Last year Joël Collin-Demers, The Channel Master, wrote a post encouraging vendors to steal his ProcureTech startup idea. Unfortunately, that idea involved the proliferation of sh!tty LLM technology and way too many vendors took him up on it.

I’m sorry to say that it was the one post I wish he hadn’t written!

Too many vendors decided to steal his idea, as evidenced by the constant proliferation of “AI” vendors believing they can wrap, or cr@p, an LLM better than the giants who have collectively spent trillions and actually deliver value.

They can’t. That’s because LLMs are fundamentally flawed. Hallucinations are core, consistency is a pipe dream (and those pipes are so dirty even Mario can’t clean them out), and you still need a considerable amount of exceptional data to get anything remotely useful out of them.

All Deepseek proved was that you don’t need to spend millions (or billions) to build an LLM — open source code and your own rack in a data center will allow you to get the same quality of results (i.e. garbage) as a mega-model if you focus it to a particular task in a particular problem domain.

The models would be small, fast, and cheap, but, just like the big models, won’t work out of the box because they are not intelligent, aren’t deterministic, and aren’t even consistent. (And let’s not overlook the fact that a subsequent iteration on a task or document might undo something they got correct in the last iteration that you approved.)

As for his examples:

• No RFX execution — draft creation, sure, but accuracy varies
• They’re more likely to enable fraud than stop it (see many SI posts)
• The contract insights they return may not be the most relevant ones (and leave you blind to million dollar risks)
• They are just as likely to make up risks as detect actual risks with new suppliers … and accuracy will vary greatly based on the data available and what you plan to use the supplier for
• Given that they can’t think, don’t understand logic, and can’t even do basic math (it has been proven, see Apple studies for e.g.), you should never use them for benchmarks (just for data extraction from hard to digest sources, providing Intern Indy reviews the data first)

Now, if you insist on riding the hype wave, knowing that failure is likely inevitable (with only 6% of companies seeing a return from AI investments), then this is the way to do it as you’ll waste the least money proving classic tech with augmented intelligence is the way to go (while doing the least harm to the environment).

Conclusion: it’s the brilliant way to go bust! 🤣 😭