Home › Hard problems

Research · open problems · free

The hardest unsolved problems, and which ones one person can test

79 problems nobody has solved yet, from the core of AI to India's own. Each one has its own page with a picture of the problem, why it is still open, an experiment one person could run in about 8 weeks, and a verdict.

79 problems in 10 groups: how many one person can testInside the model651Safety and oversight72Edge AI and serving cost42Physical AI, robots and autonomy132Data and computer systems44Chips, compute and energy33Security41Science and maths133Society and trust13India88Solo: strongSolo: possibleFollow, don't build
India has the most problems one person can test, because careful outside measurement is what India's problems lack most.

How the verdict is decided. Every problem was put through the same four questions:

  1. Can it be measured? There is a number that shows you did better, such as an error rate, a catch rate or a cost per answer.
  2. Is it cheap to try? It runs on a laptop, free open models and open data, or under about ₹50,000 of rented GPU time.
  3. Is there a public yardstick? A benchmark or dataset others use, so your result can be compared and believed.
  4. Would a small win get noticed? A lab, a funder or a customer would care about a 10% improvement, not only a full solution.

Four yeses means Solo: strong. Two or three means Solo: possible. Fewer means Follow, don't build.

Jump to: Inside the model · Safety and oversight · Edge AI and serving cost · Physical AI, robots and autonomy · Data and computer systems · Chips, compute and energy · Security · Science and maths · Society and trust · India

Inside the model: the core of LLMs (12)

Hallucination: the model doesn't know what it doesn't knowSolo: strongTraining rewards a confident answer, and the model has no reliable internal signal for "I am guessing".Reliable multi-step reasoningSolo: strongSmall errors compound: 95% right per step is about 60% right over 10 steps. Models also give reasoning that doesn't match how they reached…Thinking cost: models overthink easy questionsSolo: strongReasoning models spend thousands of hidden tokens on questions that need ten. Nobody has a reliable way to decide how much thinking a…Continual learning: learn after deployment without forgettingSolo: possibleUpdating weights on new facts damages old skills ("catastrophic forgetting"). Today's fix is retraining from scratch or bolting on…Using long context, not just fitting itSolo: strongModels accept a million tokens but miss facts in the middle and can't combine facts spread far apart.Personal memory across months of conversationsSolo: strongWhat should an assistant remember, forget, update or never store? Memories go stale, contradict each other and leak private details.Understanding long videoSolo: possibleModels describe a short clip well but lose track of what happened across an hour of video: who did what, in which order.Running out of human text; synthetic data collapseSolo: possibleGood public text is nearly used up. Training on model-written text can narrow the model ("model collapse").Evaluation itself is brokenSolo: strongBenchmarks leak into training data and saturate within months. Scores stop telling you which model is better for your task.World models and cause and effectSolo: possibleModels learn patterns, not how the world works. They fail on "what would happen if" questions outside their training data.New architectures beyond the transformerFollow, don't buildTransformer cost grows with the square of context length. Alternatives (state-space models such as Mamba, hybrids) are cheaper but not yet…Learning from far less dataSolo: possibleA child learns language from about 100 million words; models need over 10 trillion. Nobody knows how to close most of that gap.

Safety and oversight: checking AI that may be smarter than you (9)

Edge AI and serving cost: AI on small devices and cheaper answers (6)

Physical AI, robots and autonomy (6)

Data and computer systems (8)

Chips, compute and energy (6)

Security (5)

Science and maths (7)

Society and trust (4)

India: hard problems at India's scale (16)

AI in Indian languages costs more and works worseSolo: strongTokenizers split Indian scripts into many more pieces than English, so the same sentence costs about 3 to 8 times more and quality drops.…Speech recognition for dialects and code-mixed talkSolo: strongModels trained on standard Hindi fail on Bhojpuri, Marwari, Hinglish switching mid-sentence, and noisy streets and phone lines.Reading Indian documents: land records, forms, handwritingSolo: strongScanned land records and government forms mix scripts, stamps and handwriting. Text recognition for Indian scripts is far behind English.Useful AI on a ₹10,000 phone, offlineSolo: strongMemory, battery and heat limits. Most "on-device" demos run on flagship phones, while most Indian phones are budget models.Addresses that machines can't readSolo: strongIndian addresses are free text ("behind the temple, near the old well"). Deliveries, ambulances and KYC fail on them. DIGIPIN, a national…Voice-clone and UPI scams in Indian languagesSolo: strongScam calls and messages now use cloned voices in Indian languages. Detectors and scam filters are mostly trained on English.Hyperlocal monsoon and flood forecastsSolo: strongGlobal AI weather models are good at large scale, but rain for one district a few hours ahead is still poor, and floods in Assam and Kerala…Crop yield and stress for small farmsSolo: possibleSatellite pixels (10 metres) are bigger than many Indian plots, and ground truth is scarce.Groundwater running outSolo: possibleMuch of north-west India pumps more than rain refills. Wells are measured only a few times a year.Where Delhi's air pollution comes fromSolo: strongStubble burning, traffic, industry, dust and weather mix differently every day, and sources are argued over every winter.Grid forecasting with lots of solarSolo: possibleSolar and wind swing with the weather, and electricity distributors lose money balancing it.Too few doctors in rural clinicsSolo: possiblePrimary health centres often have one doctor or none. Triage tools must work with little data, in local languages, and be safe.Court case backlogSolo: possibleIndia has about 50 million pending cases. Delays come from scheduling, missing papers and adjournments.Credit for people with no credit historySolo: possibleMillions have no credit record. Models using other data can be unfair and hard to explain to a regulator.Matching informal workers to jobsSolo: possibleMost Indian workers are informal, have no CV and find work by word of mouth.Driving in Indian trafficSolo: possibleNo lanes, mixed vehicles (cows, autos, carts), and rules followed loosely. Self-driving models trained in the US or China fail here.

The pattern across all 79. The problems one person can really test are mostly about measurement: building the test, the dataset or the honest comparison a field is missing. Labs have compute; what they lack is careful outside measurement, and India's problems get the least of it.