Skip to main content
All posts
7 min read

Choosing the Right Kind of AI

Why probabilistic models, deterministic systems, and hybrid architectures solve different problems.

Artificial IntelligenceGenerative AIAI StrategyMachine LearningTechnology

Conversations about artificial intelligence keep collapsing into two extremes. One side sees only the hallucinations. The other sees a universal solution — a hammer looking for nails. Both start from the same mistake: they begin with the tool, not the job.

Start with the job instead. What must the system do. How much ambiguity can it absorb. And when the answer is wrong — because eventually it will be — what happens next.

Take something concrete: a customer support bot that can approve refunds up to $50 on its own, and has to escalate anything above that.

The $50 line isn't negotiable. It's a rule, and rules have to hold every time. But whether a customer's message actually describes a refund-worthy problem — that's read in their own words, no script, no fixed shape. Ambiguous by nature. Same bot, two jobs, two completely different relationships with error. We'll keep coming back to it.

Three patterns answer that difference.

  • Generative models explore a space of plausible outputs.
  • Deterministic systems execute a fixed logic.
  • Hybrid systems combine both — one proposes, the other verifies.

Not a hierarchy. A set of tools, each suited to its own kind of consequence.

First, a quick map

Artificial intelligence is the broad field; machine learning sits inside it; deep learning is one family of methods built from many-layered neural networks.

Generative and deterministic sit on a different axis entirely — they describe behaviour, not the field or the method.

Generative AI means a model built to produce new content.

Determinism means something else: under the same conditions, the same path, the same result, every time. A rule engine can be deterministic. So can a trained classifier, at inference time. A generative model can grow more repeatable under the right constraints without ever becoming a rule engine.

The decision was never implementing AI or not. It's what for. Do we want exploration, or we need control?.

In the refund bot: exploration for reading the message, control for the $50 line.

Generative AI explores possibilities

A language model doesn't retrieve a finished sentence from some hidden archive. It estimates, token by token, what could plausibly come next, and samples from that distribution. This mechanism gives it a way to turn learned patterns into novel sequences.

That's where the range comes from. Incomplete instructions, messy material, content that didn't exist an hour ago — a generative system can meet all of it. But the same looseness that lets it improvise also lets it lie fluently.

A confident answer to the wrong question reads exactly like a confident answer to the right one. Left alone, our refund bot might read a borderline complaint and decide — smoothly, persuasively — to approve $200.

Training itself is more layered than "unsupervised" suggests. Large-scale self-supervised pretraining comes first; supervised examples, preference data, and reinforcement learning shape it from there. OpenAI's InstructGPT work is the well-known version of this: fine-tune on demonstrations, then reinforce on human preference.

Temperature changes sampling behaviour. It doesn't make an answer more truthful...

Turn it down, output gets narrower and more repeatable. Turn it up, more variation enters. Useful for tuning a brainstorming tool differently than an extraction task — not useful as an honesty dial. Google Cloud's own documentation notes that even a temperature of zero produces only mostly deterministic output [link]. A low temperature can repeat a falsehood with total consistency. A higher one can stumble onto something true. Grounding, retrieval, evaluation, verification — that's where reliability actually lives. Temperature was never it.

Deterministic systems execute defined logic

Same input, same configuration, same output — always. Calculators, policy rules, database constraints, proof checkers, most optimisation routines.

This is the $50 line in our bot: no stylistic possibilities to weigh, just a rule, applied the same way every time it's asked.

That repeatability is valuable when consistency and reproducibility matter — a credit limit, a dosage constraint, a proof step. But control costs something too. Rules struggle the moment the world stops matching them, and someone has to keep defining the exceptions as conditions shift. A perfectly reproducible answer can still be wrong — if the rule was wrong to begin with.

Even reproducibility is less solid than it sounds. Change a model, a library, a piece of hardware, and the same prompt can drift. Real determinism needs more than a fixed prompt. It needs versioned components. Controlled inputs. Tests that actually run.

Hybrid AI separates proposing from proving

Most systems worth building need both halves.

  • One side interprets, explores, proposes.
  • The other calculates, checks, rejects what doesn't hold.

Our refund bot already lives here: the model reads the complaint and drafts a recommendation, a rule checks it against $50, and only what's under the line gets approved without a person watching. Above it, the system doesn't guess — it escalates.

The two halves can run in sequence: a model drafts a query, a database runs it; a model proposes code, tests check it; a model suggests a construction, a symbolic engine proves it or doesn't.

Or in parallel: candidates weighed against each other before anything gets returned.

No single blueprint covers this. Just one instinct, repeated: creativity where the search space is wide, verification where the error has a price.

Mathematics shows why the combination works

Olympiad geometry makes the case vividly. AlphaGeometry 2 pairs a neural model with a symbolic deduction engine — one predicts useful constructions, the other proves them by formal rule. Together: 84% of IMO geometry problems from 2000 to 2024 solved, ahead of the average gold medallist the researchers benchmarked against [link].

Gemini Deep Think took the opposite route in 2025 — five of six IMO problems, 35 points, gold-medal standard, and none of it translated into formal language first [link]. Parallel reasoning and reinforcement learning explored and combined multiple solution paths. Two systems, two entirely different ideas of where proof should live.

(Harmonic's Aristotle found a third answer again — generate the proof, let Lean 4 check it line by line — and also auto reported gold-medal-level results at the same competition [link]. One Olympiad. Three different places to put the trust.)

There's no single hybrid template hiding underneath all this. Just three ways of splitting the same labour, each one earning its result.

Choose the architecture from the consequence of error

  • Reach for generative AI where exploration pays off: drafting, brainstorming, reading a customer's own words. Several right answers can coexist. A person can still look.
  • Reach for deterministic logic where the rule already exists: a threshold, an eligibility check, a permission. Consistency beats invention, every time, here.
  • Reach for a hybrid where the input is ambiguous but the outcome can't be: let the model interpret. Let something else — a rule, a test, a person — decide.

And keep a person in the loop wherever the system touches health, money, rights, safety. Automation can gather the evidence. It still can't carry the accountability. That needs a name attached.

The useful question comes before the model

Stop asking which model is best in the abstract. A creative assistant and a fraud control and a refund bot and a clinical workflow don't share a tolerance for uncertainty — so why would they share an architecture?

Start with the decision. What can vary. What must never bend. How the output gets checked, and who answers for it. Only then choose — generative, deterministic, hybrid.

The most capable system was never the one that generates the most.

It's the one that knows where to stop.

Sources

Artificial IntelligenceGenerative AIAI StrategyMachine LearningTechnology

Val

Lead Product Designer · UX & AI Experience (AX)

All postsWork with me