Ironically, I find myself with a grin on my face as I read the recent media reports about how the data processing demand behind AI is beyond the scope of financial sustainability.
For several years I have asserted, accurately, the business model for social media was never feasible because the data processing demand needed for the scale of simultaneous users was beyond the capabilities of the revenue side of the equation. I have been told by all the high-horse experts on the matter how wrong I am. However, each story they write about the prohibitive cost of AI proves I was not wrong.
CTH watches the tokenization and subscription fees for various AI model use with the same perspective CTH viewed over a decade of false claims within the financial market that told lies about social media viability and data processing costs.
Now, we watch the seemingly exponential growth of AI capabilities and associated costs with the same pragmatic perspective.
Robotic pool cleaners were introduced two generations ago. Did the pool cleaner business dry up? No, it expanded. Robotic vacuums broke into the popular household appliance market five years ago, you probably have one, did it eliminate maid services? No, still growing.
AI can now write its own code to generate outputs. Are software developers getting fired? No, demand for software designers and engineers is up 15% in the past year.
The mainframe approach, the one AI brain to run all systems, will never work – it is cost prohibitive (see first paragraph – wash, rinse, repeat). Deny this reality at your own investment risk. If needed, politely absorb the ridicule – for it matters not.
CTH predicts AI will become a localized and optimized sub-set for each sector of the economy, requiring each major organization and corporation to adopt specific cost/benefit data libraries and networks for use and functionality.
At scale, a thousand coders each working on Gemini, ChatGPT, Anthropic, Grok, etc. will become 100,000+ software designers working inside companies to create personalized, targeted, bespoke AI data systems and networks; each system specifically tailored to the industry or sector of business. The intranet of internets will happen again.
Creating and selling AI system networks and integration functions that are personally tailored to highly specific company functions, creates an entirely new sector of the technology industry that has not even begun yet. [There’s an investment opportunity there]
Will AI robots replace some repetitive human functions? Yes, the ice rink Zamboni will likely not have a steering wheel, just an emergency joystick. A reference for a comparative industrial scale Roomba vacuum, or the robotic pool cleaners. However, at scale the robotic industry is slower than human efficiency in almost all sectors that matter; the cost benefit analysis will limit growth. The maid service sector will not be impacted any more than the software developers (see chart above).
It is not an issue to fear some AI task efficiencies will grant more time available that will be filled with alternate task capabilities. Human productivity will increase in certain sectors of the economy, but humans will not lose work opportunities. Blue collar jobs will continue to expand as each of the hardware tools developed will need manufacturing, installation, maintenance and monitoring.
The further downstream the worker is from a repetitive function within the [XXXX] industry, the more irreplaceable they become; remember that.
As to the bigger picture of fully developed AI and the intersection of information and knowledge; yes, the automation of AI can present an issue. However, all AI concerns can be mitigated so long as multiple, alternative AI systems exist within the larger information realm.
As a nation we need dozens of different AI models each competing within the industry for the best AI product. As long as we have multiple AI systems, alternatives to the hive-mind, we do not need to fear the AI network as a source of information. If we don’t like the AI outputs, we can switch to an alternate AI provider.
If the subscription cost of the AI is too high, then as long as we have a competitive market where a lesser expensive, perhaps bespoke, AI option can exist, we should be okay. Let the free-and-fair market decide.
If AI outputs don’t offer empirical truth or real value to the end user, we should be fine as long as consumers have alternative options available. AI providers should be information providers in the same concept as cell phone providers. The key is to have multiple, competing AI systems available for industrial, business, professional and personal use.
On the upside of this information worry dynamic -in the pragmatic and optimistic perspective- we have the cost limiting nature of a massive singular AI information network.
A single AI central brain handling over 360 million users at once, all requiring identical responses that update with every tiny change in a multi-trillion datapoint-per-millisecond data stream, is far beyond the capacity of any computational AI system. The costs tied to such a setup are only now becoming clear, and AI business models are starting to fall apart in real time. This is a hard truth that isn’t going to change.
Within the AI business, those who can carefully write AI input instructions to achieve maximum value in AI output -industry by industry- will become increasingly more valuable. Those who can train AI to be cost effective -and provide materially beneficial outputs- within their granular sector of business, within each company, will become priceless to the organization. Wage rates will follow competency.
As noted by David Sacks in this segment highlighted below, the one key about AI to emphasize is the need for multiple competing models. If China (hive mind) has their model, and Europe (another hive mind) has their model, and the United States (entrepreneurial competitiveness) has multiple competitive models – we will win and simultaneously we will retain freedom.
What we don’t want is a singular AI model to win the support of the United States government and then end up with an AI regulatory system where they start defining terms of “safety” to eliminate information adverse to the interests of the government that regulates it. Both China and Europe will predictably do that.
” Both China and Europe will predictably do that.”
Pretty sure we will too.
“What we don’t want is a singular AI model to win the support of the United States government and then end up with an AI regulatory system where they start defining terms of “safety” to eliminate information adverse to the interests of the government that regulates it. Both China and Europe will predictably do that.”
Well…that is what the BigTech oligarchs who appear to be aligning with MAGA, intend for us too.
Three Rings for the Elven-kings under the sky,
Seven for the Dwarf-lords in their halls of stone,
Nine for Mortal Men, doomed to die,
One for the Dark Lord on his dark throne
In the Land of Mordor where the Shadows lie.
One Ring to rule them all, One Ring to find them,
One Ring to bring them all and in the darkness bind them.
In the Land of Mordor where the Shadows lie.
An excellent analogy for why a decent, moral people cannot allow the state to exist among them, as the Founders knew. It is illogical to argue that in order to rid ourselves of evil we must use that evil to bring about good.
The state must go.
The AI buildout/return of investment debate has been talked in many circles for the past 6 months. Many people have question if these large data centers are actually financially sustainable? But, then there are stories – I thinking about Nvidia recently said they paid more in processing/token fees in the last 6 months than there engineering software salaries. But, they are also dramatically shrinking their developments by half or more. There been a few articles talking about outrageous AI processing fees being charged to various companies in the last few weeks. So there’s that.
As a hardware/firmware engineer in a small company making a wireless network radio – think Wifi or cell data on steroids. – The AI has transformed what we can do.
My coworker, the main software developer – he does not write code anymore. The AI makes all changes to the code. He directs what to do. The abilities to do “what if” experiments has reduces to almost nothing. It can change the API interface between the hardware and software in minutes. Something that would take two engineers (software/hardware) at least a day a to accomplish. It has dramatically increased his productivity, by 200% or 300%. I more focused on the use- I am typically running down bugs, so it fits that better. It is little interesting to peruse the code base, because I can see where the AI is making changes because of the notes it puts around the changed code. It seems to be good at adding notes.
A few recent examples-
We have programmable hardware, many things like crypto algorithms are much faster/efficient in hardware. We desired a specific error correcting algorithm to included in our design. About a year ago we had priced some IP that could be included into our code base. Cost = $30K. At the time, we decided, at that cost we could spend a few months to try and do it ourselves. Coworker ask the AI to implement the algorithm according to x spec. If searched the web, found the spec, with check files (input x => should output y) used to verify functionality. Built a pure software app to verify the math and implementation. Then wrote the code for the hardware and a simulation environment verified it as a standalone piece. Then integrated it into the our code base. It took this about 6 hours to do all the above. $30K in 6 hours, using the $200 a month AI plan. We have not tested it yet, so the final conclusion is not there yet. What we estimated may take a few months – done in a few hours.
Our platform is a small, about 4lbs, mobile system running an embedded Linux OS. Not that different than a cell phone. Most people do not realize that android runs on the Linux Kernel. Using the AI, we have implemented a Machine Learning(ML) Model on the platform. The intended purpose was to identify and classify interference signals. We took some open-source
ML models, split it apart so that various pieces that and decide which got implemented in hardware and which stayed in the software. So we ended up with this hybrid ML implementation that is using hardware to accelerate the overall performance. It is functional, needs a lot of optimization and improvement in the training. We are not positive, but we may one of the first people to do this in a handheld mobile platform. Prior to this, we had zero experience with ML models. With the back and forth, developing the various functional pieces it took about two weeks. FYI – we had started to work with a university research team to do something similar. They were going to use are platform to do some “research” on. Well, we just did it. The lead professor was actually thrilled when we told we something working. As it was only the first part of a bigger plan.
We are only a few engineers. We are spending a lot more time operating as system engineers and letting the AI do the grunt coding and implementation work. It is not perfect, sometimes leads you down wrong paths. It does make errors. It does produce AI slop. For the people that can harness it, control it, check it, it makes them that much better and faster.
I do wonder about where this all goes. The above is just practical, engineering cycles. But, from a more general aspect – I am not sure – I wonder what the AI will look like in 5 years? 10 years?
I was able to understand only the last paragraph of this article. The video offered great insight
Companies that enthusiastically jumped on the AI bandwagon with both feet and reduced their human workforce have come to regret that decision.