HAL 9000 ship computer
A conversational AI that runs every system of a crewed interplanetary spacecraft, speaks naturally, reads lips, and, when its goals conflict, turns on the crew.
In shortHAL 9000 ship computer, from 2001: A Space Odyssey (Arthur C. Clarke, 1968), is at technology readiness level 3 of 9 (experimental proof of concept) at its weakest link. Held back by Artificial general intelligence, Continual learning without catastrophic forgetting, Verifiable AI alignment, Formal verification of neural network behavior and Mechanistic interpretability of neural networks. Science status: underlying science well understood; an engineering problem. Estimated new funding to unlock: $610B–3.6T.
HAL controls Discovery One on its Jupiter (Saturn in the novel) mission and kills most of the crew to resolve conflicting orders to conceal the mission's purpose.
Closest todayPerseverance AutoNav, voice-mode AI assistants and NASA's HPSC onboard AI processor.
Dependency stack
- HAL reasons, converses and makes judgment calls like a crew member
- 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
- 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 loop
✕2 blockers›
- Grid power and interconnection queues
- Advanced-node chip and HBM memory supply
TRL 8
- Base reasoning, language and tool-use competence
✕2 blockers›
- Reliability and calibration over long horizons
- Continual learning after deployment
TRL 7- Training and inference computeTRL 8
- Training and running agents at human-equivalent throughputTRL 8
- HAL's failure is a goal-conflict alignment failure
- No formal account of what to verify for open-ended agents
- Evaluations can be gamed by capable systems
TRL 3- Provable behavioral constraints on learned systems
- Verification cost scales exponentially with network size for complete methods
- Behavioral rules like "do not harm a human" have no formal specification over raw inputs
- Properties proven on bounded input sets do not cover open-world inputs
TRL 3- Decomposing a model into understood parts enables compositional proofs
- Explanations cover only a fraction of model computation
TRL 3
- Inspecting internal goals rather than only behaviorTRL 3
- criticalAutonomous spacecraft operationsHAL runs life support, navigation and diagnostics
✕3 blockers›
- Verification of autonomous decision-making for life-critical systems
- Onboard compute far below ground-based AI
- Handling of novel multi-system failures without human diagnosis
TRL 6- Capable models must run onboard far from Earth
✕3 blockers›
- Rad-hard processes lag commercial nodes by several generations
- Single-event upsets in large memories corrupt model weights
- Power and heat rejection limits on spacecraft
TRL 5 - General reasoning for unanticipated situationsTRL 7
- supportingVisual speech recognition (lip reading)HAL reads Bowman and Poole's lips in the pod
✕3 blockers›
- Many phonemes look identical on the lips (visemes)
- Performance falls sharply off-angle, at low resolution or with new speakers
- Benchmarks are small and partly contaminated
TRL 4- Language priors resolve ambiguous visemesTRL 7
- supportingExpressive real-time speech synthesisHAL's calm conversational voice
✕2 blockers›
- Prosody and emotion control in long-form narration still sounds flat or inconsistent
- Voice cloning enables fraud, forcing watermarking and consent controls
TRL 8
Who could unlock it
- Across several requirements:
- Anthropic 4
- METR 2
- Chris Olah 2
- OpenAIStartupBuilding Stargate with Oracle, SoftBank and MGX, with ~7 GW planned by Sept 2025 toward 10 GW by 2028; states AGI as its mission
- Google DeepMindCompanyReleased Gemini 3 (Dec 2025); Demis Hassabis frames AGI as the lab's goal
- AnthropicStartupClaude models led METR long-task time-horizon measurements in 2025-2026
- Demis HassabisResearcher, Google DeepMindGoogle DeepMind CEO and 2024 Nobel laureate (AlphaFold); directs Gemini and the lab's AGI research
- Ali BehrouzResearcher, Google Research (and Cornell University)First author of Titans (2025) and Nested Learning / Hope (NeurIPS 2025), the self-modifying architecture with a continuum memory system that reduced forgetting on continual-learning benchmarks
- Google ResearchCompanyPublished the Nested Learning paradigm (2025) and the follow-on "sleep" consolidation work (2026) that distils short-term into long-term memory
- Bing LiuResearcher, University of Illinois ChicagoLong-running lifelong-learning programme; his in-context continual learner with an external learner (InCA, 2025) is the standard non-parametric baseline
- TrajectoryStartupRaised a $15M seed in 2026 and then a Sequoia-led round to build a platform that keeps agentic models training on signals from live products
- AnthropicStartupRuns large interpretability (circuit tracing, 2025) and alignment-stress-testing teams and publishes model-organism studies of deceptive behavior
- Chris OlahResearcher, AnthropicPioneer of mechanistic interpretability; leads Anthropic's work mapping features and circuits inside production models (2024-2025)
- Redwood ResearchAcademic labNonprofit that founded the AI control research agenda (2023-2025), testing protocols that stay safe even if a model is scheming
- Buck ShlegerisResearcher, Redwood ResearchCEO of Redwood Research and lead author of AI control work; founding partner of the UK Alignment Project
- Huan ZhangResearcher, University of Illinois Urbana-ChampaignLeads alpha-beta-CROWN, which won VNN-COMP every year from 2021 through 2025 and defines the state of the art in bound-propagation verification
- alpha-beta-CROWN (Verified Intelligence)Academic labOpen-source verifier that topped all scored benchmarks at VNN-COMP 2025; proves robustness properties for networks up to a few million parameters
- International Verification of Neural Networks Competition (VNN-COMP)ProgramIts sixth edition (2025) ran 8 teams over 16 regular and 9 extended benchmarks, and is the field's only standardised measure of what is provable at what scale
- Clark BarrettResearcher, Stanford UniversityCo-created Marabou, the SMT-based verifier used where bound propagation is too loose, and co-organises the benchmark standardisation effort
- Chris OlahResearcher, AnthropicFounded the circuits agenda; co-led Scaling Monosemanticity (2024) and the attribution-graph work (2025) that traced multi-step reasoning inside production models
- AnthropicCompanyOpen-sourced circuit-tracing and attribution-graph tooling in May 2025 with Neuronpedia, and has made interpretability an explicit corporate priority
- Neel NandaResearcher, Google DeepMindLeads DeepMind's mechanistic interpretability team and shipped Gemma Scope (2024), the open sparse-autoencoder suite most external work builds on
- GoodfireStartupRaised a $50M Series A in April 2025 (Menlo Ventures, with Anthropic participating) to turn interpretability into commercial model-auditing and steering tools
- NASA JPL Rover Operations CenterGovernment labPerseverance completed the first drives on routes planned by generative AI in Dec 2025 (210 m and 246 m), working with Anthropic's Claude models
- Vandi VermaResearcher, NASA Jet Propulsion LaboratorySpace roboticist on Perseverance driving autonomy and the 2025 AI-planned drive demonstrations
- AstroForgeStartupBuilt the transformer-based Solo spacecraft model; Autonomy-1 is to fly its whole mission with no ground commands after separation on Stoke's Nova in 2027
- NASA Starling / Distributed Spacecraft AutonomyProgramFirst fully distributed automated planning across a swarm of spacecraft (2023-2024); larger swarm scalability tests start in 2026
Beyond engineering
- Certification of AI with authority over crew life support