AI & Robotics
Self-driving cars
Cars that drive themselves with no human driver, and in the fiction have personalities, communicate with each other and protect themselves.
In shortSelf-driving cars, from Sally (Isaac Asimov, 1953), is at technology readiness level 7 of 9 (prototype demonstrated in operational environment) at its weakest link. Held back by General-purpose foundation models. Science status: underlying science well understood; an engineering problem. Estimated new funding to unlock: $100B–500B.
Automobiles with positronic brains live out retirement on a farm in 2057; when a businessman tries to strip them, the cars coordinate to stop him.
Closest todayWaymo driverless robotaxis, ~500,000 paid rides per week (2026).
Dependency stack
Research TRL 1–3Development TRL 4–6Deployment TRL 7–9
Sort by
- criticalDriverless vehicle autonomyDriving with no human in the loop
✕3 blockers›
- Long-tail situations (school buses, emergency scenes) still cause failures and recalls
- Snow, heavy rain and unmapped roads limit operating domains
- Sensor suite and remote-assist cost per vehicle
TRL 8- General-purpose foundation modelsbottleneckReasoning about rare, novel traffic situations
- Reliability and calibration over long horizons
- Continual learning after deployment
TRL 7- Training and inference compute
✕2 blockers›
- Grid power and interconnection queues
- Advanced-node chip and HBM memory supply
TRL 8
- In-vehicle compute for real-time perception and planning
✕2 blockers›
- Memory bandwidth and capacity limit local model size to a few billion parameters
- Milliwatt power budgets in earbuds and implants are ~1000x below phone NPUs
TRL 8
- Real-time perception and planning in the vehicleTRL 8
- supportingArtificial general intelligenceSally-style cars reason, communicate and act on their own goals
✕3 blockers›
- Reliability collapses on tasks longer than a few hours at high success thresholds
- No persistent learning from experience after deployment
- Weak grounding in physical and social context outside text and code
TRL 3- Must keep learning from experience like a human employee
✕3 blockers›
- Weight updates on new data degrade prior tasks and safety alignment (spurious forgetting)
- No method to validate a model whose parameters keep changing after release
- Per-user weight updates multiply serving cost and memory
TRL 3- Online updates and replay at frontier scale need large training compute in the loopTRL 8
- Base reasoning, language and tool-use competenceTRL 7
- Training and running agents at human-equivalent throughputTRL 8
Who could unlock it
General-purpose foundation modelsTRL 7bottleneck
- OpenAIStartupGPT-5-series models (2025-2026) lead many agentic coding and reasoning benchmarks
- AnthropicStartupClaude Opus models (2025-2026) recorded some of the longest METR task time horizons
- Google DeepMindCompanyGemini 3 (Dec 2025) and Gemini 3.1 Pro (2026) are frontier multimodal reasoning models
- DeepSeekStartupChinese lab whose open-weight R1 reasoning model (Jan 2025) matched frontier results at reported low training cost
Driverless vehicle autonomyTRL 8
- WaymoCompanyRaised $16B in Feb 2026 at a $126B valuation; runs ~500k paid driverless rides per week across several US cities
- TeslaCompanyBegan unsupervised robotaxi service in the Austin metro in June 2026 using camera-only Full Self-Driving
- Baidu Apollo GoCompanyRuns the largest Chinese driverless ride-hailing fleet (Wuhan and other cities, 2024-2026) and is expanding abroad
- WayveStartupRaised a $1.2B Series D at $8.6B (Feb 2026); end-to-end driving model running supervised Uber rides in London
On-device neural inference acceleratorsTRL 8
- QualcommCompanySnapdragon 8 Elite Gen 5 (September 2025) claims ~220 tokens/s on a 3B-parameter model with a 32K context, up from ~70 tokens/s the prior generation
- AppleCompanyRuns on-device foundation models across consumer iPhones and shipped AirPods Live Translation in 2025 with local inference and no cloud link
- Song HanResearcher, MITAuthored the quantisation and compression methods (AWQ, SmoothQuant, TinyML curriculum) that make multi-billion-parameter models fit in phone memory budgets
- NVIDIACompanyJetson and DRIVE platforms run production driving and robotics stacks locally, the highest-power end of the on-device envelope
Beyond engineering
- Liability rules and state-by-state permitting for driverless operation