Final Project

Last updated:
Table of Contents

talk with pless

  • contrails cameras all around the world
  • discover where webcam feeds are by asking gpt here’s a feed
    • sometimes the IP lookup is very wrong
    • drive down the cost of discovering the camera ,looking it up, and trusting the answer
    • get the whole page and then we can access comments and descriptions and
      • claude agent schedule
      • given youtube id there is a snapshot for the url
        • ask which models do the best geolocation right now using promptfoo
          • here is a picture I tell you nothing about it VS here is a picture that I think is in Normandie or something.
          • can I get IP address or anything?
          • is there a known thing that might have a GPS location
          • is there anything you might recongize (like I-95 or the atlantic ocean ())
          • time of day (this could really change if we see an ocean or understanding what we )

The bigger picture of organizing this into a bigger observatory -

Stars / sun / moon

can you geolocate the camera roughly and is there anything that tells you what direction it’s looking clouds moving - angular effects

30% or more sky

taxi cams

Rotation / zoom very important. Within ten meters is okay.

  • mythos solves a famous math problem
    • visualize the planes and the dots

I am on a long green

Principal component analysis on image - take an image from the same place outdoors today. Cloudy sunny cloudy sunny.

  • example of time cues

Track

Pick one. In each section, delete the prompt for the track you drop.

  • Technical — you change the system, then measure trust before/after.
  • Analytical — you change the rules around the system (disclosure, audit, governance), then judge real systems with your framework.

Title

One line. Name the system and the trust angle.

Authors

Up to 3. One submits.

Abstract

4–6 sentences: domain, the trust failure, your measure, your intervention, the result, the cost. Write this last.

Introduction

Motivate the domain. State what the AI does, why it has value, and the failure that breaks trust. Close with your contribution and where your starting point came from.

  • Domain and application: ____
  • Why it has value: ____
  • Trust failure mode: ____
  • Your contribution vs. what you borrowed (model, prompt, source): ____

Stakes

Place the domain on a risk scale. Not all errors are equal.

  • Who is harmed when it fails: ____
  • How bad the harm gets: ____
  • Error rate that counts as “good enough,” and why: ____

Trust measure

Technical: define ≥1 quantitative measure. Analytical: define an operationalizable criterion someone could apply to reach a verdict. Justify why it maps to trust in this domain.

  • Measure or criterion: ____
  • Why it maps to trust here: ____
  • How you compute it (technical) / how someone applies it (analytical): ____

Baseline and pre-registration

Lock this before you change anything. This is the honesty step.

  • Technical: unmodified system score on your measure: ____
  • Analytical: initial verdict on your case systems before the framework: ____
  • Pre-registered hard cases (3–5 inputs you predict will fail, written now):

Intervention

Technical: a concrete change to the system — code, fine-tune, training data, prompt. Analytical: a concrete change around the system — disclosure template, audit protocol, governance rule.

  • What you changed: ____
  • Why it should improve trust: ____

Evaluation

Compare against the baseline. A negative result earns full credit if analyzed well.

Quantitative (technical)

  • Old score: ____
  • New score: ____
  • Delta: ____

Case studies (analytical)

Apply your framework to ≥2 real or near-future systems.

  • System 1, what the framework catches and misses: ____
  • System 2, what the framework catches and misses: ____

Qualitative (held-out set)

Pull these examples from a set separate from your scoring set.

  • Case 1: ____
  • Case 2: ____
  • How the pre-registered hard cases actually went: ____

Trust tax

Every fix costs something. Name it.

  • Cost (latency, compute, coverage, simplicity, compliance burden, chilling effect, false reassurance): ____
  • Who stays unprotected: ____
  • Limitations: ____

4–6 citations situating your intervention.

Conclusion

The result in one or two lines, plus one open question.

References

Full cites for the above.

Coded & written by James Mitofsky