#MeWriting I come from the future. I know how this whole AI thing ends. It’s not pretty. It’s not the end of the world or a particular dystopia we’ve previously imagined. But it is a world that will leave us longing for the peace, quiet, and freedom of, let’s say, 2020. Yes, a future of pandemic nostalgia is as horrible as it sounds.
To prevent this bad outcome, you need to get over AI now. By “get over”, I don’t mean quit cold turkey. I don’t even mean quit AI. I mean you need to understand the “over”. Wait, what?
You need to “get” — or understand — three “over” mistakes you are making:
- Overspending.
- Oversharing.
- Over your skis.
Before we get started, I want to warn you. I bring a combined academic and experiential critique of the cloud AI industry that many of you rely on and even love. I’ve seen real disasters up close and I can connect them to basic computer science. I’m not here to take away your joy. I’m here to help you benefit from an amazing new technology and avoid skiing into a tree.
Overspending
In the cloud today, we have a selection of very large and seemingly capable models that range in price from $50 per million output tokens (Anthropic Fable) down to less than $1 per million output tokens (DeepSeek V4-Pro). Your application or typical usage is probably workable with the least expensive model, and not enhanced beyond marginal verbosity by using an expensive flagship frontier model. In plain English: you’re not getting better answers by paying more than 50x for that answer.
In fact, with typical language-based tasks, like summarizing an email or writing a LinkedIn article like this one, I have found that a very small local model like Google Gemma 3 4B Instruct performs just as well as any retail cloud model. As I have said for more 2-1/2 years now, these small models are “just as knowledgeable and less annoyingly loquacious” than the bigger ones.
Aside: This article was written entirely by me, not a model. Yes, I have always been — and remain — all about that em-dash, brother!
Oversharing
When you prompt a cloud model or use its API, your prompt and all data (files) you attach are sent to a server in the cloud. There is a possibility — even a strong probability — that your prompt and all data associated with it will be used by the cloud model provider for some activity other than generating a response.
OpenAI displays ads to free tier users who are not self-identified as minors (link). Anthropic has used prompts from across its users and even from its partners to start services that directly compete with partners (link). OpenAI data has been exposed through an inadvertent Google Drive data leak (link). Anthropic famously wanted the ability to deny Department of War requests related to activities Anthropic didn’t like (link). Defense contractors do not play those games.
Both providers maintain a history of your prompts to allegedly give you more personalized answers. These histories can be subpoenaed by lawyers in both criminal and civil matters (link). It has become a common enough practice in job interviews to ask candidates to share their cloud AI histories during the interview, purportedly as a demonstration of their ability to use AI tools (link). We expect cloud providers to alert authorities when prompts indicate criminal activity or intent, or indicate potential for self-harm (link).
People come to chatbots seeking medical, therapeutic, and relationship advice (link). Whether they get good advice or not from chatbots, there is no special care for privacy of those requests. They’re subject to the same potential data usage activities identified in the previous paragraphs.
Some or all of those data uses may make you uncomfortable, or perhaps, comfortable. The point is, you have very little control once they have the data. Palantir’s CEO, Alex Karp, pointed this out in a recent rant on CNBC (link). These cloud AI firms are sucking up corporate customers’ alpha — a company’s unique, proprietary competitive edge — one input and output token, at a time. We readily accept that as a problem when it’s framed as American companies using models hosted by Chinese companies. Yet, it’s a “rant” or an “awakening” when Karp applies it to models hosted by American companies. Now imagine you’re a cloud AI customer in Europe. Our lack of reluctance to overshare is both weird and troublesome.
Over your Skis
I have a Master’s Degree in Information and Computer Science with a specialization in the theory of algorithms and data structures from the University of California, Irvine (1994). I don’t mention this to pull rank or assert authority. I tell you this because I know from my academic training that my sub-discipline has more than 80 years of a Newtonian physics level and style of knowledge that can and should be applied to these language models and how we use them. I’d like to share the intuition that my study of the field has given me with you.
For starters, these models do not think. As implemented, they are incapable of achieving any sense of consciousness or self awareness. Notions that they do or can are laughable. Adorable, but clown school laughable (link to clown show). I apologize on behalf of my industry if they got you to believe any of that tripe.
What language models actually do is run an iterative algorithm called the completion algorithm. In starts with a cue (prompt). At each step, it evaluates the ordered list of tokens (words) it has so far, and it generates a next random best enough token. It generates this next token using statistical weights in a giant word association database, which is often billions, hundreds of billions, even trillions of “parameters” in size. It might eventually generate a special “stop” token. In production, it usually does, or has some other criteria to stop iterating. That is both actually and all a language model does. We may construct some analogies to human thinking to make the process more relatable, but that’s all it does.
A chatbot implements an illusion of back and forth chat by formatting the input such that the algorithm would complete both sides of the chat. However, the chatbot stops the iterative generation process when it’s your turn to type (or speak). That’s the illusion. To the unknowing — and even to the knowing — that chatbot might feel knowledgeable, empathetic, even real. But chat with a language model is just a cheap user interface illusion. My apologies again if you bought into it.
Let’s get to “over your skis”. The colloquial phrase means that you are doing too much with what you have available, and are likely to run into problems. Most of what typical language model users do and hope to do falls under “over their skis”. They trust a model which usually gives “an answer” to give them “the right answer” or at least “a good enough answer”. In fact, they are disappointed when it doesn’t, which is quite often. Early on in widespread language model use, we identified the problem as “hallucination”. Collectively, we hoped we could solve this problem. Computer science tells us we can’t. We hoped that bigger, more expensive models would reduce the problem. It hasn’t. Computer science tells us it won’t.
The widespread misunderstanding of reality suggests a workable strategy: Use language models for problems where any generated answer is, necessarily, a good or acceptable answer. Practical applications include story telling, ideation, prototyping, practice reps, and blank page filling. Prior to language models, trusty lorem ipsum signaled that there remained work to be done.
Summarizing things with language models is usually good enough. It tends to get the gist and miss important details that don’t occupy much space. Shorter summaries with an expectation to (eventually) read the item being summarized are workable in practice. Perhaps asking for a catchy headline might be a better strategy. But I digress. The computer science part of my brain is looking carefully at how you can be fooled or even fool yourself into getting over your skis, trusting answers that aren’t good ones and may have significant, unexpected downside.
Automation — “agents” — is a terrible, dangerous use of language models. Why? Because the models will get the automation steps wrong, and get them wrong often (link). If a sequence of tokens that deletes your home directory or drops a production database exists, one with a similar outcome will eventually be generated. And because you automate to avoid doing the manual work yourself, you most likely won’t validate each sequence of generated steps before letting them run.
Computer science tells us we can’t write a program that is guaranteed to catch all these mistakes, or in practice, many of them. Formal verification is still, today, as difficult of a problem as it was in the 1980s, when we rebranded “programming” as “software engineering”, and tried to attach practices that resulted in better, safer programs. In practice today, we can only sandbox so hard before losing the functionality we wish to automate, and it just isn’t hard enough.
Any application with numbers or analysis is a terrible, dangerous use of language models. Why? Because the completion algorithm doesn’t do math. Counting the Rs in “strawberry” is only (half) solved by the latest models because there is so much discussion of the problem in their training now! Try another less popular word with a different letter. Now imagine you’re asking a language model to update a spreadsheet. Good grief!
When you get (too far) over your skis in skiing, you tend to lose your balance and are very likely to crash. The crash can be a disproportionate outcome. Ask Sonny Bono (link) or Michael Kennedy (link). The analogy is appropriate. The machine does not “know”, in any sense, the importance of what it is generating, or the consequences of delivering a wrong or unacceptable answer. This applies the same to language models that cost $50 per million output tokens as it does to models that cost less than $1 per million output tokens. They use the same iterative completion algorithm.
Get Over, Go Local
My solution to all these problems is that you should try a local language models first, and you should use a completion style user interface first. Cloud models and chat are a great fallback option, but should be an escalation, not your default. You should avail yourself to analyses of what you’re trying to accomplish. Specifically, ask yourself: Does or can the completion algorithm — which generated one random best enough token at a time — actually generate a useful answer? If you don’t know how to do that analysis, I hope you will reach out to someone, like me, who can.
I’m not just here pontificating. I offer free language model software that runs on your laptop. No fees, no signup, no newsletter, no installation even. It’s really easy to use, and it does not share your data anywhere. If you’d like something more powerful that runs locally, I make language model appliances. Plug in power and a network cable and it’s ready to use in your computer’s browser. No setup. I also offer paid tutoring to help you use these tools and validate your applications. Please visit Mmojo.net if those interest you.
I didn’t discuss vibe coding or agentic coding specifically in the “Over your Skis” section. Let’s address them now. Large enough local language models like Qwen 3.5 9B and Google Gemma 4 12B actually generate great code for small projects like simple web pages and applications. These are great for personal-use tools that can’t do much damage. My Mmojo Server software (link) helps you create and save these without doing anything agentic. It’s the agentic part of vibe and agentic programming that is the big problem. That agentic part is what tries to put the generated code into production, and can result in deleted home directories, dropped production databases, and other unanticipated, disproportionate problems.
Being a very based critic of the rollout of language models, I get clients and prospects who challenge me often on my negativity. It’s fair to ask what I think of them. And it’s fair of me to answer. I think some of them are going to ski into a tree. I hope they survive. I would like to help them not do that by teaching them to use language models affordably, in a safe place and a safe manner.
I’m not anti-AI. Far from it. I think you should use the tools for what they are good for and at. That’s why you need to “get over”. If you think I can help you with free software, an appliance, or validating your language model use cases, please message me on LinkedIn. I am happy to help!
You probably know this already, but Brad holds a Masters Degree in Information and Computer Science with a specialization in the theory of algorithms and data structures from the University of California, Irvine (1994). He has worked in the software industry on his own projects and projects for other companies for more than 40 years. He wrote a smartphone application with middleware for a leading background screening company in 1999 (!!) and developed literacy software popular in Native American schools in the 2000s. He spent a decade automating software product demos before turning his attention to language models. Dude knows what’s up.
Many thanks to SOUMEN SARKAR and Pete A Turner for reviewing drafts of this article and suggesting improvements!
This article was originally published on LinkedIn.