Anthropic Just Quietly Changed AI Forever (68% to 5%)
ANTHROPIC JUST QUIETLY CHANGED AI FOREVER (68% TO 5%)
Four researchers at Anthropic just published the most consequential alignment paper of 2026 — and the part nobody is talking about is who actually controls the values inside frontier AI models.
In this episode:
Why the two-stage training stack (pretrain → align) was always producing "shallow alignment"
What Model Spec Midtraining actually does — and the staggering 68% → 5% misalignment numbers
The cheese-preference experiment that proves your model's soul is now a hyperparameter
Why this fragments the open source world and turns alignment-as-a-service into a puddle
Daniel Kokotajlo's "situational self-awareness" objection that could undo the whole thing
TIMESTAMPS:
0:00 — Cold open
0:08 — From 68% to 5% misalignment
0:28 — Hi I'm Jim — Subscribe / Follow CTA
1:01 — The two-stage training stack
1:34 — Why alignment was breaking
2:14 — Enter the four authors
2:35 — How midtraining actually works
3:12 — The staggering numbers
3:50 — But here's the part nobody is telling you
4:18 — The cheese preferences experiment
4:53 — Three implications
5:01 — 1: Power to whoever writes the spec
5:35 — 2: Open source fragments
6:00 — 3: Alignment-as-a-service is dead
6:27 — Skeptical counterpoint: Daniel Kokotajlo
7:32 — What to watch (3 milestones)
8:12 — Closing thought
8:38 — Outro CTA
SOURCES:
Model Spec Midtraining paper (Anthropic Alignment Science): https://alignment.anthropic.com/2026/...
arXiv preprint: https://arxiv.org/abs/2605.02087
Reference implementation: https://github.com/chloeli-15/model_s...
LessWrong discussion (incl. Kokotajlo): https://www.lesswrong.com/posts/R3Rrw...
Anthropic Agentic Misalignment background: https://www.anthropic.com/research/ag...
Claude's New Constitution coverage: https://bisi.org.uk/reports/claudes-n...
---
The Grift Podcast — Forbidden Knowledge Unlocked
New episodes every week.
SUBSCRIBE for more: https://www.youtube.com/@DigitalDream...
#Anthropic #AIAlignment #ModelSpecMidtraining #AISafety #Claude #Qwen #LLM #FrontierAI #TheGriftPodcast #TechNews