Curated watching (AI)

Short list of high-signal videos worth watching before (or while) using AI day to day. Prefer this over stuffing profile READMEs; add entries when something is fundamental enough to share across Dev-Centr.

Personal curated-learning keeps durable picks; Dev-Centr curated-watching is an index that may split into topical pages; other docs deep-link for blurbs

How this list grows

  • This page is an index. A handful of entries may live inline (as below). When the set grows, split into topical pages under explanation/ai/ (one topic, many resources) and keep only titles + one-line pointers here.

  • Other docs link to topical pages (or a single resource page), not to this index, when the link is about that topic.

  • Personal profile indexes stay separate. A GitHub profile curated list should keep its own stable video entries (title, URL, short note) and deep-link into Dev-Centr for blurbs/credits—those Dev-Centr URLs may move when topics split. Do not treat this index URL as the only durable personal bookmark.

flowchart LR
  P[Personal curated-learning.md]
  I[Org curated-watching index]
  T[Topical page]
  D[Other docs]
  P -->|deep-link blurbs| I
  I --> T
  D --> T
  P -->|title URL note Drive| YT[YouTube / Drive]

Mandatory

AI Data Centers Will Be Obsolete (Geometric Reasoning Explained)

YouTube thumbnail for AI Data Centers Will Be Obsolete Geometric Reasoning Explained
Field Value

Title

AI Data Centers Will Be Obsolete (Geometric Reasoning Explained)

Channel / publisher

Sophontic AI

Featured speaker

Julian D. Michels, PhD (founder, Sophontic)

Published

6 August 2026

Runtime

~53 minutes

YouTube

https://www.youtube.com/watch?v=4S8I22ybG2c

Project site

sophontic.ai

Local / Drive backup

Google Drive copy (Documents/Curated Learning/AI Videos/, yt-dlp archival MP4 + metadata; not a substitute for the canonical YouTube / sophontic.ai publish)
Blurb

Sophontic argues that reasoning is geometry in latent space, not an inevitable side-effect of brute-force scale. Michels contrasts statistical “paradigm of scale” training with empirically studying and cultivating internal geometries (neuro-symbolic lineage, perturbation / flip-rate evals, trainable compact reasoners). The interview frames why frozen frontier generalists and hyperscale data centers are a brittle monopoly path—and what changes if real reasoners can train and adapt on laptop-class hardware. Watch this if you use AI and want a clear counter-story to “just add GPUs.”

Credits (as published)

Interview published on the Sophontic AI YouTube channel; featured interviewee Julian D. Michels, PhD. Primary source page embeds the same video: sophontic.ai.

Claims about magnitude gains (tens to ~10³× vs larger models under their perturbation tests) are the lab’s reported measurements, not independent Dev-Centr benchmarks. The value of the piece for practitioners is the mental model: structure vs scale, frozen giants vs continual local learning, and why evals that only score memorized tests are weak.

Related