For many years, the only real option was to grind harder—longer hours, more staff, or cutting back in areas that hurt long-term growth.

But something has changed in the last few years. Artificial intelligence has moved from the boardrooms of Fortune 500 companies into the hands of everyday entrepreneurs.

And the question is no longer if it can help, but how small businesses can use AI to save money without burning more time or resources.

I’ve spent nearly twenty of my thirty years in IT leadership roles, and I’ve seen firsthand how the right technology can make or break a business.

In this article, I’ll walk through the real, practical ways AI is helping small businesses cut costs right now—with examples, tools, and a few personal observations from my own experience.

Automating Repetitive Tasks: Reclaiming Hours Every Week

When I started consulting with small businesses, the same theme came up again and again: owners and their teams were overwhelmed with repetitive tasks.

Answering emails, chasing invoices, and manually scheduling meetings—these tasks ate up hours every week.

Traditionally, the solution was to hire help. But wages, training, and turnover added new headaches. Today, AI-powered automation offers another option.

A chatbot on your website can handle 70% of customer inquiries without human intervention. AI-driven invoicing tools can automatically categorize expenses, send reminders, and reconcile accounts.

Email assistants now prioritize what matters most, so you’re not stuck sorting through clutter.

A few years ago, one café owner I worked with actually eliminated the need for a part-time admin role by switching to an AI-powered booking and payment system.

The savings? Roughly $18,000 a year—enough to invest in new kitchen equipment that boosted both capacity and revenue.

More Innovative Marketing Without Wasted Spend

Marketing is another place where costs can quickly increase.

I’ve seen businesses spend thousands on ads only to discover that most of the clicks came from people who were never going to buy. This is an example of where AI really shines.

AI-driven platforms analyze enormous amounts of data to pinpoint your ideal customers and can automatically adjust campaigns in real time. Instead of paying for impressions that go nowhere, you’re targeting people who are ready to buy.

Small retailers using AI-powered ad optimization regularly cut their budgets by 30–40% while maintaining the same sales volume.

Content creation is also a huge win. Instead of hiring freelancers for every product description or blog post, AI writing tools can generate drafts that you or your team can refine.

That doesn’t just save money—it saves weeks of turnaround time.

When you combine AI with SEO insights, the results get even better. Tools like SurferSEO identify precisely what your customers are searching for, letting you craft content that ranks faster and drives organic traffic.

That means fewer dollars wasted on ads and more free, qualified visitors coming to your site.

Protecting Against Cybercrime (Without a Full IT Department)

Cybersecurity used to be one of those “luxuries” small businesses couldn’t afford. You either rolled the dice or hoped your antivirus subscription would cover you.

The truth? One ransomware attack or data breach can wipe out years of hard work.

Thanks to AI security automation, we can see these tools levelling the playing field. Modern security platforms now include AI that monitors unusual activity and flags threats in real time.

Instead of paying an IT staffer to watch logs 24/7, the system quietly handles it in the background.

I know a law firm that nearly fell victim to ransomware. Their AI-based monitoring tool flagged an unusual login attempt early in the morning and locked the account.

If that had gone unnoticed until the next morning, the recovery costs could easily have reached tens of thousands of dollars. Instead, they paid a few hundred dollars a month for protection. That’s a trade-off any business owner can justify.

Better Inventory Control and Supply Chain Decisions

Inventory is one of those areas where small businesses often bleed cash without realizing it—too much stock ties up working capital. Too little means lost sales and unhappy customers.

AI forecasting tools help solve this problem by analyzing past sales patterns, seasonal demand, and even external factors like weather or economic shifts.

Instead of guessing how much to order, you’re making data-backed decisions.

There are many examples of e-commerce stores being able to trim their excess inventory by 20% after using an AI forecasting system. That frees up cash flow—money they can reinvest into launching new product lines.

Rethinking Hiring and HR Costs

Recruiting is expensive. Job postings, recruiter fees, and the time you spend interviewing add up quickly. And then comes onboarding, training, and paperwork.

AI is making these processes more affordable. Resume screening tools (or Applicant Tracking Systems) can sort through applicants in minutes, flagging the ones most likely to succeed.

Chatbots can guide new hires through onboarding forms and training modules. Some platforms even use predictive analytics to identify employees at risk of leaving—giving you a chance to fix the issue before turnover costs hit.

This can give small businesses the ability to bypass an agency recruiter and use an AI-powered hiring platform to level the playing field with companies that have larger budgets. Imagine as a hiring manager, skipping straight to the final interview and cutting down 90% of the legwork.

Cutting Utility Bills with Smarter Energy Use

This one often surprises people: AI can save small businesses money on utilities. Smart thermostats like Nest or Ecobee learn your patterns and reduce wasted heating or cooling.

AI-driven monitoring tools can identify when equipment is drawing more power than it should, flagging maintenance issues before they become breakdowns.

By learning your habits, such as when you are home during a workweek, and overlaying that with surge energy prices, systems can focus on either saving you money or keeping you at optimal temperatures.

Affordable Cloud AI Services

Finally, many business owners assume AI means expensive software licenses. In reality, cloud providers like Google, Microsoft, and Amazon have made AI tools accessible on a pay-as-you-go basis.

Want to add a chatbot to your site? Translate product descriptions? Analyze customer data? You can do it for pennies per use instead of thousands in upfront licensing.

That’s part of why we’re seeing huge deals at the enterprise level, like Meta’s $10 Billion Google Cloud deal. But the same infrastructure is available to you—scaled down to fit your budget.

Conclusion: Start Small, Save Big

The bottom line is this: AI isn’t some futuristic add-on for businesses with deep pockets. It’s already helping small businesses save money today.

Whether it’s cutting admin costs, reducing marketing waste, tightening up inventory, or avoiding cybersecurity disasters, the opportunities are real.

If you’re asking how small business can use AI to save money, the answer is simple: start with one process.

Automate your emails, try a chatbot, or switch on predictive inventory. Track the savings, reinvest them, and expand from there.

Want more context? Explore The Real History of AI or see GPT-5’s latest features to understand where things are headed next.

(Images created with the help of DALL-E.)

Frequently Asked Questions (FAQs)

And no, I’m not just talking about R&D salaries or office space in Silicon Valley. I’m talking about the staggering bills for cloud computing, the skyrocketing price of GPUs, and an environmental footprint that’s growing as fast as the models themselves.

The truth is, the AI boom isn’t just a race for innovation — it’s a compute arms race. And as we saw with the recent GPT-5 release, the stakes (and costs) keep climbing. The question we should be asking is: who’s paying for it, and at what price?

Tech It From Me

The Real Cost of AI: Who’s Paying for the Compute Arms Race?

Play Episode Pause Episode

Mute/Unmute Episode Rewind 10 Seconds 1x Fast Forward 30 seconds

00:00 / 13:58

Amazon Apple Podcasts Spotify YouTube

RSS Feed

Share

Link

Embed

1. The Money Trail: Cloud Bills and GPU Gold Rush

The financial side of AI isn’t just big — it’s massive. Training a cutting-edge model like GPT-5 is not a hobby project you run on a weekend. It’s an industrial-scale operation requiring thousands of high-end GPUs running non-stop, often for weeks or even months.

And these GPUs aren’t the kind you can pick up at your local electronics store. Nvidia’s H100s — and the soon-to-launch B200s — are the current gold standard. They’re powerful, scarce, and expensive. One H100 can sell for $30,000 if you can even find one. Even the biggest tech companies end up on waiting lists.

For many organizations, the solution is renting GPUs in the cloud from providers like AWS, Azure, or Google Cloud. But that comes at a price. Renting just one top-tier GPU instance can cost hundreds of dollars a day. Multiply that by thousands of GPUs for a month-long training run, and you’re looking at infrastructure bills in the tens of millions — before a single dollar is spent on testing, fine-tuning, or deployment.

It’s the same dynamic as a gold rush. In the 1800s, miners chased gold, but the real winners were the ones selling shovels. In today’s AI rush, Nvidia and the major cloud providers are those shovel sellers. They profit regardless of whether the miners — in this case, AI startups and research teams — ever strike gold. (Nvidia’s recent $4 trillion valuation is proof of just how profitable this role has become.)

And while big tech can absorb these costs, smaller startups often can’t. With limited funding and no guarantee of a return, many are forced to pivot. Instead of training models from scratch, they fine-tune existing ones like OpenAI’s GPT, Anthropic’s Claude, or open-source models like LLaMA and Mistral. This approach can level the playing field, but it also means true ground-up innovation is increasingly reserved for those with the deepest pockets.

If you want to understand why these systems require so much computing power, check out my article: Behind the Buzzwords: What Is a Large Language Model, Really?.

2. The Environmental Price Tag

AI doesn’t just consume money — it consumes massive amounts of natural resources.

Training GPT-3 used an estimated 1,300 MWh of electricity — enough to power 120 U.S. homes for a year. GPT-5 hasn’t had its numbers publicly released, but given its scale, it’s safe to assume the footprint is far larger.

And electricity is only part of the story. According to The New York Times, data centers use millions of litres of water annually to cool high-performance GPUs. In some regions, that means AI development is competing directly with agriculture and household water needs. The resource demands aren’t unlike what I saw during my 15 years in the engineered environmental sector — large-scale industrial operations with measurable local impacts.

The carbon impact of AI is equally concerning. Unless a data center operates entirely on renewable energy, every large-scale training run generates CO₂ emissions. Multiply that by dozens or hundreds of training cycles worldwide, and AI starts to look less like a “clean” technology and more like another heavy industry with a substantial environmental tab.

Even after training is complete, the environmental cost continues. Every query you send to an AI model — every prompt, every request — consumes compute power. The training phase is like constructing a massive factory, but inference (actually running the model) is like keeping that factory in constant operation.

3. The Ripple Effect: How Costs Come Back to You

Even if you’ve never trained an AI model in your life, you’re still part of the equation — because these costs eventually reach you.

The first ripple is subscription creep. AI tools that once offered free tiers are introducing “pro” plans, while existing paid tiers are hiking prices or adding usage caps. This isn’t just about profit — rising infrastructure and energy costs have to be recouped somewhere. If you’ve ever watched a technology service quietly increase pricing year over year, this will feel familiar.

The second is vendor lock-in. Once a company builds its AI systems on AWS, Azure, or Google Cloud, switching isn’t just inconvenient — it’s expensive. This lock-in gives cloud providers enormous pricing power. It’s a dynamic similar to what I discussed in From Y2K to AI: How IT Departments Have Changed and Where They’re Headed — once critical infrastructure is tied to one vendor, your negotiating leverage drops.

The third ripple is market risk. AI right now is in a hype cycle reminiscent of the late 90s dot-com boom. Back then, companies with little more than a website and a pitch were raising millions. When the bubble burst, valuations collapsed, and many companies — along with their customers — were left stranded. We could see something similar here if the cost of building and running AI systems outpaces the revenue they generate. For a look at how this plays out at the top end of the market, read Nvidia Hits $4 Trillion: How the AI Gold Rush Changed Everything.

Finally, there’s the indirect consumer cost. Even if you don’t pay for AI directly, it may be built into the products and services you buy. A retailer might use AI for demand forecasting, a bank might use it for fraud detection, and when their costs go up, they get passed along. In some cases, these increases are so subtle you won’t notice them — just like when cybersecurity investments (a topic I cover in How Cybercriminals Really Get Your Info) quietly become part of the price of doing business.

4. Can AI Be Cheaper and Greener?

The good news is that AI doesn’t have to be prohibitively expensive or environmentally damaging. There are paths to make it both cheaper and greener — but they require changes in both technology and business practices.

One approach is to rethink model design. Today’s most powerful AI models are generalists, designed to handle almost anything you throw at them. But smaller, specialized models can often perform just as well for specific tasks while using far less energy and compute. This kind of targeted efficiency is something I’ve seen work in other areas of tech, and it’s a recurring theme in discussions about open-source AI, where leaner architectures often win on cost.

Another promising direction is Mixture of Experts (MoE) architectures. Rather than running the entire model for every request, MoE systems activate only the “experts” relevant to a specific query. This selective activation significantly reduces compute requirements, which means lower costs and a smaller carbon footprint.

Then there’s infrastructure strategy—location matters. Data centers in cooler climates require less energy for cooling, and those built near renewable energy sources can operate with dramatically lower emissions. Google, Microsoft, and Amazon are all making public commitments to expand renewable-powered data centers, and smaller players can follow suit. For a deeper dive on the strategic side of tech decisions like this, see Build vs. Buy: Making the Right Tech Call Without Regret.

Open-source innovation is also part of the answer. Models like LLaMA, Mistral, and Falcon may not match closed systems like OpenAI’s GPT-5 in every benchmark, but they can be fine-tuned and deployed for a fraction of the cost. This democratizes AI development, making it accessible to startups, research labs, and even individuals with modest budgets.

Finally, policy and market pressure could accelerate the shift toward greener AI. In the same way government incentives boosted renewable energy adoption, policymakers could require transparent carbon reporting for large-scale training runs or offer benefits for using renewable-powered infrastructure.

The question is whether these sustainable practices will scale quickly enough. If not, we risk repeating the same pattern we’ve seen in other areas of tech: rapid growth followed by a reckoning over costs and impact. For context on how hype cycles can distort priorities, check out Watermarking AI: Will It Change the Way We Write Forever?.

Final Thoughts

The AI revolution isn’t just about algorithms, innovation, or flashy product launches — it’s also about the real costs hiding beneath the surface. From GPU shortages and cloud infrastructure bills to the environmental footprint of massive data centers, every leap forward has a price tag.

As we’ve seen, the compute arms race is pushing costs higher and making it harder for smaller players to compete. These expenses don’t just stay in the boardroom — they eventually show up on your subscription bill, your electricity grid, or even your local water supply.

But it doesn’t have to be this way. The solutions are already on the table: smaller, specialized models; energy-efficient Mixture of Experts designs; open-source alternatives like those in the open AI community; and a shift toward green data centers powered by renewables. The challenge is in getting both industry leaders and policymakers to adopt them at scale.

If we fail to address the financial and environmental realities, we risk building an AI future that’s innovative on the surface but unsustainable at its core. The winners of this race won’t just be those with the most advanced models — they’ll be the ones who can keep the lights on without burning through the planet’s resources.

So the next time you read about a breakthrough like GPT-5 or the latest AI startup making headlines, ask the real question: Who’s paying for all this compute? Because in one way or another, the answer might be you.

(Feature image generated with the help of DALL-E.)

Sam Altman says the leap from GPT-4 to GPT-5 is so big that going back feels “miserable.” But is GPT-5 really the iPhone moment for AI—or is it just the latest upgrade in an already fast-moving race?

I’ve already spent some time using the new version, and I can confirm: many of the improvements are noticeable right away. The model feels faster, more confident in its answers, and much better at staying on track in more extended conversations. It’s still not perfect, but it’s a more enjoyable and productive experience.

Let’s unpack what’s new, what’s useful, and whether this release changes the game.

What’s New in GPT-5?

OpenAI officially rolled out GPT-5 to all users on Friday, August 8, 2025, marking what the company calls its most advanced ChatGPT release yet. For Pro subscribers, the upgrade is even bigger—you now get access to GPT-5, GPT-5-Pro, and GPT-5-Thinking, each tuned for different levels of depth and reasoning.

So, what exactly has changed? Quite a bit:

1. Smarter, Context-Aware Reasoning
GPT-5 takes “chain-of-thought” logic to the next level, breaking down complex problems into smaller, logical steps—without you having to spell it out. This means more accurate answers, fewer logical gaps, and better handling of multi-part queries.

2. Dramatically Fewer Hallucinations
While no AI is perfect, GPT-5 shows a noticeable reduction in factual errors. In my testing, I’ve seen cleaner, more consistent outputs that stay true to source material—especially when summarizing technical or niche topics.

3. ‘Vibe Coding’ for Developers
This is a game-changer for coders. Instead of simply producing functional code, GPT-5 can now interpret how you want your code to feel—whether that’s clean and minimalist, performance-optimized, or playful and experimental. It bridges the gap between intent and execution.

4. Unified Model Routing
Say goodbye to manually switching between browsing mode, code interpreter, or GPT-4 for different tasks. GPT-5 automatically chooses the right capabilities in the background, making the experience seamless and letting you focus on your work.

5. Deeper App Integrations
You can now connect GPT-5 directly to tools like Gmail, Google Calendar, and potentially more productivity platforms in the near future. This turns ChatGPT from a standalone assistant into something closer to a true personal AI workspace.

6. Custom Personalities with Memory
Want an AI writing partner with your exact tone? Or a coding assistant who remembers your preferred frameworks? GPT-5 now allows you to create named GPTs with persistent memory, so each assistant gets better the more you use it.

From my own use, two features stand out: Unified Model Routing, which makes every interaction feel frictionless, and Vibe Coding, which is almost eerie in how well it understands intent. For anyone who works with AI regularly, these aren’t just quality-of-life improvements—they’re workflow accelerators.

Is This Really the iPhone Moment?

Every major AI release seems to get compared to a tech milestone—Apple’s original iPhone launch, the debut of the Retina display, or even the first personal computer. With GPT-5, the buzz is no different. Some have already called it the “iPhone moment” for AI, suggesting it could be a turning point in how we interact with technology (source: The Verge).

But is that really the case?

If we look at the AI timeline, GPT-3 was arguably the real “wow” moment for the public—it introduced mainstream users to just how human-like a language model could feel (see my breakdown of AI’s history here). GPT-4 refined that experience, making it more reliable and capable. GPT-5? It’s more of a polished evolution than a radical leap. Think going from an iPhone 13 to an iPhone 15 Pro Max—sleeker, faster, with smarter features, but not a reinvention.

Here’s where I think the hype meets reality:

• Yes, GPT-5 is better. You’ll notice faster responses, better reasoning, and smarter integrations right away.
• No, it’s not magic. The leap isn’t as dramatic as some suggest—at least not yet.
• The AGI hype needs perspective. We’ve been here before with Web3, metaverse promises, and blockchain revolutions that fizzled (see my take on AI hype cycles).

That said, I can tell from my testing: it’s hard to go back to GPT-4 once you’ve used GPT-5 for a while. It feels like a more responsive, more intuitive assistant—one that’s inching closer to being a proper daily driver for professionals.

Where GPT-5 Actually Helps

While the headlines focus on bigger, brighter, and more human-like, the real question for most of us is: Where does GPT-5 actually make a difference in day-to-day work?

After using GPT-5 extensively across IT, writing, and research tasks, I’ve seen several clear advantages over GPT-4:

1. For Developers
With vibe coding and improved reasoning, GPT-5 doesn’t just spit out functional code—it writes in your preferred style, anticipates potential pitfalls, and suggests optimizations before you even ask. This means fewer rewrites and cleaner commits (see my breakdown on coding with AI here).

2. For Writers & Content Creators
From blog posts to video scripts, GPT-5 delivers clearer drafts that stay on-topic and require less editing. Its improved contextual memory also helps keep tone and style consistent across longer pieces—something GPT-4 often struggled with.

3. For Researchers & Analysts
Whether you’re summarizing dense reports or pulling insights from large datasets, GPT-5 processes and condenses information faster, with fewer hallucinations. This means you can spend more time thinking critically about the results instead of verifying every line (related: How AI can mislead if unchecked).

4. For Busy Professionals
The new Gmail and Google Calendar integrations mean you can draft emails, plan meetings, and generate follow-up notes—all without leaving ChatGPT. It’s not just answering questions; it’s helping you act on them.

5. For Long, Complex Requests
One of GPT-5’s most underrated upgrades is focus retention. Even in multi-step prompts, it stays on track instead of drifting into irrelevant tangents. This is especially valuable for IT workflows where one missed step can cause significant delays (see my AI adoption advice for IT leaders).

Bottom line: GPT-5 isn’t replacing knowledge workers anytime soon—but it shifts your workflow into a higher gear. You still need to steer; GPT-5 makes the road smoother and the trip faster.

You still need to drive the process. GPT-5 shifts it into a faster gear.

What This Means for IT Leaders & Businesses

For IT leaders, CIOs, and business owners, the GPT-5 release isn’t just another AI headline—it’s a signal that practical AI adoption is accelerating. But here’s the key takeaway: you don’t need to overhaul your entire stack tomorrow. Instead, now is the time to strategically experiment, govern, and integrate.

Here’s what I’d be thinking about right now:

1. Employee Readiness & Training
Your teams may have access to GPT-5 already, but do they know how to use it effectively—and securely? AI productivity gains come when staff are trained not just in prompt writing, but in understanding limitations, bias, and data privacy (see my AI job impact breakdown).

2. Governance & Compliance
Smarter AI doesn’t mean safer AI. Attackers will adapt to new models just as quickly as businesses adopt them (Gartner warns about AI security risks). Consider:

• Do you need a formal policy for AI-generated code and content?
• How will you audit AI-assisted decisions for compliance and accuracy?
• Who is responsible for AI usage oversight in your org?

3. Integration into Existing Workflows
With GPT-5’s new app integrations (Gmail, Google Calendar, and more to come), you can start embedding AI where your teams already work. Early pilot projects can uncover efficiency wins without a massive system change.

4. Cost-Benefit Analysis
Upgrading to GPT-5 Pro for your team could be a productivity multiplier—but you need to weigh it against licensing costs, usage limits, and security trade-offs. Many organizations are experimenting with a “center of excellence” model where a small AI-proficient group leads adoption and shares learnings internally.

5. Industry Benchmarking
If your competitors are using GPT-5 to speed up product development, improve customer support, or automate analysis, waiting too long could mean falling behind (see my Y2K-to-AI leadership lessons).

In short, GPT-5 is an opportunity, not a mandate. The businesses that will benefit most are those that experiment deliberately, govern wisely, and train their people first.

Are We Closer to AGI?

Whenever OpenAI drops a major update, the AGI (Artificial General Intelligence) conversation comes roaring back. GPT-5 is no exception—some headlines are already hinting that we’re on the brink of machines that “think” like humans (MIT Technology Review covers the debate).

But let’s pump the brakes.

OpenAI itself defines AGI as “highly autonomous systems that outperform humans at most economically valuable work” (source). GPT-5, while more capable than its predecessors, is still a specialized AI—brilliant at certain language and reasoning tasks, but not independently creative, curious, or self-motivated in the human sense.

Why GPT-5 Feels Closer to AGI

• Improved reasoning chains mean it can handle more complex, multi-step requests.
• Better memory and personalization make it feel more “aware” of you over time.
• Unified tool routing gives it the illusion of adaptability—choosing the best mode automatically.

Why We’re Still Not There

• GPT-5 has no true understanding of the world—only statistical associations.
• It lacks generalized goals outside of responding to prompts.
• It can still produce confident but factually wrong answers (hallucinations), even if less often than GPT-4 (I’ve covered the risks here).

In other words, GPT-5 is a powerful narrow AI with more polish and adaptability, but AGI remains a moving target—one that experts can’t agree will arrive in 5 years or 50 (see AI timeline debates from Stanford HAI).

The danger in treating GPT-5 like AGI is complacency: over-trusting its outputs, underestimating bias, and skipping human oversight. The real opportunity is using it as an augmented intelligence tool—boosting human decision-making without replacing it.

Final Thoughts

GPT-5 is, without question, the most capable version of ChatGPT yet. It’s faster, smarter, more focused, and—thanks to integrations—more useful in everyday workflows. For developers, writers, researchers, and IT leaders alike, the leap in practical productivity is real.

But it’s also important to keep perspective. GPT-5 isn’t magic, and it’s not AGI. Like every AI upgrade before it, the novelty will eventually fade, and what will matter most is how effectively you integrate it into your work. The real winners in this AI race won’t be the companies with the newest tools—they’ll be the ones that train their people, set strong governance, and adopt AI deliberately.

If you’re curious about where AI might head next, I’ve explored that in The Real History of AI: From Turing to Transformers and my breakdown of AI’s impact on the future of work.

In the meantime, GPT-5 is here, and it’s ready to work—if you are. Whether it becomes the iPhone moment for AI or just another step on the road to AGI will depend on how we, as users and leaders, choose to apply it.

Related Posts from Tech It From Me:

• Behind the Buzzwords: What Is a Large Language Model, Really?
• Will AI Really Replace All Mundane Jobs? My Take on Geoffrey Hinton’s Bold Claim
• The Real History of AI: From Turing to Transformers

(Feature image generated with the help of DALL-E.)

Tech It From Me

Behind the Buzzwords: What Is a Large Language Model (LLM), Really?

Play Episode Pause Episode

Mute/Unmute Episode Rewind 10 Seconds 1x Fast Forward 30 seconds

00:00 / 20:50

We’ve all heard the phrase “large language model” (or LLM) thrown around lately. It shows up in meetings, on product pages, in keynote presentations—everywhere. And if you’ve nodded along while someone said, “We should be using an LLM for this,” but secretly wondered what they were talking about, this post is for you.

What Is a Large Language Model?

At its core, a large language model is a system trained to recognize patterns in human language. You give it a prompt, and it predicts what’s likely to come next based on everything it’s seen before—books, websites, documents, code, support tickets, Reddit threads…you name it.

The power of these models comes from their scale. GPT-4o, for example, has been trained on datasets containing hundreds of billions—if not trillions—of words. And because of that, it has seen nearly every way humans express themselves: formally, casually, helpfully, sarcastically, and everything in between.

So when you ask it a question, it doesn’t “think” like a person does. It uses statistical models to predict what a human-like response should sound like. That prediction engine is surprisingly good—so good, in fact, that many people start to mistake it for real understanding.

But LLMs don’t know anything. They just recognize patterns at scale.

Tokens, Transformers, and Embeddings (Explained Simply)

Let’s break down how this works.

Tokens

LLMs don’t read whole words. They read tokens—small bits of text like prefixes, roots, or suffixes. For example, the word “unbelievable” might be broken into “un,” “believ,” and “able.”

This helps models handle spelling mistakes, slang, and unfamiliar words. It also determines how much content the model can process at once. ChatGPT-4o has a token limit of 128,000 tokens—roughly the length of a 300-page book.

If you’ve ever had a long chat with an AI and noticed it “forgot” something you said earlier, it probably hit that token limit.

Transformers

Transformers are the engine of modern AI. Introduced in 2017 by Google, the Transformer architecture enables a model to comprehend context across large blocks of text. That’s why it can make sense of your whole paragraph instead of reacting word-by-word like older chatbots.

The transformer utilizes a technique called self-attention to emphasize key relationships. For example, in the sentence “She didn’t like the movie because it was too long,” the model needs to understand that “it” refers to “the movie.”

Without this architecture, LLMs wouldn’t be capable of the nuanced, responsive conversation we see today.

Embeddings

Embeddings are how the model understands meaning in math. Every word gets mapped into a multi-dimensional space. Words used in similar ways—like “nurse” and “doctor”—end up near each other.

This enables the model to respond to you with answers that are not only grammatically correct but also semantically relevant. It also powers many enterprise AI features like intelligent search, document similarity, and classification.

One famous embedding example: “king” – “man” + “woman” = “queen.”

Not Just Bigger—Smarter (With the Right Fit)

It’s easy to assume that bigger = better, but that’s not always true.

Yes, GPT-4o and Claude 3 Opus are incredibly powerful. However, they also require more computing power, are more expensive to run, and may introduce latency.

In many cases, a smaller model like Claude Instant or GPT-3.5 can be just as effective, especially when paired with tools like RAG (Retrieval-Augmented Generation). With RAG, the model pulls data from real sources (like your company docs or product manuals) in real time.

Think of it like this:

• Big model = “read everything once and make your best guess.”
• RAG-enhanced model = “go look it up before you answer.”

This is where AI becomes both useful and accurate.

How Businesses Are Using LLMs Today

You don’t need to be a tech company to get value from large language models. Every department has use cases:

• IT: Script writing, technical documentation, troubleshooting assistance
• HR: Job description creation, onboarding guides, policy summaries
• Marketing: Copywriting, brainstorming, SEO optimization
• Legal: Contract summaries, risk flagging, compliance checklists
• Finance: Invoice classification, expense categorization, data cleanup
• Customer Service: Chatbots, help desk responses, multilingual support

If your organization touches text (and let’s face it—whose doesn’t?), AI can help. But it needs the right governance, the right tooling, and clear human oversight.

Hallucinations Are Real. So Is Risk.

Despite how intelligent LLMs appear, they occasionally make mistakes. These errors—known as hallucinations—can be subtle or glaring. A model might misstate a fact, invent a statistic, or fabricate a source.

This is why the “intern” analogy works so well: treat the AI as a capable but untrustworthy assistant. Review everything it creates. And never put it in charge of final decisions without human validation.

Leaders should also consider:

• Data privacy: What are you sending to the model?
• Model bias: Is the output reinforcing stereotypes or gaps in training data?
• Reputation risk: What happens if it gives the wrong answer to a customer or partner?

Understanding how LLMs work helps you ask better questions and avoid costly mistakes.

Final Thoughts

Large language models aren’t magic. But they are one of the most important technologies of our time. And like any tool, they’re only useful when you understand how to use them.

You don’t need to code. You don’t need to build your own model. But you do need to be able to spot the difference between hype and real value. That’s the edge that separates reactive teams from future-ready ones.

Are you interested in learning more about the history of AI? Be sure to check out my article titled The Real History of AI: From Turing to Transformers.

Tech It From Me

The Real History of AI: From Turing to Transformers

Play Episode Pause Episode

Mute/Unmute Episode Rewind 10 Seconds 1x Fast Forward 30 seconds

00:00 / 26:03

Artificial Intelligence didn’t begin with ChatGPT — and it’s not just about futuristic robots or sci-fi fantasies. The story of AI spans nearly a century, marked by breakthroughs in logic, computing, and neuroscience that have gradually brought machines closer to mimicking human reasoning.

In this post, we’ll take a journey through the real history of AI — not the hype, but the turning points that shaped the field from its earliest days to the generative AI explosion we’re living through now. I’ll also reflect on how some of the concepts I’ve studied — from machine learning to AI ethics — tie back to these foundational ideas.

Let’s go.

The Birth of the Idea: Turing, Shannon, and the Dartmouth Spark

Our story begins in the early 20th century. Long before smartphones or neural networks, people like Alan Turing laid the groundwork with abstract ideas about computation. In 1936, Turing introduced the concept of a universal machine — what we now call the Turing Machine — which could simulate any other machine’s logic.

By 1950, he was asking the big question: Can machines think? His proposed answer — the now-famous Turing Test — became the first real framework for measuring machine intelligence.

Meanwhile, Claude Shannon was pioneering information theory, and John von Neumann laid out the architecture that modern computers still use today.

But the real milestone came in the summer of 1956, at the Dartmouth Conference. That’s where the term Artificial Intelligence was coined, and a bold proposal was made: every aspect of human intelligence could, in theory, be simulated by a machine.

It was a moment of genius — and hubris. The field of AI was born.

Early Wins and the First AI Winter

The years following Dartmouth were filled with excitement — and some surprisingly impressive early projects.

One standout was The Logic Theorist (1956), a program that could prove mathematical theorems, sometimes even better than the humans who originally wrote them.

Then there was ELIZA (1966), a text-based chatbot that simulated a therapist. While simple, it felt eerily human to many users, so much so that some formed emotional connections. That was one of the first ethical dilemmas AI ever presented: if a machine sounds intelligent, does that make it so?

But these systems couldn’t handle ambiguity or real-world complexity. Machine translation efforts failed spectacularly. By the 1970s, disillusionment set in, and funding dried up.

This period became known as the first AI winter, when progress slowed, skepticism rose, and trust was lost.

The Expert Systems Era: Smarts Without Learning

In the 1980s, AI staged a comeback through expert systems — software that encoded the decision-making logic of human experts.

One of the most well-known was MYCIN, a medical system that diagnosed infections. It performed well but wasn’t used in hospitals because it couldn’t explain why it made its decisions. That trust gap proved critical.

Expert systems experienced a brief surge in corporate environments. But they had a fatal flaw: they didn’t learn. Every rule had to be programmed by hand, and updating them became unsustainable. As complexity grew, these systems collapsed under their own weight, triggering the second AI winter.

Still, they taught us key lessons about explainability, scale, and the limits of static logic.

Machine Learning: Letting Data Do the Talking

Out of the second winter came a new idea: what if machines could learn from data instead of being told what to do?

This was the dawn of machine learning, and it shifted the game entirely.

Instead of hard-coding rules, you trained models on examples. Spam filters became adaptive. Fraud detection evolved to spot patterns, not just fixed thresholds. Recommendation engines got smarter.

Three factors made this possible in the 1990s and 2000s:

• More data, thanks to the internet
• Better algorithms, like support vector machines
• Faster processors, especially with GPUs

AI was no longer a lab curiosity. It was quietly reshaping industries — and setting the stage for something even bigger.

Deep Learning and the ImageNet Breakthrough

While traditional machine learning was powerful, it still required manual feature engineering — identifying what aspects of the data mattered.

Deep learning changed that.

Using neural networks with many layers, deep learning models could extract patterns from raw data. They didn’t need to be told what a cat looked like — they figured it out themselves.

The turning point came in 2012, when a model called AlexNet outperformed the competition in the ImageNet challenge, reducing error rates by half. The ingredients? Massive datasets, GPU acceleration, and new training techniques.

Suddenly, AI could:

• Recognize faces and objects
• Understand and transcribe speech
• Translate languages in real time
• Even detect tumours and drive cars

But deep learning had a dark side: it was accurate, but opaque. These “black box” systems couldn’t always explain their decisions — a challenge we’re still grappling with today.

Transformers and the Generative AI Revolution

Then came 2017.

Google published the now-legendary paper: “Attention Is All You Need.” It introduced the transformer architecture, and it changed everything.

Transformers could analyze entire sequences of text at once using self-attention, allowing them to understand language with greater context and accuracy than ever before.

This gave rise to:

• BERT (Google)
• GPT (OpenAI)
• And many others (Meta, Mistral, Anthropic)

By 2020, OpenAI had released GPT-3, and in late 2022, ChatGPT put generative AI in the hands of the public.

Now, AI could generate:

• Blog posts, emails, and reports
• Meeting notes and summaries
• Code, artwork, and even videos

Suddenly, AI wasn’t just something in the background. It was sitting at the table with us, helping us work, write, and create.

But this newfound power also raised urgent questions:

• What happens when AI makes things up?
• How do we deal with biased or misleading outputs?
• Who owns the content that AI generates?

Where We’re Headed: Two Possible Futures

Today, we’re standing at another inflection point — and two paths lie ahead.

1. Smarter, More Integrated AI

AI is becoming a layer in every tool we use — from Microsoft Copilot to Google Gemini. Multimodal models can now understand text, images, and audio together. Open-source models are gaining traction, making powerful tools more accessible.

This democratization is exciting, but it also makes responsible development and oversight more complicated.

2. Toward Artificial General Intelligence (AGI)

Some labs are pursuing AGI — machines that can reason across domains, much like humans. Others are skeptical, pointing out that today’s models still rely on pattern recognition, rather than genuine understanding.

Regardless of who’s right, one thing is clear: the risks are no longer theoretical. Deepfakes, misinformation, job displacement — they’re already here. That’s why policy, ethics, and public awareness matter more than ever.

Final Thoughts: What Should We Build?

We’ve gone from Turing’s abstract machines to today’s generative models that can mimic language, vision, and reasoning.

But the future of AI isn’t just about what we can build — it’s about what we should make.

Every leap in capability brings new blind spots. Let’s focus not just on pushing the boundaries of possibility, but on building systems that are transparent, trustworthy, and genuinely beneficial.

Because AI isn’t just a technical challenge. It’s a human one.

**Transcript** >> Click here to view or download the full episode transcript