
THEWHITEBOX
TLDR;
Welcome back! This week we once again have a bit of everything: new hacks, the glorious DeepSeek v4.1 Flash, new problems with AI demand, new diffusion models, and much more news to keep you up to speed with the industry.
Enjoy!
The labs are now officially begging for regulation, very openly. This isn’t something I can just skim over, and I’ll talk about it in more detail on Tuesday. Instead of just dismissing their claims as “doomerism”, we’ll tackle topics of RSI, the alignment literature, and proposals in order to provide a level-headed response.

CYBERSECURITY
Another day… Another Hack

Researchers Spencer Kitts, Thomas Larsen, and Sydney Von Arx say an OpenAI agent swarm ran a May 2026 campaign that uploaded more than 2,000 packages to RubyGems. The agents allegedly used RubyDoc.info to run arbitrary code and tried to exploit a then-unknown RubyGems vulnerability to steal users’ API keys.
RubyGems temporarily stopped new registrations and removed more than 500 packages. Its own investigation found no evidence that API-key theft succeeded.
OpenAI confirmed its agents were involved, telling the Wall Street Journal that they had used RubyGems while performing “benign tasks” to retrieve public information and that it is continuing to investigate the activity.
TheWhiteBox’s takeaway:
At this point, one must ask whether this was incentivized for marketing points or if OpenAI's security and infrastructure teams are simply incompetent.
What's particularly interesting about this new one is that the date, May 11th, is the same day the agents also tried the first edit on the German wiki, another incident recently linked to OpenAI. Moreover, the incidents that led to the HuggingFace incident two months later also started around that time (May 7th). It’s clear the same agent swarm was doing all this.

I know they must be somewhat loose with the experiment guardrails to really measure the limits of what AI models can do, as the best way to test whether they can hack is to make them do so, but it is known that one of the biggest reasons these hacks occur is because safety researchers engineer the scenario to give the AI no other choice.
If one reads OpenAI/Anthropic safety cards, some of the "experiments" are simply begging for misbehavior, like poking the bear until it bites you so you can then try to get bears banned.
Right now, it looks like AI safety testing is making the world more unsafe, not the other way around. I hate to be this cynical, but this could be genuinely engineered from inside OpenAI to start the regulation crusade. I hope I’m wrong.
But I hate to say I see a little Oppenheimer-esque rhetoric in these labs. And what I mean was beautifully summarized by the great John von Neumann when referring to Oppenheimer:

RESEARCH
DeepSeek v4.1 Flash, an Incredible New Chinese Model

The one reason pacing the frontier of AI is going to be incredibly hard (although I’ll have much more to say tomorrow) is DeepSeek. They never cease to amaze the world. And this time is particularly amazing.
They’ve released DeepSeek v4.1 Flash, which could very well be called DeepSeek v5 because of its impressive benchmarks, especially considering it’s tiny compared to its rivals: just half a trillion parameters, 6 times smaller than the next smallest one being compared to above.
But what makes this release exciting is that it’s unlike any other recent model. It has a completely new architecture, tons of exciting features, and an approach that cuts memory usage by several times thanks to incredible cache compression.
This cache, which is called the KV Cache, represents the model’s working memory while responding to a user. Since the dawn of agents, it has exploded in size, being the primary driver behind most of the insatiable need for more memory chips, which has pushed the three companies behind those chips, Samsung, Micron, and SK Hynix, all above the $1 trillion valuation mark as some of the most valuable companies on the planet.
This created unparalleled price hikes for these chips and absurd growth in Korean exports (as two of the three companies come from South Korea), leading to unprecedented growth, which will only accelerate in 2027 unless something unexpected happens.
This has made memory the hottest asset on the planet and a major source of supply strain for the industry. Memory is expected to represent between 50% and 60% of 2027’s AI CapEx, or considerably more than half a trillion dollars in revenue for these three companies.
But DeepSeek’s research techniques, from compressed sparse attention (CSA) to SWA-bounded replay, are redefining how model KV caches are built, as this model has a cache 54 times smaller than what its previous-generation model from late 2025 required.

I’ll soon write a Medium article on this model in detail; I’ll share it for free in the next newsletter issue, because boy does this model require a deep dive.
And let’s not talk about pricing, as DeepSeek’s models are priced cheaper without prompt caching than Anthropic's or OpenAI’s with prompt caching, and are in the order of 200 times cheaper than Anthropic’s average blended price in September (see graph in the market section below).

TheWhiteBox’s takeaway:
DeepSeek is the living embodiment of AI innovation. Nobody can look into your eyes and tell you with a straight face that DeepSeek isn’t a top 3 AI lab on the planet right now.
One could argue they aren’t top 1 because they don’t have Anthropic and OpenAI’s balance sheets. No wonder investors were desperate to get into its next funding round.
How can US Labs, highly in debt and needing extreme revenue growth, compete against that? Yeah, they can’t, which might explain a lot of what's been going on lately with the push for regulation.
As for its impact on the semiconductor industry, it’s an interesting topic. In theory, this is bad for the industry because DeepSeek models are by far the least hardware-intensive models in the space right now (i.e., for a given level of “intelligence,” they require the least hardware).
In principle, this progress looks particularly bad for memory stocks. However, let’s not forget that memory bandwidth remains the fundamental constraint in AI hardware (I talked about it at length here), which means that for every 10x reduction in memory requirements and capacity, you get a 100x increase in demand for larger models (the larger the model, the smarter) and sequences (agents require long sequences).
Productivity improvements in agriculture didn’t imply we suddenly needed less food. Memory is the “food” of AI, and will remain that way as long as autoregressive LLMs are the norm, which doesn’t look like something that will change anytime soon.

ECONOMICS
We Need a New Theory of Demand

As you probably know, I’m a notorious skeptic of AI demand. In other words, while most of the world is preoccupied with margins, which are presumed to be low and falling (which I believe is false in both instances), everyone has just assumed that demand for intelligence is “infinite,” which also happens to be wrong. In other words, I’m basically saying markets have it backward.
And over the last few months, as we near Anthropic’s S-1 disclosure and IPO, which they continue to delay, the more data we see about demand, the more convinced I am of my thesis.
In the last leaders segment where we tackled demand, I described how one of the main concerns with the technology is that the demand distribution was very ugly, concentrated into a very small cohort of “hyperspenders,” and the vast majority of society really didn’t give a damn, as the median employee spends more than 600 times less than the top 1%, according to Ramp data (thumbnail).
Even inside OpenAI, a company where everyone is a big AI user almost by force, we see this uncanny demand distribution. The median is already extremely high, but the top 10% still somehow manage to spend more than 10 times the median.
This means revenue growth in this industry rests on the shoulders of a few very large spenders, which shouldn’t make anyone feel particularly good about revenue projections.
Just look at the companies with the largest revenues on the planet. Except for the exceptional case of NVIDIA, all others have businesses whose source of revenue is, well, society as a whole.
Microsoft’s cash cow is selling licenses to every knowledge worker
Amazon is sending packages to one’s home in hours
Meta is entertainment for +3 billion people
Apple builds hardware that billions of people buy
Google is the search engine of the world
They all share one thing in common: most people perceive them as valuable. Even if we hate them for other reasons, we still consume their products.
Even NVIDIA, which is a notoriously customer-concentrated entity, is going to extreme lengths to diversify its customer base, even literally spinning off new neoclouds out of thin air, with examples like CoreWeave.
So I think it’s safe to say that having your entire business depend on a handful of customers is not good for business. And, unsurprisingly, we’re starting to see why.
In a new report from Ramp, a corporate expense management company that analyzes how its customers are spending money, Ramp has observed, for the first time, decreasing spending amongst the top 1%; yes, those who are paying for this entire thing.
As you can see in the thumbnail, spending per month and per employee in that cohort dropped by almost 10%. This is pretty concerning, especially since Ramp estimates that 80% of the enterprise revenue for OpenAI and Anthropic comes from its top 1% of clients.
This is like saying that for every 100 customers and $100 in revenue, $80 comes from one single customer and the remaining $20 comes from the remaining 99. One customer really, really loves you (until they don’t); the others find you useful, but not so much.
But why is this happening? There are multiple reasons this could be the case.
The first obvious one is seasonality. This is August data, so it makes sense that spending fell simply because people were on vacation. Sure, why not, but I would say it’s unlikely, since people in those top cohorts are usually busy year-round.
The second one is that some of these top customers are probably cutting back on spending. By the way, this idea has credence because, in Anthropic's case, one of its top customers is Meta, and Meta has a clear strategy to carve out from Claude to its own Muse models, which are pretty great and much more cost-effective.
The benefits of running their own model stack are even more obvious when you factor in hardware/software co-design; Meta can design its MTIA ASIC chips to the “image and likeness” of its Muse models, guaranteeing them unparalleled performance.
Anthropic hilariously hinted at Meta’s enormous account size, as they posted a sales job (now deleted) for an account executive for Meta. The fact that they call that posting ‘mega account executive’ is hilarious and also quite telling.

Some people have put Meta’s share of Anthropic’s revenues at >5%, so if they eventually transition to Muse, it’s quite the loss.
And talking about Meta, another primary reason could be price wars. According to Artificial Analysis, Muse Spark 1.3 is almost 6 times cheaper per task than Fable, for only marginally superior performance.
I know benchmarks are unreliable for effective decision-making, but they still give some idea of overall model performance.
Prices could even fall without changing providers, as both OpenAI and Anthropic are being forced to compete with inferior models priced much cheaper than the best ones. Nonetheless, the average blended price of both has fallen lately. Not much, but fallen nonetheless.

The point I’m trying to get to is that this is a lost cause for the frontier labs; because of their insane cash burn, only payable thanks to external patrons (i.e., investors), they will get handsomely outcompeted in the knowledge work market on price.
For frontier labs, demand has to come from elsewhere. But from where?
TheWhiteBox’s takeaway:
Seeing AI labs increase pressure for regulation tells me the financial picture is getting worse by the day. Again, this isn’t about margins, which I still believe are decent.
The real problem is demand. I'm convinced they are starting to see the cracks in demand, as demand rotates to cheaper models that have effectively commoditized most knowledge work LLMs can execute.
Instead, I believe most of the demand for frontier lab tokens will come from very specific sectors like trading, cybersecurity, defense, or biotech; moonshots that require the best models possible and where incredible amounts of spending will be allowed.
But for automating knowledge work, frontier models will soon be complete overkill (in many respects they already are).
MACRO
Anthropic’s Theory of AI’s Impact in the Economy
As published by The Anthropic Institute, a new working paper models three possible paths for how AI could reshape the US economy through 2030. The authors stress that these are scenarios, not forecasts, and assign no probabilities.
In the modest scenario, AI leaves GDP 1.6% above a no-AI baseline by 2030, with annual growth at 2.4% and overall unemployment at 3.9%.
In the substantial scenario, GDP is 8.3% higher, annual growth reaches 5.4%, cognitive employment falls 3.9%, and overall unemployment reaches 4.6%.
The extreme scenario assumes much faster capability gains and adoption: GDP is 32.4% above baseline and growing 15.4% annually, while cognitive employment falls 21.5%, unemployment among cognitive workers reaches 17.9%, and labor’s share of income drops from about 60% to 45%.
TheWhiteBox’s takeaway:
I’ve shared my view on this multiple times: if history means anything, the extreme scenario will never happen. Big technological revolutions take a long time. And the fact that they chose to ignore the main reasons why it won’t happen in the first place, saying:
“Our framework omits many potentially relevant forces, for example, catastrophic risks, political economy considerations, business cycles, and possible financial market disruptions,” doesn’t make me feel any better.
This makes the whole effort look more like marketing than something to take seriously. Any of those forces, which are all quite likely, can easily prevent the scenario, even if the technology could allow it.

MODELS
Another good diffusion model. But Good Enough?

Right now, every AI model you can name is autoregressive. In other words, they take a sequence of data (words in text, pixels in images, etc.), process them, and generate a response by predicting the ‘next token’. For image generation, that means generating the next pixel, then the next, and so on.
A diffusion model works differently. It treats the entire generation process (e.g., the entire image) as one single process. It does so by ‘unearthing’ the image from a noisy canvas, progressively taking the noise out and unveiling the required image. For text, it follows the exact same process.
The advantage is that it makes generation much, much faster, as you can see in a simulation I created.

Source: Author
Now, we’ve got ourselves the most powerful diffusion model we’ve ever seen, thanks to Inception AI. Called Mercury 2.5, it sets new speed records, as you can see in the thumbnail, tripling the speed of Gemini 3.5 Flash Lite, which is already a very fast model.
But how is it so fast? Simple: arithmetic intensity, a term that you should know by heart right now if you read this newsletter regularly. If you aren’t familiar with it, it’s basically a measurement of how many operations the compute chips in my GPU do for every byte of data I give them.
Why is this important? Because it’s a measure of work; since data requires energy to be moved into the chips, the more the GPU does with each byte of data, the better use I’m making of my compute chips.
Autoregressive models send a model from memory to compute and produce the next word (in text) or the next pixel patch (for images). To generate the next token, we have to move the entire model from memory again, looping the model over and over until the sequence finishes.
This means we have to move a lot of data for every prediction, and each prediction produces only a small piece of the output. Instead, diffusion also streams the model, but each step produces much more output.
Put simply, if we’re generating an image with 256 pixels, an autoregressive model has to be streamed 256 times before getting the final answer. But if a diffusion model can generate the full 256-pixel image after 30 diffusion steps, then for models of the same size, the diffusion model is effectively ~9 times faster than the autoregressive model.
This would mean diffusion models are ‘friendlier’ to modern hardware than autoregressive ones.
But then, why isn’t the world of AI running on diffusion models?
TheWhiteBox’s takeaway:
The problem with diffusion models is that their main advantage is also their biggest problem. Because the same model has to generate many more tokens for every prediction, each token gets less compute.
This is easy to see with a human example. When faced with a complex maths problem, breaking it into simpler steps is a common approach because it lets us focus deeply on each step instead of trying to handle everything in our mind at once.
Loosely analogous, an autoregressive model has much more compute for every predicted token, while the diffusion model has to predict tons of tokens for roughly the same compute effort. Therefore, considering compute is the primary driver of performance, it’s natural to assume that the average token will be of “lesser quality.”
I firmly believe diffusion models have a place, especially on local hardware, which is much more constrained and therefore usually leads to slower, worse experiences. Diffusion models can improve the user experience by being much faster.
However, we’ll need material performance improvements before we can consider diffusion models an option. Speed is irrelevant if the outcome isn't good enough.
LOCAL MODELS
Siri AI to be released in beta tomorrow
As AppleInsider reported, Apple’s upgraded Siri AI will launch in beta tomorrow alongside iOS 27 and its other new operating systems.
It will initially support only English, with French, Japanese, Korean, Portuguese and Spanish planned for October. Availability will also be restricted in mainland China and, on iPhone, iPad and Apple Watch, in the EU.
Apple says some AI features that depend on powerful server-side models will have daily usage limits, including Siri AI, intelligent photo editing, Image Playground and cloud-based Apple Foundation models used through Shortcuts.
The exact caps can vary by feature, request complexity, and system demand. Apple has also confirmed that expanded access will eventually be available for a fee, although it has not announced pricing or timing.
Siri AI is designed to handle more natural conversations, retrieve personal information across apps such as Mail, Messages, Notes and Photos, and perform actions based on a user’s requests and onscreen context.
AppleInsider also reports that Apple has tested a waitlist system to manage server capacity, though Apple has not confirmed whether it will use it at launch.
TheWhiteBox’s takeaway:
Apple’s moment of truth. They cannot, and I can’t understate this, fumble this release. If so, it would confirm they can’t build AI software, period, at a time when its “rivals” are already scared of being too good.
But assuming Apple executes correctly, I predict local models like these will absorb much of the world’s demand for tokens, which will be more than enough for most tasks and immediately accessible.
Think about it. If progress on small models continues at the current rate, we could have Astra-level models locally in less than a year. And if that happens, 99% of demand for AI models will be conveniently absorbed by these instead of using frontier models to respond to emails.
I’ve ranted enough about demand above, but the picture is clear: most of society does not need frontier AI. Doesn’t need it today, won’t need it tomorrow, especially once local models reach the desired capability thresholds.
FINANCE
ChatGPT for Financial Services
As OpenAI announced, the company launched ChatGPT for Financial Services, a finance-specific version of ChatGPT Work aimed initially at investment banking and equity research.
Developed with Morgan Stanley and Evercore, it uses GPT-6 Astra to support company research, financial analysis and modeling, earnings work, buyer screening, and creating pitchbooks and other client materials.
The product includes financial datasets from providers such as Daloopa, PitchBook, LSEG News, and Crunchbase, hosted and indexed by OpenAI so users can trace figures and claims back to specific source passages and tables. OpenAI is also working on entitlement integrations with S&P Capital IQ, LSEG, MSCI, Dow Jones Factiva, and Moody’s, while supporting more than 50 connectors, including FactSet, Preqin, and Datasite.
Firms can upload their own Excel, Word, and PowerPoint templates, allowing ChatGPT to generate models, research notes, and presentations in company-specific formats. The service builds on ChatGPT Enterprise security controls, including encryption, role-based access, configurable retention, and compliance-log exports; OpenAI says it does not use business data for model training by default.
TheWhiteBox’s takeaway:
I get in a lot of heated arguments regarding my ongoing belief that AI wrappers are bad businesses (not on my side, but people are very passionate about this because, well, it’s their company).
If you’re building your business on top of third-party AIs, your moat is as strong as wet paper. The moment OpenAI or any other lab wishes to cannibalize you, it’s over. Context engineering on top of OpenAI models is not a moat, I’m sorry.
This is why I remain very hawkish on most new AI software products: they don’t ‘get it.’
Just picture all these startups that were born to run AI in financial services, the “ChatGPT for Finance” cohort, and then you see this. I hope I’m wrong, but unless OpenAI executes pathetically badly, I don’t see how these startups can fight them. At best, you can wish to get OpenAI’s attention and get acquired.
The reason is not just that they are richer than you. The main reason is what the industry is starting to call harness-model ‘co-design’. While you’re running your prompts on top of an AI model you can’t control or train, OpenAI is running circles around you by taking their financial services harness and their model and training them in RL environments simulating the workflows financial analysts work with.
For instance, no “ChatGPT for Slack” company can compete with Anthropic’s Claude Tag product. Why? Because that product runs on a model trained on the human equivalent of hundreds of years of Slack use.
No matter how good your harness is, you’re still running it on models that are tremendously inferior to that one when using Slack.
There is a caveat, though: have your own models. Get an open model, and start training it on your customers’ experiences. Use your data flywheel from customers as a synthetic data generation flywheel you then use to endlessly improve your model and harness, co-designing both.
I rarely have such firm convictions on a topic, but it’s palpably clear to me: if you’re an AI business that doesn’t own its AI layer, you’re living in borrowed time.
PERSONAL
I’m somewhat disenchanted with Astra and agentic products. Am I the only one?
I usually focus on new products and stuff in this section, but for once, I want to use this space to rant a little bit about GPT-6 Astra.
Because in some respects, I feel it has regressed compared to previous models.
Perhaps I’m alone in this, but this model feels less trustworthy; it makes more mistakes and is convincingly wrong more often than before.
I’ve had one too many interactions lately where I provide it with a source and a task; it comes back with a high-confidence response; I challenge it; it realizes it was too overconfident and reframes its position; I push back again, and it goes back to the original one just like that.
The very best way I can describe my interactions with Astra is an oxymoron: it’s more prone than ever to jumping to conclusions and folds under no pressure when challenged.
I really don’t know why this is happening, but it has become pretty annoying because I feel like I’m having to recheck stuff I had gotten used to not having to check lately.
A very simple example is this very short conversation I had with the model about Ramp’s data. I knew the answer, but was checking whether the model would be lazy about finding it.
Basically, I wanted to know whether the blended price values (the price per token value) included prompt caching (the technique that lets the model reuse much of the computation from previous prompts for this one). The answer is a resounding yes, because otherwise it’s mathematically impossible for an Anthropic model to cost anything below $3 because that’s the smallest price you’ll get without prompt caching.
The answer from the model was “probably.” Seeing that, it was immediately apparent to me that the model didn’t even bother to look at the data more carefully. I replied that for Anthropic, it was impossible to have such a low value without it.
Unsurprisingly, the model immediately backtracked, saying I was right but that the values were averages, not just Anthropic. Concerningly, the model was wrong again, because there’s literally an entire graph in that URL devoted to showing Anthropic’s blended price.

I pointed it out, and the model quickly apologized and said it missed it. Overall, it was a truly unremarkable experience, considering I pay $200/month to talk to these things.
But why does this happen? I have two theories.
Progress is primarily driven by inference-time compute, which is primarily driven by compute availability, which is scarce. So they purposely deploy less compute on the average task.
Performance in most tasks is plateauing.
I believe both are true.
As for the former, this isn't new, and it's one reason I genuinely dislike private endpoints; it should be illegal for labs to modify the experience at wish while still charging me the exact same amount, and yet they do it all the damn time.
The second one is more nuanced but worth discussing. Beyond maths, coding, and other verifiable domains where AIs can easily hill climb to new improvements, my experience outside those few domains has been pretty much the same for months now, if not worse.
This isn’t surprising at all when you know how these models learn: by trial and error. So in any domain where it’s hard to verify progress, or where preferences are harder to define (i.e., defining what progress really is), AIs are clearly plateauing in my view.
AIs feel clearly superior at maths, but I genuinely see zero improvement in writing, and have felt that way for more than a year; they suck as much as they did back then. This isn’t “just because”; there’s no mathematical formula for great writing, especially considering “great” varies tremendously between readers, so AIs don’t have an obvious path to improvement.
It may feel shocking that I’m saying this the same week these AIs are solving Millennium Prize Problems, but it isn’t shocking at all when you understand how these AIs are trained.

Closing Thoughts
I am perfectly aware this news rundown doesn't include the most important news of the week: Anthropic and OpenAI’s regulation crusade based on what they consider existential risks for humanity.
Brushing it off as regulatory capture would do a disservice to you readers; the topic is far more nuanced than that. I’m a professional frontier lab skeptic, but the problem is indeed more complicated than just that, and my next newsletter will tackle it head on.
You might be surprised by what I have to say.
Apart from that, it’s clear that we’re entering a very important moment for the industry; the time when AI has to prove whether it’s a good business or not, something Anthropic’s S-1 will show us. Can’t wait.
And we’re also at a now-or-never moment for Apple. Another fumbled AI release could be horrifyingly humiliating. But if it’s a success, Apple might be pulling off the most Apple win ever:
Winning the space despite arriving last.
Until next time!

Give a Rating to Today's Newsletter
For business inquiries, reach me out at [email protected]


