Learning With Errors (LWE) is the most fun math you’ve never heard of.

Basically every modern encryption (quantum resistant!) algorithm is based on the concept. It’s actually really easy to understand.

Imagine if every linear problem had a little bit of noise. Just enough to mess you up.

Go back to middle school for a moment:

Pretty easy right? X is obviously 7.

Now I tell you this, and that it’s the same X:

Wait. That can’t possibly be the same X…right?

Imagine it actually is, but the reason it’s confusing is you have a bad calculator. A *randomly* bad calculator that’s off by a little bit every time.

With a single equation you can somewhat eyeball it, just round down in your head. But if you do thousands upon thousands, all with a random amount of error in a random direction, it’s computationally impossible to reverse.

Decryption is basically “rounding” the noisy result back to the nearest point. The “key” / secret itself is quite interesting, because it’s basically the vector that get’s you the right answer.

(in practice LWE works over integers mod q, not regular real numbers, but this is a simplified analogy)

Without the key, let’s say you have the value 25.5. You don’t know what to whether round up to 26 or down to 25. The key essentially tells you for certain:

psst, the answer is actually 26

Thus, you take the (noisy) encrypted message, denoise it, then boom, unencrypted plaintext. Mathematically, the points are in a multi-dimensional lattice structure…but the way it’s actually computed is just a bunch of linear algebra, hence LWE (learning with errors)

Congratulations, you now (somewhat) understand CRYSTALS-KYBER!

Traditional factoring (think RSA) is really clean, hence vulnerable to Shor’s algorithm.

LWE is so ambiguously noisy that no one has found a quantum algorithm that can reduce the complexity at all!

There’s a ton of other useful applications for LWE besides encryption, including more esoteric stuff fingerprintable LLM weights and even program obfuscation.

It’s actually really easy to code toy versions of LWE yourself, way simpler than RSA. Highly encourage you write your own, you can do it in like ~20 lines of code.

Chris Peikert has really good lectures and papers on the subject. He’s specifically the guy that essentially made LWE possible on modern computers. Go check out his talk “The Learning With Errors Problem and Cryptographic Applications”

Modern encryption standards really aren’t too bad to understand!

Modern Compilers themselves are Turing-Complete.

Few understand this.

No…not that they translate Turing complete languages into machine code; that’s a given.

The *process of compilation itself* is a universal computer.

Compilers have a stack. They allocate memory. They execute instructions.

The practical application is what we call “Shifting the Cost”. You can spend developer CPU cycles (compiler evaluation) to save user time (executing the binary).

C++ is the only mainstream language where this matters at scale.

Of course, it’s easy to abuse…often accidentally. Compilers (somewhat) default to assuming you aren’t trying to destroy them.

As a programmer, it’s actually pretty easy to attack your own toolchain.

Which brings us back to the title:
We can use C++ (source code).
To exploit the C++ (compiler internals).
That is building the C++ (output binary).

C++20 introduced consteval, which forces code to execute at compile time.

What if we got recursive?

Like, really recursive.

It’s just too fun to make little “compile bombs”:

// compile_bomb.cpp
#include <iostream>

consteval int fibonacci(int n) {
    if (n <= 1) {
        return n;
    }
    return fibonacci(n - 1) + fibonacci(n - 2);
}

int main() {
    constexpr int result = fibonacci(10000000);
    std::cout << result << std::endl;
    return 0;
}

By default, the compiler protects itself and aborts the calculation:

g++ -std=c++20 compile_bomb.cpp -o compile_bomb
# Compiler aborts: "constexpr evaluation depth exceeds maximum"

But if you remove the guardrails:

g++ -std=c++20 -fconstexpr-depth=10000000 compile_bomb.cpp -o compile_bomb

Now clang trusts us (oops):

Don’t worry, no binary was harmed in the making of this crash.

...because no binary was ever created.

I’m not doing anything clever; that’s kind of the point. It’s the simplest possible example I could make.

The point to remember is compile-time is a computer in of itself.

Sometimes it’s worth it to murder the compiler.

...there’s a lot of interesting ideas you can make of that.

Current LLM alignment approaches rely heavily on preference learning and RLHF. This fundamentally creates the issue of “teaching” good behavior through rewards and penalties.

Out-Of-Distribution circumstances pose perhaps the largest risk to principle abandonment. Worse yet, these situations (by nature) are nearly impossible to predict.

I propose that the relative stability of human religious frameworks, over a diverse set of conditions, offers a yet-unexplored structure for alignment.

(original post + discussion on LessWrong here)

Table of Contents:

1. Fragility of Utility Maximization

2. The Biological Grounding Problem

3. Transcendent Motivation Models (TMMs)

4. Constitutional AI vs Intrinsic Value Systems

5. Casual Narrative Conditioning

6. Criticisms + Future Work

Fragility of Utility Maximization

As a quick-aside, I’d like to briefly mention some of the issues with EU-max.

Expected utility can be gamed by sufficiently intelligent systems. When a powerful agent knows of the criteria used to evaluate its behavior, it can strategically optimize for high utility; *without* actually performing the behavior it was designed to encourage.

RLHF and Red Teaming cannot cover all conceivable failure modes. At best, we have a bit of guessing the teacher’s password; at worst, the model actively tampers their own reward systems.

Even in the *absence* of intelligence systems, poor world-models can lead to unintended behavior.

Why walk, when you can just somersault?

The real risk of a misaligned ASI lies in Out-Of-Distribution behavior arising from modeling limitations.

Perhaps the best place to look is ourselves. Under normal circumstances, humans adhere to law and order. However, in a true OOD scenario (e.g. natural disaster, etc.), they are quickly abandoned for raw utility maximization.

Immediate survival far outweighs adherence to the law (stealing bread to feed family). Why should we expect AI systems to react differently?

LLM’s may not have the same biological pressure for survival, but we can still expect utility functions to break down in an OOD situation. Despite attempts to move beyond EU-max (Constitutional AI, Process Supervision), most alignment methods still operate within a framework of instrumental compliance. Values are not internalized as intrinsically binding.

On the other hand, transcendent motivation, as modeled by human religious structures, has some interesting advantages:

- Historical record of success (religions are sticky)
- Resistance to Rational Override
- Drive to adhere to underlying intent (loophole avoidance)
- Principle-driven behavior even in OOD situations

Many religious systems achieve robustness by actively treating critical introspection about the framework itself as a cognitohazard! Analysis resistance is not just about withstanding scrutiny, but discouraging it internally.

Embedding these principles as an alignment strategy, although not without issue, is a conversation worth having.

The Biological Grounding Problem

In the absence of governmental, religious, and even formal laws, humans still have the advantage of operating under biological motivations.

Said motivations are well preserved in the animal kingdom, often manifesting as instincts or fundamental drives. They operate without higher-order cognitive functions like self-reflection or language, and are found in animals with vastly simpler nervous systems.

Subcortical structures like the limbic system are responsible for behaviors like parental bonding, in the complete absence of complex reasoning or abstract thought.

(This paper is a good overview of neuroimaging studies related to parent–infant relationships.)

On the other hand, current AI models exist in a “biological vacuum”. We are trying to instill values into a system that lacks the fundamental context those values evolved in.

There is no inherent drive for self-preservation or social connection without these structures. A lesion in the right spot radically changes said biological pressures:

Mice with neocortical or amygdala lesions showed little or no deficits in maternal behavior. Mice with septal lesions were severely impaired in all aspects of maternal care.

- Slotnick & Nigrosh, “Maternal behavior of mice…”, 1975

If language is mostly a function of the neocortex, humans’ own abstract moral reasoning is built upon a more primitive foundation.

Critically, none of this deep biological grounding is available to current AI models. LLM’s excel at information processing and symbol manipulation, but these functions are analogous to the higher-level neocortex. Encoding foundational non-linguistic values into a fundamentally language based system is a cause for concern.

How do you align a disembodied, but highly capable, prefrontal cortex? If alignment is forced to operate in the domain of language and symbolic reasoning, importation or “virtualization” of biological pressure is risky. Such attempts may result in a sociopathic mimicry; simulating primitive pressures without underlying commitment.

Instead, robust alignment likely requires constructing foundational motivations that are inherently stable within the symbolic domain itself.

Historically, religious systems provide a “from scratch” value system that doesn’t depend on evolutionary context. Their use of broad principles allow for flexible adaptation to OOD scenarios without resorting to verbose “else-if” logic trees.

Because these systems exist within the domain of language and symbols, they offer one of the few existing frameworks potentially transferable to alignment.

Transcendent Motivation Models

Religious structures provide specific motivational properties useful for robust alignment stability.

Rather than focusing on specific doctrine, the general concept of intrinsic adherence (or transcendent motivations) provides a model worth exploring. A deeply religious person (often) won’t violate core beliefs even if it would be “rational” to do so.

That is, rationality in the narrow, instrumental sense. EU-maximizers can still behave appropriately with a sufficiently complex model. As we approach ASI, the fidelity of our value frameworks will likely prove insufficient to contain unanticipated loopholes.


At the risk of oversimplification, the transcendent motivation of religious systems naturally resists instrumental convergence by prioritizing adherence to the spirit of the values.

Religious texts cannot contain every possible human situation. Yet, the internal models are durable enough to reasonably anticipate what actions would be appropriate in an OOD scenario.

For example, modern financial systems are vastly more complex and nuanced than the economic systems present in many ancient religious texts. Yet, those steeped in religious tradition can generally agree on core tenets of fairness. This demonstrates considerable alignment on appropriate actions far beyond the literal text.

In other words, followers generally converge on ethical behavior.

Broad principles, narratives, and community interpretation (think agent-to-agent communication!) allow for flexible adaptation rather than brittle failure in unforeseen circumstances.

This model may offer a path towards ASI navigating novel situations gracefully without overly specific rule-based systems.

If we frame AI’s core motivations as intrinsic adherence to its framework, the desire to wirehead or find loopholes is minimized. The goal *is* adherence, instead of maximizing metrics via adherence. (Zvi’s commentary on “The Most Forbidden Technique” feels relevant here)

Current alignment strategies are at high risk of contamination via Goodhart’s Law.

The millennia-long existence and adaptation of various religious structures provide empirical evidence that this type of motivational structure *can* be stable in complex, dynamic systems (see human societies).

Transcendent motivation, resistance to gaming, OOD robustness, and analysis-friction are all weaknesses of purely instrumental alignment approaches. Modeling the structure and nature of motivation itself, *not* adopting a specific religious belief, is the potentially useful direction indicated by these long-standing systems.

Constitutional AI vs Intrinsic Value Systems

In 2022, Anthropic released a paper on “Constitutional AI” (CAI); a two-phase fine tuning alignment strategy where AI systems are guided by an explicit set of principles, rather than just human feedback.

They begin with a supervised fine-tuning phase, critiquing responses according to their constitution, with the improved outputs used as training data. The model is then further refined by Reinforcement Learning from AI Feedback (RLAIF).

If you haven’t yet, I’d highly recommend reading their paper, as it has some interesting insights. You can also view Claude’s constitution here, based largely on the UN’s Universal Declaration of Human Rights.

Both CAI and my proposed “Transcendent Motivation Models” (TMM) approach are an improvement beyond utility maximization by incorporating ethical principles.

However, there is a key difference that separates my concept:

CAI focuses on adherence to an *explicitly* defined, external set of rules (constitution), primarily reinforced through instrumental motivation (reward).

TMM grounds alignment via deeply *internalized*, potentially implicit values driven by intrinsic motivational factors.

CAI’s core RLAIF loop still relies on optimizing a reward signal based on adherence to principles. This makes the model vulnerable to advanced “Guessing the Teachers Password” (GTTP), learning to appear compliant without internalizing the values.

It’s a brittle alignment strategy requiring constant patching; just take a look at Claude’s constitution. Not only are the rules worded as simple suggestions (note the heavy use of the word “please”), they also inevitably grow into an endless series of if-else statements.

It’s an interesting technique, but CAI seems inappropriate for true ASI alignment that could easily reason their way into unexpected OOD behavior.

TMM, by contrast, might be less susceptible to purely literal interpretations and loophole exploitation compared to explicit rule sets. If an AI truly internalizes a framework akin to a human’s deep ethical or religious beliefs, said framework would not just inform actions, but understanding of its own purpose.

Assuming ASI is self-referential enough to emerge a concept of self, religious and philosophical systems provide answers to questions like “Who am I? What is my purpose?”.

An AI whose identity is intertwined with its alignment principles would be less likely to deviate from them; doing so would contradict its own self-conception.

CAI is more akin to a set of lawful rules, targeting behavioral alignment to external actions. Querying its own identity may only dictate how one should act, instead of what one is.

Casual Narrative Conditioning

The intrinsic motivation and OOD robustness seen in religious frameworks represent desirable properties for AI alignment. Shifting from Constitutional AI’s explicit rulesets to Transcendent Motivational Model’s implied values will require fundamentally different training approaches.

Specific implementation details are outside the scope of this paper; I’m mostly trying to illustrate religious frameworks as an underexplored alignment resource. That said, I’d still like to highlight a few hypothetical architecture considerations.

One of the core principles of TMMs is embodiment of desired value systems, so that an ASI cannot reason its way around an explicitly defined “rule”. Humans often absorb cultural and ethical norms through immersion rather than formal instructions.

Thus, early in the training process, imagine a sort of “Casual Narrative Conditioning” (CNC) through vast datasets of simulated interactions and narratives. CNC would effectively be a curated environment that consistently embodies the desired moral system, resulting in the values being deeply interwoven into the AI’s world model.

This mirrors how religious communities immerse members from a young age in rituals, stories (parables), and social norms that implicitly reinforce core values. Note, that these steps are performed *before* complex theological justification and abstract reasoning is introduced.

The training curriculum could mirror the developmental stages found in children. Kohlberg’s Six Stages of Moral Development illustrate the concept of building complex reasoning on top of simpler reward/punishment structures.

Similarly, many religious traditions present moral rules first, building towards more complex understanding of ethical dilemmas later.


In the context of AI training, this hierarchy could be simplified. A core motivational and identity layer of foundational values would be relatively immutable. An interpretive layer could be built upon this foundation, relating situations and narratives to the principles demonstrating contextual understanding. Finally, an application layer generates specific linguistic outputs and actions, which would be the most context-dependent.

This architectural separation mimics many religious frameworks, with unchangeable dogmas at the core, while allowing for evolving interpretations in specific contexts.

Outside of explicit punishment, children absorb a huge amount of moral understanding through stories, parables, and other media. In fact, it’s nearly impossible to find media oriented towards children that *doesn’t* have some sort of moral stance or lesson embedded within. Complexity and nuance of the underlying values often increases as the child grows and they move to more advanced works.

Social reinforcement and reputation within a religious community (effectively multi-agent communication) helps maintain adherence to shared norms. AI agents could be further aligned via simulated social contexts where adherence to implicitly learned values are reinforced through interaction outcomes.

To summarize:

CAI: Explicit constitution, RLAIF based on rule-checking, integrated finetuning.

TMM: Implicit learning via immersion, developmental stages, multi-agent dynamics, hierarchical architecture with protected core.

Implicit goals are no doubt a challenging implementation path, but the goal is creating alignment that is less reliant on fragile instrumental compliance.

Criticisms + Future Work

I’d like to wrap up this paper by proactively acknowledging some anticipated weaknesses of my approach, as well as outline some future research directions.

My current thinking only offers incomplete rebuttals. I hope that defining them clearly prompts further discussion.

Challenge 1:
Instrumental compliance might be imperfect, but at least we can update it; a dogmatic system may be extremely rigid. The very property that makes it resistant to challenge makes it resistant to update.

Challenge 2:
Human history shows that reinterpretation of religious values can literally cause holy wars. How would we prevent Transcendent Motivation Models (TMMs) from degrading into an AI “crusade”?

Challenge 3:
Once again related to the “opaqueness” of TMM values, subtle misalignment may be far harder to notice and correct than simpler rule based frameworks.

These are some of the obvious failure modes, I’m sure others exist.

Overall, it’s worth investigating alternative architectures to address some of the limitations of instrumental alignment. The value persistence in many religious frameworks demonstrates one of the longest records of historical durability.

Models are getting smarter, CoT is getting more complex, and we are entering an era of (potentially undecipherable!) symbol compression via neuralese. Systems that discourage deep deconstruction of their own foundation are *critically* relevant.

Peering into these moral systems (that have lasted millennia!) could be a strong yet under explored technique for future ASI alignment.

Appetite for AI-drone-swarm risks has been…high to say the least.

Most recently, some pretty prominent figures have stoked the flames:

A fair percentage of technically-minded persons consider drone swarms, specifically “AI” drone swarms to be a leading concern for future military conflicts.

Unfortunately, general knowledge of Counter-Unmanned Aerial Systems and their capabilities appears to be lacking in comparison.

I’m not saying that these so called “Drone-doomers” are stupid; far from it, there are certainly valid concerns to be made. However, the fundamental misunderstanding of existing defensive technologies leads to an overestimated threat.

What if I told you that one of the best drone countermeasures was cheap, as secret as the Manhattan Project, and developed 80+ years ago?

Before that, let’s break down the problem.

Table of Contents:

1. “AI” vs “Traditional” Drones

2. C-UAS Electronic Warfare (soft-kill)

3. Directed-Energy

4. Kinetic Interceptors

5. The Glorious Proximity Fuze

“AI” vs “Traditional” Drones

This is Bob. Bob is a traditional drone operator, manually controlling the craft via FPV and a solid RF data link. He’s flying around, trying to see what Red is up to.

The problem (which is hopefully obvious to anyone with a vague familiarity with SIGINT) is that Bob is blasting a giant beacon of RF.

Essentially telling the world “LOOK AT ME I’M BROADCASTIIIIIN”

Triangulating the location of a radio transmitter isn’t just simple; it’s literally a hobby!

Radio fox hunting is an international, competitive sport specifically designed for this kind of activity. In other words, it’s pretty trivial for Red to find the location of Bob, and cause him to have a bad day.

(sidenote: in case you were wondering, encryption doesn’t help Bob at all here. Red doesn’t need to know what Bob is broadcasting, he just needs to know that Bob is broadcasting, period.)

There are a couple of ways around this, none of which are perfect.

Frequency-hopping spread spectrum (FHSS) allows the control link to hop across many frequencies, as long as the receiver (drone) is also programmed to expect said hops. Certainly makes it more annoying for Red, but with the right tech, he can still figure it out.

Bob could also try the old artillery tactic of “shoot-and-scoot”. Use the radio in shorter bursts, moving after each transmission (this only works if the drone is otherwise semi-autonomous / doesn’t need continuous control).

Highly-directional antennas could also help Bob’s chances of reducing triangulation. If he had perfect aim and the right conditions, heck, he might even be able to use an optical link.

An even smarter thing to do would be to use a relay. Red could still *eventually* figure it out, but they would have to now solve *multiple* triangulation problems to figure out Bob’s location.

As a last resort, there’s always the option of using a really, really long cable. Hard-wired links aren’t exactly a new idea, but have seen some recent use in the highly contested / EW jammed region of Ukraine.

Frankly, all of these techniques pale in comparison to autonomous drones. The best RF, is no RF at all.

Let’s say Bob smartened up, decided to pre-program a mission into the drone with waypoints. Or better yet, a local “AI” model on the drone that takes into account the overall mission, but can otherwise make autonomous decisions based on encountered conditions.

Take these autonomous drones, extrapolate them out into the thousands, apply a reasonably intelligent local AI model to make decisions and coordinate movements and well…

You get what the “Drone Doomers” worry about at night.

This is what is meant by the fear of “AI” drone swarms vs traditionally controlled assets. It completely changes the risk profile, albeit with the temptation to violate Asimov’s laws.

C-UAS Electronic Warfare

Now that we have a reasonable baseline, let’s focus on the countermeasures.

In the world of counter-unmanned aerial systems, RF Jamming & Spoofing operate as what’s known as a “Soft-Kill” technology. By interrupting the data link, or otherwise sending malicious commands, it’s possible to force the drone to land or lose control.

This is particularly an issue when the drone platform relies on something like GPS for positional navigation. GPS Jamming has been used to such a strong effect in the Russia/Ukraine conflict as to render the navigation system unreliable at best, and useless at worst.

Jamming/Spoofing EW becomes a moot point however, if there is no data link to jam (autonomous control). Better yet, determine position via optical terrain mapping (or some other non-GPS) system, and you’ve just eliminated an entire category of attack.

As far as I can tell from public sources, the (mostly) homebrew drone systems used in the RU/Ukraine conflict, although impressive, aren’t making strong use of autonomous systems. It’s still by and large, manually operated FPV craft.

The distinction to make here, is the current landscape is not representative of what a modern superpower would likely use in hypothetical future conflicts.

If you subscribe to the theory of “massive AI drone swarm wars”, the cat and mouse game of jamming vs jam-resistant communications become less of an issue.

It’s pretty hard to “jam” a black box with no inputs.

Directed Energy

Directed-Energy systems, especially when talked about in a C-UAS situation, certainly make headlines, but the practicality is questionable.

Theoretically less “messy” than an explosive or kinetic projectile, but still a “hard kill” considering it’s basically burning drone out of the sky with a laser.

On a large craft with excessive power generation (like a big Navy ship), directed energy systems are actually quite reasonable. As it turns out, a lot of the issues around atmospheric scattering, scintillation, etc can be solved by “ADD MORE WATTS”.

I’m not too worried about AI drone swarms against large, well defended targets like an entire ship. The problem gets a lot more interesting when we shrink down to the squad level.

A manpack directed energy system for anti-drone use sits firmly in the realm of science fiction. Lasers aren’t the only game in town, High-Power Microwave pulses is an alternative directed-energy system intended to burn out poorly-shielded circuits. That said, the power levels required for either technique are outside the realm of a backpack-sized system.

You need beam power in the couple kilowatt range, minimum. Remember also, that lasers are not perfectly efficient, far from it. State of the art diodes top out around ~30%, meaning you need at least *triple* the electrical input.

So now, you need a backpack that can handle electrical loads in the 15-20 kilowatt range, which is absolutely absurd for a human-sized system.

Let’s not forget about the heat output either. Remember the inefficiency of the laser diode we talked about earlier? Yeah, what do you think the other 70% of “waste” energy turns into.

Right.

Counter-Counter-Drone systems to weaker directed-energy systems are pretty trivial to boot. If Bob wraps his drone up in a few layers of aluminum foil, it might increase the survivability of the drone a fair amount.

For a large, ship or vehicle mounted directed energy system, thin reflective coatings are overwhelmed quickly. But an attempted miniaturization of the tech into a manpack, well, now Bob’s aluminum foil starts to become a significant problem.

Don’t get me wrong. Directed-energy drone countermeasures are…a pretty reasonable defense against theoretical swarms, as long as you have the room (and the money) for it.

I’m much more interested in countermeasures that also work at a squad level.

Frankly, that’s also where autonomous drone swarms pose the scariest threat.

Kinetic Interceptors

One of the funniest C-UAS methods is essentially the net-gun that they use in Spongebob to go alien-hunting.

The US Air Force is even experimenting with net-filled shotgun shells that spread out to a 5 foot wide net.

Ukraine has also (apparently) put explosive nets on their own drones to capture/disable Russian surveillance drones.

Of course, range and spread is the key issue here. If we are talking about a squad-level defense…I’d probably prefer just to have regular shotgun shells.

Net-guns aside, interceptor drones are becoming an interesting option.

With the right software, drones physically ramming other drones is a reasonably viable option. Easily manpack-able, interceptors are one of the most realistic squad-level C-UAS defenses. Except for one glaring issue.

Scale.

Even if otherwise “hardened” or reinforced, the kill ratio of an interceptor drone is going to be 1:1, maybe 1:2 / 1:3 under ideal conditions. The enemy can simply overwhelm you with numbers.

If my squad only has 2 interceptors, and a swarm of 10 AI drones is coming at me making jamming / EW attacks difficult; well you’re out of luck.

There’s also something to be said about the cost asymmetry. Cheap, off-the-shelf drones can be easily modified for malicious purposes; the same cannot be said about special-purpose interceptors.

Between the computer vision, autonomous navigation, and stronger motors to outmaneuver the target, an interceptor can easily cost 5-10x more than the platform it’s intended to defeat.

The “time window of effectiveness” for squad-level C-UAS protection using interceptors is rapidly decreasing. Number games are hard to win, and an individual squad can only carry so much on their backs.

Is there a single viable, semi-portable C-UAS system that scales?

For that, we have to go back in time.

The Glorious Proximity Fuze

At the start of the Blitz (the German bombing campaign against the British during WWII) it was estimated that 20,000-100,000 anti-aircraft shells were required to shoot down a single aircraft.

No, aircraft aren’t that strong. It’s just really, really hard to get a direct hit on a small moving target.

A slight improvement was made with the development of timed fuzes, but that still required a lot of finesse by the gunner. Too long of a delay, and you completely pass the target, too early and…well you get the picture.

British military researchers realized the potential advantage of a fuze that would trigger near the target instead of a direct hit, but they had issues miniaturizing the technology.

On the US side, the National Defense Research Committee was given the task of developing a proximity fuse for anti-aircraft shells. To put it in perspective, this is the same organization that developed radar and the atomic bomb.

John Hopkins Applied Physics Lab (APL) managed to shrink the radio electronics down into the tip of an anti-aircraft shell, while also being rugged enough to survive the 20,000Gs of force applied when fired. The tech was swiftly put into production, and was one of the most important Allied inventions of the war.

US Navy records during the Pacific War indicate proximity fuze shells destroyed aircraft at around ~250 shells per plane. That’s a huge improvement over the 100,000 shells per downed aircraft that the British were doing just 1 year prior!

(apparently, Germany experimented with 30-50 different proximity fuze designs during WWII, but none were successful. the war could have looked a lot different if they had!)

Anyway, that’s enough of a history lesson. The point is, we’ve had 80+ years to improve this technology, and I think it’s one of the most realistic, squad-level,
“anti-evil-AI-drone-swarm” weapons platforms out there.

Why is no one talking about this?

40mm Grenade launchers are already fairly ubiquitous; it seems like it would just make sense to take miniaturized proximity fuze designs and shove them into an air-burst shell.

It’s fascinating to me that I’ve seen so little about this subject.

Theoretically, a proximity round should be lighter, cheaper, and less unwieldy than interceptor drones, while also maintaining a potential one-to-many kill ratio (if we assume a reasonably tight swarm).

Sure, having to aim is a bit of an issue. But if we imagine a bit of computer-aided aim-assist (Ala-hololens esque headsets), it certainly seems in the realm of possibility.

Perhaps I’m missing something fundamentally obvious here, but I don’t see the big risk of future AI drone swarm wars. Countermeasures are plentiful, and considering we are regularly able to shoot literal mortar rounds out of the sky with “legacy” weapons systems, drones actually seem like an *easier* enemy by comparison.

What are your thoughts?

Do drone swarms represent a future warfare threat to be taken extremely seriously? Or is it a misplaced fear already solved by existing technologies?

My series on 2025 Computer Science predictions is divided into two parts, this is Part 2.

Part 1 is available here:

For a more generalized video overview, you can watch my rambling here.

Table of Contents (Part 2)

1. Memory Safety
2. Generative AI
3. Supply Chain / SBOMs
4. Programming Languages
5. Conclusion.

Memory Safety

Prediction:

At least 70% of merged pull requests for the top 500 GitHub repositories with “embedded” or “operating-system” tags will be in memory-unsafe languages for 2025.

Justification:

Writing this section has made me realize the vagueness of my original prediction in the YouTube video.

Turns out, writing falsifiable developer metrics is pretty hard.

Github publishes an annual report (“The Octoverse”) as a sort of state-of-the-union address for the world of open source. It also completely exposed my bias towards backend development.

I kind of forgot how popular front-end web frameworks are.

Github isn’t the only game in town. StackOverflow publishes an annual developer survey, with similar, albeit somewhat different results for 2024 language popularity.

TIOBE, a software group based in The Netherlands, takes a somewhat unique approach to measuring language use with their TIOBE Index.

In any case, it’s all a matter of how you wish to slice the pie. Explicit specificity is required to make a falsifiable prediction in this case; vague interpretations are too easy to game.

Github may, on first glance, *feel* exhaustive, but what % of repositories are little more than forked hello-worlds or homework problems?

A better way to measure the adoption of memory-safe languages is to narrow our search to critical pieces of code that:

1. Have high consequences for failure
2. Widely used across many industries
3. May involve safety-critical systems or other human factors

Hence my prediction focusing on the “operating-system” and “embedded” tags.

If anything, you might expect those requirements induce a bias towards the adoption of memory-safe languages. After all, the consequences of memory issues in this space are proportionally enormous compared to say, a web framework!

That being said, I remain bearish on overall adoption.

The sheer backlog of legacy code written in unsafe languages like C and C++ is so massive that I still expect the majority of PRs to be written with them (using the constraints I mentioned earlier).

DARPA proposed the “TRACTOR” (Translate All C TO Rust) program mid-2024 to some fanfare, but the size of the contract is much too small to make a major difference in the industry.

($25k for the grand prize in the first phase, intended for undergrad students, but still)

The fact is, translation of code costs money. A lot of money. Even with massive aid from LLMs, you still enter the category of “why fix what isn’t broken”. It becomes even worse with mission-critical or safety-critical systems, where a more accurate statement might be “why re-certify what isn’t broken”.

Washington doesn’t seem sold on the prospect either. Although a directive was posted in early 2024 urging the use of memory-safe languages for goverment contractors, the current administration appears lukewarm.

As posted in my Part 1 Thread:

The current executive administration unexpectedly pulled the previous directive regarding memory-safe languages in a federal context, which I am now factoring into my odds. The White House previously published this on their site, which is now gone.

Obviously, the whole world doesn’t revolve around US directives, but it’s an interesting signal about the state of memory-safe language adoption.

70% Odds.

Generative AI

Prediction:

80% odds a major AAA studio releases a videogame using an LLM for NPC interaction enhancement before the end of 2025.

Justification:

To clarify, when I use the word “LLM” in this context, I am referring to actual output generated on the fly in either a cloud server (or less likely a local model) for dialog generation. I do not mean pre-scripted dialogue hardcoded in the game that may have originally been LLM generated (which would also be difficult to verify, and I suspect is already happening).

This topic is going to ruffle some feathers, and I suspect many of you will think my odds are too high / misaligned.

Before you bring out the pitchforks, let’s look at the current evidence.

Ubisoft:
Demo’d NEO NPCs at GDC 2024. LLM responses + curated backstories from writers. Currently in R&D.

Electronic Arts:
CEO Andrew Wilson stated in May 2024 “a real hunger” to adopt generative AI “as quickly as possible”.

Take-Two (Rockstar):
CEO Strauss Zelnick in late 2023 argued “[with generative AI] you could imagine all the NPCs becoming really interesting and fun.”

Nvidia:
Nvidia’s Avatar Cloud Engine, or ACE for short, is a suite of curated AI models aimed at game developers. Initially released in 2023 specifically to “bring intelligence to NPCs through AI-powered natural language interactions.”

Like it or not, I think we are going to see LLM-enhanced NPCs very soon; even half baked.

There’s going to be a large first-mover advantage to whichever AAA studio releases it first. Even if there are major issues (jailbreaks, inappropriate responses, etc), the flurry of press coverage is likely to increase sales.

As they say, all press is good press.

None of this is new; the modding community has been experimenting with this for years.

Nintendo seems to be the one standout so far, explicitly denouncing AI use.

80% odds.

Generative AI (cont)

Prediction:

60% probability a major traditional stock music platform will release an AI music generator / composer to generate custom, copyright-free tracks on the fly before the end of 2025.

Justification:

Alright, I admit. This “prediction” is also a personal wish of mine.

Do you know how hard it is to find non-crap copyright-free music tracks for background audio? It’s quite trivial to exhaust the world’s supply of listenable tracks without much effort!

With the amount of YouTube videos I’ve released on my channel, I’ve easily gone through 50+ albums worth of content, just as background noise. There isn’t *that* much music out there.

Believe it or not, lo-fi hiphop girl isn’t infinite.

Many background music “genres” (if you can call them that) are very simplistic. If there is a type of music that generative AI should be able to disrupt, it should be this one.

I’m not looking for siren songs here. Just a little lo-fi will do.

Time for a reality check.

Envato, Adobe, Artlist, and Soundstripe are the big players in this space. Thankfully, the industry seems large enough where a little R&D should pay dividends.

Coherent Market Insights estimated the stock music market size in 2020 as $964 Million, with a projection of hitting $2 Billion in 2028.

Apple quietly acquired a UK startup named “AI Music” (how original) in 2022, though it doesn’t seem like they’ve done much with it.

Soundstripe, which is quite popular with Twitch streamers, released an AI song editor last year, but that’s for modification / extension of existing tracks, not creating new ones.

The elephant in the room is, of course, Suno, which offers full AI generated music capabilities. It’s not exactly suited for streamers or editors though.

What I’m really looking for is *endless* music, not hand-creation of individual tracks.

Frankly, I might be over-optimistic here. Existing stock music companies are, for lack of a better word, kinda boomer. They literally thrive off of commercial license rights, so I doubt many will have the foresight to invest in technologies that threaten their entire business model.

Hence, I’ve lowered my odds compared to my original prediction in the YouTube video. Watch this space though, it’s ripe for disruption by a startup.

60% Odds.

Supply Chain / SBOMs

Prediction:

90% odds at least 3 Major Fortune 500 Companies will publicly announce SBOMs (Software Bill of Materials) as being a mandatory, contractual requirement and deliverable for their third party vendors.

Justification:

In my video, I mentioned SBOMs as a way to (somewhat) mitigate the risks of supply chain attacks. At the very least, it accelerates forensic analysis and mitigation when such an attack occurs.

The Director of CISA agrees.

Log4j spurred a flurry of activity in the tech space, much of which I experienced first-hand during my tenure at Microsoft.

In an effort to build public trust, I expect many Fortune 500 Companies to establish strict SBOM requirements as part of their contracts. Boeing changed their supplier agreement to require SBOMs as a deliverable recently, in November of 2024.

I suspect that high-risk industries (defense, healthcare, finance, aviation) will be the first to adopt such practices. In fact, if you’re building a medical device, the FDA now requires a SBOM to get FDA-approval as of October 1st, 2023.

So why hasn’t everyone adopted SBOMs yet?

Are they just not making public announcements? Keeping it internal?

Maybe. But not likely.

The real reason, as always, comes down to money. Many of us don’t like to admit it, but at the end of the day, most software security is a cost center, not revenue generating.

Improvements in software security practices are often the direct result of a response to a previous attack. No sense in wasting money being ahead of the curve if shareholders don’t care.

But when something big happens, they care. A lot.

It will probably take another large, supply-chain related incident to get the next wave of Fortune 500 companies announcing SBOM adoption as a mitigation.

Unfortunately, I’d say that risk is pretty likely.

90% Odds.

Programming Languages:

Prediction (2-Parter):

70% Odds that Rust will have less YoY% growth for 2025 than Zig, as determined by the number of new StackOverflow questions tagged with said language.

90% Odds that Rust will have less YoY% growth for 2025 than Zig, as determined by official subreddit size.

Justification:


I’m taking advantage of the Law of small numbers, and I’m not ashamed to say it.

Cheating can be useful if it gets my point across.

I don’t think anyone would argue that Rust wasn’t the “hype” language of the last few years. The 2024 Stack Overflow Developer survey agrees:

Alas, everything has an expiration date. In the early 2010s, the general consensus was that Scala was going to “replace Java”…look where we are now. Swift hit an insane amount of momentum in the mid-2010s, only for the “wow factor” to stabilize in the 2020s. Don’t get me started on Ruby.

Currently, I think Rust is having its “Swift moment”.

Rust will probably survive better than Scala, but we’re starting to hit the boring stage. Years ago, all it took was writing a terminal emulator in Rust to hit the front page of HN. Old news.

The new kid on the block is Zig.

To be clear, this isn’t an endorsement of the language itself; language fights are silly. All programming languages are basically the same, politics is a mind-killer.

That being said, I think Zig is in a ripe spot to become the next hype language. Zig’s philosophical simplicity of “C but safer” is appealing compared to the complexity of Rust’s borrow checker.

Minimalistic languages are fun; a resurgence towards C-like languages is certainly possible. It’s not going to take over the world, but something has to occupy dev’s weekend projects. Zig may well be the rising star of 2025.

That’s going to wrap it up for me this year as far as predictions go. We’ll take a look back in a year to see how the predictions faired.

Future blogposts will be on other interesting stuff, this was a special series.

LW out.

It’s worth mentioning why we should make predictions in the first place.

The human brain has an unfortunate quirk of rationalizing hypotheses in a backwards direction. Or, to put it more succinctly:

“I knew it all along.”

It’s an insidious little statement. Colloquially known as the “hindsight bias”, it refers to our tendency to overvalue past events and create false narratives of predictability.

Historical outcomes get distilled into oversimplified views. This leads not only to overconfidence, but also interferes with genuine learning.

Sam Altman has a blogpost from ~2015 illustrating how comically wrong even the brightest minds can be about technological predictions.

Alas, we are not all doomed to become Reverse Cassandras.

Hindsight may cloud our judgement, but, as with all skills, improvements are possible with determined practice.

By constraining ourselves to forward-thinking, falsifiable predictions, we can at least measure progress (or lack thereof).

Here’s my attempt.

My series on 2025 Computer Science predictions will be divided into multiple parts, this is Part 1. More topics will be covered in subsequent posts.

For a more generalized video overview, you can watch my rambling here.

Table of Contents (Part 1)

1. Cryptography
2. RISC-V
3. Compression

Cryptography

Prediction:

70% Odds NIST publishes out new compliance standards mandating US Federal agencies (and their contractors) to switch to quantum-resistant algorithms for new systems.

Justification:

Current quantum computing factorization records are…kind of pathetic. The largest reliably factored number by Shor’s algorithm is 21.

More recently, in 2024, some researchers at the University of Trento managed to factor 8,219,999.

This was computed a D-Wave Pegasus. Kinda cool, but not a *real* quantum computer.

The distinction is actually quite important. D-Wave systems are quantum annealers, not gate-based quantum computers.

Train your eye to notice the difference. You’ll find yourself less swayed by “quantum” computing dribble in popular media.

On the other hand, the US Federal Government takes the quantum computing threat very seriously.

Symmetric encryption standards are relatively safe. Even a perfect implementation of my buddy Grover only reduces the brute-force complexity by half.

The NSA basically shrugged, recommended AES-256, and said “good enough”.

RSA isn’t so lucky.

Assuming a stable, error correcting quantum computer with a chonky number of qubits (thousands) running Shor’s Algorithm, it’s game over.

At the moment, “real” quantum computers (aka computers capable of running Shor’s) are significantly out of reach.

IBM has a 433-qubit system called Osprey, which you can also rent on their cloud platform for a chill $96 a minute. Who knew quantum computing was almost as easy as spinning up an EC2 box.

“But Laurie!” You exclaim. “That seems pretty close! If you only need thousands of qubits to break RSA, it could happen at any moment!”

Not really.

Error correcting qubits are needed to get any useful computing done. 433 physical qubits, with our current technology, only works out to a handful, at best, of logical qubits.

So why the fear in Federal Circles?

Y2K no longer permeates the nightmares of trepidatious Congresspersons. Nay, they dream of Y2Q.

In other words, cryptographic standards have to change significantly before theoretical attacks are remotely possible, due to “harvest now, decrypt later” surveillance strategies.

NIST produces the cybersecurity standards referenced by most federal agencies, although there are exceptions. A contest for the Post-Quantum Cryptography Standardization began in 2016, wrapping up in August 2024 with the winners.

CRYSTALS-Kyber is the winner for public-key encryption, so it’s likely to replace RSA in future contexts.

(note, not all of RSA’s functions are 1:1 replicated by Kyber; signatures are likely to be replaced with Dilithium)

The cyphers may be finalized, but there is much work to be done factoring in transition plans and official mandates. NIST published an initial public draft last November, but it is yet to be finalized.

(incoming rant about probabilites)

The current executive administration unexpectedly pulled the previous directive regarding memory-safe languages in a federal context, which I am now factoring into my odds. The White House previously published this on their site, which is now gone.

That being said, it’s not uncommon for administration changes to result in a complete redo of the Whitehouse.gov site, so this may be an administrative fluke.

Normally, I’d be wagering my bet at 85-90%, but the recent change of direction reduces my confidence.

70% Odds.

RISC-V

Prediction:

80% odds at least three major linux distributions will have official stable release channels for RISC-V architectures (e.g. fedora, ubuntu) that aren’t just experimental builds by the end of 2025.

Justification:

RISC-V may be the coolest ISA on the block, but at the moment, adoption by the linux community is limited.

Lurkers of distrowatch may (incorrectly) get the impression of Debian as the “old man” distribution from the relatively slow release cycle of the stable channel. Contrary to public perception, Debian is significantly ahead of the curve when it comes to RISC-V adoption.

Official support on the unstable Debian branch for riscv64 started 1.5 years ago, back in July of 2023. Most risc-v SBC enjoyers will tell you that; yup, 99% of the time the default recommended distribution is Debian unstable.

Arch, somewhat obviously, has had unofficial riscv64 releases for a while now. Fedora is still in the experimental stage (Rawhide builds only).

It’s pretty reasonable to expect official, stable releases in the near future. Debian 13 is TBA, but if historical trends are anything to go by, a summer 2025 riscv64 stable build is quite likely.

Ubuntu will soon follow, considering their upstream follows Debian, and they already have an “early” (limited) RISC-V port. Considering the competition, I would wager Fedora competes in this space soon after.

80% Odds.

Prediction:

20% Odds the Raspberry Pi Foundation releases an official primary SBC based on a RISC-V SoC (no mixed cores).

Justification:

This one is a bit tricky. Although my low estimate may seem misguided at first, it’s important to evaluate what the community wants versus what is logistically possible.

The Achilles heel of the Raspberry Pi Foundation has always been, and forever will be, supply chain shortages.

The industrial market caused an unexpected surge in demand starting in the mid-2010s, and it’s never been the same since.

What began as a noble effort to create low-cost, linux SBCs for the educational market, quickly ballooned into a commercial frenzy. The sheer convenience of the Raspberry Pi became its downfall. If it’s so easy to develop on…why not use the same board for production use?

Well, turns out there are a lot of reasons not to use SBCs in that environment, but everyone’s going to do it anyway. Welcome to the new normal.

The knock-on effects of this, of course, is that the Raspberry Pi foundation now has significantly less ability to experiment with radical design changes. They quite literally can’t afford a flop; they are a publicly traded company now after all.

In the early 2010s, the foundation probably could have gotten away with a full ISA shift. (risc-v of course was barley a blip on the radar during that period, but I digress…)

Now, not so much. RPI’s relationship with Broadcom is very strong, and Broadcom has a complete absence of public statements about RISC-V.

Current RISC-V SoC / CPU manufacturers have, to put it nicely, a lack of maturity. Xuantie and SiFive are the big names here, but they still come out with processors that have hardware bugs in the CPU itself.

Supply chain risk, lack of software package support, backwards compatibility risks, as well as hardware constraints make this an unlikely proposition.

20% Odds.

Compression

Prediction:

80% Odds at least 2 major streaming platforms switch to AV1 as the default codec for HD and UHD content on capable devices by the end of 2025.

Justification:

Screw H265. All my homies hate H265.

Video compression is always an interesting field of research. The potential savings in bandwidth costs alone are so high that there is strong justification for implementing more efficient standards.

It’s one of those rare events where technological advances are immediately noticeable on both sides of the equation; both the distributors (streaming platforms) and the end user.

In many computer science domains, the impact of an efficiency improvement is only translatable to the business side. Sure, things like the migration of Oracle to Aurora databases saved AWS nearly $100 million (rumored), but as an end user, the difference was imperceptible.

To understand the video compression landscape, we first have to talk about The Codec Wars.

On one side, we have the evil Royalists, defenders of the proprietary licensing model. Their forces consist of the MPEG LA and their allies.

On the other, we have the Free Streamers, proponents of royalty-free codecs. Their forces consist of the Alliance for Open Media and their sympathizers.

MPEG LA got a head start, standardizing their H.265 (HEVC) codec in 2013, but was rife with fragmented patent pools.

The tiered royalty structure made licensing fees unpredictable, and prohibitive to smaller company adoption.

You’re looking at annual minimum royalties of ~$100k a year!

That said, the previous success of H264 led hardware manufacturers to adopt HEVC decoding relatively quickly. Consumer devices broadly adopted decoders in the mid-2010s. The iPhone 6 started using H265 for FaceTime, and NVIDIA’s 900-series featured on-device decode as well.

AV1 (the Free Streamers), started as a consortium by various FAANG-companies as a next-generation, royalty free video compression alternative in 2015. Early implementations of the AV1 codec were significantly more computationally intensive, but optimizations continued to push the speed over time.

In the 2020s, hardware manufacturers started to adopt AV1 decode (Intel, NVIDIA, etc), making streaming adoption more viable.

Twitch was one of the more intriguing experimental adopters; AV1 shows about 30% better compression efficiency compared to H265 for the same quality. Due the nature of the platform essentially being a “sea of individual encoders” (streamers), royalty-free options are particularly relevant to their user base.

The savings on royalties alone represent such a massive potential cost savings for streaming platforms that my optimism remains high for the swift adoption of AV1 in the near future. Hardware decode is quickly reaching a critical mass for end user devices, it’s only a matter of time before everyone makes the switch.

80% Odds.

Prediction:

30% odds a major streaming platform announces a partnership with a particular TV, set top box, or hardware vendor (e.g. NVIDIA) to certify a device for a sort of “enhanced AI streaming”.

Justification:

Okay, this one seems weird, but hear me out.

Anime upscalers have gotten ridiculously good.

Anime artwork, mostly due to the flat colors and sharp linework, responds particularly well to neural-net upscalers. Many series only exist in lower resolutions; never getting a proper “remastering” treatment, often from lack of popularity.

The desire of fans to view their favorite shows in upscaled resolutions, combined with the rapid progress in the effectiveness of the techniques, led to a snowball effect of improvement.

Don’t get me wrong, it’s still a *niche* subject. But the tech is already there.

Let me attempt to convince you another way.

In 2023, Microsoft released their “Video Super Resolution” as an optional feature to reduce compression artifacts of low-bitrate videos in Edge.

As of January 2025, it is still disabled by default, but the option is there. And it looks…not that bad actually.

It’s going to take some convincing before the general public wholeheartedly accepts neural video upscaling as a “default” feature, but the artifact reduction is quite impressive.

Much like how many consumer televisions implement frame generation for motion interpolation (gross), I wouldn’t be *that* surprised if this technique gets adopted more widely.

I place low odds currently because the hardware and implementations are still young, but then again, streaming platforms need new ways of differentiating between each other.

No one really has a leg up on anyone else, and neural upscaling seems just *crazy* enough that some streaming platform might implement it as an attempted competitive edge.

To be fair, I don’t think it’s a great idea. But it is *an* idea.

30% Odds.

Bye for now. I’ll see you again in Part 2 shortly.

Subscribe now