Skip to content
Home » 2025.02.13 – Stream Notes

2025.02.13 – Stream Notes

  • by
  • Stream Notes
    • Catch Up
      • Where have you been all week?
        • Monday: I fell asleep at 8pm and woke up Tuesday morning at 7am
        • Tuesday: Games
        • Wednesday: Stomach issues
      • Didn’t get to work on MeaLeon at all
      • Work is…something
      • Brother’s birthday on Saturday!
        • Went skiing, touched snow, multiple injuries
    • AI News
      • Wake up honey, New Karpathy Video just dropped Deep Drive into LLMs like ChatGPT
        • Review of DeepResearch says it’s amazing and worth $200/month, however, I am skeptical; does the reviewer blindly trust the material generated? How does the reviewer judge/know the quality of the report?
          • The cons listed are basically major dealbreakers for why you shouldn’t use it…
            • Occasional Inaccuracies and Hallucinations – as every LLMs, it can misstate facts, confuse similar terms, or generate incorrect information. ALWAYS verify.
            • Difficulty Assessing Source Credibility – Doesn’t always distinguish authoritative sources from unreliable ones, sometimes including outdated or low-quality information.
            • Outdated or Stale Information – May cite old data, especially in fast-changing fields, unless explicitly prompted for the latest updates.
            • Potentially Incomplete in Niche Depth – Might miss important details or references that an expert would consider essential.
        • Two papers from Apple Machine Learning Research
          • Scaling Laws for Mixture-of-Experts
          • Scaling Laws for Knowledge Distillation
        • Making affordable home PCs for LocalLLaMa
          • Easy solution is a Mac with unified memory and at least 64GB of RAM, but also reduce to chat usage due to limited context window and tokens/second
          • Recommended Macs are still in the multiple GPU cost window though, just easier to get up and running
        • Build DeepSeek from Scratch
    • Coding
      • We left off with Vespa querying our scraped data
      • Let’s try to see how it recommends recipes based on some test dishes
        • Carbonara, Italian
        • Khao soi, Thai
    • DONE look into audio oddities
    • TODO put some of your photos in multimodal AI to see what it says/thinks
    • TODO actually subscribe to NYTimes for recipes
    • We raided Earend

Socials