Personal development using an AI development method for studying incarcerated youth.

Context

I love how dedicated the LessWrong community is to improving AI… but I don't see much application of all that science to personal development. Of course, evolutionarily, it may be cheaper to change the external than the internal. But it's worth paying the price for me, and for you? Haha.

I'm building an app—and I want to study its method with young people in prison—to vectorize goals and values using methods inspired by AI development. Does that make sense or is it crazy? Haha. I'm sharing the full proposal below.

Summary
KeyMetas It is an operational, open-source framework that links personal goals, latent vectors, and extreme moments. It is used to update beliefs based on evidence from observable inputs (available hours, resources, experience, past signals, etc.) and compare them with pre-established latent values to avoid self-assessment errors (Goodhart).

We present:

1) the operating method;

2) a reproducible mini-pilot experiment;

3) safeguards against Goodhart and failure modes;

4) a mini-dataset and example notebook.

I will publish the code on GitHub and send the apk to those interested.

1. What problem are we addressing?

Many people have goals, but may lack a systematic way to compare them, estimate the likelihood of success and cost, and prioritize in a way that incorporates both personal effects and external impacts.

KeyMetasIt is about:

1) make explicit the implicit assumptions about why a goal matters;

2) infer latent characteristics (motivation, real cost, external dependence) from observable signals;

3) help prioritize and monitor without turning those metrics into dictators.

If we had to explain in 20 seconds why this goal is worth your time, what evidence would you cite? This body of evidence is the raw material for inferring latent vectors.

2. Brief structure of the (operational) framework

Dimensions(each one is operationalized numerically):

• Impact: a mix of expected reward and risk. We propose a relative benchmark scale: we define a "Reference goal of benefit and cost = 100%" (e.g., completing a technical course that historically creates jobs, helps me and others, but also costs 100%) and we score other goals as a percentage relative to that benchmark. In addition, we maintain a subjective probability of success (0–100%).
  
   
• AddressWhere should we intervene to increase probability? Chart showing the proportion of inner or outer edges.
  
   
• Scope (level of dedication required): categories ordered by level of possibilities (Elemental, Emotional, Informational, Social) compared proportionally with 100% graphs.
  
   
• Latent value: outcome in inferred necessary values that correlates goal with values (e.g., care, recognition, effective altruism, empathy, tactics, strategy, execution, analysis, simplification, communication, cooperation). It is calculated using simple Bayesian inference on subjective (self-reported) inputs and prior observables.
  
   
• Hierarchy of goals, by benefit, cost and cost-benefit and latent values
  
   

Suggested observable inputs(for Fermi-type estimation):

• Available hours/week, self-proclaimed.
  
   
• Relevant experience or skills (0–5).
  
   
• External resources (none / low / medium / high).
  
   
• Past signals (previous attempts, % success).
  
   
• Dependence on third parties (0–1).
  
   
• External risks (quick list: financial, legal, health).
  
   
• Current subjective priority (0–10).
  
   
• Reference Result: Is this goal equal to, greater than, or less than the Reference Goal (100%)?
  
   
• 3. Inference method (three simple steps)
  
   

1. Feature vector: we transform the inputs into a meta-normalized vector.
   
    
2. Latent inferenceWe apply PCA/factor analysis or a simple Bayesian model to identify latent dimensions (e.g., actual cost, external dependency, value alignment). For the initial version, PCA + cross-validation is sufficient.
   
    
3. Scoring and comparison: We project each goal onto the latent dimensions and calculate composite scores (Expected Impact × Risk-Adjusted Probability of Success). We show uncertainty (bands) and relative comparison to the Reference Goal.
   
    

Practical note: We don't provide a single, definitive number. We always show intervals and the contribution of each input to the score.

4. Mini-pilot experiment (reproducible plan)

Aim:detect signs of improvement in calibration and progress toward goals in 12 weeks.

Design:

• Pilot N ≈ 40–60 participants; randomized: Estimat (app + brief guide) vs control (traditional goal list).
  
   
• Proximal measurements (3 months):
  
   
  • Progress towards milestones (% complete).
    
     
  • Probability calibration (Brier score between estimated P and observed result).
    
     
  • Self-report of motivation and well-being (brief pre/post questionnaire).
    
     
  • Contextual metrics (workshop participation, relative behavior in institutional settings if applicable).
    
     

Analysis:Simple comparisons and hierarchical models to separate individual effects; open reporting of methods and anonymized data.

5. Possible failure modes and mitigations (ready to paste)

Failure modes:

1. Goodhart: convert a proxy (score) into a perverted target.
   
    
2. Mis-specification of the model: inferred latent values that do not represent actual values.
   
    
3. Overconfidence / illusion of certainty: Users rely too much on scores with high uncertainty.
   
    
4. Ethical problems in vulnerable populations(e.g., in prisons).
   
    

Mitigations:

• Multi-metric (not a single digit).
  
   
• Show uncertainty and penalize extremes (regularization).
  
   
• Human-in-the-loop: periodic review and override possibility.
  
   
• Validation with holdouts and correlation monitoring between proxies and outcomes.
  
   
• Pre-registration and ethics review for pilots in prison settings.
  
   

6. Visualization and accessibility.

 

• Large bars with intervals, clear icons, and brief explanations next to each metric.
  
   
• "Story mode" (1–2 sentences summarizing what the punctuation means) for later AI API support.
  
   
• Future conversational interface that repeats what the user says in text and audio mode
  
   

7. Proximal (easy to collect) metrics — definitions and how to calculate

1. % of milestones completed:(milestones achieved / milestones planned) × 100.
   
    
2. Probability calibration (Brier score):asks for P(achieve goal) at the start and compares it with outcome (1/0) at the end; lower Brier score = better calibration relative to latent values.
   
    
3. Monthly motivation self-assessment:0–10 pre/post scale; compare mean change.
   
    
4. Well-being (WHO-5 or short scale):pre/post; report mean change and % with improvement ≥ X points.
   
    
5. Objective engagement within the center:workshop attendance (%) and active participation (session count).
   
    
6. Institutional indicators (if possible):reduction of disciplinary incidents per participant (if ethical/possible to use).
   
    
7. Academic results / courses completed(e.g. % who complete technical course).
   
    
8. Long-term measures (if possible):post-release employment, readmission/relapse (report only if authorized and with ethical considerations).

8. Suggested title and meta-description

Title:"Vectorizing Goals: KeyMetas — An Open Framework, a Reproducible Pilot, and How to Avoid Goodhart"

Meta-description (for LW):We propose KeyMetas, a framework for representing goals as latent vectors inferred from observable evidence. We present an operational method, an example mini-dataset, a pilot design, and safeguards against failure modes.

9. Next steps (immediate action)

1. Upload notebook + mini-dataset a GitHub.
   
    
2. Prepare a post in Markdown (copyable): introduction, method, experiment, expected results, failure modes, links.
   
    
3. Request community review (internal peer review) before publishing.
   
    

attachment
Onboard 

keyMetas

Estimate your goals by AI development model.

Score your goals based on impact, direction, scope, and underlying values. How closely do your goals relate to the factors that shape you?

Sometimes, achieving goals can be helped by more than just positive; it can be interesting to compare evidence, a more rational approach, that could be help to be more positive.

keyMetas break down your goals into key features and apply a similar approach to the latent vectors in AI models: inferring unobserved features from observed data and correlating them with latent values to avoid self-assessment errors (Goodhard).

Compare your goals to each other to make probabilistic estimates of them and score them on:

• Impact– Compare risk and reward between your goals and compare evidence.
  
   
• Address– Identify where intervention is most needed: in yourself or in your environment.
  
   
• Scope– Determine the level of commitment:
   • Elemental
   • Emotional
  • Informational
   • Social
  
   
• Latent value– Correlate your goals with your latent values to increase motivation and avoid errors in specific assessments (Goodhart).

With this, we have a more solid basis for evaluating key moments to determine your priorities and how your goals truly relate to all the evolutionary factors that shaped you as a person. Sometimes you need to shift priorities among your goals, or sometimes you need to find factors that contribute most to satisfaction and motivation in your goals.

 

Each step has an information icon (“i”) that explains the logic and evidence behind it.

The full open-source project on GitHub, along with the theoretical framework, scoring formulas, and practical examples, are published separately and linked.

keyMetas is supported by Jacominesp to help you discover and appreciate yourself in both senses.(I don't think so! A play on words, to estimate in love and to approximate in calculations?).

Definitions of latent values with information theory 

It has some upgrades but the latest one published is at:
https://www.lesswrong.com/posts/F7oBMckEz3TxKukZy/8-latent-values-a-simplified-construction-from-maxent 

Personal experience
 

I have experience applying this in my own life (I've been trying it out since dealing with corruption in my former profession as a military firefighter for 11 years):
https://www.lesswrong.com/posts/NuZjME7u5goHiCnai/emotional-theory-for-a-disorder-manual-on-how-not-to-freeze 

https://www.lesswrong.com/posts/zzMbNKYaGqazwBnjF/when-the-wannabe-rambo-comedian-cried 

 and I have some evidence that it can help others looking for more targeted ways to improve their motivation and personal satisfaction.