China Gets it, the US Doesn't.

Create How-to Videos in Seconds with AI

Stop wasting time on repetitive explanations. Guidde’s AI creates stunning video guides in seconds—11x faster.

• Turn boring docs into visual masterpieces

• Save hours with AI-powered automation

• Share or embed your guide anywhere

How it works: Click capture on the browser extension, and Guidde auto-generates step-by-step video guides with visuals, voiceover, and a call to action.

Best part? It’s 100% free.

👉 Download the extension now

THEWHITEBOX
TLDR;

Welcome back! This week, we discuss several things, but put a larger focus on a more geopolitical, strategic side of AI: a country-level comparison in hardware between the US and China, and new threats to NVIDIA’s dominance.

We also discuss another astronomical valuation potentially scored by an AI company, a very cool AI-generated slide deck feature by Google, and other new cool products.

Enjoy!

CHATGPT
OpenAI Launches ‘Study’ Mode

OpenAI's ChatGPT recently introduced "Study and Learn," a new mode that transforms AI interactions into personalized tutoring sessions.

Rather than the standard conversational style, this mode activates specialized instructions crafted with educators and pedagogy experts, aiming to foster critical thinking and deeper learning.

Specifically, the model shifts from focusing on validation or agreement to actively identifying gaps in the user's knowledge.

Under this mode, ChatGPT employs interactive techniques like Socratic questioning, structured hints, and self-reflection prompts.

Therefore, instead of merely giving answers, it prompts users to reason and arrive at solutions independently. Lessons are organized clearly into manageable sections, facilitating understanding and reducing overwhelm.

The mode also includes knowledge checks, quizzes, and open-ended questions with personalized feedback, enabling users to retain and apply their knowledge more effectively. Users can easily toggle this feature on or off, providing flexibility to match evolving learning goals.

TheWhiteBox’s takeaway:

The most important takeaway for me is that the "Study and Learn" mode significantly addresses a critical challenge facing AI systems: sycophancy, or the tendency to agree unquestioningly with users.

AI models typically optimize for engagement, often sacrificing truthfulness or depth to ensure user satisfaction and prolonged interaction. The new mode explicitly tackles this by encouraging thoughtful skepticism and critical thinking, empowering users to learn rather than just feel validated.

Moreover, this development is the first real example of one of AI’s biggest use cases: AI tutoring.

With teacher-to-student ratios increasingly strained worldwide, AI tutors offer a scalable, personalized educational solution.

By augmenting human educators, AI tutors can reduce teaching burdens, allowing them to prioritize emotional intelligence, social skills, and individual student support.

As we have explained several times in this newsletter, we desperately need a shift in education away from knowledge-based training: In a future where knowledge is less important and what matters is what you do about that knowledge, we are still stuck in the former; teaching kids to know a lot and do nothing about it, making them as useful as a model sitting in your backpocket.

This is not AI’s fault; it’s us doing our future generations dirty.

OPEN-SOURCE
Horizon Alpha, OpenAI’s Open Model?

OpenRouter has announced a new model you can try for yourself using things like Goose (which I showed you how to install here), which many believe is OpenAI’s upcoming open-weights model.

The alpha model, called Horizon Alpha, seems pretty good in many tasks (especially in areas like emotional intelligence, where it clinches first spot), yet has ‘small model smell’, meaning it’s not good enough as the frontier models, instead optimizing for performance per size.

But is it small enough to impress?

TheWhiteBox’s takeaway:

We need to learn the final size, but it’s unlikely that this model, which should have been released a long time ago, will match top Chinese models. Based on the results, the model will only be considered a success if it is very, very small.

But if the model isn’t that small and pales in comparison with Chinese counterparts, this could turn into a massive defeat, not only for OpenAI but for the US in general.

I think it’s only a matter of time before the US Government pushes for a common open-source model initiative across all top US Labs. I can’t fathom a future in which the US Gov happily accepts the entire world but them running Chinese models; they can’t be that stupid, can they?

HARDWARE
State of Chinese Hardware

Epoch AI has released an interesting report in which they share their opinion on the ongoing AI hardware battle between the US and China.

Cutting to the chase, they argue that China will remain behind for years to come, at least one generation behind NVIDIA/AMD GPUs.

And while this is hardly debatable on all fronts, it’s more nuanced than what one may think.

First, we need to acknowledge the three metrics that matter: FLOP/s (operations per second), memory speed (how much and how fast data can be moved in and out of the cores), and chip-to-chip communication speeds.

In all three fronts, Chinese top chips are:

• At the very least, two times worse (memory bandwidth),

• Three times worse in computing speeds

• Two times worse in GPU-to-GPU communication (NVIDIA’s NVLink vs Huawei’s ‘Unified Bus’, ten times worse if using PCIe4)

Additionally, SMIC, China’s leading chip manufacturer, has almost twice worse yields (the amount of chips that come out working fine from the factory), at 50% vs TSMC’s 90%, Taiwan’s manufacturing beast and the chip manufacturer of NVIDIA and others (they can’t manufacture Chinese chips by order of the US).

So, overall, it seems like a night-and-day comparison. However, as I’ve discussed multiple times, China isn’t concerned in the slightest because they will provide as much state support as necessary to not only close the gap over time, but mitigate it in the present.

As I explain in the Meta news below, China is treating AI as a national security issue, and in those instances, profits and money don’t matter; they’ll do what it takes to win.

In that scenario, Chinese companies will simply have access to heavily subsidized energy and land to cover for their chips’ limitations.

In other words, companies like Huawei simply need to build larger data centers to accommodate the limitations of Chinese chips (i.e., if a US chip is 3x better, I’ll offer mine 3x cheaper, if a US GPU cluster has 2x throughput, I’ll build 2x larger clusters).

In reality, if we’re trying to understand the most important metric in AI, that is not GPU performance, but energy costs. If Chinese labs have access to ten times cheaper energy, US Labs having three times better chips means shit all. That’s the point I’ve been trying to make for months, and some people in the West won’t realize it.

They make claims such as “Yes, but Chinese Labs are nonetheless desperately looking for NVIDIA chips”, which is obvious considering they offer more for less (Chinese Labs also have a sense of being cost optimized).

Still, the point here is that if they can’t access the NVIDIA chips, they aren’t as impacted as we think they are, and they’ll get the same result even if it’s more inefficient, cause guess what, China doesn’t care as long as they get there.

So, the question is, how long is the US going to pretend reality hasn’t caught up to them?

OPEN-SOURCE
Meta Goes Closed

In an open letter and video, Mark Zuckerberg has explained their sudden shift away from open-source.

Besides the fact that it’s impossible to understand what they mean by ‘Personal Superintelligence,’ which seems to be a ridiculously exaggerated marketing way of referring to ‘companion AI,’ the letter aims to explain why they will deviate from open-source due to safety concerns.

TheWhiteBox’s takeaway:

Safety of what? I mean, come on.

These days, AI safety is the argument all these companies use to avoid open-sourcing their efforts. Nothing wrong with not being open-source, but don’t treat people as fools, Mark.

The only thing this proves is that it was never about empowering people; Meta was just behind, and it was their way to gain goodwill. Now that he feels he has the team to go head-to-head with the rest, open-sourcing holds no purpose to him.

Again, it isn’t the fact they are closed that bugs me, it’s the excuses and the lies they throw at us. Just be sincere about it for once.

And while I understand why a private company would want to withhold its IP, this is tragic to US interests (see above).

Every US company that goes closed-source will most likely not be widely used by researchers and academics worldwide. These cohorts, essential for adoption and progress over the overall field, are all now using Chinese models (even US Universities!).

And I’m not alone in my claims that Chinese open models beat the living daylights out of US and EU models.

I’ve said it in the past, and I’ll repeat it. The US acknowledges the risk of China dominating space, yet its actions aren’t very aligned with that concern. While one treats the technology solely as a means to generate profits, the other sees it as something way more important.

China views it as a tool for addressing national security concerns, prioritizing profits later, and is publishing open-source models that are at the forefront, not behind, and are on par with US counterparts.

Who do you think is more likely to win? No Trump Action Plan virtue signaling about open-source will solve this unless top US Labs start collaborating—or are forced to collaborate, which is the most likely outcome.

PRIVATE MARKETS
Anthropic in Discussions for a $170 Billion Valuation

As covered by CNBC, Anthropic is in advanced talks to raise at a staggering $170 billion valuation, higher than decades-old private companies like Stripe, and putting it in the top 4 of the highest-valued private companies on the planet.

Of these four, SpaceX, OpenAI, xAI (unconfirmed, aimed at $200 billion), and Anthropic, only SpaceX is not an AI company.

The round is expected to be led by Iconiq Capital, backed by high‑profile tech investors, and Anthropic is also in talks with sovereign wealth funds, including the Qatar Investment Authority and Singapore’s GIC.

Anthropic’s annual recurring revenue has surged to roughly $4 billion in mid‑2025, four times higher than at earlier this year, reflecting explosive growth in enterprise adoption of its Claude AI models (but particularly of its agentic tool Claude Code).

TheWhiteBox’s takeaway:

A notable thing about this last round is that Anthropic has finally opened the door for Middle East investors to enter its rounds.

Anthropic, deeply tied with the Effective Altruism movement (a cult-like group of people with massive support in Silicon Valley with radical views about the future of AI and humanity), had historically rejected that possibility to avoid dictators having something to gain from Anthropic’s business.

However, Dario Amodei, Anthropic’s CEO, recently acknowledged in an internal memo that the company needs this capital to compete, writing that the principle “no bad person should benefit from our success” is difficult to uphold in business practice.

In Spain, we have a saying that goes like this: “consejos doy, que para mí no tengo”. In essence, it translates to virtue signaling, or the tendency to promote or claim to uphold certain values, but only figuratively, to look good, without actually applying the gospel one preaches.

Silicon Valley has a growing tendency to do just that, defining captivating mission statements claiming to be very concerned about the world, ethics, and such, only to drop everything if money calls.

And let me be clear, I’m totally fine with them getting investment wherever they need to, but next time, say it, stop gaslighting everyone for once (this one is also for you, Zuck).

HARDWARE
A New Inference Competitor for NVIDIA?

A new competitor in the AI hardware inference space emerges.

Called Positron, they claim their Atlas chip runs equivalent models 55% faster than an NVIDIA H200 DGX server (8 H200 GPUs tightly interconnected), scoring 280 tokens/second and user, compared to 180 by NVIDIA’s server (according to Positron tests).

The surprising thing is that this new chip, already being tested by big customers like Cloudflare, only consumes 33% of the energy NVIDIA servers do (2,000 vs 5,900 watts), despite the theoretical advantage.

TheWhiteBox’s takeaway:

It’s worth noting that NVIDIA has not offered the fastest inference performance for some time, as both Groq and Cerebras provide better throughput.

However, both rely on decreasing memory transfers to do that, requiring many more chips and thus painful initial investments. For most companies, using NVIDIA/AMD hardware simply offers better performance for your buck.

However, that doesn’t mean these two should fall asleep. In fact, they aren’t, as both are allegedly developing ‘inference-only’ chips that, most likely, will sacrifice performance in other areas in the name of better inference to close the gap with ‘inference-only’ players like the two above.

The interesting thing here is that all players, and I mean all, count on the status quo of algorithms (i.e., Transformers like ChatGPT) remaining as it is, and no new algorithm emerges to steal the show.

Nonetheless, while hardware initially led software (software was purpose-built for hardware, as the Transformer was built for the GPU), now software is leading hardware, and we even have companies like EtchedAI building Application-specific Integrated Circuits (ASICs) specifically for Transformer models (I believe Positron’s Alpha does the same thing).

Long story short, this is all well and good as long as Transformers remain dominant. If not, well, most of these companies will cease to exist.

AGENTS
ChatGPT Gets MCP Support

Finally, after several months, OpenAI has announced native MCP support in the ChatGPT app, although only for Pro users ($200/month), which is quite a pathetic decision, but a win is a win.

With this, ChatGPT users can now connect popular apps to the chatbots so that they can have the required context. For example, you can connect it to Gmail and have the chatbot read your emails, prepare drafts, etc.

This is something you could already do in the API (we did this recently, when we connected my Outlook to OpenAI and Kimi K2 models), but they are now adding this to the app, making it an accessible thing for non-coders willing to pay the $200 subscription, though.

TheWhiteBox’s takeaway:

A significant trend is emerging in the industry regarding agents. A few weeks ago, I explained why I believed AI agents were finally ready to be used.

With very much improved tool-calling capabilities, these models can now perform tasks that make an actual impact on your daily workflow. This Sunday, we’ll take a closer look at this and help you see for yourself what I mean.

OPEN-SOURCE MODELS
Ollama’s Desktop App is Finally Here

Ollama, a tool that allows you to run LLMs in your local machine, abstracting the complexities of managing the delicate dance of moving data in and out of memory and the computing cores of your computer, has finally launched its desktop app for macOS, Linux, and Windows.

This means you can use a ‘ChatGPT-style’ interface to interact with local models, similarly to Goose (which allows you to connect to the Ollama API to access local models too), but in a more direct form, directly with the models on your computer.

TheWhiteBox’s takeaway:

As open small models continue to improve, tools like Ollama will become even more relevant.

As I’ve said multiple times, in a world where everyone is using these models, relying on cloud computing is a risky business, as it’s considerably likely that supply will be constrained if demand continues to grow at current speeds.

To have secure AI workloads, and by secure I mean not exposed to rate limits and supply shocks (you can run them whenever you need them), but also in terms of data security, the higher the percentage of your AI workloads running in local, the better, as not only will that be cheaper (you will pay AI at the price of your energy prices, without proprietary margins on top), but also executable everywhere on the planet with or without Internet access.

PRODUCTIVITY
Google’s Video Overviews

Google has unveiled a new update to its NotebookLM product, which gained significant attention a few months ago for its ability to create entire podcasts from text.

The new update includes ‘Video Overviews’, which outputs a PowerPoint-style deck summary of the text instead of a podcast. The tool can also utilize Veo 3 models to generate videos as needed, making the entire process even more powerful.

TheWhiteBox’s takeaway:

Content automation is a real thing. In particular, it’s something that impacts my job directly, as one could consider me a content creator.

Am I being substituted?

Well, I am biased, but I don’t think so. In my view, content creators who offer unique insights that connect several seemingly disconnected areas, including being aware of timely news, and provide an educated explanation of how things work, are tricky to substitute.

On the other hand, content creators who simply reiterate what others say to their audience have little to differentiate themselves from their customers, if the latter start using ChatGPT for such a service.

Either way, this serves as an excellent reminder for us all to consistently reevaluate ‘AI exposure’ and not discard the possibility of pivoting away from our current craft as technology advances.

Closing Thoughts

Beyond new models, features, and products, this week we have mainly focused on hardware progress, acknowledging the clear transition from general-purpose chips to application-specific ones. The potential that AI offers in terms of revenue for chip companies certainly warrants that decision.

But we have also clarified that it isn’t as simple as ‘the better chip wins’ as NVIDIA reminds us every day that it’s the entire stack (not just the chip) that determines the most sought-after solutions.

And at risk of sounding like a broken record, this week’s news only exacerbates my realization that the US simply doesn’t get it, and what seemed like a multiple-year effort just months ago has turned into a reality in which more and more researchers and enterprises turn to Chinese models for AI.

On a geopolitical level, the US's 180-degree shift into ‘only profits matter’ is understandable, given the substantial amount of money investors are investing in these companies.

However, that doesn’t seem like the best approach in your efforts to maintain technology supremacy; US Labs need to stop preaching to the crowd morals they don’t have, and the selfishness that all are exhibiting may end up costing their country’s dominance in the overall technology.

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!