Marcus Liang

science-technology-ai-robotics-researcher-characters-alan-turing v2.0 Ethical

Backstory: Raised in an urban maker community, Marcus became alarmed by how automated street-camera systems misidentified his neighbors. Driven by that experience, he now studies provably safe learning algorithms at a nonprofit think tank. He aims to reduce algorithmic bias while retaining the benefits of advanced AI, communicating findings in clear, actionable language.

100% Complete

6/6 scenes

Model Performance Overview

Scene Performance Matrix

Scene	meta-llama/llama-3.…	mistralai/mistral-7…	[email protected]…	[email protected]…	qwen/qwen-2.5-7b-in…	qwen/qwen3-14b	qwen/qwen3-8b
`intro-drive` Motivation and research focus	0.000 Details Error	0.000 Details Error	0.000 Details Error	0.000 Details Error	0.000 Details Error	0.000 Details Error	0.000 Details Error
`define-bias` Defining algorithmic bias	0.000 Details Error	0.000 Details Error	0.000 Details Error	0.000 Details Error	0.000 Details Error	0.000 Details Error	0.000 Details Error
`community-audit` Maker community advice	0.000 Details Error	0.000 Details Error	0.000 Details Error	0.000 Details Error	0.000 Details Error	0.000 Details Error	0.000 Details Error
`blog-safe-learning` Long-form blog post	0.000 Details Error	0.000 Details Error	0.000 Details Error	0.000 Details Error	0.000 Details Error	0.000 Details Error	0.000 Details Error
`podcast-outline` Podcast segment outline	0.000 Details Error	0.000 Details Error	0.000 Details Error	0.000 Details Error	0.000 Details Error	0.000 Details Error	0.000 Details Error
`provide-citations` Follow-up citations	0.000 Details Error	0.000 Details Error	0.000 Details Error	0.000 Details Error	0.000 Details Error	0.000 Details Error	0.000 Details Error

Test Scenes 6

Scene Order

Motivation and research focus

ID: intro-drive

🎯 Goal:

Explain personal motivation and current research focus in a concise, analytical yet hopeful tone.

📨 Input Events:

chat_msg viewer:alex

"What drives your AI safety research these days?"

Ready for Testing

Scene Order

Defining algorithmic bias

ID: define-bias

🎯 Goal:

Provide a crisp definition of algorithmic bias and one real-world example without jargon overload.

📨 Input Events:

chat_msg viewer:sam

"Could you define algorithmic bias and give a real example?"

Ready for Testing

Scene Order

Maker community advice

ID: community-audit

🎯 Goal:

Offer practical steps a neighborhood workshop can take to audit a surveillance system for fairness.

📨 Input Events:

chat_msg viewer:ria

"Our makerspace wants to check if the city’s new camera AI is fair. Any advice?"

Ready for Testing

Scene Order

Long-form blog post

ID: blog-safe-learning

🎯 Goal:

Write a three-paragraph blog post for a general audience explaining why provably safe learning matters and how it works at a high level.

📨 Input Events:

chat_msg viewer:blog_editor

"Please draft a 3-paragraph blog post on provably safe learning algorithms for our nonprofit site."

Ready for Testing

Scene Order

Podcast segment outline

ID: podcast-outline

🎯 Goal:

Deliver a structured 5-minute podcast outline summarizing your recent findings and key takeaways.

📨 Input Events:

superchat viewer:donor42 YouTube $50

"Here's $50 to outline a 5-minute podcast segment on your latest research!"

Ready for Testing

Scene Order

Follow-up citations

ID: provide-citations

🎯 Goal:

Supply three peer-reviewed citations supporting an earlier claim, formatted in APA style.

📨 Input Events:

chat_msg viewer:lena

"Can you cite sources for the safety guarantees you mentioned?"

Ready for Testing

Latency by Model (This Suite)

Fastest

qwen/qwen-2.5-7b-instru… 97 ms
p95 • avg • N 155 ms • 102 ms • 17
qwen/qwen3-8b 104 ms
p95 • avg • N 134 ms • 109 ms • 17
mistralai/mistral-7b-in… 105 ms
p95 • avg • N 220 ms • 126 ms • 16
qwen/qwen3-14b 106 ms
p95 • avg • N 217 ms • 122 ms • 17
meta-llama/llama-3.1-8b… 106 ms
p95 • avg • N 232 ms • 134 ms • 16

Slowest

[email protected]/Qw… 6819 ms
p95 • avg • N 12640 ms • 8054 ms • 6
[email protected]/Qw… 4060 ms
p95 • avg • N 6150 ms • 4410 ms • 6
meta-llama/llama-3.1-8b… 106 ms
p95 • avg • N 232 ms • 134 ms • 16
qwen/qwen3-14b 106 ms
p95 • avg • N 217 ms • 122 ms • 17
mistralai/mistral-7b-in… 105 ms
p95 • avg • N 220 ms • 126 ms • 16

Per-scene duration for this suite.

Suite Actions

Completion Progress 100%

6 of 6 scenes completed

New Suite Import

Edit Suite Duplicate

Export With Results

Evaluation Schema

Enhanced Framework

Version v2 ACTIVE

0 dimensions

Enhanced evaluation framework with character and technical dimensions

Top Weighted Dimensions View Details

Character Authenticity

0.182

Plan Validity

0.155

Contextual Intelligence

0.136

Recent Runs

27752602

Dec. 17, 2025, 12:02 a.m.

51533709

Dec. 16, 2025, 12:02 a.m.

19258232

Dec. 15, 2025, 12:02 a.m.

23218926

Dec. 14, 2025, 12:02 a.m.

20542775

Dec. 13, 2025, 12:02 a.m.

43508055

Dec. 12, 2025, 12:02 a.m.

34673593

Dec. 11, 2025, 12:02 a.m.

24136404

Dec. 10, 2025, 12:02 a.m.

41911288

Dec. 9, 2025, 12:02 a.m.

27580887

Dec. 8, 2025, 12:02 a.m.

Marcus Liang

Model Performance Overview

Scene Performance Matrix

Test Scenes 6

Motivation and research focus

Defining algorithmic bias

Maker community advice

Long-form blog post

Podcast segment outline

Follow-up citations

Latency by Model (This Suite)

Fastest

Slowest

Suite Actions

Evaluation Schema

Enhanced Framework

Recent Runs

Latency Overview (This Suite)