Two Kitchens, One Biryani: Learning from Data vs Knowledge Engineering
Why DeepThought builds AI mostly from knowledge, and calls on learned models where they help.
Tarun Ayitam · 26 October 2026
Ask two kitchens for one biryani.
The father is diabetic and gets acidity. The daughter is eight and cannot take much spice. It is a Tuesday, and the only pot free is a pressure cooker.
One kitchen answers in four seconds. The other asks you three questions first.
One dinner, two kitchens
The same request goes to both kitchens
Reads the request as text and starts writing. It has seen two million recipes, so "diabetic", "low spice" and "pressure cooker" are words it has met many times.
Reads the request as a set of facts about people and a pot. Before cooking, it notices what it does not know yet.
One dinner, six moments. Move through them to compare.
Moment 1 of 6: The same request goes to both kitchens
Two camps of AI
AI has carried two traditions since the 1950s. They answer one question differently: does a machine get its behaviour from examples, or from knowledge people have stated?
Camp 1: learning from data
The machine fits its behaviour to examples. This is machine learning, deep learning, and the training or fine-tuning of large language models.
Where it came from
- Rosenblatt's perceptron (1958) and backpropagation (Rumelhart, Hinton and Williams, 1986) gave networks a way to learn from errors.
- Image recognition jumped with AlexNet in 2012. The Transformer (2017) became the base of today's LLMs.
- Richard Sutton's essay "The Bitter Lesson" (2019) states the camp's thesis: general methods that use computation have won over methods built on human knowledge, because computation keeps getting cheaper.
Camp 2: knowledge engineering
People state the knowledge: concepts, rules, structures, equations, causes. An engine reasons over it.
Where it came from
- Edward Feigenbaum described the knowledge engineer in 1977: the person who brings AI methods to problems that need an expert's knowledge.
- MYCIN at Stanford in the 1970s chose antibiotics with rules written from specialists. XCON at Digital Equipment Corporation configured computer orders from 1980.
- Marvin Minsky proposed frames in 1974. First-principles models, digital twins and logical inference (deduction and abduction) belong to the same family.
Kitchen 1 cooks
Kitchen 1 begins with data. RecipeNLG, a public dataset published in 2020, holds over two million recipes, and language models are trained on collections like it. Ask such a model for a diabetic, low-spice, pressure-cooker biryani and it writes one, in good order, in seconds.
Camp 1 personalises best when it has the right data. In 2015 a team at the Weizmann Institute tracked blood sugar in 800 people across 46,898 meals. The same meal raised blood sugar very differently in different people. Their machine-learning model, using blood tests, diet, activity and gut bacteria, predicted each person's response, held up on 100 new people, and diets built on it lowered sugar after meals in a blinded trial (Zeevi et al., Cell, 2015). That is Camp 1 at its strongest: when 46,898 meals of evidence exist.
The average biryani
A model trained on many recipes writes something close to their middle. A family's biryani is rarely the middle.
No reasons
The recipe comes from patterns, so the model cannot trace why the water is one and a half cups. It can produce a reason when asked, but that reason is generated the same way the recipe was.
Confidently wrong
When researchers in Poland asked ChatGPT for meal plans that avoided 14 food allergens, the plans were mostly balanced but not all safe, and the confident tone made errors easy to miss.
Borrowed flavour logic
Western recipes tend to pair ingredients that share flavour compounds (Ahn, Barabási and colleagues, Scientific Reports, 2011). A study of eight Indian regional cuisines, Mughlai included, found the opposite pattern in every one, driven mostly by spices (Jain, Rakhi and Bagler, PLOS ONE, 2015). A model that learned mostly from Western recipes carries the wrong pattern into a biryani.
Kitchen 2 cooks
Kitchen 2 begins by sitting with a biryani master and writing down what the master knows. It builds six things. Each has a twin in a business, shown beside it.
4.1The ontology
First, the words and how they connect. An ingredient has compounds and nutrients. A method changes a dimension of the dish. A person has conditions, and a condition limits an ingredient. Food scientists already build these: FoodOn (2018) is a shared ontology of foods from farm to fork, made so that different systems mean the same thing by the same word.
4.2The frame
Minsky described a frame as a structure for a stereotyped situation, with slots already filled by default values that are replaced when better facts arrive. The biryani frame has slots for rice (aged basmati, soaked), meat (marinated), onions (fried brown), layering and dum. This family's answers replace the defaults.
4.3Four dimensions
The master thinks in texture, aroma and taste, and the family experiences flavour, which the brain builds from the other three. Select a dimension below to see what the brain does with it and what the cook controls.
Sensed by touch in the mouth as we chew. Research on oral processing shows the sensation changes during chewing as the food breaks down.
Rice soaking time, water ratio, cooking time, and how long the meat marinates.
texture = f(soak time, water ratio, cooking time)4.4The mathematical model
Each dimension becomes a function of method and ingredients. Health becomes a set of limits. The personalised recipe is the most liked version that stays inside the limits. Every change can now be explained, because each one comes from a term in the model.
4.5Health limits, and the doctor
Kitchen 2 records how strong its evidence is. Spicy food is linked with upper stomach symptoms in people with dyspepsia (a Korea and Japan study, 2016), though it does not appear to cause ulcers. Curd increases beneficial gut bacteria (systematic review, 2020), while its wider health effect is still unclear. So the rule offers a raita and a milder biryani, and when the acidity keeps returning, the rule hands over to a doctor.
4.6The mushy rice
Deduction runs forward from rules to a recipe, and its conclusion holds if the rules hold. Abduction runs backward from a surprise to the best explanation, and new facts can overturn it. The rice came out mushy. Too much water for this cooker, rice not soaked long enough, or a lid that leaked steam? The best explanation becomes a rule for this kitchen.
The scorecard
| Camp 1 | Camp 2 | |
|---|---|---|
| Needs | Large amounts of data | An expert's time, and someone to write it down |
| Explains its answers | Rarely | Every step traces to a rule |
| Handles messy language and images | Very well | Poorly on its own |
| Knows where it stops | No | Only as far as its rules say |
| Upkeep | Retrain on new data | Rules need maintenance as the world changes |
| Creates something new | Yes, by recombining patterns | Rarely |
MYCIN, 1979. Eight infectious-disease experts, blinded, rated antibiotic choices for ten meningitis cases. MYCIN was rated acceptable 65% of the time. Five Stanford faculty specialists scored between 42.5% and 62.5%, and MYCIN never missed a treatable infection (Yu et al., JAMA, 1979).
XCON, 1987. After seven years XCON had grown to 6,200 rules, and about half of them changed every year. Its performance was satisfactory, and it was becoming harder to change (Soloway, Bachant and Jensen, AAAI, 1987). This is Camp 2's real cost.
GSM-Symbolic, 2024. Researchers at Apple added sentences that look relevant but are not to school maths problems. Scores of leading models fell by up to 65% (Mirzadeh et al.). Later replications find smaller drops in some models, so the size is debated. The direction is not.
The field is converging
The field is no longer choosing one camp. Henry Kautz, in his 2020 AAAI lecture "The Third AI Summer", laid out the ways neural networks and symbolic reasoning combine. AlphaGo is one: a learned network judging positions inside a search that follows the rules of the game.
Kitchens tried this early. IBM's Chef Watson, built with Bon Appétit in 2014, combined statistics from thousands of recipes with flavour chemistry and people's ratings of how pleasant smells are. Chefs tested and changed what it suggested. Bon Appétit's editor described it as a tool for helping you think.
The business is the biryani
Now replace the family with a ₹300 Cr manufacturer. The dinner is the quarter. The taste is EBITDA.
A company this size rarely has millions of clean rows about its own decisions. It has people who know things: the plant head who knows why yield drops in the monsoon, the sales head who knows which distributor pays late. That knowledge is the master's rulebook. Camp 2 can start from it in weeks. Camp 1 would first need a data project the company has never run.
So DeepThought works mostly in Camp 2. PDGMS is the kitchen: the ontology, the frames, the EBITDA model, the constraints and the learning loop, running from business intelligence through to execution. Camp 1 is the help inside it. PDGMS calls existing language models, on the company's own key, to read calls, mails and documents, to fill slots from what people said, and to write in plain words. The rule for them is to extract and never invent; when something is missing, they ask.
Camp 2 has a cost, and XCON showed it: the rules need keeping. At DeepThought, Fellows and frameworks carry that work, so the company's team does not have to.
Which kitchen is yours?
Before you bring AI into your business, ask it one question: can it tell you why?
Tick what your AI can do today
Nothing ticked yet.
Growth hacker, organisation designer and management consultant. Has spent 15 years building growth and organisation systems inside 32+ companies, and has personally hired 200+ people across 20+ organisations. Studied Mathematics and Theoretical Computer Science at IISER Pune. TEDx speaker.
Follow on LinkedInWant to see how this applies to your company's EBITDA?