My Cheatsheets

Updated Guide Sept 2025

Core Principles

• LLMs are predictors, not logic engines. They generate plausible answers, not guaranteed truths.
• Still like a brilliant, distractible assistant. Fast, knowledgeable, but can lose track, misremember, or follow your lead even when you’re wrong.

Tokens & Context

• Token windows are expanding rapidly:

  • GPT‑5: ~256k tokens (standard), higher in some API tiers.
  • GPT‑4.1 / GPT‑4o-mini: up to 1M tokens in enterprise contexts.
  • Claude 4 / Claude Sonnet 4: ~200k–1M tokens depending on variant and plan.
  • Gemini 2.5 Flash/Pro: ~1M tokens input, ~65k output (Flash documented).
  • Qwen 2.5: ~32k tokens for 72B Instruct.
  • Perplexity Sonar: ~128k tokens context reported.

• Context drops still occur when limits are exceeded—older messages get truncated.

• Best practice: Don’t rely on infinite context. Chunk long work, summarize checkpoints, and use external docs.

Model Behavior

What’s Stable

• Sessions need focus.
• Breaking tasks into small steps works best.
• Reinforcing key info improves consistency.

What’s Changed

• Memory features are rolling out broadly. Many models now:

  • Remember preferences and style between sessions.
  • Can be put into “private” or “incognito” mode.

• Safety tuning is stronger: models hedge more, avoid risky content, and may refuse edge cases that older versions answered.

Confidence vs Accuracy

• Models remain agreeable—may mirror your errors.

• They are less likely to argue, more likely to hedge.

• Rule of thumb:

  • Trust for brainstorming, summarization, code scaffolding.
  • Verify for math, logic, critical facts.

Choosing the Right Model ( Sept 2025 Landscape)

Important caveats:

• Many limits apply only to paid/enterprise tiers.
• Output often capped lower than input windows.
• UI versions may expose far smaller windows than API.
• “Million token” contexts are real but not always reliable in practice; performance can degrade.

File & Data Handling

• Uploading many files increases confusion and token load.

• Best practice:

  • Work with 2–3 files at a time.
  • Use phased comparisons (A vs B, then compare with C).

• Modalities now matter:

  • Images: OCR quality impacts parsing.
  • Audio/video: transcription may drop detail.

Math & Logic (Still Weak Spots)

• Good at pattern math, bad at symbolic manipulation.

• Multi-step logic still fragile without explicit scaffolding.

• Better option:

  • Ask for code to compute results.
  • Use external math engines for reliability.

Practical Tips

• Be explicit about output form (“full response, no summary”).
• Use structured prompts with headings/checkpoints.
• Offload long-running work to shared docs (Google Docs, Notion, etc.).
• Reset threads if drift/contradictions appear.

Memory & Personalization

• Memory is now active by default in many systems.
• Check settings: you can view, edit, or wipe memory.
• Use memory deliberately (preferences, style notes, recurring project context).
• Prune old/irrelevant data regularly.

Considerations

• Costs: Longer contexts = higher cost. Be conscious of diminishing returns.
• Privacy: Many providers now use chats for training unless you opt out. Review policies.
• Safety: Guardrails can change without notice—what worked last month may be blocked today.

Appendix: Tokenization Nuances

• Tokenization still varies by model and platform.

• Special characters, hyphens, and punctuation can inflate counts.

• Symptoms of hitting limits:

  • Truncated or partial code.
  • Lost references to earlier variables.

• Mitigation:

  • Use explicit requests for “complete code”.
  • Provide earlier snippets manually.
  • Break large codebases into modules.

Bottom Line

• Keep chats focused, chunk work, and verify important details.
• Know your model’s current limits, tools, and memory features.
• Treat AI as an assistant—not a calculator, not a database, not an oracle.