to detect the Chinese Distillation Attacks in real time before they do too much damage. Or use the same boring off-the-shelf, non-AI tools for online fraud detection that everyone else uses.

Unfortunately, since it appears that the Frontier Security Models aren’t up to the task, it looks like Anthropic and OpenAI may be forced to go back to the 1990s for an admittedly painful solution:

Business Process Reengineering

If we examine Anthropic’s publicly disclosed Federal H1-B Foreign Worker Visa Applications, we see that Anthropic pays a median base salary of about $226,000 to H1-B applicants, with top technical roles reaching up to $1.38 million.

Jack Welch completed his Ph.D. in chemical engineering at the University of Illinois Urbana-Champaign in 1960, with a doctoral dissertation titled “Microscopic Study of Dropwise Condensation.” The Condensation referred to in the Dissertation relates to condensation in steam turbines, not condensation in chemical distillation.

Keep in mind that the H1-Bs are typically at the lower end of the pay scale. These workers make excellent candidates for Jack Welch’s famous 20-70-10 ‘Vitality Curve’ to rank employees:

• The Top 20% is Rewarded

• The Middle 70% is Nurtured and Developed

• The Bottom 10% is Ruthlessly Eliminated

With these kinds of pay scales, a little bit of Jack Welch will go a long way in bringing Anthropic and OpenAI’s cost structure in line with the Chinese Competition.

Welch was named Manager of the Century by Fortune Magazine in 1999. After his retirement in 2001, GE’s stock price lost about 80% of its value over the next 20 years, calling into question many of his leadership ideas.

If the Frontier AI Firms don’t address this problem now, they will soon be in very deep trouble once they start facing the brilliant and frugal Vietnamese Entrepreneurs who are waiting in the wings and already gutting the Chinese in tech manufacturing.

Here is how my all time favorite Vietnamese Technology Vendor explains it:

Labor here is dirt cheap compared to the West (even compared to China). A skilled CNC operator makes under $30/day. A family of four lives fine on $20/day. So yeah, I’ve got a highly skilled workforce at a fraction of the cost. And no, it’s not “new slavery,” it’s just local economics, thanks for asking.

This vendor also eliminates risk by pre-selling everything. No inventory at all. Then uses this Lower Risk to generate the all-important Razor Thin Margins:

My segment is crowded with big names, and if I want to make some noise, I’ve got to keep the price stupidly attractive. Once we’ve proven ourselves, expect the price to creep up—still competitive, just not “too good to be true” cheap like now.

Although the classic Jack Welch approach and the updated and much improved Vietnamese Corporate Strategies will (eventually) bail out our AI Friends, it is probably too late help the Government Regulators with their distilling problem … the one that just seems to come back no matter what..

1920s Distillation Gear

Disclaimer : All Content on the Nuclear Option Substack is for Education and Information Purposes only. It is not a solicitation or recommendation to buy or sell any security. Before trading, consult with your Professional Financial Advisor and read the booklet Characteristics and Risks of Standardized Options Contracts, available from the Options Clearing Corporation.