Human oversight of AIs could occur at different timescales: Slow (days-weeks)[1] and Fast (seconds-minutes)[2].
The community has mostly focused on Slow Human Oversight. This makes sense: It is likely that weak trusted AIs can perform all tasks that humans can perform in minutes.[3] If so, then clearly those AIs can replace for humans in Fast Oversight.
But perhaps there are cases where Fast Human Oversight is helpful:
High-stakes decisions, which are rare enough that human labour cost isn’t prohibitive.
Domains where weak trusted AIs are insufficiently robust.
For Fast Human Oversight, reading text won’t work, because humans process text too slowly. Instead, we will need to leverage the visual cortex — it has higher bandwidth, better parallel processing, better anomaly detection.
Here are some techniques we could use:
Emoji logging
Prompt/finetune models to annotate their outputs with emojis indicating their current activity (🔍 for searching, ⚠️ for uncertainty, 🔄 for backtracking, etc).
Probe-colored tokens
Display generated tokens with colors based on probe values (e.g., deception probe). Use 2D color space (hue + saturation) to show two different probes simultaneously.
A chat interface with Llama-3.3-70B with tokens coloured via a deception probe. Source: Parrack et al.
Activation-to-image
Convert model activations to images representing the internal thoughts, e.g. using a standard text-to-image model. For example, when the model thinks of “dog” then a picture of a dog appears.
Activation-to-video
Convert a stream of activations into a video stream, providing a continuous view into the AI’s internals.
If we have hundreds of AIs operating on a single project simultaneously, then we could convert the list of activation streams into a single video stream.
Multi-agent swarm visualization
For monitoring many AI agents simultaneously, display as particle swarms with color-coded states. Interactions between AIs are visualized via a network.
You get used to it. I don’t even see the code. All I see is blonde, brunette, redhead.
Collect logs/transcripts/activations from suspicious behavior, run controlled experiments on model internals, understand failure modes, hold stakeholder meetings, negotiate safety measures, decide whether to halt deployment or modify training procedures.
See Richard Ngo’s t-AGI framework, which characterizes the ‘capability’ of a model by the duration of tasks:
A 1-minute AGI would need to beat humans at tasks like answering questions about short text passages or videos, common-sense reasoning (e.g. Yann LeCun’s gears problems), simple computer tasks (e.g. use photoshop to blur an image), justifying an opinion, looking up facts, etc.
See also METR’s Time Horizon’s evaluations, which I think vindicates Richard’s framework at least in specific domains.
Visual Cortex in the Loop:
Human oversight of AIs could occur at different timescales: Slow (days-weeks)[1] and Fast (seconds-minutes)[2].
The community has mostly focused on Slow Human Oversight. This makes sense: It is likely that weak trusted AIs can perform all tasks that humans can perform in minutes.[3] If so, then clearly those AIs can replace for humans in Fast Oversight.
But perhaps there are cases where Fast Human Oversight is helpful:
High-stakes decisions, which are rare enough that human labour cost isn’t prohibitive.
Domains where weak trusted AIs are insufficiently robust.
For Fast Human Oversight, reading text won’t work, because humans process text too slowly. Instead, we will need to leverage the visual cortex — it has higher bandwidth, better parallel processing, better anomaly detection.
Here are some techniques we could use:
Emoji logging
Prompt/finetune models to annotate their outputs with emojis indicating their current activity (🔍 for searching, ⚠️ for uncertainty, 🔄 for backtracking, etc).
Probe-colored tokens
Display generated tokens with colors based on probe values (e.g., deception probe). Use 2D color space (hue + saturation) to show two different probes simultaneously.
Source: Parrack et al.
Activation-to-image
Convert model activations to images representing the internal thoughts, e.g. using a standard text-to-image model. For example, when the model thinks of “dog” then a picture of a dog appears.
Activation-to-video
Convert a stream of activations into a video stream, providing a continuous view into the AI’s internals.
If we have hundreds of AIs operating on a single project simultaneously, then we could convert the list of activation streams into a single video stream.
Multi-agent swarm visualization
For monitoring many AI agents simultaneously, display as particle swarms with color-coded states. Interactions between AIs are visualized via a network.
Collect logs/transcripts/activations from suspicious behavior, run controlled experiments on model internals, understand failure modes, hold stakeholder meetings, negotiate safety measures, decide whether to halt deployment or modify training procedures.
Monitor real-time outputs, recognize deceptive patterns or dangerous trajectories, determine if immediate intervention needed, execute killswitch or block outputs.
See Richard Ngo’s t-AGI framework, which characterizes the ‘capability’ of a model by the duration of tasks:
A 1-minute AGI would need to beat humans at tasks like answering questions about short text passages or videos, common-sense reasoning (e.g. Yann LeCun’s gears problems), simple computer tasks (e.g. use photoshop to blur an image), justifying an opinion, looking up facts, etc.
See also METR’s Time Horizon’s evaluations, which I think vindicates Richard’s framework at least in specific domains.