One Million Digital Humans

Ooooh, he should not have said that…

Or that, or even that…

That’s all I could think as I was watching a podcast on Sunday morning of an xAI engineer being interviewed on the inner workings of xAI.

Source: xAI @ti_morse

I didn’t expect much when I started watching, but the discussion got very interesting quickly, as the xAI engineer shared a lot about the inner workings of xAI and some of the company’s operational and technical strategies.

It was the kind of information that should never be discussed outside of xAI – least of all on a very public podcast. Much of what was discussed was proprietary to xAI and part of what gives xAI a strategic advantage. All I could think as I was listening was “he’s in trouble now.”

My suspicions were right, five days – and roughly 4.7 million views on X – later, the engineer posted this:

Which makes the podcast that much more interesting. He didn’t “leave.” His employment was terminated for what he made public. For anyone interested, you can find the entire podcast here. I don’t know how much longer it will be up, as I’m sure that xAI would like it taken down.

As for what he shared, that’s where it gets extremely interesting…

One Million Macrohard Agents

From the discussion, it was clear that the xAI engineer was directly involved in xAI’s Macrohard efforts. Macrohard – announced last August and discussed in my issue, Is Microsoft Doomed? – is a tongue-in-cheek parody of Microsoft (Macrohard >> Microsoft). But in all seriousness, the goal of Macrohard is to be the very opposite of Microsoft.

Musk’s goal with Marcohard is to develop an entirely AI software company whereby an entire software business can be run by AI agents (agentic AI) – artificial intelligence with agency/self-governance and empowered to carry out human-given directions according to how it thinks it should, with minimal or no human intervention.

The AI agents will be fundamentally powered by xAI’s frontier AI model – Grok.

The ramifications, if successful, are deeply threatening for the software industry. Naturally, anything that Microsoft products can do today will be easily within the scope of Macrohard’s AI agents. And this will be true of nearly any software business in operation today.

xAI has gone so far as to paint the name Macrohard on the top of its Colossus data center in South Memphis as an indication of its commitment to succeed in this technological breakthrough.

The importance and strategic significance of what the Macrohard team is doing is precisely why the engineer got fired. This isn’t just some experimental research being conducted by xAI. This is a product that could potentially put Microsoft out of business and turn xAI into a trillion-dollar company. In other words, it is not something that any xAI employee talks about in public.

He revealed on the podcast xAI’s current initiative to develop one million “human emulators.” Specifically, they are developing a system for “pre-training one million digital humans.”

The idea is that by creating a training environment with a million human emulators, you can train these emulators on any and every task that a human currently undertakes digitally.

Things like data input, data analysis, moving a mouse/cursor to interact with the internet, reading, writing, filling out forms, coding, you name it.

The challenge, as the engineer discussed, is how to spin up one million digital humans efficiently. Is it one million GPUs? One GPU per human emulator? As it turns out, that is a very inefficient approach.

Five Minutes of Fame – A Lifetime of Regret

One of the more interesting ideas that was shared was using consumers’ Teslas, those that have hardware version 4, to run the human emulators.

This is an idea under consideration, whereby xAI would pay Tesla owners to “lease” the computational power of the Teslas. It would be like running the emulators in the cloud, with the cloud being Teslas all around the country.

The benefit is that it is efficient, cheap, and fast to do so, without putting any strain on Colossus, which is improving Grok as quickly as it can in the race to artificial general intelligence (AGI).

The engineer said explicitly that xAI is “pushing very hard on Macrohard” and that it is relying more and more on agents for programming. “Agents are a force multiplier.”

He stated that the AI agents are already “looking like it is 8 times faster than a human’s capability for work” and that the “initial assumption was 1.5 times faster.”

Let’s think about that. The brilliant team at xAI originally predicted they could make agentic AI systems that were about 1.5 times faster than humans, and now they have systems that are already eight times faster.

It will scale to even higher levels of productivity in the coming months.

He also mentioned that the AI agents are often capable of performing a task flawlessly without any specific pre-training. No training required, “they” can just figure a new task out. It’s not perfect…yet…but just imagine where Macrohard’s agents will be in six months, especially with xAI on the fast track toward achieving AGI within the same time frame.

The xAI engineer also shared a nuanced secret of how xAI was getting around local laws to use a large number of backup natural gas-powered generators to deliver more power to Colossus.

He shared that xAI was operating under an exception to the town’s regulations for the use of backup generators that was designed for carnivals, which needed a temporary boost of energy beyond what the grid was able to provide. Some clever xAI lawyer found the loophole to enable xAI to scale its Colossus data center so quickly with 80 backup generators.

Another fascinating disclosure was that xAI and Tesla found that they improved compute efficiency by going with small AI models, which led to unexpected efficiencies.

He specifically mentioned that when Tesla migrated to a smaller AI model with Tesla’s full self-driving AI (FSD), Tesla was able to iterate much more quickly and improve the performance of the model. This explicitly explains the radical jump in performance of FSD when Tesla released version 14 of FSD.

The podcast was so revealing.

I can’t believe that such a clearly intelligent person was so stupid to speak publicly about what was evidently inside information. And he lost his dream job as a result of his stupidity… something he’ll regret for the rest of his life.

What he shared has jaw-dropping implications. xAI is so much farther along than anyone realized (except me and my predictions about xAI). The information shared is proof of this.

Macrohard Is Coming for the Software Industry

xAI and Macrohard are going to hit the software industry like a Mack truck.

The cost of using xAI’s technology for a business or consumer will be reduced to simply the cost of compute, plus a margin for xAI. The pricing will result in something that is a fraction of what software companies license their technology today.

Don’t believe it?

SpaceX launch services are provided at a fraction of the cost of the competition (per kilogram to low Earth orbit). Musk could charge a whole lot more than he does and still have the best pricing in the market.

He’s playing the long game, though. He wants to accelerate the space economy, which can only be done using the most competitive pricing possible.

Lower prices and lower margins create a much larger market for SpaceX. Musk did the same with Tesla. And the same with The Boring Company.

Every industry he enters reinvents the underlying technology, offers products and services at the most competitive prices in the industry, and builds the company to scale as quickly as he can.

This will be precisely the same with Macrohard, only it will be faster than anything we’ve seen, because it is software, and adoption can happen far more quickly.

This is the way.

Human emulators to accelerate agentic AI systems.

Most still have no idea what’s coming.

Jeff