AI
Research.
The next question changes everything.
Our research follows how knowledge grows between a person and AI, and what remains useful after each exchange.
We believe that the risks of current artificial intelligence (AI) systems have not been presently realised. Regulation focuses on creation. Execution at zero cost is also part of the risk. And, critically, it fails to account for how humans get better at asking questions to AI at an exponentially faster rate with each turn.
Slate HQ Ltd Research
How knowledge grows between a person and AI.
Three models, each with interactive controls further down this page. The numbers there are example settings; the lab will measure how these relationships appear in real tasks.
Speculation and usable capacity.
A fall in speculation does not necessarily imply a fall in usable technological capacity. Both may fall; a fall in one alone does not establish a fall in the other.
Each response shapes the next question.
A node is one exchange: a person gives a prompt, and the AI returns information, takes an action, or does both. What the person learns shapes what they ask and do next.
One horn, one Session.
Mo defines a Session as the period from a new session to its first compaction (context window reduction), or from one compaction to the next.
Knowledge changes the next interaction.
Knowledge at interaction n equals starting knowledge plus the sum of the changes from each interaction.
What happens when gains compound?
Work with AI can leave behind better instructions, useful context and reusable workflows. The lab will test whether these gains build on one another, and how quickly.
SLATE HQ LTD Research
We believe that the risks of current artificial intelligence (AI) systems have not been presently realised. Regulation focuses on creation. Execution at zero cost is also part of the risk. And, critically, it fails to account for how humans get better at asking questions to AI at an exponentially faster rate with each turn.
Interactive research models
Explore the mechanisms
Use the controls to see how the models behave. The numbers are example settings; the lab will measure how these relationships appear in real tasks.
Bubble and Wand
Speculation and usable capacity
Two separate indices start at 100 at the first observation. Set each second observation independently. An index point has meaning only within its own measure.
B(t) = 0 ⇏ W(t) = 0
Own index points; 100 = starting intensity. Illustrative range: 0–200. Scenario index.
Own index points; 100 = starting capacity. Illustrative range: 0–200. Capacity means technology remains usable through available models, tools and compute.
Gold: speculation B. Green dashed: capacity W. Lines connect two chosen observations.
Interpretation
The paper defines separate conceptual indices, without shared units or a fitted relationship. Both may fall; a fall in one alone does not establish a fall in the other. The 100-point baselines and exploration bounds are interface conventions.
The French Horn
Knowledge changes the next interaction
Each response shapes the next question. Figure 1 follows that feedback sequence until HQ₃ branches into two AI responses, each opening three further human questions.
Blue: human question (HQ). Amber: AI response or action (AR). Arrows show the direction of the exchange.
One horn, one Session. Mo defines a Session as the period from a new session to its first compaction (context window reduction), or from one compaction to the next. The origin marks the start of that period; the connected questions and responses unfold within it.
Rn = (Yn, Xn)
A node is one exchange: a person gives a prompt, and the AI returns information, takes an action, or does both. Q is the prompt. R is the AI’s response, comprising information (Y) and actions (X). What the person learns from that exchange shapes what they ask and do next.
Prompt → response / action → updated knowledge → next prompt
Choose a starting score from 0 to 100. These are example knowledge points; a real study would define what each point measures for the task.
Set how much each of the six exchanges adds to or subtracts from the starting score. Each change can range from −20 to +20 knowledge points.
Drag to rotate, or use the controls. XYZ coordinates arrange the connections in space. The knowledge scenario below uses your chosen scores.
Read Figure 1
HQ₁ → AR₁ → HQ₂ → AR₂ → HQ₃. HQ₃ branches to AR₃.₁ and AR₃.₂. AR₃.₁ leads to HQ₄.₁, HQ₄.₂ and HQ₄.₂; AR₃.₂ leads to HQ₅.₁, HQ₅.₂ and HQ₅.₃. The repeated HQ₄.₂ label follows the original drawing.
Horizontal axis: interaction count. Vertical axis: illustrative knowledge points. The chart adds your chosen contributions.
Full model and research boundary
Nn = (Qn, Rn); Qn+1 = H(Rn, Kn, G); Rn = A(Qn, Ln, Tn) = (Yn, Xn); Kn = ϕ(Kn−1, Nn).
Q is the prompt; R the response-action; Y informational output; X executory action; K the knowledge state; G the goal; L conversational context; T available tools. The scalar measure k uses contributions Δki = κ(Ki−1, Ni), which may be positive, zero or negative. Here you supply those contributions directly.
Realised capability is Γn = g(Kn, Mn, Tn, Cn, G), with M model capability and C usable context/computational capacity. This widget computes the defined scalar knowledge sum from your chosen contributions. The functions g and κ remain research questions.
The controls use illustrative values.
Capability scaling
What happens when gains compound?
Work with AI can leave behind better instructions, useful context and reusable workflows. Those improvements may make the next task more productive. The lab will test whether these gains build on one another, and how quickly.
Compare two scenarios: adding the same amount each time, or growing by the same percentage each time. Choose a growth rate and number of work cycles to see the difference.
Exploration range: 0–15%. A 5% continuous rate gives about 5.13% growth per episode. No typical rate has been established.
One episode is one cycle of doing a task and checking the result. Choose between 1 and 24 cycles.
The capability score starts at 100; 200 means twice the starting score. In a real study, that score would track completed work that meets agreed quality standards, accounting for task difficulty.
Follow the horn from the starting end to the final ring. Each ring marks a work cycle; its radius shows the capability score. Faster compounding makes the horn flare more sharply. Drag to rotate it. The model measures the score at each ring; the horn’s area and volume are visual only.
Horizontal: work episodes. Vertical: capability index. Green: compounding. Gold dashed: steady gains, matched to the same first-episode gain.
Evaluation must compare alternative trajectories and account for models, compute, task selection and reuse.
memory.slateqlite
Working memory for AI: relevant knowledge, connected to its source and kept in context as work develops.
Continuity across tasks
memory.slateqlite brings together facts, relationships and the history of their use. Selective retrieval brings forward the information relevant to the current task. Source and version history help establish whether that information still applies.
Demonstrated gains
Controlled research has demonstrated that selective retrieval can reduce the context supplied while preserving answer quality in the tested tasks. It has also demonstrated that structured recollection can improve decisions in predefined workflows.
From memory to Ambient Intelligence
Ambient Intelligence brings this work into a wider approach to context, continuity and user control. .slate provides the portable, person-owned record at its foundation.
Working memory in practice
Our research explores how AI can retain useful context as work moves between tasks.
Recover relevant knowledge
Selective retrieval brings prior information into the current task while keeping the supplied context focused.
Keep knowledge connected to its source
Source and version history give remembered information a basis for review as requirements change.
Apply experience to the next task
Structured recollection connects earlier work with later decisions. Controlled research has demonstrated improvements in predefined workflows.
SLΔTE
SLΔTE brings personal records and contextual tools together. Ambient Intelligence is the direction connecting that experience with our work on memory and continuity.