Category Archives: AI

There is NO Infinite Compression – The Latest DeepSeek Paper is BullCr@p!

Every decade or so, some idiots who never studied Huffman coding or Information Theory believe they have cracked the problem of infinite compression, and this linked paper is just the latest example of this lunacy. I really hope this was a joke paper authored by AI because it’s all bullcr@p!

On average, a text token in a LLM should require 20 bits or less (as 17 bits support a 129,000 word vocabulary) while a vision token can be 16,384 bits (based on 1024 dimensional continuous vectors) — because it takes a lot of bits to represent pixelation of a square in a 2-D image! This says you can store about 820 text tokens in the same space it takes to store one vision token. Or, you can store the entire text (lossless) in 48K, versus the 4M it would take to store the 250 vision tokens (using very lossy compression) that are required in the paper. Looks like a LOT of people can’t do basic math if this is being praised as revolutionary!

Moreover, the raw text, which maintains the full context if the tokens are kept in order, is not only fully lossless, but can be compressed using a modified Lempel-Ziv algorithm to take up an average of less than 2 bits per character (and achieve up to an 80% compression rate). Given that the average length of a word in average text is 5 characters, and a space is one character, 2500 words would be 15,000 characters, storable in 30,000 bits or a mere 4K! In other words, this paper is trying to pass off a ONE THOUSAND FOLD increase in space requirements as space saving! Pure lunacy!

In other words, if someone is claiming something too good to be true, it is! Don’t fall for it or the sure to follow claims that DeepSeek OCR is revolutionary because of this. (Since every document is different, you can’t imagine the true loss with a 90% vision token reduction!)

CEOs are hugely expensive. Why not automate them?

As per Will Dunn, as published on The New Statesman

Especially when hiring a CEO who doesn’t understand what makes the business profitable loses Billions:

Starbucks Loses 30 Billion

and doesn’t understand what is critical to the company product to the point costs can never be cut no matter how high those costs may look on the spreadsheet because the net result is not only product failure, but grounding/banning of your product and expensive lawsuits that costs Billions:

Boeing lost 11.8 Billion in 2024

After all, if we’re hiring CEOs without any relevant experience, actual business intelligence, or even logic, then why not use Artificial Idiocy? It’s not like the occasional hallucinations will be any worse that an average CEO’s these days (who believes investing Billions on empty promises is a good idea) … and the actual compute costs, even if in the six figures, will still be a tenth (or [much {much}] less) of what a CEO salary and benefit package actually costs!

So if you insist on creating fictional “AI Employees”, why not kick off 2026 by starting with a job that, sadly, Gen-AI agents can actually do?

Here’s why you DO NOT want Agentic Buying and you DEFINITELY DO NOT want AI Employees

buying for you!

An AI Vending Machine lost hundreds of dollars!

Just imagine what AI is gonna lose on your multi-million dollar categories?!

And when you demand a certain savings that’s unachievable, it’s going to find a loss that equals the savings amount, multiply it by -1, and tell you that’s the savings.

< Stanford, Anthropic, Redwood, Meta, etc. studies on negotiation games, competitive scenarios, and goal-seeking behaviours, etc. >

So unless you’re looking to LOSE money …

Stick with classic automation and point-based AI where the automation runs everything for you, does all the verifications and data checks that can be automated, does all the standard analysis for raking and recommendations, and gets rid of 90%+ of the tactical time-consuming work, freeing you up for the manual review, safety checks, and strategic decisions where you, as a human, can check and find obvious supplier misunderstandings, frauds, and bad decisions for the long term because the system does the grunt work and pre-does all the standard analytics, freeing up 80% of your time to do more sourcing, more relationship management (to prevent problems and loss), and more decision making (when it’s hard to make the right decisions on numbers alone or its impossible to satisfy all the goals and choices must be made).

Who’s Funding Your ProcureTech Vendor?

This question is more important now than ever! Not only is the RCD (Relative Corporate Debt) of many FinTech companies too high right now (See: Calculating RCD), signalling a decline in customer service and potential abandonment, if not outright vendor failure down the road, but the ongoing viability of many VC and PE firms, or at least their ability to support their investments, is also in question.

Many firms are too heavy on AI plays that are still losing as much as $4 (or more) for every $1 of revenue they take in, requiring massive ongoing investments to maintain. Even big PE funds only have so much cash to burn, and the only way they can do this is to liquidate assets and holdings if they can, or, in the worst case, simply write off losses (and associated future costs) of those holdings they can’t liquidate.

Softbank’s end-of-year investment in OpenAI really puts this into perspective, as chronicled by Mr. Klein of Curiouser.AI and Berkley in this LinkedIn post.

As far as I am concerned, this is bad news for any of SoftBank’s FinTech holdings that may require funding in the next few years, and a warning to make sure you don’t select / continue / depend on any of their FinTech holdings where they have a large or majority stake until verifying those holdings are profitable and likely to stay that way! (Now, SoftBank has traditionally had very good investment chops, so it’s likely the majority of holdings are profitable …)

However, they aren’t the only firm making huge over-investments in AI and weighting the portfolio down with companies that might never see a profit. This means that this warning also applies to many other Tech investment funds, starting with Thrive, Dragoneer, Altimeter, and Coatue who also have large stakes in OpenAI. They could all end up in the position where they are going to have to sell off / dump assets to maintain the ridiculous losses OpenAI is seeing, and any holdings not performing well will likely be the first to go / get dropped. (Remember that the average age of the first three of these groups is 15 years, and they are [becoming] modern SaaS/AI heavy, whereas Softbank Capital has been investing for 30 years, and is a lot more diversified. Softbank may be able to weather a complete crash in OpenAI valuation if it occurs. But these other firms may not!)

But, as we noted, the real warning is not for SoftBank or these other mega funds (in the significant 8 and 9 digit range) that have funds to weather a storm. It is for the smaller funds, especially those less than 1 Billion, that are too AI heavy.

As a result, when selecting any FinTech platform, you need look at the portfolio of any investment player with a substantial majority stake. If a large segment of the portfolio of a significant/majority investor is “AI” companies losing money hand over fist, then the vendor of that FinTech platform cannot be considered a stable vendor if it is not profitable. This is because you can’t count on the fund having the resources to support the vendor to profitability, even if vendor is a fund darling. This is the case even if the RCD calculation looks good! A lot of the smaller funds can’t afford an AI crash given the AI-heavy focus of their SaaS portfolio.

(Face it. An AI crash is coming. Too much valuation against too little return, and investors only have so much patience. The only thing we don’t know is how severe the crash is going to end up being. Is it going to be a minor drop across the tech markets or a major crash like the 2008 housing crash or the 1999/2000 dot com crash?)

A Review of The October Diaries (in 4 Parts)

Part I

The October Diaries is a supernatural drama centred on the interaction of the protagonist, Jon W. Hansen, a distinguished analyst with a 40 year career in tech and, in computing years, an AI RAM Model 5 based on centuries of development. His work becomes increasingly complicated as other models continually challenge his and self-proclaimed AI Experts continually threaten our space from the shadows. The book chronicles the complex relationships as Jon tries to find new ways to preserve the truth and protect …

Oh wait, that’s the plot archetype for the Vampire Diaries. Did I read the right book?

Yes I did. But I just made you think, and that’s one of the primary goals of Jon’s book and one of the key points I have to make.

Every Influencer, Consultant, and Analyst needs to read this book, but 99% won’t learn anything if they don’t think and question everything they read. (And that’s one of the unwritten reasons Jon says you’ll have to read the book two and even three times.) If they don’t come to suspect the truths on their own before Jon exposes most of them in later chapters. If they don’t understand that this is not a guide or manual for success or the answer to all their problems (as there is none) …

It’s a book designed to make you do what we don’t do enough of in the age of AI: think, and, most importantly think in a way that will, in time (may not today, tomorrow, or even next year) allow you to actually use modern AI tools productively and extract value in real time.

Gen-AI efforts are failing across all the board, from large scale corporate projects down to small scale individual efforts to extract useful content for reasons that include:

  • lack of focus
  • lack of verified data & reinforcement training
  • lack of knowledge
  • lack of skill

You see, for success, you need to have

  • focussed domain models
  • deep context
  • deep domain knowledge to know when the output is good, ok, and bad
  • appropriate skills to utilize the models effectively

Jon gets at this with his six skills of conversational fluency, which is his name for the methodology he uses to train the models to do what computers do best (identify patterns, surface them, and draw correlations) while he does the strategic thinking humans do best.

As well as his five common mistakes that are one of the reasons the vast amount of human prompted content generated is AI slop.

But he also goes deeper into what is truly required for long-term success. Which may shock many of you who aren’t from the old-school we are, but, like Billy Idol, you have to deal with the shock to the system it will give you and push forward.

Discuss Part I on LinkedIn

Part II

Every Influencer, Consultant, and Analyst needs to read this book, but not for the reasons they think. It’s because they need to think deeply about AI, and that’s what this book forces them to do. It may be framed as a step by step guide to take you from zero hero, but that’s just to psychologically convince you that this is the guide for you — and if you want to understand AI, it is!

Most people are using AI wrong. More specifically, they are using the A.S.S.H.O.L.E. to sh!t out plagiarized slop that is turning the internet into massive sewer that is likely making Jon Oliver rethink his Facebook is a Toilet rant (from 2018) (because now the entire Internet is a sewer).

While that is one of the few things that LLMs can actually do, that’s NOT what they should do. They might be lying, hallucinating, soulless algorithms that will happily tell you to commit suicide, suppress life saving alarms while you’re locked in a server room on fire, or even ignore your shadow and have the self-driving car run you over, but they have their uses.

While they can’t do 94%/95% of what the firms selling them advertise (or we wouldn’t have 94%/95% failure rates, as per McKinsey and MIT), they can do four things very well, with reasonably high reliability when appropriately trained and deployed, that we can’t. The first two, as I keep promoting, are:

1) large corpus search & summarization
2) natural language processing

The third, as Jon makes clear in this book is

3) deep pattern detection and surfacing

But only if you know how to get the algorithm to do it!

You see, all these systems are trained to deliver direct responses to direct requests. As a result, when you give them a typical direct request in your “carefully calibrated prompt“, they give you what they think you are asking for, and that’s it. But that doesn’t help you, or me, or anyone, especially if they weren’t trained on the right data or it’s not available and the only way they can give you what you want is to make sh!t up.

Sure it might spit out 2997 characters for your Linkedin post 10 times faster while addressing the seven points you wanted, but is that really helping you when you have to read it, edit it, and copy and paste and verify it? That takes time — and even worse, it’s not productive time. If you’re not thinking about the 5 Ws, not only are you not sharing anything valuable, but you’re not advancing your thinking. (Right now, the only edge we have over machines is our ability to think critically and strategically — so what happens if we lose that?)

But if you can learn how to work with the technology, instead of getting bland plagiaristic derivations, you can get it to surface patterns across related bodies of work, document progressions over time, and use that to more quickly validate your instincts and formalize your ideas, allowing you to advance your own abilities while ensuring you can serve your customers faster and better by speeding up research and delivery efforts by multiplicative factors.

Discuss Part II on LinkedIn

Part III

Today we continue with our review of the supernatural drama that chronicles the interaction of the protagonist, Jon W. Hansen, and the RAM Model 5 that we’re sure you’ll find more thrilling than the pages of the Vampire Diaries we thought we were reviewing (due to the similarities in plot archetypes). You might not have the love triangle, but I’m sure the dollar signs will be more than enough to get your attention. (What dollar signs? Well, you’ll have to read it.)

In part one we said that you need to read the book because it will make you think (if you’re reading it right).

In part two we said you need to read the book because it helps you understand the power of LLMs is not its ability to create watered down plagiarized slop 10 times faster than the drunken plagiarist intern ever could but uncover patterns that you might never uncover on your own due to lack of time.

Today we’re giving you a third reason — and that reason is that it helps you understand why you are invaluable in the age of AI. While it has been true since the introduction of computers that monkeys could do all back office jobs if they knew what buttons to push, the reality is that AI, which should be called Artificial Idiocy, still doesn’t know what buttons to push, it’s just able, in many situations, to compute what button to push with high probability. But it DOES NOT know. Only YOU know! (You see, what AI really stands for is Algorithmic Improvement, as it is the label that is consistently applied to any algorithm that is an advancement over a previous algorithm, and that has nothing to do with intelligence.)

Now, it does mean that if your job is simply tactical data processing then you’re out of work, and it does mean some of your peers who aren’t as good and efficient as you are also out of work since the tech will make those who know how to use it up to 10 times as efficient at some tasks, but if you’re a skilled expert, then you are more desperately needed than ever because, as per our last post, only you will be able to detect the very convincing inaccuracies, lies, and hallucinations it returns.

But understanding is not enough, you need to be able to explain it, and when pressed, demonstrate it. That is what the book, after a few reads, will help you do. Use AI in a way that demonstrates you are what’s needed to make AI effective and make sure the organization isn’t part of the 95% failure statistic.

Part IV

In Part I of our review of Jon W. Hansen’s October Diaries, his take on the modern thriller, I told every Analyst, Consultant, and Influencer (ACI) that they need to read it because it will force them to finally think — deep — about AI.

In Part II of our review I told the ACI they need to read it because it will help them use LLMs properly and surface patterns they might not ever find on their own due to time constraints.

In Part III of our review I told the ACI that it will help them defend their positions in the “Age of AI” purge that is coming. (Since it’s a new excuse to fire people so the organizational shareholders can [temporarily] get richer!)

Now, in Part IV, I tell most of the ACI that I’m sorry. You shouldn’t read it. You want a quick fix and an easy solution to your relevance problem and this isn’t it. In fact, for some of you, it won’t even be worth the cost of the minuscule amount of storage it takes up on your hard drive.

Because it makes a few assumptions.

1) You have, or are willing to build (with your own hands), a deep archive of unique, human authored content to augment the models with.

2) You are willing to take the time to not only ensure the models are trained on this, and only this, archive but to learn how to both use the models appropriately and get them to retain and access relevant context across multiple sessions over days, weeks, and months, which is a skill that goes beyond creating executable ChatGPT prompts.

3) You have, or are willing to develop, the expertise necessary to know when the model is 100%, 95%, 90%, 50%, and 0% right, no matter how convincing the words are that it returns, and how to correct it and guide it to 95% every time (so you can make the corrections faster than doing the work from scratch), which could take minutes, hours, or days for any particular request you throw at it.

But let’s face it.

1) Most of you don’t have the archive, unless you work for a consultancy that has been delivering projects for at least five years, and preferably 10. Jon and I remember the early days with hundreds of blogs, and the 3/3/3/3 rule. Up to 90% of wanna-bees would quit after 3 posts/3 days, then the next batch by 9 posts/3 weeks, then the next batch by 27 posts/3 months, and the majority by 3 years would say “hey bloggie, I’m packing you in“. The hundreds of blogs I chronicled on the now-defunct SI resource site were down to a few dozen by the 2010s.

2) You won’t put in the months necessary to get the model and your skills to the point you are getting close to what you want every time. And it will be months!

3) Not only do you have to keep learning tech, you have to be constantly seeking out experts to learn your trade. That’s also a lot of work. When you’re Bowling for Soup, you know that High School Never Ends!

In an age where founders want to vibe code and flip companies within 3 years, you want instant gratification, but you’re not going to get that!

All it will give those of you starting out is a way to build a skill that is sustainable for life. But the vast majority of you will have to wait for the good things to come. And I don’t think you will. Sorry.

But if you want to prove me wrong, get the book!