Category Archives: AI

Phil’s new HfS Services-as-Software FlyWheel Is Right On the Mark From a Customer-Centric Viewpoint

… but hides the full support required on the back-end!

This is important to point out for two reasons:

  • Gen-AI Hype-mongers will use this as another excuse to claim most white-collar functions will be entirely eliminated when, in fact, it strengthens the need for true back-office white-collar workers and real software engineers
  • Expert human support becomes more critical at each stage of the process (while bit pushers became less and less useful)

But let’s backup. In his most recent piece where he (re-)introduced the SaS Flywheel, Phil made one critical statement which is constantly overlooked by the industry: Stop treating FDE as optional: Your AI Flywheel will not spin without it.

As Phil astutely points out: the hard question nobody is answering is this: who actually wires AI into your live systems, governs it in production, and makes it keep working when the AI software vendors leave the room. The answer is, of course, your Forward Deployed Engineer (FDE) — and if your transformation strategy does not have it, you are building an AI theatre, not an AI operating model. (Which, FYI, is what most companies are building — and, as Stephen Klein astutely points out, putting on puppet shows. Great for entertainment, but not so great for getting anything done. Especially since they all overlook what AI can actually do.)

Now, a forward deployed engineer alone will not get you out of pilot purgatory, but it is an essential condition — just like you can’t climb out of a deep wide hole with smooth 90° vertical surfaces on all sides without a rope or a ladder, you can’t fly your way out of a pilot without a working plane, which you don’t have without an engineer to keep it running.

As Phil continues, FDE is not implementation – it is the engineering layer that makes AI governable this is because FDE teams build ontologies that reflect how the enterprise actually operates, wire models into real data with real permissions, and design the governance architecture that keeps autonomous systems accountable, which will, and for quite some time into the future, wire in non-overridable human oversight, approval, and review.

Phil goes on to list a few key things that LLMs cannot do on their own. (It’s in no way a complete list, but hopefully enough to get executives questioning all the AI-BS form the AI-Hype-mongers presenting grandiose claims that likely won’t be a reality within most of our professional life-times. Even better, Phil points out that Agentic AI without FDE governance is not transformation. It is risk accumulation!, and points out five key requirements of workable AI that can’t be achieved without an FDE. (There are more, but again, these should be enough key points to help executives realize that not only are LLMs sorely insufficient for almost every task they are being promoted for, but they aren’t even usable at all without the help of a FDE team.)

Phil also does us a great service by pointing out that while vibe coding creates velocity, FDE prevents it from becoming chaos — which is what happens every single time you employe vibe coding without FDEs (and a real engineering team — but we’ll get to that).

Vibe coding is simultaneously one of the biggest boons to software development and the greatest destructors, especially since it is almost universally misunderstood and misapplied. For example, while Phil’s statement that business analysts can express intent and receive working agent code in return is technically correct, it’s not practically correct. That’s because vibe coding produces code that is insecure, inefficient, and not appropriate for enterprise software. In fact, just about every startup that tried to launch an enterprise app on vibe-coding alone have lost hundreds of thousands (or more) attempting to do so — see this great post from Alex Turnbull.

Vibe Coding is super useful because, with the help of an FDE team with a good business analyst, the end user organization can quickly create functional prototypes that demonstrate precisely what they are looking for, which are much more powerful functional specifications than traditional functional specification documents with text descriptions of required functionality and powerpoint mockups. Plus, these prototype specifications can be created in a fraction of the time. But that’s all they are, prototypes. Real applications still need to be built by real software engineering teams who can build optimized, bug-free, secure code — vs. unoptimized, buggy (especially at the boundaries), and insecure code regularly generated by AI-based vibe coding tools (where, depending on what source you access, 53% to 78% of code generated has serious security issues).

In other words, it’s a great article, from a customer-centric viewpoint and written for customer executives. From a back-end, provider perspective, it’s missing one key step — the development step that takes vibe coding prototypes and produces real (AI-backed) enterprise applications.

Moreover, it centralizes the FDE activities when, in reality, they are ongoing throughout the entire cycle.

  1. they activate, and put the foundation in place
  2. they train the users on how to properly use the LLMs for accelerated research and are always on call for help
  3. they maintain the orchestration layer, and improve (and correct) it as necessary
  4. they work with the end users to vibe code prototypes
  5. they work with the development team to build the next generation (or iteration) of the enterprise apps in the SaS model

In other words, AI can enhance SaS, but it cannot replace the need for skilled humans on the provider side (for development, implementation, maintenance, and improvement) or the buyer side (for process definition, improvement, decision criteria, etc.).

At the end of the day, AI can only replace bit-pushers who do tactical data processing tasks which should have been automated by machines 30 years ago (when it was promised), but it can’t replace anyone who needs to make a (strategic) decision. This is true regardless of the model, and the right model, like Phil’s SaS flywheel, actually exemplify the need for the right, skilled, talent.

Dear Graduate, Don’t Skip the Internship … You Need a Gateway to an Apprenticeship!

A number of AI enthusiasts are advising soon-to-be and recent graduates to skip the internship and instead become proficient with AI because that’s how they are going to get a job. And, as you should know by now, it’s bullcr@p. Being able to write a prompt for a Gen-AI LLM that will return a convincing (but not necessarily sound) result is not going to get you a job. The only skill that’s going to get you a job is competence!

As with every over-hyped tech-du-jour that came before ([predictive] analytics, the fluffy magic cloud, SaaS, the WWW, etc), AI is not a silver bullet that’s going to solve all of an organization’s problems and grant magical status to those who have mastered it.

The only thing you’ll master with Gen-AI is the art of the con since whatever it spits out is so well written (compared to the average literary skill of an average high school, and even University, graduate these days) and so convincing that, without expert guidance, an average person is convinced that it must be right when they don’t know better. But that’s not a skill most organizations are going to hire you for (outside of sales and marketing), even if the organization is known for questionable ethics.

Organizations don’t need clueless idiots. They need experts who can assess situations, determine options, decide on the best option, and implement the decision. Someone who knows the analysis to run, the data to collect, the tools to use, the reports to create, the logs to keep, and the contracts to write.

And while you can’t graduate an expert, you can graduate with the skills to start you on the path to becoming one — the traditional skills of math, logic, critical reasoning, project planning, project management, and relevant domain knowledge — not creative crafting of perilous prompts for a flakey LLM that will eventually fail you no matter how much time and effort you put into that prompt.

And if you get get an internship and prove yourself, maybe that will lead to full time job where you can apprentice under a master in the real world and gain the experience you need to go from an adept (with the core knowledge and skills but not the wisdom needed to succeed in the real world) to practitioner (who has gained enough wisdom and experience to manage standard tasks and functions on their own, and who only needs guidance for new or complex situations not yet encountered) and, eventually, to expert where you become the new organizational mentor and the one that new hires turn to for help.

And organizations need (future) experts because only an expert knows when

  • it only has wrong/incomplete data (which will prevent an AI from ever working)
  • an analysis/outcome is wrong based on math fundamentals
    (and when an LLM-based AI multiplied by -1 because you told it to deliver savings vs. find the best opportunities based on price variability, lowest price, market trends, and differential analysis)
  • reasoning is correlative, not causative (which is a failure of not just LLMs, but many people as well)
  • an analysis is incomplete (because only they have specific insight that was not available to the machine or another analyst)
  • etc.

That’s why, if you want to become a true master of your craft, you need to forget the AI mastery and instead land an internship where you can apply the mastery of the real skills you learned in your degree program to stand out, get an apprenticeship, and learn how things work in the real world and acquire the real world mastery you need to get the job you want. Only then will you be able to work your way up to becoming the leader, and expert, you want to be.

There is no Artificial Intelligence (just Artificial Idiocy) and organizations will always need top talent. Automation, and well designed applications that solve real problems efficiently and effectively, will reduce the number of back-office employees that an organization needs and any employee who’s only skill is pushing bits will be eliminated. However, the need for talented employees will only increase to not only oversee the tools and handle the exceptions, but correctly analyze increasingly complex real-world situations and make the right decisions.

At the end of the day, AI tool mastery is meaningless if you can’t logically and holistically analyze the outputs with respect to math fundamentals and a real-world scenario!

This Should Be Obvious But Expert in the Loop …

… is Human in the Loop. Not another (AI) system in the loop, no matter how specialized that system is or how well it is trained!

The future is Augmented Intelligence, NOT Artificial Intelligence (which doesn’t exist and won’t exist any time soon until brilliant researchers come up with a few more insights that get us closer to understanding

  1. what intelligence actually is and
  2. modelling it.)

The algorithms might be getting more accurate in average use cases, but the illusion of intelligence, no matter how grand, is still NOT intelligence. (And, even worse, The Wizard of Oz has been replaced by a very poor digital facsimile.)

Done right, Augmented Intelligence will still let your organization reduce its non-value-add tactical workforce by 80% to 90% because the right tools will enable the strategic experts to be 3, 5, 7, and even 10 times as productive and oversee all the tactical work that needs to be done using an exception based approach where every instruction that is given forms a rule that allows the system to automatically deal with the same, and similar, exceptions should they arise again in the future in a predictable and repeatable fashion.

Instead of having to oversee a team of tactical grunts that just take up space (because they don’t have the education, experience, or raw capability required to make good strategic decisions, manage projects, and identify value), a strategic expert can instead focus her time on value-centric activities and training a protege or two who will be one that posses the right mix of EQ and TQ to grow into, and take over, her expert role (when she moves on and up).

In the near future, there will be no more bodies in seats just to push bits around, because that’s what software does best. Number crunching and thunking. NOT analyzing strategically and thinking. (I admit most humans don’t do that well either, especially these days, because they are too attracted to the principle of least action and/or enjoying the cognitive decline from ChatGPT, but those willing to practice strategic thinking daily still do it way better than a machine ever will based on our current approaches to AI). [And while there might be fewer of us each year that are willing to think, there are still enough of us to get the job done if you let us select tools that work. Not necessarily AI. Tools that work.]

How You Know Your Education System Is Broken!

Only 40% of employees say they’d be fine NEVER using AI again! (As per a recent Section AI survey in the Wall Street Journal of 5,000 white collar workers, as reported in a recent post by Stephen Klein who also noted that the majority of employees say it only saves them 2 hours or less per week. Furthermore, he also mentioned a Workday study that reported every 10 hours “saved” by AI resulted in 4 hours being lost due to required error corrections, flawed output revision, and necessary verifications, which means there aren’t much savings at all. [Specifically, for an average employee to actually save 10 hours, they’d have to save almost 16 hours, which would take them two months to achieve!])

Gen-AI is failing 94% of the time. It’s causing serious cognitive apathy and decreasing our IQs far beyond what Twitter achieved on its introduction (where it reduced our collective attention spans to that of a goldfish). It’s direct and indirect costs to run 8 hours a day are often more than to just hire another person (due to compute requirements that are 20X to 200X that of Google for a basic query, and the extreme amount of energy and water [for cooling] required on grids that are already stressed and ecosystems where fresh water is running out).

Chat-GPT. Claude. Grok. Rufus. Gemini. Meta. DeepSeek. Perplexity. Co-pilot. Poe. Le Chat. They’re all over applied due to over promises when they all have fundamental issues (like hallucinations) that cannot be trained out (as the issues are a result of their core design and programming), limited data sets (and now that AIs are being used to generate additional training data, performance is getting worse), limited guidance, and no guardrails.

There’s always been a time and a place for proper AI, but it’s not now, it’s not everywhere the investors losing Billions on Open AI and competitors are telling you, and it’s not the “AI” they are pushing.

Every time a new advancement in tech comes along, we always forget how long it takes to get from prototype to safe for unmonitored regular industrial and home use, be it hardware or software. With AI, it’s always been about two decades between a new algorithm being invented, and a production ready system with known performance, limits, and guardrails being ready for the mass market. In other words, this tech shouldn’t even be out of the research labs yet! We definitely shouldn’t have every major consultancy trying to push it as the cure-all for every problem throughout your entire enterprise. (Or new start-ups claiming they can offer you AI Employees!)

How many more examples of (silicon) snake oil do we need before we accept there is no panacea for all your ailments — be they physical, mental, or industrial — abandon this current iteration of Gen-AI, and go back to the targeted, mature, solutions that were finally ready for prime time (as we finally had enough processing power, data, and research behind us to deploy them with confidence)?

And even though the technology might work as much as 12% of the time, as per a PwC study that found that 12% of 4,454 CEOs surveyed reported both revenue gains and cost reductions, that’s not much of a validation of the technology — especially since those gains and cost reductions could have nothing to do with AI at all (and the pilot success of 6% from a recent McKinsey is a much more reliable metric here).

If you want real success, find a (A)RPA solution that works, lie its AI and buy it while you wait another decade for this technology to mature to the point its reliable, guarded, and safe for mass market adoption and widespread application. (Or wait for an AI-enabled SaS provider to come along who will do the 24/7/365 human monitoring required for you and make its software is usable and safe through this monitoring. Because all the current generation of LLM[-powered Agentic AI] tech is doing is increasing the need for human monitoring, not decreasing it.)

Without Human Smarts, There Will Be No (Usable) AI!

And I’m so happy I’m not the only one pushing this theory. Mr. Stephen Klein recently published a great post on The Age of Pretend.

In the post he notes that:

“Everyone assumes AI’s biggest bottleneck is compute. … That assumption is wrong. The real bottleneck … is architecture, specifically, a design decision made in 1945. … The real constraint: the von Neumann bottleneck. Modern computers separate memory and processing. Data has to move back and forth between them. For most software, that’s fine.
For AI, it’s catastrophic.

Some numbers the industry rarely highlights:

  • Accessing off-chip memory consumes ~200× more energy than the computation itself
  • Roughly 80% of Google TPU energy goes to electrical connections, not math
  • A 70-billion-parameter model moves ~140 GB of data just to generate one token”

LET THAT SINK IN. Us old timers remember “640K out to be enough for anyone”! The Apollo Guidance Computer — you know, the one that was installed on each Apollo Command Module and Lunar Module in the Apollo Missions, had 2K Core RAM Memory and a 36K ROM. Even today, unless you have an iPhone 17, your phone probably only has 128 GB of storage. That means, even with the processing power of your phone (that dwarfs most computers us old timers have ever owned), you can only process ONE token. (Now do you understand why the data center [energy] demands for your Gen-AI chat-bots are destroying the planet? Anyway, we digress …)

This means that (Gen-)AI has hit a wall. Computer Architecture supports massive compute at scale, massive storage at scale, but not massive transfers at scale.

So what does this mean?

Do you remember the days of RAM drives? Not only did it speed things up, but it kept your machine cooler because, as Stephen noted, less energy accessing data in RAM than on disk.

And do you remember the fun of Assembly? (Okay, that’s sarcasm!) Once you learned to maximize register usage (i.e. re-sequencing processing so that you minimized reads from, and writes to, memory), your code got faster still (and machines stayed cooler longer, which was obvious by the lack of noisy fans spinning up).

We’ve known about this problem for decades. (Eight decades to be exact!) It’s too bad today’s students don’t study the basics and understand it’s not strength that determines computational speed and energy requirements, it’s data scale — whether the data fits in memory or not, whether “significant” chunks fit in the onboard GPU memory or not. (And specifically, can you scale the data down enough for the efficiency you require?)

But this is still the key point in Stephen’s article:
“The next major improvements will likely come from smarter algorithms.”

We might need brute force to detect patterns we can’t (yet) see, but the only way to truly advance is to understand those patterns and code optimal, light-weight algorithms that exploit fundamental rules to allow us to process data quickly and efficiently.

Until we figure that out. You’ll never have usable AI (and definitely never have REAL AI as not only will it never be intelligent, but it will never, ever, get anywhere close).