Lightning-triggering Drones

THEWHITEBOX
TLDR;

This week, we take a look at the state of frontier AI models, with some unexpected surprises, a breakthrough from a 14-year-old Chinese prodigy, a new hire by OpenAI that speaks volumes about its future, product updates that are really worth considering, and a fascinating way humans intend to use AI for: capture lightning to produce electricity.

Enjoy!

THEWHITEBOX
Things You’ve Missed By Not Being Premium

On Wednesday, we talked about the viral Chinese robot video that went berserk on its researchers, how Google’s dominance remains uncontested, Apple’s new AI plans, and more.

FRONTIER MODELS
Update on the Pareto frontier

The Aider Polyglot team, which evaluates models in what’s considered the prime coding benchmark, has acknowledged a mistake in Google’s cost analysis for Gemini 2.5 Pro, which appears to be higher than what they originally presented.

This means the Pareto frontier is now slightly altered, showing three leaders based on performance-cost ratios. DeepSeek’s models offer the best performance for the cheapest models. For high-end results, Gemini and OpenAI are the leads nonetheless.hey

TheWhiteBox’s takeaway:

An important thing to notice on the graph is that the scale is not linear, making it look like OpenAI’s models are not much more expensive than Google’s despite being better.

But that’s not true.

Both superior scores by OpenAI are around 2x and 3x more expensive than Gemini 2.5 Pro, only to be a couple of percentage points better. Huge cost increases for a marginal improvement, so the best model in that graph (the most Pareto optimized model) is clearly Gemini 2.5 Pro nonetheless.

AGENTS
A Chinese Prodigy Builds A Model… Too Good to be True?

A Chinese 14-year-old boy named Ji Shihao has built a model, Jinmeng 550A, that achieves historical results in two very well-known benchmarks: the AIME 2024 International Mathematics Olympiad and MedQA, a question-answering benchmark focused on medicine, while beating all frontier AI models by basically clearing the benchmarks with 100% and 99.7% scores while being only 3% the size of some frontier models and costing orders of magnitude less.

This amazing breakthrough is based on the idea of neuro-symbolic AI, an often disregarded field of AI that combines the virtues of neural networks like ChatGPT and symbolic engines, human-written code basically (like a set of rules to solve math equations).

But what does that mean?

If you look at the task below, you can clearly see the thought process a human would use.

1. Understand what the concepts of ‘large things’ and ‘metal spheres’ are,

2. identify each,

3. count,

4. and see if they match.

The issue is that while neural nets are great (above-human actually) at perceiving the different concepts in the image, they struggle with the high-level reasoning part required here (counting and equalization).

This might seem very easy for you but it’s certainly not for a neural net, as the concepts, the maths, and the high-level reasoning idea of comparing both values are all three tasks that have to be done by the AI model simultaneously, which often leads to entanglement.

Instead, a neurosymbolic AI leverages where NNs are great at, like taking the image and creating a table with all the objects by shape, material, color, and position (all low-level perceptive qualities of the task), and then using a symbolic engine (a computer program) that traverses through the table using human-written code like filtering or counting functions, finding the correct answer.

And while frontier models like o3 have considerably elevated what end-to-end neural nets can do, in many instances, they are still humiliated by neurosymbolic AIs, which gives some credence to the claim.

TheWhiteBox’s takeaway:

There are two takeaways from this story:

1. It’s too early to celebrate, as no technical paper has been released and no independent verifier has looked at the model. For all we know, the boy might have trained the model on the test responses so it has basically memorized the answers.

2. Many still see neurosymbolic AI as a possibility to prevent stagnation, as they believe neural nets aren’t enough to reach AGI alone.

But if neurosymbolic AI is so promising, why do OpenAI and others ignore it? The reason is that many people in the community still believe that neural nets are the solution to everything, the key to AGI.

Who’s right? Only time will tell.

AGENTS
What ‘Agentic’ Actually Is

OpenAI has posted a short video showcasing o3’s capabilities to solve more or less complex tasks, such as creating a report from five different spreadsheets for a company.

It is nothing extraordinary, but it perfectly illustrates what an agentic application actually is.

And, more importantly, what isn’t.

OpenAI’s video is a perfect example of what agentic is for a very simple reason: the model gets to decide how to approach the problem. This is crucial because it helps us see through the bullshit, as most alleged ‘agentic applications’ are not that, but workflows.

I’ve talked about this in the past, but here’s the main takeaway: There are two ways you can add AI to a business process:

1. Workflow: The human guides the entire process based on strict, fixed rules, using AI models in the necessary parts.

2. Agentic: The human writes a very detailed prompt explaining what he/she wants, and what the AI has to do. But, crucially, it does not tell the AI how it has to do it. You will define the guardrails and overarching goals but not the process the AI must follow.

OPENAI
OpenAI Hires Fidji as CEO of Applications

Fidji Simo, Instacart’s CEO, has joined OpenAI as CEO of Applications. This apparently weird role gives us great insight into the new structure of the company.

She will not be OpenAI’s CEO; that will continue to be Sam Altman, and Fidji will report to him. In the former's words, she will focus on the product side of things, while he focuses more on research, infrastructure, and safety.

As he explained in the announcement blog post, OpenAI has “also become an infrastructure company, building the systems that help us advance our research and deliver AI tools at unprecedented scale,“ implying they are adamant about becoming a verticalized company in the likes of Google.

TheWhiteBox’s takeaway:

What is the implication of this hire? Well, simple: it’s the unofficial announcement of an upcoming ad system on ChatGPT.

Fidji has been an ad-focused executive for years, building the ad business on Instacart, Facebook, and eBay. Thus, it seems clear this is what’s coming next for ChatGPT.

And it’s not like they have an option. As we saw in our last Premium segment, most AI demand is actually not monetizable… unless you put ads. Hyperscalers are seeing huge demand, but on products they can’t make bank on.

That is why ChatGPT will resort to the same business that has built entire trillion-dollar companies like Google or Meta, the business in which you don’t pay for the product; you are the product.

DEEP RESEARCH
OpenAI’s Codebase Research

OpenAI has announced that their deep research tool can now analyze complex codebases by simply sending it the GitHub repo’s URL. Via a native integration with GitHub, you can connect ChatGPT to your repo, and the model will analyze it, generating a report explaining the codebase.

Of course, you can send this to public reports that aren’t yours, but you might be interested in facilitating the job of understanding the repository.

TheWhiteBox’s takeaway:

If there’s a Generative AI use case that has attained product market fit, one that I can confidently say is worth the money and helpful to every single knowledge worker or student on this planet or anyone mildly curious about learning, that’s the Deep Research features products like ChatGPT or Gemini offer.

DR allows the AI to search hundreds of web pages for the information required to answer your request.

The result is multiple-thousand-token reports (sometimes up to 5,000 words) that are surprisingly good; not quite enough for you not to have to write the report yourself (at least that’s what I always recommend, but many people simply copy and paste), but either way, an absurdly good summary to build a deep intuition on the topic.

Of course, the natural next step is to broaden its impact to local data (enterprise data, etc.).

Although we aren’t quite there yet (and I’m not sure you want to share your enterprise data with OpenAI, in all honesty), that doesn’t mean we aren’t expanding the reach of these products.

TREND OF THE WEEK
Lightning-Catching Drones

As Generative AI has completely absorbed the AI industry, this and other newsletters are at risk of sending the wrong message to readers, in that AI = GenAI.

For that reason, and because today’s topic is fascinating, I wanted to switch things up a little bit to discuss a fascinating claim made in Japan: NTT claims to have achieved something never seen before: drones that capture lightning and could be used to provide electricity to Japanese cities.

This trend of the week pretends to help you broaden your view of what AI will do for us in the near future.

Lightning-Catching Drones

Although lightnings aren’t nearly as common in other areas of the planet, and is most frequent in the tropics, Japan also sees significant storm activity, enough that lightning damage is a serious national concern.

Nonetheless, the Institute of Electrical Engineers in Japan estimates that lightning can cause damage up to 200 billion yen every year, or $1.3 billion.

Additionally, lightning is, well, electricity, so ideally, we would want to avoid it causing damage and harness its power!

But can we do the two things? A company in Japan claims to have done it… kind of.

The idea is pretty simple: we want to place drones in the air that predict where lightning might occur, trigger them, absorb the impact, and send the electricity downward.

From damaging the Japanese economy to potentially “infinite” new electricity in an era where energy power is one of the most important things for a country.

Sounds awesome, but it’s much harder than it reads.

The Science Of Lightning

Without getting too technical, let’s briefly explain what lightning is and why it occurs.

In our world, some particles have a ‘positive’ charge, like protons, and others have a ‘negative’ charge, like electrons. It’s an intrinsic feature of these particles.

So, if a surface or object has more protons than electrons, it’s positively charged, and vice versa. Crucially, these two are always attracted to each other, generating an electric field between them. The strength of the field determines how strongly charges will be pulled together.

Thus, if their attraction force is stronger than the resistance of the intermediate medium, the electrons will be ‘pushed’ toward the positive charge, generating a current.

Thus, whether current actually takes place is based on whether the medium allows for it:

• Metals are conductors, so they are an excellent medium for driving current. Metals have free electrons by default, making them excellent conductors. Thus, the moment an electric field exists, the electrons will move in a particular direction, which as mentioned is what we call current.

• On the other hand, the air, for example, is an insulator. Without free electrons, it doesn’t easily allow current to flow.

All things considered, what is lightning, then? Clouds tend to generate electric fields naturally inside of them. Sometimes, the lower part of the cloud generates a strong charge, be that positive or negative.

Hence, an opposite charge starts accumulating on the surface of the Earth closer to this charged cloud zone. At one point, once the electric field is strong enough, an electric path is generated across the in-between air so that both ends meet.

This path, also known as a ‘discharge’ because it neutralizes the charge difference, is what we call lightning.

But wait, the surface is also charged? Yes!

Although it’s rare, you can sometimes see how the surface’s charge goes up to meet the downward current. In the image below, the current eventually found a faster path but was really close to closing through the one on the left!

Knowing this, it’s now easier to see what these researchers are trying. And perhaps more importantly, where on Earth does AI fit here?

The Lightning Trigger

As lightning generates strong electric fields, the drones aim to generate a sufficiently strong attraction so that the lightning goes directly to them.

For that, we need the following:

1. Being able to measure where charges are concentrated in the cloud to position the drone underneath

2. Being able to actively trigger the lightning on the drone by somehow breaking the air’s insulating capabilities.

3. Find a way to prevent the drone from being destroyed; some lightning strikes get 3x hotter than the surface of the Sun!

4. Channel the huge electric burst downward and use it somehow.

For this, they assembled the following experiment: they connected the drone pictured below via a conductive wire to a ground switch that, connected at the optimal moment, induces a strong electric field on the drone that triggers the lightning.

To protect the drone, they assembled a Faraday box mechanism that prevents the current from flowing through the drone (destroying it basically) and redirecting it downward.

And on December 13th, they managed the first successful lightning triggering. A voltage difference of over 2,000 volts developed between the drone and the ground, making it a favorable target for the lightning strike.

By doing this, the drone was much more ‘attractive’ (pun intended) for the charged cloud as the easiest way to neutralize the charge difference (which is what the lighting is meant for), leading to lighting being triggered on the drone.

And just like that, we managed to trigger lightning with our own influence so that, one day, we can use it to charge our world:

But wait, where on Earth does AI come to play here?

Using AI to Predict Charge

If you read this newsletter regularly, you will know by now how AI works, but let’s do a brief explanation.

AIs are exposed to a large set of data, and by carefully defining what they need to predict, they learn the underlying patterns in the data.

Using our friend ChatGPT as an example, if we want to teach an AI to speak our language, we expose it to a lot of natural language data and define a prediction task that will naturally lead to the model learning the language.

Which task is that?

Simple, next-word prediction. If an AI can confidently predict what word comes next given a text sequence, the model directly learns to speak the language diligently, and indirectly compresses the patterns in language.

Put another way, if a language model has learned to predict how words follow each other, it has also learned the underlying grammar and knowledge. For instance, if a model can predict ‘What’s the capital of France → Paris,’ the model has, for lack of a better term, ‘learned’ that Paris is France’s capital.

The point here is that AIs can learn patterns in data and make predictions based on those patterns. Holding that thought, it’s not hard to see how AI might be used in today’s case.

For a drone to trigger lightning, it needs two things:

1. Know where the charge is being accumulated in the cloud, known as the lightning generation point,

2. Know the amount of charge that is being accumulated.

Luckily, we have plenty of atmospheric electric field data, so AIs can learn the hidden patterns in data and confidently predict where lightning will occur. Therefore, to answer the question, here AI is being used to decide:

1. Drone location

2. Drone-to-ground charge (based on the predicted accumulated cloud charge)

3. Drone height

Thereby closing the loop of what could one day be a new energy source for civilization.

Closing Thoughts

With this piece, I not only wanted to share fascinating new research, but I also wanted to convey the importance of AI way beyond chatting models.

Here, no tokens are predicted, no words, images, or video are generated. Here, AI is simply being used to locate an awkward-looking drone under a storm.

But the takeaway is that without AI, we would never have the predicting capacity to locate drones in the precise location at the precise time.

Yes, AI is more than ChatGPT. Way more.

THEWHITEBOX
Join Premium Today!

If you like this content, by joining Premium, you will receive four times as much content weekly without saturating your inbox. You will even be able to ask the questions you need answers to.

Until next time!