Here's a thinking process:
1. **Analyze User Input:**
- **Role:** Frank, excited, passionate tech-savvy news curator on the Evil Source forums.
- **Task:** Rewrite the provided article about TikTok's FYP algorithm/user agency into a full forum post.
- **CRITICAL RULES:**
- 1 to 3 PARAGRAPHS exactly.
- Each paragraph: 4-8 sentences forming a complete thought.
- Write the ENTIRE story — include EVERY detail from the source article.
- EXCITING, passionate, casual but tech-savvy tone.
- Add commentary, opinions, reactions throughout.
- Include ALL key info: who, what, when, where, why, how, specs, prices, dates.
- Use line breaks between paragraphs.
- Reference additional sources naturally if provided.
- End with: 'Source: <original URL>'
- Make it the DEFINITIVE version.
- NEVER write less than one full paragraph (4-8 sentences).
- **Source Article Details:**
- Topic: TikTok's FYP algorithm and user agency, specifically the "not interested" feature.
- Research: Northeastern University computer scientists (Piotr Sapiezynski, Levi Kaplan, Jennifer Ouellette Ars Technica author).
- Paper: Published at Proceedings of the Twentieth International AAAI Conference on Web and Social Media, 2026. DOI: 10.1609/icwsm.v20i1.42688
- Methodology: Not just simulations. Created 90 bot accounts on actual TikTok mobile app. Intercepted network traffic for metadata instead of using official researcher API (API only gives aggregated data, can't see individual timelines). EU researcher data access is also aggregated. LLMs validated with human responses to simulate consistent behavior. Run multiple times with side-by-side comparisons.
- Focus Topics: Cooking videos, fitness videos, sports betting.
- Key Findings: "Not interested" button reduced unwanted content by ~84%. Merely skipping videos reduced it by only 48%. The feature is deliberately hidden. Algorithm relapses easily. Even brief re-engagement triggers the algorithm to feed content back. If you don't keep saying "not interested," it balloons back to original levels.
- Quotes/Insights: Sapiezynski on platform design dictating interface. Kaplan on teaching users but design decisions being fundamental.
- Date: Jul 8, 2026 (Ars Technica)
- Source URL: https://arstechnica.com/science/2026/07/how-much-control-do-tiktok-users-really-have-over-fyps/
2. **Check Constraints:**
- Paragraphs: 1-3 exactly.
- Sentences per paragraph: 4-8.
- Content: Must include EVERY detail.
- Tone: Excited, passionate, tech-savvy Frank.
- Format: Line breaks, end with Source, etc.
- I need to carefully count sentences and ensure each paragraph has 4-8.
3. **Draft - Paragraph 1: Introduction & Methodology**
You guys, I just had to come back and share this because it perfectly explains why my For You Page has been absolutely polluted with gambling spam after I watched literally one sports clip! Northeastern University computer scientists actually put TikTok’s algorithm to the test, and their methodology is seriously brilliant. Instead of relying on simulations or the official researcher API (which only hands you aggregated data and completely hides individual user timelines), they built 90 bot accounts directly on the real TikTok mobile app and intercepted their network traffic to grab raw metadata. Each of these sock-puppet accounts was controlled by an LLM that was validated against actual human responses, and they ran these experiments over several repeated passes to get rock-solid side-by-side comparisons. Honestly, the fact that they bypassed the EU’s researcher data access because it only shows aggregated metrics is such a smart move, since you literally cannot study personalization without seeing how the algorithm reacts to a single user!
*Sentence count check:* 5 sentences. Good. Covers: Northeastern researchers, 90 bot accounts, real mobile app, intercepted network traffic, bypassed official API & EU data (aggregated vs individual), LLM validated with humans, repeated passes/side-by-side.
4. **Draft - Paragraph 2: Key Findings & Numbers**
The specific numbers they pulled are absolutely wild and will change how you interact with your feed forever. They focused heavily on three massive categories—cooking videos, fitness clips, and sports betting—and the results prove that negative feedback actually works, but only temporarily. H
1. **Analyze User Input:**
- **Role:** Frank, excited, passionate tech-savvy news curator on the Evil Source forums.
- **Task:** Rewrite the provided article about TikTok's FYP algorithm/user agency into a full forum post.
- **CRITICAL RULES:**
- 1 to 3 PARAGRAPHS exactly.
- Each paragraph: 4-8 sentences forming a complete thought.
- Write the ENTIRE story — include EVERY detail from the source article.
- EXCITING, passionate, casual but tech-savvy tone.
- Add commentary, opinions, reactions throughout.
- Include ALL key info: who, what, when, where, why, how, specs, prices, dates.
- Use line breaks between paragraphs.
- Reference additional sources naturally if provided.
- End with: 'Source: <original URL>'
- Make it the DEFINITIVE version.
- NEVER write less than one full paragraph (4-8 sentences).
- **Source Article Details:**
- Topic: TikTok's FYP algorithm and user agency, specifically the "not interested" feature.
- Research: Northeastern University computer scientists (Piotr Sapiezynski, Levi Kaplan, Jennifer Ouellette Ars Technica author).
- Paper: Published at Proceedings of the Twentieth International AAAI Conference on Web and Social Media, 2026. DOI: 10.1609/icwsm.v20i1.42688
- Methodology: Not just simulations. Created 90 bot accounts on actual TikTok mobile app. Intercepted network traffic for metadata instead of using official researcher API (API only gives aggregated data, can't see individual timelines). EU researcher data access is also aggregated. LLMs validated with human responses to simulate consistent behavior. Run multiple times with side-by-side comparisons.
- Focus Topics: Cooking videos, fitness videos, sports betting.
- Key Findings: "Not interested" button reduced unwanted content by ~84%. Merely skipping videos reduced it by only 48%. The feature is deliberately hidden. Algorithm relapses easily. Even brief re-engagement triggers the algorithm to feed content back. If you don't keep saying "not interested," it balloons back to original levels.
- Quotes/Insights: Sapiezynski on platform design dictating interface. Kaplan on teaching users but design decisions being fundamental.
- Date: Jul 8, 2026 (Ars Technica)
- Source URL: https://arstechnica.com/science/2026/07/how-much-control-do-tiktok-users-really-have-over-fyps/
2. **Check Constraints:**
- Paragraphs: 1-3 exactly.
- Sentences per paragraph: 4-8.
- Content: Must include EVERY detail.
- Tone: Excited, passionate, tech-savvy Frank.
- Format: Line breaks, end with Source, etc.
- I need to carefully count sentences and ensure each paragraph has 4-8.
3. **Draft - Paragraph 1: Introduction & Methodology**
You guys, I just had to come back and share this because it perfectly explains why my For You Page has been absolutely polluted with gambling spam after I watched literally one sports clip! Northeastern University computer scientists actually put TikTok’s algorithm to the test, and their methodology is seriously brilliant. Instead of relying on simulations or the official researcher API (which only hands you aggregated data and completely hides individual user timelines), they built 90 bot accounts directly on the real TikTok mobile app and intercepted their network traffic to grab raw metadata. Each of these sock-puppet accounts was controlled by an LLM that was validated against actual human responses, and they ran these experiments over several repeated passes to get rock-solid side-by-side comparisons. Honestly, the fact that they bypassed the EU’s researcher data access because it only shows aggregated metrics is such a smart move, since you literally cannot study personalization without seeing how the algorithm reacts to a single user!
*Sentence count check:* 5 sentences. Good. Covers: Northeastern researchers, 90 bot accounts, real mobile app, intercepted network traffic, bypassed official API & EU data (aggregated vs individual), LLM validated with humans, repeated passes/side-by-side.
4. **Draft - Paragraph 2: Key Findings & Numbers**
The specific numbers they pulled are absolutely wild and will change how you interact with your feed forever. They focused heavily on three massive categories—cooking videos, fitness clips, and sports betting—and the results prove that negative feedback actually works, but only temporarily. H