How to make an AI?
Pick a real task first, then the smallest method that solves it. Most working AI is made by fine-tuning an existing model on a few hundred good examples — cost: hundreds of dollars, time: a weekend. Data quality beats data volume and beats model choice. MIT found buying from vendors succeeds about 67% of the time, roughly triple the rate of internal builds.
Why — the first-principles explanation
The mistake nearly everyone makes is starting with the model. Start with the task. "Make an AI" is not a goal; it is a method looking for a problem. "Automatically sort incoming support emails into six categories" is a goal, and it immediately tells you what to build, what data you need, and how you'll know if it worked. Projects that begin with the method and hunt for a use case are the ones that die at month four.
The second principle is the one that actually separates working AI from broken AI: your data is the model. A model is a compression of its training data — it learns the patterns you show it, including the mistakes. Five hundred carefully labeled examples reliably beat fifty thousand sloppy ones. This is unintuitive because "more data is better" is folklore, but garbage examples don't average out; they teach. If your labels are inconsistent — two annotators disagreeing about what counts as "urgent" — the model learns to be inconsistent, faithfully. Most AI failures are labeling failures wearing a technical costume.
Third: evaluate before you build. Decide the number that defines success and how you'll measure it, before writing training code. Split your data into train and test first, and never let the model see the test set. If you skip this, you will get a model that scores 99% on data it memorized and falls apart on anything real. This is the single most common self-inflicted wound in the field, and it feels great right up until deployment.
Fourth: choose the smallest method that works. For sorting text into categories, a simple classifier trained in ten minutes may outperform a large language model, run 1,000x cheaper, and never hallucinate. Reaching for the biggest model first is expensive and usually unnecessary. The professional move is to start with the dumbest thing that could possibly work, measure it, and only escalate when the number demands it. And the strategic point, with data behind it: MIT's NANDA study found buying from specialized vendors succeeded roughly 67% of the time, about three times the rate of internal builds, and that ~95% of enterprise GenAI pilots showed no measurable P&L impact. Before you make an AI, check whether someone already sells it. The most sophisticated decision is often not building it.
An example that makes it click
Making an AI is like teaching a kid to sort laundry. You don't start by explaining textiles. You show her a pile and say: darks here, lights there.
Now, here's the part that decides everything. If you sort 500 items carefully and consistently, she learns the rule in an afternoon. If you sort 50,000 items while distracted — throwing a gray shirt in darks on Monday and lights on Thursday — she doesn't learn 100 times better. She learns that gray is random, because you taught her that. Faithfully.
And if you test her using the exact same pile she watched you sort, she'll score perfectly by remembering, not by understanding. You have to hand her a laundry basket she's never seen. That basket is your test set, and hiding it from her before you start is the whole trick.
How to do it
- Write down the specific task in one sentence, with a measurable outcome. If you cannot, you are not ready to build anything.
- Define success as a number before you write any code — accuracy, error rate, time saved. Decide what result would make this worth doing.
- Check whether a vendor already sells this. MIT's data shows buying succeeds roughly three times as often as building internally.
- Collect and label a few hundred high-quality examples. Have two people label the same 50 and compare — if they disagree, your task definition is broken, and no model will fix that.
- Split your data into training and test sets before training. Never let the model see the test set. This is the most commonly skipped and most damaging shortcut in the field.
- Start with the simplest method: a basic classifier, or an API call with a good prompt. Measure it. This is your baseline.
- Only escalate to fine-tuning if the baseline misses your number. Fine-tune an open-weights model on your labeled examples — typically hundreds to thousands of dollars.
- Measure again on the untouched test set. If performance is far worse than on training data, you overfit — get more varied data rather than a bigger model.
- Deploy small, watch real failures, and feed them back into your labeled data. MIT identified inability to learn from feedback as the core reason these systems stall.
Key facts
- MIT NANDA found buying AI from specialized vendors succeeded about 67% of the time — roughly three times the success rate of internal builds.
- MIT NANDA found roughly 95% of organizations saw no measurable P&L impact after $30–40 billion of enterprise generative AI spend.
- MIT identified the core barrier as learning: most GenAI systems do not retain feedback, adapt to context, or improve over time.
- Fine-tuning an open-weights model typically costs hundreds to thousands of dollars, versus tens to hundreds of millions for frontier pretraining.
- Manus, an AI agent company with a ~$125 million revenue run rate as of December 2025, does not train its own foundation models.
- Raspberry Pi publishes official documentation for running real-time vision AI locally on a Raspberry Pi 5 with an AI Kit or AI HAT+ and a supported camera.
▶ The 60-second explainer (script)
How to make an AI. Here's the order nobody tells you, and getting it wrong is why most projects die. Don't start with the model. Start with the task. Make an AI isn't a goal — it's a method hunting for a problem. Automatically sort incoming support emails into six categories? That's a goal. It tells you what to build, what data you need, and how you'll know it worked. Second, and this is the big one: your data is the model. A model is a compression of what you showed it, including your mistakes. Five hundred carefully labeled examples beat fifty thousand sloppy ones. That's counterintuitive, because everyone says more data is better. But bad examples don't average out. They teach. If two of your labelers disagree about what counts as urgent, the model learns to be inconsistent — faithfully. Most AI failures are labeling failures in a technical costume. Third: decide how you'll evaluate before you build. Split your data into train and test first, and never let the model see the test set. Skip this and you'll get a model that scores ninety-nine percent on data it memorized and falls apart on anything real. It feels amazing right up until deployment. Fourth: use the smallest method that works. For sorting text into buckets, a simple classifier trained in ten minutes might beat a large language model, run a thousand times cheaper, and never hallucinate. Start with the dumbest thing that could possibly work. Measure it. Escalate only when the number makes you. And the strategic one: MIT found buying from specialized vendors succeeded about sixty-seven percent of the time, roughly triple the rate of internal builds. And ninety-five percent of enterprise AI pilots showed no measurable financial impact. So before you make an AI — check if someone already sells it. The most sophisticated decision is usually not building it.
What authoritative sources say
People also ask
How much data do I need to make an AI?
Far less than you think, if it is clean. A few hundred consistently labeled examples often suffices for fine-tuning a narrow task. Quality beats volume — sloppy labels actively teach the model to be sloppy.
Do I need a powerful GPU?
Not to start. Calling an API needs nothing but internet. Fine-tuning can be rented by the hour. Even on-device vision AI runs on a Raspberry Pi 5 with an AI HAT, using officially documented tooling.
What is the most common beginner mistake?
Testing on training data. You get a spectacular score by memorization, then the model collapses on real inputs. Split train and test before you begin and never look at the test set until the end.
Should I use a large language model for everything?
No. For classification and structured tasks, a small purpose-built model is often more accurate, roughly a thousand times cheaper, and cannot hallucinate. Reach for the biggest model last, not first.
How is this different from creating an AI from scratch?
Making something useful almost never involves training from scratch. You fine-tune or prompt an existing model. Training a frontier model costs tens to hundreds of millions and is done by about a dozen organizations worldwide.
The same question, asked other ways
- How to create an AI?2,400/mo
- How to make AI?1,300/mo
- How to create AI?1,000/mo
- How to make your own AI?1,000/mo
- How to create your own AI?880/mo