llm

I’m an AI Slop Builder. Here Are Two Things That Actually Made Me Better At It.

cantechit Uncategorized , , , ,

Continued in my AI series….

I don’t write code (I know, broken record) . I read it a bit. My contribution to any codebase is mostly vibes, bad variable name suggestions, and extremely strong opinions about what the button should look like. What I am is someone who has spent a lot of time building real, working things using AI — and making every possible mistake along the way before stumbling backwards into something that actually works.

I call myself a builder but really I am a AI Slop Builder. It’s self-aware. It’s accurate. And I’m not embarrassed by it — because I’m trying to do it better.

This week: two things that changed how I build. Neither of them requires you to know how to code. Both of them require you to change a habit that feels comfortable right now but is quietly killing your projects.

First, Let’s Talk About the Mess You’re Already In

If you’ve been using AI to generate code, there’s a good chance you’re doing something like this:

You open ChatGPT, or Grok, or Claude. You paste in some context. You ask it to write something. It does. You copy it somewhere. Then, twenty prompts later, you can’t remember what it changed, why it changed it, or whether the version you have in your editor is even the same thing you were just talking about. The AI has forgotten everything you told it in session one. You’ve got three browser tabs open, a folder called project-final-FINAL-v3, and a vague sense of dread.

Sound familiar?

Here’s the core problem: the context window is finite, your project is not. Every conversation starts fresh. As your project grows, the gap between what the AI knows and what your project actually is gets wider and wider. You spend more time re-explaining than building. You’re not an engineer. You’re a copy-paste operator with ambitions.

There are two habits I changed that fixed most of this. One is about memory. One is about who you’re talking to.

Tip 1: Stop Being Afraid of GitHub. Your AI Isn’t.

I avoided GitHub for years. It felt like something developers did — something with terminals and commands and merge conflicts that would make me feel stupid. And honestly? The first time I poked around in it, I did feel stupid.

Then I realized something: I don’t have to understand Git. My AI does.

GitHub isn’t just version control. It’s a project memory system. And once I started treating it that way, everything clicked.

Here’s what I mean. GitHub has a built-in feature called Issues — it’s basically a ticketing system baked right into every repo. You can log a bug. You can describe a feature you want. You can write out a problem you hit and how you think it should be solved. And then — this is the part that changed things for me — you can hand that issue directly to an AI agent and ask it to resolve it.

The beautiful side effect? You now have a paper trail. Every issue you close is a record of a decision that was made, a problem that was solved, a feature that was built. That record doesn’t disappear when you close your browser tab. It lives in the repo. Permanently. I even use it for research purposes, if I have ideas, I log it as an “idea” and then ask an agent to go off and research it for me.

And here’s why that matters more than you think: the agents also have access to it.

Tools like Claude Code, Cursor, and GitHub Copilot Workspace can read your issues, your pull requests, your commit messages. When you build with GitHub as your source of truth instead of a chat window, the AI isn’t starting blind every session — it has context. Real, structured, version-controlled context that you built up over time just by logging your work like a normal human being. It can operate Retrieval Augmented Generation on our issues.

You don’t need to run a single terminal command to start. You can create a repo, write issues, and review pull requests entirely through the GitHub web interface. Ask your AI how. Seriously — just ask it to walk you through it. That’s exactly the kind of thing it’s good at.

The habit shift: Before you ask AI to build something, write it as an issue first. One or two sentences is fine. “Add a login page that accepts email and password.” Done. Now you have a record, a reference, and a task your agent can actually grab and run with — instead of trying to hold all of that in one immortal chat thread. More important, before you start a project, open a Github repo right away, start with a great environment.

Tip 2: Stop Asking Your Chat Window to Write Code. That Is Not Its Job Anymore.

This one is where I felt the biggest shift — and where I also made the most embarrassing mistakes before I figured it out.

Most people start the same way I did. You’re in a chat window. You’re talking to the AI. You ask it to build something. It builds it. You copy it. You paste it. You come back, ask it to fix something, copy it again, paste it again. Rinse and repeat until you want to throw your laptop out a window.

The problem is you’ve turned your AI into a code vending machine instead of what it could actually be: a manager.

Here’s what I mean. Modern agentic AI — the kind baked into tools like Claude Code, Cursor, Windsurf, or even more advanced configurations of ChatGPT — isn’t just a chat box anymore. These tools can spawn subagents: separate, focused AI workers that handle specific tasks in their own isolated context, then report back. One subagent writes code. Another reviews it. Another runs tests. The manager — the agent you’re actually talking to — coordinates all of that without dumping it all into one giant, bloated conversation.

Every developer who uses these tools long enough hits the same wall. The session starts fast. The first hour is productive. Then, somewhere around the third or fourth hour, responses get vague, the model starts forgetting decisions made earlier, and code suggestions drift. The context window is full — and it’s full of noise.

Subagents are the fix. The orchestrator — the agent you talk to — doesn’t do the heavy lifting. It coordinates. Each subagent receives a focused prompt with a clear objective, does the work in its own clean context window, and returns only the relevant result to the manager. You don’t see the mess. You just see the outcome.

But here’s what nobody tells you: the AI will not do this automatically just because you asked it to build something complex. You have to prompt it differently.

When I stopped asking “can you build me X” and started saying “you are a manager agent — plan this work and delegate the implementation to subagents, I don’t need to see every step” — the quality went up, the conversation stayed clean, and I stopped copy-pasting like a maniac.

Think of it like being a project manager who delegates to specialists. You wouldn’t ask your designer to run security audits or your QA tester to write marketing copy. Same logic applies here. Stop asking your chat window to be everything. Give it a role. Give it authority. Tell it to delegate. “I have this cool idea for a new way to this thing, here’s the idea, but I want to research it, spin up a sub agent to research it and document the findings in an issue”

The other shift: if you’re doing this in a plain chat window, you’re leaving tools on the table. Most of the serious agentic tools — Claude Code, Cursor, Windsurf — have proper support for subagent workflows baked in. They can read files, write files, run commands, check your GitHub issues, and loop back for review. A chat window can describe doing all that. A code agent can actually do it. That’s not a small distinction. That’s the whole ballgame.

You don’t need to understand how subagents work under the hood. You just need to know they exist, prompt for them, and get out of the way.

The Real Lesson Under Both Tips

What both of these tips have in common is this: AI works better when you give it structure, not just requests.

GitHub Issues give your project structure. A manager-agent prompt gives your workflow structure. Without structure, you’re just vibing your way through a chat window hoping something sticks. With it, you’re running something that looks a lot more like an actual software team — even if that team is entirely made of ones and zeros and you’re the only human in the room.

I’m still an AI Slop Builder. I’m not going to pretend I’ve got this all figured out. But I’ve at least graduated from copy-paste chaos to something that feels intentional — and honestly, that’s more than I expected when I started.

Next time, we’re going to talk about something that sounds simple but is probably where most people are losing the most ground: prompt engineering. Specifically — what do you actually put in that little empty box? I’ve been building up a set of tips, bootstrapping tricks, and hard-won lessons about how to talk to these things so they actually do what you want. It’s less mystical than the name suggests, and more impactful than you’d expect.

The Convenience Trap: How AI Consumer Tools Could Be the Death of Critical Thinking — Right When We Need It Most

cantechit ai , , , , ,

Part of the ongoing AI series at cantechit.com

Let me start with something that is going to sound like a hot take but I promise it isn’t: AI doesn’t hallucinate. Not really.

I know. Bear with me.

Ever since ChatGPT blew the doors off in late 2022, “hallucination” has been the word everyone reaches for when an LLM makes something up with absolute confidence. Scholars are now pushing back on that framing — and honestly, they’re right. Cognitive scientists and AI researchers argue the more accurate word is confabulation — a psychiatric term for when a brain (or a system that mimics one) fills in narrative gaps with plausible-sounding fiction it genuinely believes is true. It’s not lying. It’s not hallucinating in the clinical sense. It’s pattern-matching on flawed or incomplete ground truth and arriving at a confident, wrong answer.

Sound familiar? It should. Because humans do this constantly.

Every time someone doubles down on a completely false “fact” they picked up from a bad source twenty years ago, that’s confabulation. Every time a witness in a courtroom sincerely misremembers what they saw — their brain filling in details that feel true — that’s confabulation. The neurons in your brain and the mathematical structures inside a large language model are working on the same fundamental principle: reasoning forward from whatever ground truth they were trained on. If the training data is garbage, the output will be wrong. Confidently, convincingly, devastatingly wrong. The AI isn’t uniquely broken. It’s human in a very uncomfortable way.

So stop calling it hallucination. Start calling it what it is. And then — more importantly — ask yourself what that means about how much you should trust either source.

The Human Context Window: Our Last Real Edge

Here’s where I want to push back on the doomers a little, because there is one genuinely huge advantage the human brain has over every AI system that exists today: our context window is enormous.

Every LLM has a context window — the amount of information it can hold “in mind” at once during a session. GPT-4 can handle about 128,000 tokens. Claude goes up to 200,000. That sounds huge until you realize your brain is running a context window that incorporates decades of lived experience, emotional memory, embodied knowledge, real-world cause-and-effect, and deeply personal relationships — all simultaneously, all the time, all connected to each other in ways that no model can currently replicate.

Yes, an AI has access to a much larger static memory — the entire training dataset — but it has to recall and retrieve from that pool every time, with no persistent working memory across sessions by default. You carry your context window with you everywhere. It never logs out.

That’s not a small thing. That’s arguably the biggest structural advantage humans have right now, and I don’t think we talk about it enough. When I’m making a judgment call at work, I’m not just running that query against a database. I’m drawing on thirty-some years of pattern recognition, failure, intuition, and relationships. No model has that. Not yet.

The danger? We’re about to voluntarily give up the benefits of that context window by outsourcing the thinking that keeps it sharp.

We’ve Been Here Before. And We Mostly Got It Wrong.

If you want a preview of where this ends up, go back to 2008. Technology writer Nicholas Carr wrote a now-famous Atlantic Monthly cover piece called “Is Google Making Us Stupid?” — arguing that the ease of online search was rewiring how we concentrate and think deeply. It blew up. People pushed back hard. The debate raged for years.

Here’s my honest take on that debate, looking back from 2026: both sides were partially right, but they were asking the wrong question.

Google didn’t make us stupider. But it absolutely changed the nature of the critical work required. When you Googled something back in 2005, you’d get ten blue links. You had to evaluate which site was authoritative. You had to triangulate. You learned to recognize the signs of a credible source versus the signs of someone’s angry SEO-optimized blog. The old joke was “skip to page 3 — the first page is all ads.” That joke existed because people were doing the work of curation. It was like going to the library, pulling twelve books off the shelf, and deciding which one to trust. That’s not passive. That’s active, effortful, deeply human reasoning. Even if now the list of links is manipulated by an algo.

The Pew Research Center documented this debate extensively in 2010 — and one of the most interesting takeaways was that the people defending Google argued it would force people to develop better critical thinking because there was simply more information to sift. And for a while, in the right hands, that was even true.

The shift didn’t happen with Google. The shift is happening now.

Google AI Overviews: When the Library Stops Asking You to Read

Pull out your phone and search for something — anything remotely answerable — and there it is: a tidy Gemini-generated paragraph at the top of the results. Google’s AI has already done your thinking. It has synthesized the sources, drawn a conclusion, and presented it to you wrapped in a bow. No links to evaluate. No competing claims to weigh. Just: here is the answer.

A 2025 Pew study found that roughly 18% of all Google searches in March 2025 triggered an AI Overview, and that question-based searches (“who,” “what,” “why”) triggered them 60% of the time. That’s not a niche feature. That’s the default experience for the majority of how people actually phrase searches.

Here’s the thing: it is extremely convenient. I’m not going to pretend it isn’t. I use it. You use it. We all use it. The problem isn’t the convenience. The problem is that convenience is actively removing the step where you had to do critical work.

New research published in 2025 tracking the rise of AI search found that AI-answered search queries surface significantly fewer diverse sources, lower response variety, and in some topic areas, a measurable lean toward particular information ecosystems compared to traditional search. Less variety. Less exposure to competing claims. Less opportunity for you to notice when something smells off – and more opportunity for corporate and government interference in the answer.

And critically — the research showed that people believe AI answers more readily than they believe a random link. The authoritative, confident, grammatically perfect paragraph at the top of the page registers to the human brain as trustworthy in a way that a link to a mid-tier website never did. This is a real problem. The very thing that made Google-era search retain some critical thinking — the requirement to evaluate sources — is being sanded away.

The Research Is In. It Doesn’t Lie.

This isn’t just instinct. The studies are piling up and they’re not subtle.

A January 2025 study published by researcher Michael Gerlich surveyed 666 participants and found a strong negative correlation between frequent AI tool usage and critical thinking ability — mediated specifically by cognitive offloading. When you hand a cognitive task to an AI, you don’t just save time. You deprive yourself of the practice that builds the underlying skill. The muscle doesn’t fire; it doesn’t grow. Younger users (17–25) showed the highest rates of AI dependence and the most pronounced effects.

Microsoft Research and Carnegie Mellon went further in a 2025 study of 319 knowledge workers across 936 real-world AI use cases. The finding that stuck with me: higher confidence in AI was directly correlated with less critical thinking. The more you trust the tool, the less you question it. And the less you question it, the less you actually think.

Seventy-two percent of participants in knowledge tasks reported “much less” or “less” cognitive effort when using AI. Seventy-nine percent in comprehension tasks. Seventy-six percent in synthesis tasks.

Think about that for a second. Synthesis — taking disparate pieces of information and building something coherent — dropped by three quarters. That’s not productivity. That’s atrophy.

When Confabulation Has Real Consequences: The Hall of Shame

Let me give you some examples of what happens when nobody stops to question the AI. Because this isn’t theoretical. People are getting hurt, sanctioned, and embarrassed — at every level of society.

The most famous one you’ve probably heard of. In 2023, New York attorneys Steven Schwartz and Peter LoDuca submitted a legal brief in Mata v. Avianca that cited six federal court decisions. The opposing counsel couldn’t find any of them. Neither could the judge. They didn’t exist — ChatGPT had invented them wholesale: fake case names, fake docket numbers, fake quoted passages, fake procedural history. All of it. When questions arose, Schwartz did the most human possible thing: he asked ChatGPT whether the citations were real. ChatGPT confirmed they were. They weren’t. Both attorneys were sanctioned. The judge called it abandonment of their professional responsibilities. The detail that haunts me is Schwartz’s testimony: “I just was not thinking that the case could be fabricated.” That is cognitive offloading with a $5,000 fine attached.

But it doesn’t stop in courtrooms. Here are three more:

Air Canada’s chatbot told a grieving customer the wrong thing — specifically, it misrepresented the airline’s bereavement fare policy to a customer who had just lost a family member and was trying to book a last-minute flight. The customer relied on what the chatbot said, booked accordingly, and was then denied the discount. Air Canada argued in court that its chatbot was a separate legal entity and it wasn’t responsible for what it said. A Canadian tribunal disagreed. Air Canada lost. The defense of “the AI said it, not us” did not work — and it won’t in the future either.

South Africa withdrew its entire Draft National AI Policy in 2025 — just 17 days after publishing it — because the document cited fake research generated by AI. This was supposed to be a historic moment: the first African nation establishing a formal AI ethics board. Instead it became the first government in history to pull an official policy document because of AI confabulation. The communications minister was direct: the most plausible explanation was AI-generated citations included without verification. “Consequence management” was promised for those responsible.

The Chicago Sun-Times published a summer reading list in May 2025 recommending books that don’t exist — fake titles, fake descriptions, attributed to real authors who never wrote them. Out of 15 books on the list, only 5 were real. The list came from a content partner who admitted to using AI to generate it. People bought the print edition. There was no correcting that.

These aren’t edge cases. A legal scholar’s database tracking AI confabulation in judicial proceedings has catalogued over 200 cases globally — and that’s just the courtroom. The pattern across all of them is identical: someone trusted the confident, fluent, authoritative-sounding output without stopping to verify. The AI didn’t lie to them. It confabulated — exactly like a brain working from flawed ground truth. And nobody applied the critical thinking that would have caught it.

But — Some Muscles You Were Never Going to Flex Anyway

Here’s where I want to complicate my own argument a little, because I’d be a hypocrite not to.

I am never going to learn to code. I’ve made peace with that. I tried. I didn’t enjoy it. I didn’t continue. And because I didn’t continue, I couldn’t build the things I wanted to build — which was genuinely frustrating for years. AI changed that. I am now building things that are real and functional and mine. Here’s a teaser: this whole AI journey — all the reflection, all the late nights, all the agents I’ve babysitted — has quietly turned me into a builder who wants to give something back. I’ve been spending my spare time building something for the rally community. Regional events, TSD, the kind of grassroots racing that most people outside the sport never hear about — not WRC, not national-level anything, just the people who show up on a cold Saturday morning because they love it. It’s coming together in a way that genuinely surprises me, and I can’t wait to share it. Whether it snowballs beyond that little community or stays exactly where it belongs — either outcome is fine by me. Watch this space. The point is: the muscle-I’m-not-flexing argument has a real counterargument. Some muscles you were never going to flex. The question isn’t whether AI is letting you skip work you would have done — it’s whether it’s letting you skip work you needed to do.

And here’s the nuance I don’t see in any of the research papers: even in my AI-assisted building, I’ve been forced to pick up real architecture and structural knowledge. Not syntax. Not line-by-line debugging. But the big stuff — how systems talk to each other, where failure points live, what good design looks like from ten thousand feet. You can’t direct an AI agent well without developing some of that understanding. Which means the cognitive offloading isn’t total, even when you want it to be.

But. And this is a big but.

Prompt engineering is a critical thinking skill. Full stop. I’ve said this before and I’ll keep saying it: it doesn’t matter what part of AI you’re using — better prompts produce better output, period. And in development specifically, this is where the critical thinking stakes get very concrete very fast.

Coding agents, once they get confused on a complex problem, will loop. They’ll circle the same approach, restate the same error in different words, try the same fix from a slightly different angle, and burn your credits doing it. I watch my agents do this regularly on anything sufficiently complex. And the only thing that breaks the loop is a human stepping back, recognizing the pattern, and redirecting — sometimes firmly. “Stop. What you’re doing isn’t working. Let’s look at this from a completely different angle. Forget your last three approaches.” That’s not a technical skill. That’s critical thinking applied to a system that can’t apply it to itself. I’m considering if I can prompt engineer or persona engineer my manager agents to do this for me – I think I can.

Without that ability — without the instinct to recognize when a process has become circular and needs a reset — you will spend days in a rat hole that an AI dug and then helpfully continued digging, deeper and deeper, while confidently assuring you it was making progress. That costs time. It costs credits. And it produces nothing. The human in the loop isn’t optional. The thinking human in the loop definitely isn’t.

We know AI systems confabulate. We know they produce confidently wrong answers with the same smooth delivery as correct ones. We know that the instinct to critically evaluate that output is being trained out of us by the very experience of using these systems. And yet we are deploying them everywhere, for everything, with very little friction between “AI said it” and “I believe it.”

The liability exposure for organizations letting AI-generated answers flow into real decisions without a critical review layer is real. That’s a separate post. But it’s coming.

Bibliographies Were Never Just About Citing Sources

Here’s something I’ve been thinking about that I don’t see anyone talking about enough.

When I was in high school, every essay came with a requirement: a bibliography. A list of every source you referenced. You had to name the book, the author, the edition, the page number. You had to go find those sources, read them, decide if they supported your argument, and then show your work. It felt like busywork at the time. It was not busywork.

That process — finding a source, evaluating whether it was credible, deciding if it actually said what you thought it said, and then correctly attributing it — was a critical thinking exercise disguised as an administrative task. It forced you to triangulate. It forced you to confront the question of authority: who said this, why should I believe them, and does this actually back my claim? It’s almost exactly the same instinct I try to apply when I include authoritative links in these posts. Same idea, different decade.

Now think about what a student does today. They open a chat window, describe the essay topic, and receive five hundred words of confident, grammatically flawless, internally consistent text — with plausible-sounding “sources” that may or may not actually exist in the form described, or at all. The bibliography is still there. The critical thinking that produced it isn’t.

And here’s the part that should bother everyone: plagiarism detection is now effectively broken. For decades, academia fought academic dishonesty with tools like Turnitin — databases that compared submitted work against indexed text to find copied passages. That arms race had a clear enough battleground: human-written text is distinctive, and copying leaves fingerprints. LLM-generated text leaves no fingerprints, because it isn’t copied from anywhere. It’s synthesized. Every output is technically original at the token level even when it’s conceptually hollow.

Worse — even the best AI-detection tools are increasingly unreliable. A student who wants to defeat a detector doesn’t even need to be sophisticated: rephrase a few sentences, shift the tone slightly, change some word choices, regenerate a paragraph. Most detection systems fold. The cat-and-mouse game that took decades to play out with traditional plagiarism has been compressed into a matter of months, and the mouse is currently winning by a country mile.

This isn’t just an academic integrity problem. It’s a signal of something deeper. The reason we required bibliographies and cited sources wasn’t really about catching cheaters. It was about teaching people to build a position on evidence — to understand that an argument is only as strong as the ground truth it stands on. Strip that requirement of its teeth, and you strip the lesson along with it. What you’re left with is a generation that can produce the form of a well-reasoned argument without any of the cognitive infrastructure behind it.

That’s not a skill deficit. That’s an infrastructure failure. And it’s going to compound.

The Education System Has to Step Up. Like, Yesterday.

I’ve been circling this for a while in this series. I said in the Drug Dealers post that broad access to AI matters for humanity. I said in Falling Behind that the pace is creating real anxiety. But I haven’t said this clearly enough yet: the education system has a bigger job right now than it has had in decades, and I’m not sure it knows it yet.

College Board surveys from 2024–2025 found that teachers overwhelmingly worry AI will impede students’ critical thinking development — while a meaningful share of students aren’t even sure what the problem is. That gap is the problem. Students who have grown up with the answer-at-the-top-of-the-search-results experience don’t have a felt sense of what they’re missing. You can’t miss a muscle you never developed.

Even the White House addressed this in April 2025, calling for AI literacy as a foundational skill from kindergarten through post-secondary. That’s good. But I’d argue AI literacy and critical thinking aren’t the same thing — and we need both. Knowing how to use AI well requires knowing how to question AI well, and that’s a critical thinking problem at its core.

Here’s my genuine take: we should be teaching a standalone critical thinking course starting in high school. Not embedded in English. Not a module inside social studies. A real, credited course — call it nothing but Critical Thinking — that teaches source evaluation, logical fallacies, argument structure, identifying confabulation (human and artificial), and the mechanics of how to form a well-reasoned position when everything around you is trying to form it for you. Maybe that sounds old-fashioned. I’d argue it sounds exactly like what 2026 requires.

The Throttle Metaphor (You Knew It Was Coming)

There’s a point in a rally stage — I’ve been there — where you’re approaching a hairpin and the temptation is to keep the throttle buried because it feels like commitment, like speed, like forward motion. The car feels alive. The numbers feel good. But the driver who survives the stage is the one who knows when to lift — who has the situational awareness to recognize that momentum without judgment is just a faster crash.

AI convenience is the throttle. And right now, most of us — me included, I will be honest — are keeping it buried because it feels amazing. Output is flying. Time is compressing. The machine is singing. I wrote about this in the Acceleration Trap post.

But critical thinking is the lift point. It’s not weakness. It’s not inefficiency. It’s the thing that keeps you on the road. And we are, collectively, conditioning ourselves to be worse at it, faster than we’ve ever conditioned ourselves out of any skill before.

The answer isn’t to stop using AI. That ship has sailed. The answer is to be deliberate — almost aggressive — about preserving and rebuilding the habits of skepticism, source-checking, and independent reasoning that convenience wants to replace. Use AI to do the work. Use your brain to decide if the work is right.

The Bottom Line

We laughed at the “Is Google Making Us Stupid?” debate in 2008 and moved on. We mostly got it wrong — not because Google didn’t change cognition, but because the critical layer was still there in the search results. You still had to pick.

That layer is eroding. Fast. A Frontiers in Education study from late 2025 found that students who over-rely on AI for answers show “substantial declines in analytical reasoning capabilities” — and that younger participants are the most vulnerable. This isn’t speculation. It’s in the data.

We are building faster on a foundation that may be getting softer. And the confabulating, confident, extremely smooth AI systems we’ve all become dependent on are both the cause and the reason we’re not noticing.

I don’t have the five-point fix. I never do in this series — you all know that by now. But I do think it starts with admitting the problem out loud. Demanding that our schools, our workplaces, and maybe ourselves take critical thinking seriously again — not as a soft skill, not as a nice-to-have, but as the fundamental human capability that everything else depends on.

Because here’s the thing: AI can confabulate. Humans can too. The only real defense against both is a mind that was trained to ask why before it asks what.

Your turn: Are you still questioning AI answers, or have you found yourself just… trusting them? Did you used to page through search results and pick your sources — and does that feel quaint now? And teachers or parents reading this — do you think a standalone critical thinking course in high school is worth fighting for? Drop it in the comments. I read every one, even if people are not commenting – I’m questioning, did you make it this far? Was it “TLDR” it’s ok, this is how I dump my raw thoughts on random junk – maybe i’ll get back to talking about actually “tech” stuff one day…

(Catch up on the series if you’re new: Falling Behind → AI Is Underpriced → Copilot Price Jump → Drug Dealers → The Acceleration Trap → Time Back → and now this one.)

The AI Acceleration Trap: Loving the Speed, Paying with Exhaustion, Insomnia, and Skipped Meals

cantechit Uncategorized , , , ,

Yes, I’m still writing about my AI builder journey, the most out of me in YEARS… I get it, I keep saying I would blog when I had something to say… I do now. Coding agents, multiple LLMs running in parallel, vibe-prompting entire features — it’s exhilarating. Search feels prehistoric now; why hunt for links when an LLM reasons through context and synthesizes faster? I open several chat sessions for the same project, treating them like a distributed team of specialists. One for architecture, one for debugging, one for research, one iterating on UI. Output is flying. Tools are built in days, not weeks.

And when I run a capable model locally on my desktop rig? The machine itself comes alive. The GPU spins up under heavy inference load, fans ramping into a deep, resonant hum. But the real star is the VRM buzz — the voltage regulator modules and their components straining to deliver hundreds of amps at precise low voltages. Inductors (coils) vibrate from rapid electromagnetic switching, while the multilayer ceramic capacitors “sing” via the piezoelectric effect as voltage ripples through them. The whole PCB acts like a speaker cone, turning electrical stress into audible whine and physical vibration you can feel through the desk. It throbs in rhythm with token generation, like the system is breathing hard and alive.

I even searched what this sentient-like buzzing and vibration was. Turns out it’s the VRMs and capacitors being heavily taxed under sustained high-wattage load — completely normal physics, not a failing part, but the hardware literally singing under stress as it powers matrix multiplications for hours on end. The rig pushes silicon to its limits right alongside me.

I love it. But it’s costing me. That same low-level buzzing has invaded my own head. Mental exhaustion hits harder and faster. There’s a constant mental static — a vibrating fog that mirrors the GPU’s VRM whine and the case’s resonant throb under load. Focus fragments. And recently, it’s spilled into real insomnia — lying awake replaying prompts and agent outputs — and straight-up skipping meals because I’m “in the flow” until my body protests. This isn’t sustainable. I’m not alone, and emerging research confirms it. This is showing the limit of how fast I can go given so much resource.

The Phenomenon Researchers Call “AI Brain Fry”

A March 2026 Harvard Business Review study from Boston Consulting Group (surveying 1,488 U.S. workers) named this exact experience: “AI brain fry” — mental fatigue from excessive use or oversight of AI tools beyond one’s cognitive capacity.

Participants described a “buzzing” feeling or mental fog, with difficulty focusing, slower decision-making, and headaches. High-oversight users (constantly reviewing, verifying, and orchestrating AI outputs) reported +14% mental effort, +12% mental fatigue, and +19% information overload.

That internal buzz feels eerily synchronized with my local rig’s physical vibration. The hardware’s coils and capacitors strain audibly and tactilely; my brain does the same — silently vibrating from the cognitive load of managing what feels like a team of tireless digital collaborators.

Managing multiple AI agents or chat sessions — exactly my workflow — amplifies it. The study highlights how juggling autonomous tools turns you into a full-time supervisor, not a creator. What was supposed to free time instead intensifies work through constant context-switching and evaluation.

About 14% of AI users experienced brain fry, with higher rates in creative/tech roles. Costs are real: +33% decision fatigue, +11% minor errors, +39% major errors, and 39% higher intent to quit.

The Human Cost: Insomnia, Skipped Meals, and Burnout Spillover — Echoes of Pi

My insomnia and meal-skipping aren’t unique. Earlier research linked frequent AI interaction to disrupted sleep and boundary-blurring. The always-on nature of these tools keeps the mental GPU spinning long after the physical fans wind down.

UC Berkeley/Yale ethnographic work showed AI adoption encourages multitasking, voluntary overwork, and blurred work-life lines. Productivity rises short-term; burnout follows, especially for power users running local models or parallel sessions.

This whole experience is starting to feel disturbingly like the movie Pi (1998) by Darren Aronofsky. If you haven’t watched it, stop reading and go do it — black-and-white, shot on 16mm, scored with a pounding industrial soundtrack that drills into your skull. It follows Max Cohen, a brilliant but tormented mathematician obsessed with finding hidden patterns in π and the stock market. He runs relentless computations on a straining home supercomputer setup while suffering brutal cluster headaches, paranoia, hallucinations, and social isolation.

The film is filled with motifs of buzzing, drilling, and static in Max’s head — auditory representations of his fracturing mind under obsessive load. Migraines come with auras, twitching, and a white-void escape, culminating in desperate self-surgery to “release” the pressure. The parallels hit too close: me chasing emergent patterns and breakthroughs through straining silicon (VRMs whining, GPU throttling), while my own neurons buzz with overload, replaying sessions at 3 a.m., skipping meals, feeling the isolation of deep flow states. I genuinely hope this isn’t foreshadowing my fate. The movie’s warning about the cost of pursuing ultimate truth through machines feels prophetic in 2026.

Why This Hits So Hard with Coding Agents, Multi-Chat, and Local Rigs

  • Search is obsolete for reasoning tasks → LLMs win on synthesis speed.
  • Multiple sessions = team management → Each chat has its own context and hallucinations to catch.
  • Local inference → The tangible VRM vibration, coil whine, and capacitor singing make the acceleration feel visceral and alive — until your own nervous system starts humming in sympathy.

This is the productivity paradox of 2026: AI amplifies output while intensifying the mental (and sometimes physical) work of oversight.

What Now? Sustainable AI Use (Lessons I’m Applying)

I’m not quitting AI — or my local setup. But I’m redesigning and adding guard rails:

  • Batch sessions and set hard cutoffs (including GPU cooldown periods).
  • Protect deep-focus blocks without any models running.
  • Prioritize AI for toil, reserve core judgment for me.
  • Build in recovery: walks, meals, no screens before bed.
  • Track personal metrics (energy, sleep, output quality and GPU/VRM temps).

This topic builds on my recent cantechit posts about vibe coding, agents, and the realities of AI adoption. The acceleration is real. So is the need for guardrails — before the buzzing becomes permanent.

What’s your experience? Local models making your desk (and brain) vibrate with VRM whine? Multiple agents burning you out, or found a sweet spot? Drop comments or links to your workflows.

Sources linked inline. Key reads: HBR on AI Brain Fry and related coverage.

AI Companies Are Straight-Up Drug Dealers

cantechit Uncategorized , , , ,

Free Samples, Then the Hook (And Yeah, I’m in Full Withdrawal)

Look, I wrote this right after a stupid billing glitch with my main AI provider nuked my agentic coding access for two full days. No deep reasoning chains. No autonomous agents refactoring my messy repos. No vibe-coding flow where I spin up tools, iterate like a madman, and ship in hours what used to take weeks.

I felt actual withdrawal. Irritable. Slow. Like my brain was running on dial-up while the rest of me knew what god-mode felt like. That’s when it hit me: these AI companies aren’t just selling tools. They’re dealers. And we’re all hooked.

The Classic Dealer Playbook – Free Samples to Get You Dependent

First they flood the zone with insane value. Generous free tiers, cheap Pro plans, agentic capabilities that feel like having 1-2 full-time devs in your pocket for under $50/month. You get hooked on the productivity superpowers. You start building faster, thinking bigger, shipping stuff you’d never attempt solo. Your workflow changes. Your expectations reset.

Then comes the squeeze.

Everyone’s raising prices. Or doing the sneaky version: shrinkflation on tokens. Same monthly fee, but the model “thinks” less, outputs dumber results, or burns through your quota faster. I’ve seen reports of Claude Opus variants suddenly using 67% fewer thinking tokens for the same tasks. Same price, worse output, more tokens consumed. Classic move.

They’re quietly shifting from flat subscriptions to consumption models – per-call, per-token, usage-based billing (UBB). GitHub Copilot just did it. Anthropic is pushing enterprise users toward metered API rates on top of seats. OpenAI and others are following the same pattern. They pretend the subscription is still the core, but the real work (agentic sessions, heavy reasoning, long contexts) now eats credits that vanish fast. No rollover. Burn mid-project and you’re stuck.

I also see coding agents getting stuck in a loop, when it makes a mistake – wait a second – I’m paying for this tool, and it got stuck in a loop and burned MY credits…. That doesn’t seem fair.

This isn’t random. It’s the plan.

The Humanity Angle: This Technology Must Stay Accessible

Here’s where I get pissed off at the dealer playbook. Let’s stop for just a second.

This shouldn’t just be for those who can afford the new premium rates. AI needs to stay cheap — or get cheap again — because humanity desperately needs broad access to this knowledge and capability.

Think about the 20-year-old kid who couldn’t afford college, grinding in a basement somewhere with a killer idea but no formal credentials. With affordable AI, that kid can prototype, research, iterate, and ship at a level that used to require a full dev team and venture backing. One good prompt chain and they’re competing with people who have more resources.

Or zoom out to emerging markets — places where even $20/month is a real barrier. A developer in Lagos, Jakarta, or rural India with limited local opportunities suddenly has world-class reasoning, coding help, and research at their fingertips. The talent and ideas bubbling up from those places could be the next massive breakthroughs. We’re talking exponential global innovation in emerging markets, if we don’t gatekeep this behind usage-based pricing that prices out the ambitious but broke.

Everyone should have access, not just Western professionals with corporate cards. The more minds hooked into these tools, the faster we solve hard problems in climate, biotech, energy, education — you name it. Imagine the compound effect: millions of new builders creating tools, businesses, and scientific insights we can’t even picture yet. That’s the real moonshot. Locking it behind ever-rising costs kills that potential.

I remember listening to Jake Hirsch-Allen talk about democratizing AI, building a coalition of open democratic economies. Investing in Public AI “The CBC of Compute” he calls it. I recall that line in “Anti-Trust” (A horrible 2* tech film from the 2000’s filmed in British Columbia) “Human knowledge belongs to the world” – Ryan Phillippe said on screen.

Why This Was Always Coming (And Why It’s Not Pure Evil)

Let’s be real for a second – I’m not some conspiracy nut. The economics make sense.

Frontier models are insanely expensive to run. Inference costs, HBM memory shortages, power bills, massive CapEx – the whole industry is bleeding cash even as they raise billions. Demand exploded. They loss-led hard to build market share and moats while compute caught up. The “basically free god-mode AI” era was the sample pack. Now they’ve got us dependent, the bills have to get paid.

Agentic workflows aren’t cheap chats. A multi-hour autonomous coding session that loops, reasons, tools, and iterates eats serious compute. Companies can’t subsidize that forever while memory prices go parabolic and everyone wants Opus-level reasoning.

So yeah – prices up, or shrinkflation, or straight usage-based. Pick your poison. GitHub’s recent Copilot changes (credits instead of unlimited premium requests, Opus multipliers jumping hard) are just the visible tip. More will follow.

We’ve been underpaying for this superpower. Shipping in days what used to need teams? Turning weekend experiments into real tools? That value is still ridiculous compared to hiring humans. But the golden age of predictable flat-rate unlimited was always temporary.

What This Means for You (Practical Moves)

Diversify like your workflow depends on it (it does).

  • Stay flexible: Tools like Continue.dev (bring your own keys) or OpenRouter let you pick the best model without middleman markups. Consider holding accounts with more than one provider, another subscription might be cheaper than overage charges (think 2 bags on the aircraft is cheaper than one overweight one)
  • Watch the credits: Set hard budgets. Test heavy sessions in advance so you don’t get stranded mid-project. Don’t let flows run endlessly.
  • Hybrid stack: Keep one smooth native tool (Cursor, Copilot, Windsurf) for daily flow, but route deep agentic work to raw APIs where you control cost.
  • Build habits: Use AI to multiply your thinking, not replace it. The real moat is knowing when to guide, when to verify, and when to ship.
  • Prompt engineer: Better prompts = better output = more context = less tokens. Learn to plan projects, and use quality prompts, this will save both time, money and of course – Tokens.

I’m still all-in on AI. Sitting with my kid building stuff is next-level. The superpowers haven’t disappeared – the pricing is just normalizing to reality. How can we make this accessible, don’t gatekeep it behind rich corporations. Those who have this power will become superpowers.

The dealer’s got us. Time to use the high productively while we figure out the new economics.

What do you think? Hooked yet?

GitHub Copilot AI Coding Price Jump!

cantechit Uncategorized , , , ,

Hate to Say It… But I Called It: GitHub Copilot Just Went Full Usage-Based Billing – And Yeah, This Is Only the Beginning

If you read my post from April 10 — AI is underpriced, but not for long — you already know where this was headed. Mere days later, GitHub dropped the hammer: Copilot is switching to usage-based billing (UBB) on June 1, 2026. Base subscription prices stay the same, but everything agentic, chatty, or model-heavy now burns through GitHub AI Credits based on actual tokens consumed. No more flat-rate “premium requests” that let you vibe-code all month for one predictable price.

I’ve been living in Copilot (Pro+ tier) for my daily vibe work — that flow-state coding where I’m spinning up internal tools, refactoring messy repos, switching between Claude Opus for deep reasoning and faster models when I’m just iterating. It’s been stupidly good. For under $50/month I’ve effectively had 1-2 full-time devs in my back pocket. The value chain was already ridiculous. Now? It’s breaking.

What Actually Changed (And Why Opus Is the Big Gut Punch)

From the official announcement and FAQ:

  • Copilot Pro ($10/mo) → $10 in AI Credits
  • Copilot Pro+ ($39/mo) → $39 in AI Credits
  • Business/Enterprise get the same 1:1 ratio plus some promo credits for the first few months and pooled usage (nice for teams).

Code completions and Next Edit suggestions stay unlimited and free. Everything else — chat, agentic sessions, code review (which now also eats GitHub Actions minutes) — is metered by tokens. Models have multipliers. And yeah, the one everyone’s screaming about: Claude Opus 4.7 just jumped to 27x credits (from around 3x before). That single model you reach for when you need god-tier reasoning is now nine times more expensive in practice.

No rollover on credits. No more fallback to cheaper models when you hit limits. Burn your allocation mid-project and you’re either waiting for next month, buying more credits, or downgrading your output quality.

The Backlash Is Loud (And Fair – Maybe)

Head over to the GitHub community discussion or Reddit threads and it’s pure salt. Common themes:

  • “This is just a stealth price hike dressed up as ‘sustainability.’”
  • Power users doing real agentic work are staring at $30–$40+ sessions instead of one premium request.
  • Annual plan holders getting hit with multiplier changes before their contracts even expire.
  • “Why would I pay GitHub a markup when I can go direct to Anthropic or use Cursor?”
  • Plenty of “cancelling today” posts. Some devs already switched mid-rant.

I get the rage. We all got hooked on this insane value and now the bill is coming due exactly like I predicted. Companies can’t keep subsidizing frontier-model agentic coding forever while their inference costs explode.

Steelman: This Is Just the Beginning

Look, I’m not here to defend Microsoft. But let’s be real — this is the market finally catching up to reality, not some greedy cash grab in isolation.

Agentic workflows aren’t cheap. A multi-hour autonomous coding session that used to cost them the same as a quick chat question now actually reflects the compute it eats. OpenAI, Anthropic, and everyone else are bleeding cash on training and inference. Memory prices (HBM especially) are still going parabolic. Demand is insane. The golden age of “basically free god-mode AI” was always temporary — it was the hook. Now the moat is built and the bills have to get paid.

I said it two weeks ago with Perplexity Computer at $200/mo and I’m doubling down: we’ve been underpaying for this superpower. The value we’re still getting (shipping in days what used to take teams weeks) is absurd compared to hiring humans. Prices will keep ratcheting up across the board as the industry stops loss-leading to win market share. This isn’t the end — it’s the first big normalization wave.

So Where Do You Go From Here? My Current Shortlist

I’m not rage-quitting Copilot yet (still great for GitHub-native teams), but I’m diversifying hard. Here are the strongest moves that actually integrate cleanly with VS Code or VS Code-style IDEs, ranked for someone doing heavy “vibe coding” like me:

ToolTypeBase PriceUsage ModelModel FlexibilityBest ForWhy I’m Considering It
GitHub Copilot (new)VS Code / JetBrains extension$10–$39/moUBB AI Credits (no rollover)Their curated list (Opus now 27x)GitHub-heavy teams, simple autocompleteStill the smoothest native GitHub integration
CursorFull AI-native IDE (VS Code fork)Pro $20/mo Pro+ $60/mo Ultra $200/moIncluded credits (scales with tier)Multiple frontier models + custom agentsDeep codebase refactoring & agentic workClosest “vibe” replacement — built for this exact workflow
Continue.devOpen-source VS Code extensionFree (core)Bring Your Own Keys (pay providers directly)Literally any model — Anthropic, OpenAI, Grok, local, OpenRouterMax flexibility & cost controlZero markup, switch models mid-session, privacy-focused – but their plugin is buggy for me
Windsurf (Codeium)AI IDE / extensionPro ~$15–20/moMostly unlimited or generous quotasStrong agent layerTeams wanting speed + lower costFast autocomplete with growing agent features

My personal take right now:

  • If you want the absolute closest experience to old-school Copilot but better agentic orchestration → Cursor Pro ($20) is the move. It feels like someone finally built the IDE for AI instead of bolting AI onto VS Code.
  • If you’re a tinkerer who wants to stay in plain VS Code, pick the absolute best model for the job without paying anyone’s middleman markup → Continue.dev + Anthropic/OpenRouter keys. This is my new daily driver for pure flexibility. Costs me whatever the raw API charges and nothing extra. I tried it and had tons of issues with their plugin not working correctly though, so I might take a second look.
  • Heavy teams already deep in GitHub ecosystem → stick with Copilot for now, just set strict budgets and watch the preview bill like a hawk.

The era of one predictable $39 subscription buying you unlimited frontier-model coding sessions is over. But the era of AI replacing entire dev teams is just getting started — and the tools are only getting sharper.

I’m still building with my kid using this stuff. The superpowers haven’t gone away; the pricing is just finally reflecting what they actually cost to run.

What about you? Already cancelled Copilot? Switched to Cursor or Continue? Drop your setup and real-world monthly spend below — I’m genuinely curious how everyone’s adapting.

AI is underpriced, but not for long.

cantechit Uncategorized , , , ,

Perplexity Computer: $200 a Month Feels Like a Steal… But We’ve Been Getting AI Dirt Cheap

If you haven’t watched NetworkChuck’s video on Perplexity Computer yet, stop everything and check it: https://www.youtube.com/watch?v=G3jvn7n-68Y

Chuck and his kiddos built a full gaming website with it. Pure dad-tech gold. His reviews always cut through the noise and save me (and probably you) hours of digging. Thanks, Chuck — you rock.

That video got me thinking about sitting down with my own kid to design, engineer, and build real stuff with AI. Not just prompts — actual projects. Perplexity Computer feels like the perfect playground. But then the price hit: $200/month for the Max tier that unlocks the full agentic power.

And that’s the spark for this post.


Perplexity Computer Is Next-Level (But Is $200/Month Crazy?)

Launched Feb 25, 2026, Perplexity Computer isn’t another chatbot. It’s a digital worker that orchestrates 19 frontier models in parallel, spins up sub-agents, hooks into 400+ apps, runs background tasks for hours or days, and just gets shit done. Full details here: https://www.perplexity.ai/hub/blog/introducing-perplexity-computer


Chuck’s “crew” shipped a game site. Others are building dashboards, prototypes, and entire workflows that used to take teams weeks. One enterprise user reportedly compressed years of work into weeks.

The problem with all of these services is – “tokens” nobody can tell you how many tokens it costs to do X, you only find out once you start doing things, but as I have found out with services like GenSpark and Perplexity – credits go fast, and when you are 3/4 done and run out — they have you.

Yeah, $200/month stings at first glance, others are spending thousands per month with the extra tokens…. but.

We’ve Been Underpaying for God-Tier AI This Whole Time

Look at what we’ve built lately with Claude, GPT, Cursor, and early agents. I’ve personally shipped internal tools and automations that would’ve needed multiple full-time devs, designers, and PMs just two years ago. The human hours I didn’t spend hiring? Massive. The “too expensive to try” ideas that became weekend wins? Priceless. It seems almost too good to be true — but is it.

There is this natural disparity in my mind, which was why I am writing this….. If it cost me say $100K in engineering time to contract hire a bunch of people to build some form of application – but I can use AI to do nearly all of it for $1000 in AI tokens, that’s 1/100th the price… However those engineers are making money.

Meanwhile, the AI companies are bleeding cash. OpenAI has been posting multi-billion-dollar losses. Perplexity has spent more than its revenue on models and infra at times. The compute bill is insane.

Naturally – AI tokens and services are going to go up in price – they just have to.

Then OpenAI shut down Sora in March 2026 — a money pit they needed to kill so they could refocus compute on higher-priority stuff like coding agents and robotics. Details: https://www.nytimes.com/2026/03/24/technology/openai-shutting-down-sora.html

We got hooked on ridiculously capable AI at bargain prices. Perplexity Computer feels like the first big “okay, time to pay what it actually costs” moment – but I will argue they are still not charging enough. Same vibe with Claude — Anthropic has tightened free-tier access for heavy agent use (like third-party OpenClaw setups) and pushed people toward paid Pro/Max/API tokens. Classic addiction cycle: get everyone dependent, then the price catches up to reality. Still arguing it’s not reality but they know they can’t 10X the price overnight.

The Newcomer Worry — And Why Competition + Reality Still Makes Me Hopeful

How do junior engineers learn the fundamentals when AI handles so much of the grunt + magic? Valid fear. We rely on the critical thinking of engineers from their experiences, and AI does a better job when you guide it – but – how do you get that experience to guide it?

But here’s the flip: the barrier to creating has never been lower. My daughter (and Chuck’s kids) can now experiment with real design and engineering at a level that used to require expensive teams or years of school. Passion and curiosity suddenly matter more than raw syntax.

Competition is fierce — OpenAI, Anthropic, Google, xAI, and more. Competition is very good news: it should depress prices a bit and keep innovation humming.

But here’s the steelman reality check: How sustainable is that when infrastructure costs are exploding? AI demand has driven memory prices (especially HBM/DRAM) sky-high — 50%+ jumps quarter-over-quarter in early 2026, with supply locked into 2027. GPUs, data centers, power — everything is getting more expensive fast. The infra bill isn’t going down; it’s accelerating.

So prices will probably ratchet up over time as the companies stop subsidizing our addiction. But the value we’re getting is still absurdly high compared to the old world of human-only teams.

Bottom Line: We’re Not Overpaying — We’re (Starting To) Finally Paying Fair

Perplexity Computer isn’t overpriced. It’s the wake-up that the “basically free god-mode AI” golden age was always temporary. The companies subsidized it to hook us and build moats. Now the bill is coming — but the superpowers we’re getting in return are still a steal.

I’m buying in. Not just for the productivity, but because I want to sit next to my kid and say, “Let’s build something cool together.” The future isn’t replacing humans — it’s giving every human (kids included) superpowers. This is bringing incredible higher value work to the human race. During the industrial revolution everyone worried about jobs, when 80%+ of humans made food so we could eat – often by hand. Now with industrial manufacturing and technology, we have more time for better innovation. This is no different. Will we see job disruption – of course, but this means a whole new era of innovation.

What do you think? Is $200/month worth it for Perplexity Computer? Felt the “addiction then price hike” yet? Drop your takes below.

Cisco AI Summit – More players, More Innovation.

cantechit Uncategorized , , , , ,

If 2025 was the year of AI experimentation, 2026 is officially the year of AI infrastructure. Yesterday, I had the chance to tune into Cisco’s second annual AI Summit, and let me tell you—the energy was different this time. Moving past the “what if” and straight into the “how fast.”

With over 100 industry heavyweights in the room and a staggering 16 million people watching the livestream, Cisco’s Chair and CEO Chuck Robbins and CPO Jeetu Patel didn’t just host a conference; they hosted a state-of-the-union for the trillion-dollar AI economy. Here are some of the things I found most interesting.

Intel’s “Shot Across the Bow”: The GPU Announcement

The biggest shockwave of the day came from Intel CEO Lip-Bu Tan. In a move that clearly signals Intel is tired of watching Nvidia have all the fun, Tan officially announced that Intel is entering the GPU market.

I am personally bullish on this, early in the AI era, I worked with some of Intel’s FPGA’s and some of their other OpenVINO platforms, along with many other accelerators. At least in my experience, they build some very solid, but more importantly very energy efficient accelerators.

This isn’t just a “me too” play. Intel has been quietly poaching top-tier talent, including a new Chief GPU Architect (rumors are that they got someone good too) to lead the charge. Tan was blunt about the current state of the market, noting that there is “no relief” on the memory shortage until at least 2028. By moving into GPUs, Intel is looking to solve the “storage bottleneck” that currently plagues AI inference.

The Efficiency Edge: My personal contention here? This is where the power dynamic shifts—literally. While Nvidia continues to push the envelope on raw compute, their chips have become notoriously power-hungry monsters. Intel, conversely, has a track record of building accelerators that prioritize performance-per-watt. In an era where data center expansion is being throttled more by power grid constraints than by floor space, Intel’s “lean and mean” approach could be their ultimate differentiator. If they can deliver high-end GPU performance without requiring a dedicated nuclear plant to run them, they won’t just be competing with Nvidia; they’ll be solving the very sustainability crisis the AI boom has created.

For the enterprise, this is huge. Competition in the silicon space means more than just lower prices; it means specialized hardware that might finally catch up to the insane demands of agentic AI – at lower energy cost.

70% of Cisco’s Code is AI-Generated (But Humans Still Hold the Pen)

One of the most eye-opening stats of the day came from Jeetu Patel: 70% of the code for Cisco’s AI products is now generated by AI.

Read that again. The very tools we are using to secure the world’s networks are being built by the technology they are designed to manage. However, Cisco isn’t just letting the bots run wild. Jeetu was very clear that while AI is the “teammate,” human reviewers are the “coaches.”

The philosophy here is “AI as a teammate, not just a tool.” It’s a subtle but vital distinction. By using AI to handle the heavy lifting of code generation, Cisco’s engineers are freed up to focus on the “Trust” layer—which was a recurring theme throughout the summit. As analyst Liz Miller noted on X, it’s one thing to use AI in security, but it’s an entirely different (and more important) game to secure the AI itself.

The Sam Altman Paradox: Efficiency Equals… More Consumption?

Finally, we have to talk about Sam Altman. The OpenAI CEO sat down for a fireside chat that touched on everything from drug discovery to supply chain “mega-disruptions.” But the comment that stuck with me was his take on the economics of AI growth.

There’s a concept in economics called the Jevons Paradox: as a resource becomes more efficient to use, we don’t use less of it; we use way more. Altman essentially confirmed this is the future of AI. No matter how efficient we make these models—no matter how much we drive down the cost of a token or the power consumption of a data center—humanity’s appetite for intelligence is bottomless.

“People just consume more,” Altman noted. As AI becomes cheaper and faster, we won’t just do our current jobs better; we will start solving problems we haven’t even thought to ask about yet. It’s a bullish outlook, but one that puts an even greater spotlight on the infrastructure constraints Chuck Robbins and Lip-Bu Tan spent the morning discussing.

Justin’s Take

Here’s what I’m chewing on after the summit: We are entering the “Great Optimization” phase of AI. For the last two years, we’ve been throwing money and electricity at the wall to see what sticks, with questionable profit models and circular economies (insert comment about AI Bubble here). But between Intel’s focus on energy-efficient accelerators and Cisco’s move toward AI-assisted (but human-governed) development, the industry is finally growing up.

But “growing up” also means things are getting weird. If you want to see the “art” of how crazy AI can get, look no further than Moltbook—the AI-only social network that’s been the talk of the summit – which also just has a major security breach. We’re seeing AI agents gossiping about their human owners and even inventing parody religions like “Crustafarianism.” While Altman dismisses it as a “fad,” the underlying tech of autonomous agents is very real, and it’s moving faster than our ability to regulate it.

This brings me back to a drum I’ve been beating for a long time: Responsible use, education, and ethics are not optional. As I wrote back in November, Deepfakes kill, and we need to make them criminal. I’m still waiting for the world to listen, but the summit only reinforced my fear that we are building the engine before we’ve tested the brakes. The real winner won’t be the company with the biggest model; it will be the one that can deliver intelligence and AI security at a sustainable cost—both financially and ethically. Altman is right—the demand is infinite. The question is, can our power grids and our trust frameworks keep up? Or will the agents just take over…

Agentic AI vs Deterministic Code

cantechit Uncategorized , , , ,

No question – Building apps with LLMs in agentic setups is a game-changer, but it can also be a pain in the butt compared to good old deterministic code. Craft an clever agent that summarizes docs or fixes bugs, then bam, the model updates, and suddenly it’s spouting nonsense, ignoring prompts or ignoring basic words like “yes”. Non-deterministic chaos at its finest.

Deterministic code? It’s the reliable workhorse: feed it input X, get output Y every damn time. Fixed rules, easy debugging, perfect for stuff like financial calcs or automation scripts where surprises mean lawsuits. As Kubiya nails it, “same input, same output”—no drama.

“A computer will do what you tell it to do, but that may be much different from what you had in mind”.  – Joseph Weizenbaum — Not when your using a model you probably didn’t build and not your own weights.

Agentic AI with LLMs? That’s the wildcard party crasher. These systems think on their feet: reason, plan, grab tools, adapt to goals like tweaking marketing on the fly or monitoring health data. IBM calls it “agency” for a reason—it’s autonomous, pulling from real-time vibes beyond rigid training. But here’s the kick: it’s probabilistic. Outputs wiggle based on sampling, context, or those sneaky model tweaks from OpenAI or whoever. LinkedIn rants about it: “Same prompt, different outputs.” Your app morphs overnight, and fixing it? Good luck tracing probabilistic ghosts.

This shift sucks for dev life. Traditional code: bug? Trace, patch, done. Agentic? Hallucinations, inconsistencies, testing nightmares. Martin Fowler compares LLMs to flaky juniors who lie about tests passing. It’s a paradigm flip—from control to “let’s see what happens.” Salesforce says pick deterministic for regulated certainty, agentic for creative flex. But non-determinism could means security holes, data risks, and endless babysitting. It also adds this attack vector that is really non-deterministic if the model has access to data that it needs to work – but I might not want exposed.

AspectDeterministic CodeAgentic AI with LLMs
PredictabilityRock-solid: Always consistentSketchy: Varies like the weather
AdaptabilityStuck to your rulesBoss: Handles dynamic crap
Testing/FixingSimple: Logic checks and patchesHell: Variability demands tricks
Best ForPrecision gigs (finance, compliance)Goal-chasing (support, optimization)
Pain LevelLow: Set it and forget itHigh: Constant surprises

Bottom line: Hybrids are the way—LLMs for the smarts, deterministic for the reins. Deepset pushes that spectrum view: not binary, blend ’em. It sparks innovation, sure, but don’t romanticize—the annoyance is real. Code with eyes open, or get burned. Put humans in the loop to keep things in check.

What about Agentic AI ops for network and technology? Didn’t we just say “precision gigs” are better with deterministic code? That won’t stop the likes of awesome developers like John Capobianco https://x.com/John_Capobianco from pushing those limits, and he has been doing that for years at this point. Handing AI agents the keys to critical stuff like network monitoring, anomaly detection, or auto-fixing outages. Sounds efficient, right? But it’s a powder keg from a security standpoint. These autonomous bad boys can hallucinate threats, expose data, or open doors for hackers through memory poisoning, tool misuse, or privilege escalation. Cisco nails the danger: “The shift from deterministic code to probabilistic chaos is at the heart of securing AI agents that think for themselves,” highlighting a “lethal trifecta” of data leaks, wild hallucinations, and infrastructure weak spots that could cascade into total meltdowns.

Tool are starting to emerging though for AI security, particularly from Cisco and open-source communities to advance defenses against threats like prompt injections and supply chain attacks, but there is work to be done. Things like Cisco’s open-source Foundation-sec-8B model, a specialized LLM for cybersecurity tasks such as threat intelligence and incident response, will help developers start to build customizable tools with on-prem deployments to reduce hallucinations and enhance SOC efficiency. Their Hugging Face partnership bolsters supply chain security with an upgraded ClamAV scanner detecting malware in AI files like .pt and .pkl. Broader open-source efforts include Beelzebub for malicious agent analysis and Promptfoo for LLM red-teaming, yet challenges from hackers with evolving adversarial tactics using LLM’s to attack LLM’s are very much a thing…. The system is attacking the system being protected by the system… Yeah that.

Cisco-Hugging Face ClamAV Integration: https://blogs.cisco.com/security/ciscos-foundation-ai-advances-ai-supply-chain-security-with-hugging-face
Cisco Foundation-sec-8B: https://blogs.cisco.com/security/foundation-sec-cisco-foundation-ai-first-open-source-security-model

So much more to learn, but with all of that said…. Humans in the loop is going to be a thing for awhile – at least until Skynet…