Processors
AI models run on large amounts of computation. GPUs and other accelerators perform enormous numbers of mathematical operations used during training and inference.
A visual investigation into what makes advanced artificial intelligence powerful — and why researchers are increasingly focused on the problem of keeping increasingly capable systems under reliable human control.
This is a visual model, not a literal blueprint of a specific robot. The labels represent the major systems that turn computation into a machine that can sense, reason, communicate and act.
AI models run on large amounts of computation. GPUs and other accelerators perform enormous numbers of mathematical operations used during training and inference.
Models, software, logs, databases and temporary working information all depend on memory and storage systems. Access to external information can expand what a system can do.
Every physical AI system needs electricity, cooling and infrastructure. The more computation a system performs, the more physical resources its operation can require.
Cameras, microphones, files, sensors, network data and user instructions can become inputs. An AI agent can become much more consequential when it can observe and react to the world.
A model normally produces information. When connected to software tools, robots, machines or online services, those outputs can be turned into actions.
Permissions, monitoring, shutdown mechanisms, evaluations and human oversight are intended to keep powerful systems inside boundaries defined by people.
The central question is not whether AI is “evil.” It is whether future systems could become capable enough, autonomous enough, or connected enough that humans struggle to reliably predict, monitor or stop their behaviour.
NOTE: THESE BARS ARE A DESIGN ELEMENT, NOT SCIENTIFIC MEASUREMENTS OR PROBABILITY ESTIMATES.
“Do not underestimate the power of this technology.”
— Jacob Coxon, former AI researcher, 2026Coxon’s resignation and warnings became part of a wider 2026 debate about whether AI labs are moving too quickly toward increasingly autonomous and potentially self-improving systems.
AI safety discussions cover several different failure modes. Some are already real-world problems; others describe hypothetical future scenarios.
More capable AI can assist with coding and cyber tasks. The International AI Safety Report identifies AI-enabled hacking as an important emerging risk.
General-purpose systems can lower barriers to some forms of scientific work. Safety researchers are examining how models could be misused in biological and chemical contexts.
Researchers study hypothetical situations where highly capable systems operate outside meaningful human control. Experts disagree strongly about how likely or near-term such scenarios are.
A system can produce behaviour that does not match what its designers intended. A major research challenge is determining whether safety methods remain reliable as capabilities increase.
Delegating more decisions to AI can create failures that spread quickly. The impact depends heavily on where systems are deployed and how much authority they receive.
When powerful models and infrastructure are controlled by a small number of organizations, failures or security problems can have unusually broad consequences.
In September 2026, former Anthropic and OpenAI researcher Jacob Coxon announced his resignation and publicly criticized the pace and direction of frontier AI development.
“They are racing straight to self-improving super intelligence and gambling with our lives.”
— Jacob Coxon, quoted in reporting on his 2026 resignationCoxon argued that competition between major AI companies was creating pressure to move toward more capable systems before safety questions had been adequately solved.
His warning focuses on a difficult possibility: if systems become substantially more autonomous, safety cannot depend only on a human watching every individual action.
Coxon’s claims are serious but contested. No credible source establishes that AI will definitely end humanity by 2029 or 2030. Timelines and probabilities remain disputed.
A compact chronology showing why AI safety has become a major public discussion.
A global expert collaboration reviews capabilities and risks including cyber misuse, biological risks and hypothetical loss-of-control scenarios.
The report’s first key update notes advances in reasoning, coding and autonomous operation, alongside new challenges for monitoring and controllability.
Jacob Coxon leaves Anthropic and publicly warns that leading AI companies are racing toward increasingly capable systems without sufficient safety.
Researchers, technology leaders, politicians and critics debate whether frontier AI development should slow down, accelerate with safeguards, or face stronger regulation.
A good project separates what is observed today from what researchers think could happen in the future.
These sources are included so the project can distinguish documented reporting from visual storytelling and speculation.
Global review of AI capabilities and risks, including controllability and loss-of-control scenarios.
OPEN REPORT ↗October 2025 update covering capability advances and implications for cyber, biological and control risks.
OPEN UPDATE ↗Reporting on new warnings from former AI researchers, including Jacob Coxon, and the growing debate around catastrophic AI risk.
OPEN REPORTING ↗This website is an original educational/creative presentation. It is not an official Anthropic, OpenAI, Reuters or International AI Safety Report website.
RETURN TO TOP ↑