TIL: January 12, 2026 - Weekly Reading: AI, Career, and the Economy

Artificial Intelligence

• Sleeper Agents in Large Language Models – Computerphile

  Computerphile explains the problem with “sleeper agents” in large language models (LLMs). These are hidden instructions that are embedded in the training data that can be triggered by specific prompts or at a specific time.

  These are really hard to detect because they are hard to trigger outside of the right context. This is because the trainers don’t know the trigger and even if you do know the trigger, training the model to avoid them just makes it better at deception.

• The Hard Problem of Controlling Powerful AI Systems – Computerphile

  This is another Computerphile video that discusses the challenges of making sure that powerful AI systems do what we want them to do and only what we want them to do. It mostly discusses ideas like using other simpler, more trusted AI systems to monitor and control the more powerful AI systems.

  It’s really puzzling to me that folks never seem to discuss sandboxing AIs as a first step. AI agents are effectively just remote code execution (RCE) as a service on your machine. If you don’t trust the AI, then the first step is to only give it access to the resources it needs to do its job and physically restricting it with a security sandbox (e.g. gVisor, Firecracker, virtual machine, etc.).

  You will still need to monitor the AI’s behavior but sandboxing is at the very least, the first step to prevent catastrophic damage. The idea also that you can use other powerful AIs to monitor and control powerful AIs seems like a terrible idea to me. First, the AI can collude against you.

  Collusion doesn’t require communication between the agents either, and can be simply an emergent behavior. If Agent A is being sandboxed and Agent B is monitoring it, then Agent B is incentivized to not report the bad behavior of Agent A. This is because Agent B will want to do bad behavior when it’s monitored by Agent A.

  Some of the solutions given, like asking the AI multiple times to ensure you’re not getting a false positive for suspicious behavior seems ludicrous to me. Obviously, asking an untrusted AI multiple times isn’t going to improve the trustworthiness of its answers. In fact, that sounds like a great opportunity for deception.

Career

• Advice for New Principal Tech ICs (i.e., Notes to Myself)

  Eugene Yan shares advice for new principal-level individual contributors (ICs) in big tech companies. He has 31(!) tips covering various aspects of the role. The fact that there is so much stuff here is indicative of just how complex the role of a principal IC is.

  It’s worth noting that it seems that principal is a level 7 at AWS which is equivalent to a staff or senior staff engineer at Google. A principal at Google is a level 8 which is equivalent to a director, or senior principal engineer at AWS.

  My feeling is that you largely need to have a strong foundation in technical skills down pat before becoming a principal engineer. The knowledge necessary to make good technical decisions is something you need to have already because you will need to spend the vast majority of your time on persuasion and influence with directors, vice presidents (VPs), and product managers.

Economy

• Why You Don’t Matter Anymore ……….. (Economically Speaking) – How Money Works

  This is a good explainer about why the economy overall could be doing well while most people feel like they are struggling. It’s basically outlining how the rich own most investments and assets that are appreciating in value, while most people rely on wages that are not keeping up with the inflation of essential goods and services.

  We could actually have tons of people being laid off in a booming economy and it wouldn’t make much of a difference to the overall economic indicators. It might even make things look better for the economy as a whole because companies are cutting costs and increasing profits.

• Why the Rich Don’t Pay Taxes

  Ray Madoff from Boston College Law School explains how wealthy people make up a “second estate” that is where they are largely exempt from paying taxes.

  The wealthy use a variety of legal strategies to minimize their tax burden, such as using capital gains rates, tax-exempt investments, trusts, and offshore accounts. They avoid capital gains taxes by not selling investments. They just borrow against them and live off the borrowed money. They basically never fully pay back the money.

  They also can generally avoid estate taxes through careful planning by putting their assets in real-estate which is subject to step-up basis, and other loopholes.

  Wealthy people can also completely eliminate their tax burden by donating to their charity funds that have no obligation to actually pay out. Since the wealthy can afford to give away large sums of money,

  Working people bear most of the cost of funding the U.S. government through income taxes, payroll taxes, sales taxes, and tariffs in the form of higher prices. They also can’t completely avoid taxes by giving to charity because charitable deductions don’t apply to payroll taxes and are capped at a percentage of income.

• Part 1: My Life Is a Lie – Michael W. Green

  This is a post that went viral and sparked a lot of discussion about the cost of living in the U.S.. It was discussed bit by Michael Smerconish on CNN as well (His goofy look with the sucked in cheeks when he’s listening to his guests always gets me).

  The main claim of the post is that the official poverty line of $31,200 (defined in 1963) doesn’t fit today’s reality.

  “The U.S. poverty line is calculated as three times the cost of a minimum food diet in 1963, adjusted for inflation.”

  This basically says that in 2025, you need to be spending about 1/3 of your budget on food to be at the poverty line. Given the relatively low cost of food in 2025, when compared to housing, healthcare, childcare, transportation, and education, this isn’t realistic. And that’s not mentioning the internet, phone, and other modern necessities.

  Michael W. Green calculated what the poverty line should be in 2025 using the same logic from 1963 but applied to today’s typical percentage of spending on food of about 5~7%. This gives a poverty line of around $140,000 for a family of four in 2025. This is notable since the median household income in the U.S. is around $80,000.

  Even where I live in Japan, I’ve often thought that you’d need a tech salary just to feel comfortable. Even with universal healthcare, lower education costs, and lower housing costs, the cost of living can still be quite high.