Thank you as well for your excellent review. I really want to ask three questions: 1. are we back to compute? I thought it was only a few weeks ago DeepSeek showed us processor power is not everything (did I not get the memo :)? 2. Can we trust Musk? He has been consistently overselling both his cars, space adventure, social media space and more. He is a marketing man knowing what stories people want to hear. 3. Does the world want to trust their data and decision making to the owner of X and the leader of DOGE? Sales of Telsa are down up to 40% in some European markets January 24 compared to 25.
Unlocking the next generation of AI models—ones that will make today’s models look like clumsy, primitive tools—will require an order-of-magnitude increase in compute. Only a handful of well-resourced players can compete at this level because it demands vast capital expenditure and cutting-edge engineering.
This is why Elon Musk is saying XAI will build another cluster that’s 5x more powerful—and why Sam Altman is focused on securing long-term AI compute supply. (Stargate - $500bn.)
The significance of DeepSeek R1 is that it demonstrates a massive **efficiency** breakthrough. The DeepSeek team managed to achieve cutting-edge capabilities at a fraction of the cost. That’s a game-changer. But even if we continue making AI models more compute-efficient, we will still need compute.
Why?
1. There may be no scaling wall. If intelligence has no ceiling, then we will always need more compute to reach the next level. The only hard limit might be physics itself—how many data centers and superclusters can actually be built?
2. Second, compute won’t just be needed for training new models—it will power the entire AI ecosystem. The cheaper and more efficient models become, the more compute will be required to run AI across countless applications. This won’t just be software—it extends to embodied AI (robots) and real-world automation. Inference isn’t static—AI will function more like a self-iterating organism, constantly learning and adapting. That means it will need to be powered by 24/7 compute.
***
The players building AI infrastructure will own the most valuable commodity of all: compute itself. Right now, they’re using it to build better and better models—which may continue infinitely. But the same compute infrastructure will also be the resource everyone else will depend on as AI becomes cheaper, more ubiquitous, and embedded into everything.
What’s clear is that AI’s next leap will demand vast infrastructure and cutting-edge engineering -- and those who 'own' the supply chain for compute stand to gain A LOT.
1. Yes and no. The big players and the cutting edge on research are in an arms race, so spend and compute remain a big deal. However, investing to be first might mean nothing if everyone arrives at the finish line more or less together.
2. No, if you mean is he doing this all with the focus on improving your life. He's doing it for him, but he is a showman (spoiler: most people get weary of the show eventually).
3. No. See 2.
All of the above are fragile states. 1. is a big financial risk if whatever is achieved can be replicated more cheaply. DeepSeek shows that the costs do not add up unless these investments can be turned to an income stream. 2. Musk is not only burning through cash, but leaving a lot of damage on the way. Daily, he demonstrates his unsuitability to be trusted or lauded. 3. [As DeepSeek has shown] there is plenty of mileage in being a beta to the alpha AI-egos. While the alphas scrap and fight and perform their courtship dances in front of each other, the betas can focus on where the value is. The risk with having huge wealth is that the focus on cost is lost. Not so with poorer folks, who optimise with reduced resources often to far greater effect.
Thanks for the early morning review — it arrived here in Germany just before breakfast. One thing that “Herr Musk” has proven again and again is that he can assemble teams to solve extreme engineering challenges in record time. Of course, it doesn’t hurt to have semi-infinite deep financial pockets. So when will we see a really good LLM based agent called “Deep Finance”?
My jaw was dropping when he was describing the complexity of the engineering challenge. I think people underestimate just what an incredible (superhuman even) feat they pulled off. Deep Finance will be coming soon, I see that Perplexity is already trying to do this, and the Grok-3 Demo showed how you can use X to follow companies and track market trends. My view is that some kind of super accessible AI 'Bloomberg terminal' is coming soon.
I agree with you that all sorts of interesting specialized agents will appear this year: translator, programmer, finance specialist, stock trader, tutors of various types, consultants-in-a-box, and more. It will take a while longer until we will see legal advisors and medical advisors, due to the regulatory thickets. But they will come. For someone like me who started programming Fortran IV on punched cards in 1970 before my fist semester of math and physics in Germany, and who was an avid reader of pulp and higher quality sci-fi, the developments of the past 15 years are astounding, exciting, and clearly accelerating, much more so than the PC, the Internet, cloud computing, and mobile computing ever were. Being able to formulate a complex 1-page prompt and have the “OpenAI deep research” agent produce a high quality report of 30 pages was initially mind-bending until by now I’m already used to it. I have no idea what life will be like past the next 18 to 24 months.
- Its own data center with the largest number of colocated GPUs (200k) in the world
- Power generators and Tesla batteries to store and distribute power efficiently
- A custom-designed advanced water-cooling system to prevent overheating and ensure every component of the hardware stack operated in perfect synchronization
I have reservations about the potential capabilities of Grok 3 and am concerned that the significant investment in compute resources may not yield the desired outcomes. While I appreciate the overview provided, I question whether further amplifying Elon Musk's profile is beneficial at this time.
Furthermore, relying solely on Elon Musk's statements regarding the model's capabilities may be overly optimistic. Claims of Grok 2 being 'maximally truth-seeking' seem overstated based on my initial explorations through prompting. I was also able to elicit humorous and arguably undesirable outputs from Grok 2 by prompting it with Elon Musk's publicly available Twitter posts, raising questions about its intended behavior.
In my personal network, I haven't encountered individuals who actively use Grok as a primary or secondary LLM. Therefore, I am uncertain about Grok's primary target audience beyond the Twitter/X user base."
I understand that Elon Musk is a highly politicized figure, but I analyze AI and politics. If commenting on the most important developments in these areas is considered "amplifying Musk," so be it. Ignoring key players isn't analysis—it's willful omission.
Not trusting his statements on Grok’s performance is fair. But AI models aren't built in a vacuum and left untested. They are vetted on independent benchmarks—that’s the entire point. Anyone can claim they have the best model, but everyone gets to test it.
"I don’t know anyone who uses Grok" is not an argument. That tells us nothing about its technical significance—only that you don’t know anyone who uses it.
The bigger story here isn’t Grok-3 itself—it’s what this level of compute and engineering signals for the future of AI. As I lay it out:
- AI’s next breakthroughs won’t come from just better models.
- More compute and brilliant engineering are the real bottlenecks.
- What they’re building will make today’s frontier models look like rudimentary tools.
This all points to an order of magnitude increase in compute to unlock the next level of AI. That breakthrough may not even be based on transformers or LLMs, as Yann LeCun has suggested.
What’s clear is that AI’s next leap will demand vast infrastructure and cutting-edge engineering.
Seems to me like Elon Musk will be a central figure in that story - whatever you think of him.
appreciate the thoughtful response! i had to get google Gemini to make my initial thoughts on this a bit more a little less unhinged, haha.
yea, i agree that elon has really accelerated this race to build the biggest cluster possible for LLMs, with the first stage of the Memphis colossus project, which i believe resulted in grok-2.
regarding the safety work for grok, this data source to me indicates xAI is not interested in any kind of safety work for their models, which is what is triggering for me since they want to push ai capability without even acknowledging ai safety or alignment work.
willing to change my mind on xAI design process for their Grok models if you have any work highlighting that. for reference, i prefer anthropic level of safety testing and alignment work.
Thank you as well for your excellent review. I really want to ask three questions: 1. are we back to compute? I thought it was only a few weeks ago DeepSeek showed us processor power is not everything (did I not get the memo :)? 2. Can we trust Musk? He has been consistently overselling both his cars, space adventure, social media space and more. He is a marketing man knowing what stories people want to hear. 3. Does the world want to trust their data and decision making to the owner of X and the leader of DOGE? Sales of Telsa are down up to 40% in some European markets January 24 compared to 25.
I wouldn’t bet against compute
Here's why:
Unlocking the next generation of AI models—ones that will make today’s models look like clumsy, primitive tools—will require an order-of-magnitude increase in compute. Only a handful of well-resourced players can compete at this level because it demands vast capital expenditure and cutting-edge engineering.
This is why Elon Musk is saying XAI will build another cluster that’s 5x more powerful—and why Sam Altman is focused on securing long-term AI compute supply. (Stargate - $500bn.)
The significance of DeepSeek R1 is that it demonstrates a massive **efficiency** breakthrough. The DeepSeek team managed to achieve cutting-edge capabilities at a fraction of the cost. That’s a game-changer. But even if we continue making AI models more compute-efficient, we will still need compute.
Why?
1. There may be no scaling wall. If intelligence has no ceiling, then we will always need more compute to reach the next level. The only hard limit might be physics itself—how many data centers and superclusters can actually be built?
2. Second, compute won’t just be needed for training new models—it will power the entire AI ecosystem. The cheaper and more efficient models become, the more compute will be required to run AI across countless applications. This won’t just be software—it extends to embodied AI (robots) and real-world automation. Inference isn’t static—AI will function more like a self-iterating organism, constantly learning and adapting. That means it will need to be powered by 24/7 compute.
***
The players building AI infrastructure will own the most valuable commodity of all: compute itself. Right now, they’re using it to build better and better models—which may continue infinitely. But the same compute infrastructure will also be the resource everyone else will depend on as AI becomes cheaper, more ubiquitous, and embedded into everything.
What’s clear is that AI’s next leap will demand vast infrastructure and cutting-edge engineering -- and those who 'own' the supply chain for compute stand to gain A LOT.
I have some thoughts:
1. Yes and no. The big players and the cutting edge on research are in an arms race, so spend and compute remain a big deal. However, investing to be first might mean nothing if everyone arrives at the finish line more or less together.
2. No, if you mean is he doing this all with the focus on improving your life. He's doing it for him, but he is a showman (spoiler: most people get weary of the show eventually).
3. No. See 2.
All of the above are fragile states. 1. is a big financial risk if whatever is achieved can be replicated more cheaply. DeepSeek shows that the costs do not add up unless these investments can be turned to an income stream. 2. Musk is not only burning through cash, but leaving a lot of damage on the way. Daily, he demonstrates his unsuitability to be trusted or lauded. 3. [As DeepSeek has shown] there is plenty of mileage in being a beta to the alpha AI-egos. While the alphas scrap and fight and perform their courtship dances in front of each other, the betas can focus on where the value is. The risk with having huge wealth is that the focus on cost is lost. Not so with poorer folks, who optimise with reduced resources often to far greater effect.
Thanks for the early morning review — it arrived here in Germany just before breakfast. One thing that “Herr Musk” has proven again and again is that he can assemble teams to solve extreme engineering challenges in record time. Of course, it doesn’t hurt to have semi-infinite deep financial pockets. So when will we see a really good LLM based agent called “Deep Finance”?
My jaw was dropping when he was describing the complexity of the engineering challenge. I think people underestimate just what an incredible (superhuman even) feat they pulled off. Deep Finance will be coming soon, I see that Perplexity is already trying to do this, and the Grok-3 Demo showed how you can use X to follow companies and track market trends. My view is that some kind of super accessible AI 'Bloomberg terminal' is coming soon.
I agree with you that all sorts of interesting specialized agents will appear this year: translator, programmer, finance specialist, stock trader, tutors of various types, consultants-in-a-box, and more. It will take a while longer until we will see legal advisors and medical advisors, due to the regulatory thickets. But they will come. For someone like me who started programming Fortran IV on punched cards in 1970 before my fist semester of math and physics in Germany, and who was an avid reader of pulp and higher quality sci-fi, the developments of the past 15 years are astounding, exciting, and clearly accelerating, much more so than the PC, the Internet, cloud computing, and mobile computing ever were. Being able to formulate a complex 1-page prompt and have the “OpenAI deep research” agent produce a high quality report of 30 pages was initially mind-bending until by now I’m already used to it. I have no idea what life will be like past the next 18 to 24 months.
xAI built
- Its own data center with the largest number of colocated GPUs (200k) in the world
- Power generators and Tesla batteries to store and distribute power efficiently
- A custom-designed advanced water-cooling system to prevent overheating and ensure every component of the hardware stack operated in perfect synchronization
- The top AI model in the world today: Grok-3
Only Elon could pull this off.
I wonder if the ebullient AI narrative is still the same. Looking like it's all gone or going even further off the rails.
I have reservations about the potential capabilities of Grok 3 and am concerned that the significant investment in compute resources may not yield the desired outcomes. While I appreciate the overview provided, I question whether further amplifying Elon Musk's profile is beneficial at this time.
Furthermore, relying solely on Elon Musk's statements regarding the model's capabilities may be overly optimistic. Claims of Grok 2 being 'maximally truth-seeking' seem overstated based on my initial explorations through prompting. I was also able to elicit humorous and arguably undesirable outputs from Grok 2 by prompting it with Elon Musk's publicly available Twitter posts, raising questions about its intended behavior.
In my personal network, I haven't encountered individuals who actively use Grok as a primary or secondary LLM. Therefore, I am uncertain about Grok's primary target audience beyond the Twitter/X user base."
I understand that Elon Musk is a highly politicized figure, but I analyze AI and politics. If commenting on the most important developments in these areas is considered "amplifying Musk," so be it. Ignoring key players isn't analysis—it's willful omission.
Not trusting his statements on Grok’s performance is fair. But AI models aren't built in a vacuum and left untested. They are vetted on independent benchmarks—that’s the entire point. Anyone can claim they have the best model, but everyone gets to test it.
"I don’t know anyone who uses Grok" is not an argument. That tells us nothing about its technical significance—only that you don’t know anyone who uses it.
The bigger story here isn’t Grok-3 itself—it’s what this level of compute and engineering signals for the future of AI. As I lay it out:
- AI’s next breakthroughs won’t come from just better models.
- More compute and brilliant engineering are the real bottlenecks.
- What they’re building will make today’s frontier models look like rudimentary tools.
This all points to an order of magnitude increase in compute to unlock the next level of AI. That breakthrough may not even be based on transformers or LLMs, as Yann LeCun has suggested.
What’s clear is that AI’s next leap will demand vast infrastructure and cutting-edge engineering.
Seems to me like Elon Musk will be a central figure in that story - whatever you think of him.
appreciate the thoughtful response! i had to get google Gemini to make my initial thoughts on this a bit more a little less unhinged, haha.
yea, i agree that elon has really accelerated this race to build the biggest cluster possible for LLMs, with the first stage of the Memphis colossus project, which i believe resulted in grok-2.
regarding the safety work for grok, this data source to me indicates xAI is not interested in any kind of safety work for their models, which is what is triggering for me since they want to push ai capability without even acknowledging ai safety or alignment work.
https://futureoflife.org/document/fli-ai-safety-index-2024/
willing to change my mind on xAI design process for their Grok models if you have any work highlighting that. for reference, i prefer anthropic level of safety testing and alignment work.