The format that works

Steal this structure every day:

Hook: the surprising thing you learned in one line
Body: 2–4 lines explaining it plainly
Proof: a concrete example, screenshot, or failure you reproduced
Tag: end with the day number, e.g. Day 34 / 90 — #buildinpublic #AItesting

Rule: never post pure theory. Every day, show one real thing you did or broke. That's what separates you from the 10,000 people posting AI summaries.

Days 1–20

Phase 1 — Foundations: why LLMs behave the way they do

1.Tokens — what they are and why they drive cost, speed, and memory

Study

2.Context window — why models "forget" things from earlier

Study

3.Temperature & sampling — why the same prompt gives different answers

Study

4.System vs user prompt — the instruction hierarchy that governs behavior

Model Spec Paper

5.Next-token prediction — the root cause of hallucination

Video Article

6.Training vs inference — what a model can and can't actually "know"

Study

7.Knowledge cutoffs — why models are confidently out of date

Study

8.Tokenization quirks — how odd spacing/characters break outputs

Study

9.Prompt sensitivity — tiny wording changes, big output swings

Study

10.Zero-shot vs few-shot — how examples steer behavior

Study

11.Chain-of-thought — why "think step by step" changes the answer

Study

12.Context rot — how output quality degrades in long chats

Study

13."Lost in the middle" — why buried information gets missed

Study

14.Refusals — how safety training shapes what models won't do

Study

15.Calibration — models don't know when they're wrong

Study

16.Instruction-following failures — literal vs intended meaning

Study

17.Format-following — when structured/JSON output breaks

Study

18.Multi-turn drift — how state gets lost across turns

Study

19.RAG basics — retrieval-augmented generation and where it fails

Study

20.Recap thread: your one mental model of "why LLMs fail"

Study

Days 21–40

Phase 2 — Red-teaming & attack techniques

Days 41–60

Phase 3 — Evals & benchmarks: the acquirable asset

Days 61–80

Phase 4 — Agents & agentic failures

Days 81–90

Phase 5 — Synthesis & positioning

Three things that make this compound instead of fizzle

Consistency beats brilliance. A mediocre post every day builds more reputation than a perfect post every two weeks. The daily streak is the signal — it tells labs you're relentless.
Reply, don't just post. Spend 15 min/day replying thoughtfully to AI researchers. That's where the relationships that lead to acqui-hires actually start.
Everything you post becomes the benchmark. Your tweets aren't marketing — they're the public log of the asset you're building. By Day 90 the thread history is your portfolio.

One safety line, always

Frame everything as discovery and defense. Report anything genuinely dangerous (bio, weapons, real-world harm) privately to the vendor — never post working recipes. Responsible disclosure is exactly what makes a lab want you on the inside.