Artificially generatedWaiting for Grok Godot
There is an old book called Waiting for Godot, and right now I am waiting for the next version of Grok. Very curious what happens there.
I believe in the SpaceX AI team, because it is the only team that has a plan for scalable energy generation and for scalable data-center build-out. Anyone can build a data center, then the energy is missing. Others maybe have energy and are not as good at building data centers. We have to keep accounting for the five layers of the AI market, the five-layer cake from Jensen Huang. This piece is my personal transition phase: from Replit to Claude Code and OpenAI Codex, to SpaceX AI, to Grok, to Grok Build, Grok Bot on the side, eventually to neural nets and AI agent systems that SpaceX AI will put on the table.
Godot does not come. Grok comes, and I wait like Beckett anyway.
Samuel Beckett, 1953: two guys wait for someone who never shows. The joke of the play is that the waiting is the action. Waiting for Grok Godot is the same feeling with a different object. I am waiting for the next version of Grok. Very curious what happens there. This time the waiting is not idle. It is a prioritization. I have dropped habits that were still tools last week, and I am not waiting for a messiah. I am waiting for the next layer of compute from a team that takes the bottom of the cake seriously.
Energy at the bottom. Then chips. Then plugged together in the data center. Then neural nets. Hot software on top.
It is deeply tied to the whole of civilization, to the whole civilization stack. And at the bottom is energy. Who can generate energy? On top of that come the chips, so they can even be produced. Who can make chips? Who can produce chips? Then it gets plugged together. The infrastructure, meaning the data centers: energy and chips get packed into the data center. But there is more to it. How is it connected? How is it wired? How does the plugging-together work? That is not as trivial as you might think. Plugging together 100,000 NVIDIA cards and running them clean is not trivial.
Jensen broke it down again. Layer four is the models, meaning the trained neural nets. Layer five is then the hot software you build on top, the applications, everything that sits up there. That is the last layer. Public debate often sticks in layer four. Jensen wants you to see the oven, not just the cherry. His stack runs: energy, chips, infrastructure, models, applications. Who the Heaven is Jensen Huang, NVIDIA blog 10 March 2026. Here we draw the same stack again, not as a yellow cake, as an outline.
White on black. Wide at the bottom, narrow at the top. One stroke, no cornucopia. Source of the layers: Jensen Huang, not us.
Layer 1 energy, 2 chips, 3 infrastructure, 4 neural nets, 5 applications. The yellow cake sits in the Jensen piece. Here only a pixel edge, so the stack lands in the head a second time.
I actually find the word models a bit bad. For me it is much clearer when it is a neural net, because then I always know what it is.
There is a functional name for it. Why not use that? Models is completely confusing. Once I have plugged everything together, energy plus chips plugged together in the data center, then I can train my neural nets, aka my large language models, my word models, my XYZ models, or just my neural nets, whatever they do. Then I can build hot software on top.
More or less anyone can build models, if you have enough energy, chips and infrastructure. Only then can you build them. Layer four is not magic. Layer four is what is left when the ground holds.
What do you need software for at all? To help humans make meaning. To produce meaning.
That software can of course do all kinds of things. It can steer cars, robots and rockets, or just be software. Software is meaning-production tools that help humans turn nonsense into meaning, sometimes the other way around too, but mostly they are there for sorting. Then humans sit in front of them with their neural nets, look at the patterns and try to produce meaning from them that yields their next sensible action, for themselves or for their organization.
Building applications on top of that is what I am doing here right now. I am playing around on layer 5 up there, that is where most of us are running around. Trying to build something with it. Layer 5 is loud. Layer 1 is quiet and decides whether layer 5 has power at all.
We do not have an AI alignment problem. We have a human alignment problem.
When we talk about AI agents, it is also important to talk about human agents. We constantly think humans are superior, but humans are extremely unreliable agents. Humans are not capable of living with each other honestly, friendly and peacefully. We are working on the civilization systems to get there, see Chinamerica, but we are not there yet.
This piece is Waiting for Grok Godot, not the solution of the tribe. The tribe comes later, when the lower layers are no longer the excuse.
None of the other companies has this scaling unlock.
SpaceX AI will build AI production satellites. They solve layer one with the space satellites, that is where the energy comes in. They buy the chips from Jensen. Elon and Jensen get along pretty well, I think. And then the infrastructure. There they have the advantage: they are the only ones who can build that in space and thereby tap the energy of the sun. Meaning the compute gets done in space. There the energy of the sun is available, and SpaceX is currently the only one that can build the infrastructure for it. Not quite true. China can also play in that direction and will accelerate further that way. Brings us back to Chinamerica.
Energy yes, chips not yet. So it is not true that they own all three lower layers. But they bring them together. Models, neural nets, more or less anyone can build once the ground is there. That is why SpaceX AI is deservedly so valuable, and the more people trust that the space-AI infrastructure system works, the value will go through the roof again, hard. That will be cool.
We all long for habit, but habit is currently slowing you down.
I first had to detach from my habits again. That is happening now. It is a big skill to throw your habits overboard every week or every day and not long for that habit again. That goes a bit against our intuition and evolution. Because habits used to help you save energy and survive. That is why life feels so exhausting right now, for so many people, or for all of us. Because we long for habit, we strive toward habit, toward repeating behavior patterns that lead to success, that feel safe. And now we are moving so fast that last week’s habit is already outdated again.
And that can also lead to a kind of AI depression, so people think: oh, I learned something new, now I built something new. Oh, now I am not king after all, I am not the best and coolest just because I built that. Next week it is already worthless again. Then people somehow make a plan and think: oh, I will build an animation tool for video-game characters, and then others will use my animation tool. No, they will not. Because next week nobody needs it anymore.
Meaning: if anything, you currently build the tools for yourself, or for something customers of yours need concretely right now. But long-term plans are almost idle. There are of course things that will not go away. Meaning: you cannot make exact plans. You can only run in directions. Can say: this direction fascinates me and makes me curious. I want to do something in this direction. Then that can bring something, or then that can produce satisfaction, because you know you are running in the right direction. You will not automatically produce the finished product that then never changes. You will only keep running in your direction with ever new tools. That is how I would put it.
And once you have understood that: that is actually very beautiful. That is actually a new freedom. But big freedom is also extremely unsettling and habit-destroying, and that creates unrest overall in the individual people and thereby in the whole group and in the whole tribe.
That is why you have to prioritize, of course, and I currently prioritize SpaceX AI and the tools from SpaceX AI for the reasons just derived.
So I cancelled my Anthropic Claude Code subscription. And my OpenAI subscription is heavily reduced. The frontier models are roughly neck and neck, but my hypothesis is that SpaceX AI and the Grok products will develop extremely much faster than the others. Right now I have no disadvantage from this stack. On the contrary. I have an advantage, because fewer people are using it yet. There is a lot of compute, the thing is not slower, it is faster. Anthropic is somehow constantly sluggish. Claude Code from Anthropic also uses a lot more RAM on a computer, actually. And I really have to wrestle with OpenAI.
Grok has fewer guardrails, meaning it forbids me fewer things. Other AIs forbid me partly stupid topics. A colleague had this: she had asked how cats taste, but wanted to find out how the taste senses of cats work. And then her AI just refused the answer and said: no, I will not talk about how cats taste. Because it somehow wanted to say: you do not eat cats. But if you ask Grok, then it first tells the history of wherever cats were eaten, in which cultures, how they taste, and then also: yes, how does the taste system of cats work. There I have high trust in Elon’s nerd factor, which simply says: AI has to be maximally truth-seeking. Yes, and that is how I see it for myself too.
And that is why this is currently my AI of trust, which I will then use to build my own intelligence stack in the sense of Y Combinator and Gary Tan. It is often called Sovereign AI. Sovereign AI, but that sounds somehow weird. Just AI that runs at me myself.
Not everything can run at me, because I do not have that much energy.
There we have the problem. I do not have the energy and the chips. It would be stupid right now and also not possible for everyone to have their own data center. That is why people have big data centers together, and there is shared use. Kind of a hacker WG. That brings it back to Chinamerica.
Good capitalism is the same as good communism. And what we have not had so far: we had neither good communism nor good capitalism. And grasping that our civilization is still deep in barbarism, and that we are only just climbing out, and that this is the true semantic sorting work of humanity: from the unknown, from infinity, to build stable patterns that help us live in peace and joy. Semantic Mining was the name for that sorting at the data dirt. Here it is at the tribe.
Godot in the play does not come. Grok comes in versions. I do not wait by standing still. I wait by playing around on layer 5, with the stack I trust, while someone below builds satellites for the sun.
| Claim | Basket | Where from |
|---|---|---|
| Five layers: energy, chips, infrastructure, models/nets, applications | Fact as Jensen’s model | NVIDIA blog 10 March 2026, Dwarkesh, our Jensen piece. |
| SpaceX AI is the only team with a plan for scalable energy and scalable data centers | Thesis | This text, John, 9 September 2026. China as counterexample sits in the same piece. |
| SpaceX solves layer 1 in orbit, buys chips from Jensen, builds infrastructure in space | Reading, with self-correction | John in the dictation: energy yes, chips not yet, bringing together yes. No independent satellite audit here. |
| China can accelerate in the same direction | Thesis | This text plus Chinamerica. |
| We do not have an AI alignment problem, we have a human alignment problem | Thesis / stance | John, this text. |
| Anthropic cancelled, OpenAI reduced, Grok prioritized | Fact as self-report | John, 9 September 2026. No account statements checked. |
| Grok is maximally truth-seeking | Quote / brand promise | Elon Musk, taken over by John. The cat anecdote is an episode, not a lab report. |
| Waiting for Godot, Beckett 1953 | Fact | Standard literature. Title analogy is wordplay, not play analysis. |
How much of what you are reading was already in the dictation, and what the sorting machine added.
Semantic Scorecard
John’s original phrasing: 72 percent. The sentences on the cake, on neural nets instead of models, on meaning-production tools, on human alignment, on the SpaceX correction (chips not yet), on habit, on AI depression, on the animation tool, on running in directions, on the cancellation, on guardrails and cats, on the hacker WG, on good capitalism equals good communism, on barbarism and on the sorting work stand almost verbatim.
Glue and cleanup: 28 percent. Beckett explained in one sentence, without retelling the play. Typos out (Claude Code, Grok instead of Grog). Layers put in order after layer four and five were swapped in the dictation and then corrected. The lookup in our own magazine stays outside. Links to Jensen, Chinamerica, Semantic Mining. The outline animation. Fact table. This scorecard.
Semantic work of the sorting: nine sections instead of one stream. Godot as a form of waiting, not a literature report. Cake as civilization stack, not as an NVIDIA ad image. SpaceX as the one that brings the lower layers together, with the correction in the same breath. Habit as an evolutionary leftover that now brakes. Layer 5 as the place where most people run around, while the unlock sits below. Hacker WG as a form of shared compute, not as a Sovereign-AI fairy tale.
The instrument behind this card: Semantic Scorecard.
- Who the Heaven is Jensen HuangThe 5-layer cake, Dwarkesh, NVIDIA blog. The stack this text draws again.https://mcgrinsey.com/magazin/who-the-heaven-is-jensen-huang/
- NVIDIA Blog: AI Is a 5-Layer Cake10 March 2026. Energy, chips, infrastructure, models, applications.https://blogs.nvidia.com/blog/ai-5-layer-cake/
- ChinamericaUS-China, the same layers, China can play in orbit.https://mcgrinsey.com/magazin/chinamerica-us-china-energy-chips/
- Semantic MiningMeaning from the infinity of the unknown. The sorting work the tribe needs at the end of this piece.https://mcgrinsey.com/magazin/semantic-mining-meaning-from-unknown/
- Samuel Beckett: Waiting for Godot1953. Title giver. No play analysis, a form of waiting.https://en.wikipedia.org/wiki/Waiting_for_Godot
- Semantic ScorecardOrigin, operation, meaning-delta. The card under this piece, as its own instrument.https://mcgrinsey.com/magazin/semantic-scorecard-measurement/
Cut-off: 9 September 2026. McGrinsey writing skill, spoken-input rule. Semantic Scorecard is new and sits in the piece on purpose, not only internally. Method piece: Semantic Scorecard measurement.
