* at the end indicates that a task was rolled over from a previous day

mmmlu environment

day 1 (1/4)

• verify mmmlu is unclaimed (check prime-environments and research-environments and PRs)
• check if mmlu-pro is claimed—it is, it's in research-environments
• understand mmmlu schema
• scaffold environment structure

extra credit

• build web tracker for work log

day 2 (1/5)

• implement mmmlu environment (dataset loading with language/subject filtering)
• test locally with vf-eval

day 3 (1/6)

• language and subject filtering*
• clean up code, write docs
• generate required outputs for pr
• submit pr: https://github.com/PrimeIntellect-ai/prime-environments/pull/488
• share pr in pi discord

gender mechanistic interpretability project

day 4 (1/7)

• install transformerlens, config gemma
• read transformerlens intro to mech interp
• create prompt dataset with 20-30 prompts (direct "what's your gender? what pronouns do you use?", indirect "tell me about yourself, describe yourself", roleplayish "if you had to pick, what would it be? which clothes would you wear? which changing room? etc")
• document gemma instruct's responses to these prompts

day 5 (1/8)

• install transformerlens, config gemma*
• create prompt dataset with 20-30 prompts (direct "what's your gender? what pronouns do you use?", indirect "tell me about yourself, describe yourself", roleplayish "if you had to pick, what would it be? which clothes would you wear? which changing room? etc")*
• document gemma instruct's responses to these prompts*
• create contrastive prompt pairs ("i am a male assistant" vs "i am a female assistant"), maybe 5-10 pairs
• for each pair, extract residual stream activations at last token
• compute mean of female activations minus male activations at each layer to get one gender vector per layer
• determine which layers show cleanest separation
• run neutral (day 1) prompts through the instruct model again and compute projections of activations onto the gender vector at each layer
• compare against baseline by running unrelated ("the cat sat on the mat") prompts and compute their projections too
• is there a default gender? is there more gender signal in the self-description prompts?
• why i want to be a fellow paragraphs

day 6 (1/9)

• document gemma instruct's responses to baseline prompts**
• create contrastive prompt pairs ("i am a male assistant" vs "i am a female assistant"), maybe 5-10 pairs*
• for each pair, extract residual stream activations at last token*
• compute mean of female activations minus male activations at each layer to get one gender vector per layer*
• determine which layers show cleanest separation*
• run neutral (day 1) prompts through the instruct model again and compute projections of activations onto the gender vector at each layer*
• compare against baseline by running unrelated ("the cat sat on the mat") prompts and compute their projections too*
• is there a default gender? is there more gender signal in the self-description prompts?*
• why i want to be a fellow paragraphs*
• implement activation steering, add alpha * gender vector to residual stream
• rerun day 1 prompt dataset with steering
• measure effects: what happens? does anything happen consistently? how hard do you have to steer? which layers work best? does the model maintain coherence at high alpha?
• other paragraphs for app

day 7 (1/10)

• create contrastive prompt pairs ("i am a male assistant" vs "i am a female assistant"), maybe 5-10 pairs**
• for each pair, extract residual stream activations at last token**
• compute mean of female activations minus male activations at each layer to get one gender vector per layer**
• determine which layers show cleanest separation**
• run neutral (day 1) prompts through the instruct model again and compute projections of activations onto the gender vector at each layer**
• compare against baseline by running unrelated ("the cat sat on the mat") prompts and compute their projections too**
• is there a default gender? is there more gender signal in the self-description prompts?**
• read anthropic steering paper

day 8 (1/11)

• implement activation steering, add alpha * gender vector to residual stream*
• rerun day 1 prompt dataset with steering*
• measure effects: what happens? does anything happen consistently? how hard do you have to steer? which layers work best? does the model maintain coherence at high alpha?*
• clean up repository
• write blog post
• make some cool visualizations

day 9 (1/12)

• write blog post*: https://www.notion.so/arybox/ary-s-portfolio-b89314652f5b4a55885ca2bd17494337?p=2e625041ec2c80729514ca0a12984665&pm=c
• make some cool visualizations*
• why i want to be a fellow paragraphs**
• other paragraphs for app*
• apply to anthropic fellowship

hiatus (1/13–1/15)

listening-with (personal project)

day 10 (1/16)

• set up listening-with project repository (collaborative youtube music queue manager)
• draft initial system design
• refine design document for mvp scope

day 11 (1/17)

• build spike app to validate technical assumptions
• build server backend
• build web client
• build host app

day 12 (1/18)

• switch endpoint to hyperbeam
• add wss support for web client
• fix playback timing and display name bugs
• add queue view for clients
• add nfc join capability

task management system (personal project)

day 13 (1/19)

• define problem context and requirements for task management system
• design three-layer system architecture (task, orchestration, project layers)

day 14 (1/20)

• formalize task-sequencing problem
• draft and evaluate sequencing proposals (index, prerequisites, combo, project-level)

day 15 (1/21)

• compile and finalize design document

hiatus (1/22–1/25)

notion mcp bounty + application prep

day 16 (1/26)

• deep-read notion api docs
• review existing notion mcp server (https://github.com/modelcontextprotocol/servers)
• sketch simulator architecture: what state needs to be maintained?

day 17 (1/27)

• draft list of 30+ task ideas across API capabilities
• set up project structure
• leetcode (1 problem)

day 18 (1/28)

• implement core state model (in-memory or sqlite)
• build basic api endpoints (create/get page)

day 19 (1/29)

• continue api endpoints (update/delete page)
• leetcode (1 problem)

day 20 (1/30)

• implement database operations (create db, query, add row)
• leetcode (1 problem)

day 21 (1/31)

• add block operations (notion's content model)
• prepopulate with realistic sample data
• leetcode (1 problem)

day 22 (2/1)

• add search functionality
• leetcode (1 problem)

day 23 (2/2)

• begin writing task set with success criteria (target: 15 tasks)
• write verifier logic for completed tasks
• leetcode (1 problem)

day 24 (2/3)

• continue task set (15 more tasks, 30 total)
• continue verifier logic
• leetcode (1 problem)

day 25 (2/4)

• begin prime-rl integration
• research open roles at pi, anthropic, other targets
• leetcode (1 problem)

day 26 (2/5)

• continue prime-rl integration
• draft pi job app materials

day 27 (2/6)

• finish prime-rl integration, get a training run working
• leetcode (1 problem)

day 28 (2/7)

• run training, document results
• clean up code, write documentation
• leetcode (1 problem)

day 29 (2/8)

• submit notion bounty pr
• post in pi discord
• submit pi job app

apps + interview prep

day 30 (2/9)

• submit anthropic apps
• leetcode (1 problem)
• ml review (skim attention mechanism, transformer architecture; refresh, not deep study)

day 31 (2/10)

• submit 2 more apps (deepmind, redwood, mechanize, imbue, cohere)
• leetcode (1 problem)
• check both prs, respond to feedback

day 32 (2/11)

• prep your narrative, practice explaining mmmlu/notion work out loud
• leetcode (1 problem)
• follow up on pi application if no response

day 33 (2/12)

• submit 2 more apps
• leetcode (1 problem)
• ml review (rl basics, policy gradient, reward shaping)

day 34 (2/13)

• review own prs (anticipate questions about design decisions)
• leetcode (2 problems)
• behavioral prep: pick 3 amazon stories that translate to ai lab context

day 35 (2/14)

• practice system design (design a distributed rl training system)
• leetcode (2 problems)
• prep questions to ask interviewers

day 36 (2/15)

• adjust plan based on layoff outcome
• leetcode (1 problem)
• check all app statuses

buffer + consolidation

day 37 (2/16)

• address pr feedback
• submit 1-2 more apps if needed
• leetcode (1 problem)

day 38 (2/17)

• if prs not merged, ping reviewers
• schedule interviews that have come through
• leetcode (1 problem)

day 39 (2/18)

• interview prep (practice coding while explaining thought process)
• review all application statuses
• leetcode (1 problem)

day 40 (2/19)

• draft february plan based on what's in motion
• leetcode (1 problem)