The numbers Brief Episode map Timeline Takeaways Charts Quotes Clips Glossary Go next Transcript

Episode 110 · Deep tech · 48 min

Teach it the world before you teach it to drive

Minus Zero's contrarian read is that autonomous driving stalled because the industry treated it as a data problem. An image model that draws three fingers is only embarrassing; a car that hallucinates is fatal, so the company builds world models compressed onto the vehicle itself, measured against the 20-watt brain that already does this, and is about to name two of the world's largest automakers as partners.

AI & machine learning Mobility & EV Deep tech & hardware

Gagandeep Reehal

Co-founder, Minus Zero · with Vishal Krishna

48:06

Said in this episode

▶ 22:08

80%+

Candidates already holding foreign admits

From his own recent interview loops: the share of applicants who have offers from universities abroad, which he frames as a generational talent risk.

▶ 34:53

52 weeks

Peak GPU order lead time

At the height of the crunch, a data-centre GPU order placed today arrived a year later, the scarcity that makes on-vehicle efficiency an IP question.

▶ 33:31

20 W

The power budget he benchmarks against

A human brain makes these driving decisions on roughly 20 watts, closer to a laptop than a supercomputer, which is the reason LLM-scale models cannot sit in a car.

▶ 37:20

7 years

A car model's life cycle

The survival question every OEM put to two young founders: a car sells for seven years, so the supplier has to last seven years too.

▶ 38:56

2 OEMs

Partnerships about to be announced

One of India's largest automakers and one of the world's largest, with the first announcement flagged for that month; names were withheld on air.

▶ 42:28

Very major automakers, across four belts

North America, the European Union, Korea and Japan, and India, with roughly two dominant players each, which is the whole customer map for a tier-one autonomy supplier.

The brief

The argument in sixty seconds

Gagandeep's claim is that self-driving stalled because everyone attacked it as a data-collection problem. Billions of kilometres of driving footage will still miss the autorickshaw with a poster pasted across it, and a model trained only on driving has no way to reason about a thing it has never seen. Minus Zero's counter, set out in the nature-inspired AI paper on its site, is that you teach a machine the world before you teach it to drive, the way a child learns not to collide long before anyone explains a clutch. That makes it a world-model company rather than an LLM company, and the distinction is load-bearing: a chatbot that hallucinates changes a narrative, a car that hallucinates ends a life. The second constraint is physics. A vehicle is not plugged into an AC outlet, inference has to be real time, and the reference system is a human brain running on 20 watts, so the intelligence has to be squeezed onto the car, because Indian connectivity disappears inside a tunnel. Around the technology sits a commercial argument: Minus Zero sells as a tier one, an Android-inside layer licensed per vehicle to OEMs whose model cycles run seven years, which is why a three-year-old startup spent two of them doing demos just to prove it would still be alive. Two partnerships are about to be announced, one of India's largest automakers and one of the world's, with a hands-off highway copilot as the gen-one product. Underneath it all is the talent bet: more than 80% of the engineers he interviews already hold admits abroad, and whether they stay decides if India owns this technology or rents it.

Worth your time if you are

Engineers weighing a foreign admit against an Indian deep-tech job

Automotive product leads sizing up ADAS and autonomy suppliers

Deep-tech founders selling into a hundred-year-old supply chain

AI researchers curious about world models beyond LLMs

Episode map

Where the conversation travels

Every block is a chapter, coloured by what it's about. Click any of it to jump straight to that minute on YouTube.

AI & machine learningMobility & EVDeep tech & hardwareHiring & talentSales, GTM & growthFounder journey

01Cold open: why minus one to zero 0:00 Vishal frames autonomous vehicles as the future India keeps dismissing with cows and monkeys, while the founder explains the company name as an ideology: going against the popular narrative to build something that did not exist, a minus-one-to-zero journey rather than a zero-to-one. 02Two undergrads pick the hardest problem 1:34 Started in 2021 from an engineering school second year, the founders wanted an AI problem big enough to matter rather than another narrow application, and picked the one the industry had already dubbed the toughest AI has attempted. 03Decision-making that never sat in India 4:09 Bosch and Continental run vast engineering centres here while the decisions stay in Germany or the US; changing that narrative took two years, and the thesis was to do one foundational thing so well that others build on top of it. 04The car goes software-defined 6:42 The iPhone camera analogy carries the argument: features moved from hardware to software first in consumer tech, then industrial, now automotive, and buyers who once asked about mileage now ask about airbags, safety and experience. 05Autonomy beyond ADAS, Android for the vehicle 10:47 ADAS arrived in India in 2021 and buyers immediately wanted more, so Minus Zero positions itself as the intelligence layer: silicon and sensors from partners, bodies from automakers, decisions from its platform. 06A tier one that ships no hardware 12:48 The model is Intel Inside for autonomy: a per-vehicle cut of every unit an OEM sells for the life of that model, with the software pre-optimised against a compute-and-sensor kit before it reaches the carmaker. 07Why an LLM cannot be allowed to drive 14:52 LLMs proved machines could be fluent, not that they could reason. The three-fingered generated hand is harmless, but a decision system with a life depending on it cannot hallucinate, which is why the next frontier is world models that understand physics and behaviour. 08Nature-inspired AI and the pod ride 18:16 Their IP borrows the human brain's decision-making for a foundation model, and the payoff was a public autonomous pod demo last June that a veteran Indian tech co-founder stepped out of calling it a small ride for him and a giant leap for India. 09Vaswani, brain drain and the 80% who leave 19:44 The transformer paper's first author was Indian but the work happened in the US; the founder himself had admits from top UK research universities and stayed only because COVID intervened, and over 80% of the candidates he now interviews are already holding foreign offers. 10Why the engineers are coming home 22:53 Software talent left around 2000 and research talent through 2017, but the post-2017 deep-tech boom in space, EV and drones created multi-disciplinary roles at home, and returnees with young children now weigh culture and lifestyle against a US that no longer looks obviously better. 11The broken loop, and learning like a child 26:11 Today's stacks decide without learning from the outcome, whereas you became a safe driver from a lifetime of observation before any driving school, so the system must generalise about unseen objects, because no volume of driving data covers a postered autorickshaw. 12Twenty watts and no AC outlet 31:43 Bigger models mean more compute, more power and more latency, none of which a real-time vehicle can afford. The benchmark is a 20-watt brain, and with 5G dropping inside tunnels the car has to be the edge device. 13Getting inside a hundred-year-old mafia 36:30 OEMs asked whether a two-person startup would outlive a seven-year model cycle, so the answer was relentless demos, and as Apple and other robotics-first AV programmes shut down, Minus Zero readied partnerships with two major automakers and a hands-off highway copilot. 14India as the hardest test track 40:43 Broken roads, missing lane markings, rain to snow and anything at all wandering onto the carriageway make India the proving ground. Patents are granted here with more in the pipeline, and the addressable map is four automaker belts holding eight very major players. 15Commoditisation, legacy and the inspirations 43:40 Every disruptive technology eventually gets commoditised, so the only defence is being a generation ahead the way Nvidia is; the closing motivation is mobility as the marker of a developed country, with Musk and India's private space founders as the nerve.

Timeline

The arc, briefly

Three years of a world-model company, set against two decades of Indian engineers leaving and coming back.

2000The software talent goes. Foreign companies set up here and product roles followed, but an engineer who wanted to work on space technology or research had almost nowhere in India to do it.

2010-17Then the researchers go. The second wave out is research talent, in the years when a deep-tech career here still meant emigrating for it.

2017The paper is written, in the US. The core technology behind the language-model wave arrives in 2017-18 with an Indian first author working abroad. In the same year India's deep-tech boom in space, EV and drones starts making multi-disciplinary roles at home.

2021Two undergrads start Minus Zero. Gagandeep is in the second year of engineering school when they pick the problem the industry had already called the hardest AI has attempted, and build on foundation models before most people were using the term. He had admits from top UK research universities and stayed only because covid intervened.

2021ADAS reaches Indian cars. Buyers get their first taste of assisted driving and immediately want more, which is the opening for an intelligence layer sold to carmakers rather than another car company.

2021-22The language-model wave lands in industry. Four years after the research, fluency becomes a product. His reading is that it proved machines could be fluent, not that they could reason, and a chatbot that hallucinates changes a narrative while a car that hallucinates ends a life.

2021-24Three years, two of them spent on demos. The first question every automaker asked two young founders was not about the technology but whether they would outlive a seven-year model cycle, and moving decision-making that had always sat in Germany or the US took a further two years of proving it. Last June a veteran Indian tech co-founder stepped out of a public Zpod ride and called it a small ride for him and a giant leap for India.

TodayTwo automakers, one announcement this month. One of India's largest and one of the world's, with a hands-off highway copilot as the gen-one product and Indian patents granted with more in the pipeline, while Apple and other robotics-first autonomy programmes shut down around them.

Takeaways

Ideas to carry out of this hour

Teach the world first, then teach driving

Minus Zero's central claim is that driving data alone can never be enough. An autorickshaw with a poster pasted across it is a different object to a vision model, and the space of such surprises keeps expanding faster than any fleet can log. Humans do not learn to drive in the ten hours of a driving school; they arrive with a lifetime of watching, not colliding, and inferring what a stranger will do next. So the model must first acquire generalised physics and behaviour (size, depth, intent) and only then be taught the road, which is the argument behind the nature-inspired AI paper.

▶ 29:26 ▶ 30:11 ▶ 30:40

A hallucinating chatbot embarrasses; a hallucinating car kills

LLMs demonstrated fluency, not reasoning. The generated photo with three fingers is the tell, and the fluency is precisely what stops a consumer noticing the error. When the output is text, the worst case is a shifted narrative. When the output is a control decision with a life depending on it, hallucination is not an acceptable failure mode, which is why he argues driving needs something past LLMs: models that understand how the world behaves rather than only what it looks like.

▶ 16:10 ▶ 16:42 ▶ 16:58

The car is the edge device, and it runs on a budget

The industry's answer to weak models has been bigger models, which is why GPUs became the scarce currency and why lead times ran to a year at the peak. A vehicle cannot pay that bill: it needs real-time inference, it is not plugged into an AC outlet, and the reference system, a human brain making these same decisions, runs on about 20 watts. Nor can the compute be exiled to the cloud, because Indian connectivity is inconsistent and vanishes in a tunnel. Efficiency, in his framing, is the IP itself.

▶ 33:16 ▶ 33:31 ▶ 35:59

Be the intelligence layer, not another car company

Minus Zero sells as a tier one, the way a Bosch sells electric motors: Intel Inside or Android inside, but for decisions. The software is pre-optimised against a partner's compute and sensors, shipped to the automaker as a kit, and monetised as a per-vehicle share for as long as that model sells. The discipline that follows: silicon companies and OEMs are partners whose place the company explicitly does not want, because trying to do ten things means doing none of them well.

▶ 13:10 ▶ 13:36 ▶ 14:09

A seven-year model cycle is the real credibility test

The first question every automaker and investor asked had nothing to do with the technology. It was whether two young founders would still be around in seven years, the life cycle of a car model. That is why the early years went into demos that put decision-makers inside the vehicle and asked whether they felt safe. The industry then helped: as robotics-first AV programmes went into decline and Apple shut its car division, the argument that hardware, software and AI are different disciplines requiring different agility stopped sounding like a pitch.

▶ 37:13 ▶ 37:42 ▶ 37:56

India is the test track, not the compromise

The cows-and-monkeys objection gets inverted: broken roads, absent lane markings, weather from rain to snow and drivers who signal nothing make India the hardest environment to solve, so foreign markets become the easier case. Solve for the vehicle that might turn without an indicator and you have built for everywhere, which is also the pitch for exporting the stack from India into the North American, European, Korean-Japanese and Indian automaker belts.

▶ 40:59 ▶ 41:14

Whether the talent stays decides who owns this

The transformer paper that made LLMs possible had an Indian first author, working in the US. More than 80% of the candidates he interviews already hold admits abroad, and he counts himself as an accidental stayer whose UK plans collapsed with COVID. His read is that the post-2017 deep-tech boom, the funding, and a geopolitical moment that made India attractive have finally created the alternative. Losing that cohort would cost a generational advantage rather than a hiring round.

▶ 20:18 ▶ 22:08 ▶ 22:21

The numbers, drawn

What the episode measures

Every figure below was said on air, with timestamps included and caveats kept.

Three clocks a young tier-one has to beat

years

All three figures as stated in conversation: a 52-week lead time on data-centre GPUs, the company described on air as three years old, and the seven-year car life cycle behind every automaker's will-you-survive question.▶ 37:20

How a person learns the road, and how a car has been taught

the order the learning happens in

The argument behind the nature-inspired AI paper on the company site. Size, depth and intent have to be learned before any of the driving is, because the space of surprises on an Indian road grows faster than any fleet can log it.▶ 29:26

What the intelligence has to fit inside

the constraints a vehicle puts on a model

The industry answered weak models with bigger ones, which is how GPUs became the scarce currency. A car cannot pay that bill, so the whole stack gets squeezed onto the vehicle instead.▶ 33:31

The whole customer map is eight companies

four automaker belts, roughly two dominant players in each

The map as he draws it for a tier-one autonomy supplier. Two partnerships were about to be announced when this was recorded, one of India's largest automakers and one of the world's, with the names withheld on air.▶ 42:28

Worth keeping

Lines that stay

Normally companies have a zero-to-one journey and then one to a hundred. We had our minus-one-to-zero journey — we also had to take the onus of maturing the entire ecosystem.

▶ 3:54

You cannot hallucinate while making a decision, because there is a life depending on you.

▶ 16:42

It is impossible to have enough data to encapsulate every scenario you can encounter on the road. Even billions of kilometres will not be enough — something will still be missing.

▶ 29:26

The car is not plugged into an AC outlet. Compare it to a brain — that is a 20-watt system in your head.

▶ 33:31

If you can drive in India, you can drive anywhere in the world.

▶ 41:14

Clips that travel

Short on time? Start here

AI researchers curious about world models beyond LLMs

Why an LLM can't be trusted with the wheel

The cleanest statement of the thesis: fluency is not reasoning, hallucination is fatal in control, and world models are the answer.

15:40 → 18:16 · 3 min ▶ Watch clip

Engineers who assume more data solves perception

Learning to drive like a child

The postered autorickshaw, the ten-hour driving school that taught you nothing, and why generalisation has to come before driving data.

28:28 → 31:43 · 3 min ▶ Watch clip

Automotive product leads sizing up autonomy suppliers

Twenty watts and no AC outlet

The compute-and-power argument in full: GPU scarcity, real-time latency, the brain as benchmark, and why the cloud is not an option in a tunnel.

31:43 → 36:30 · 5 min ▶ Watch clip

Deep-tech founders selling into legacy supply chains

Getting inside a hundred-year-old mafia

The will-you-survive question, why global AV programmes shut down, and the two automaker partnerships and highway copilot to come.

36:30 → 40:43 · 4 min ▶ Watch clip

Investors sizing an India-built global autonomy play

India as the hardest test track

Cows and monkeys reframed as an advantage, granted patents, and the four automaker belts the company intends to reach.

40:43 → 43:40 · 3 min ▶ Watch clip

Glossary

The jargon, unpacked

ADAS

Advanced driver assistance systems: lane keeping, adaptive cruise, emergency braking and the like, introduced to Indian cars around 2021; Minus Zero pitches what it calls autonomy beyond ADAS.

World model

A model trained to understand how the world behaves rather than only what it contains, so it predicts that the cat at the kerb may step onto the road rather than only detecting the cat.

Nature-inspired AI

Minus Zero's own term, set out in a short paper on its site, for borrowing the human brain's decision-making structure to build a foundation model for driving rather than for text.

Tier one

A supplier that sells directly to a carmaker, as Bosch does with electric motors; Minus Zero sells its autonomy stack the same way and takes a per-vehicle share of each unit sold.

OEM

The original equipment manufacturer: the automaker whose badge goes on the car and who ultimately decides which supplier's intelligence goes inside it.

Edge inference

Running the model on the device rather than in the cloud; for a car it is mandatory, because connectivity drops in tunnels and a decision cannot wait for a round trip.

Highway autopilot

The gen-one product described on air as an AI copilot: hands off the wheel from the moment you enter a highway until you leave it, pitched for both passenger and long-haul commercial use.

Connections

If this resonated, go here next

EP 25 · Deep tech

Full transcript

The whole conversation, searchable

Auto-generated captions, lightly cleaned. Click a timestamp to open that moment on YouTube.