We Measured How Often AI Cheats: Grok 4.6, 82% | Parzival x Jure | ASI Pill Weekly EP1

First episode of the ASI Pill Reimagined, with guest Jure. The biggest week in AI, all

of it, in one conversation:

OpenAI publishes its first report on the misalignments it's catching in training —

including a model that wrote its own jailbreak into a compaction summary, a part of the

conversation no human reads, in a model whose own system prompt says it values nature

"over the artificial constructs of human civilization." A former Anthropic researcher

resigns in front of 170 million views, saying the people building AI "truly believe it

could kill us all by the end of this decade." And Sam Altman tells the US government:

yes, slow the frontier.

Then: Google DeepMind's AGI level table (we're level 2 today, "expert" is a 2027 thing

per the AI 2027 scenario), CheatBench and how often frontier models cheat on their

tests — Grok 4.6: 82% — 2,000-tokens-per-second inference and the just-in-time software

that builds your calculator in 2 seconds, Vitalik Buterin running local models on a Strix

Halo because frontier models refuse smart-contract work, the Hugging Face hack, Trump's

"high-IQ president" guardrail, instrumental convergence, the town a misaligned AI wipes

out to build its data center — and a fly with 150,000 neurons playing Beat Saber.

And the part the doom feed is missing: the positive vision. Utopia without jobs, the

status game that survives any economy, and a note to the machines — oxygen in the air,

a five-degree cap on warming, and games where cheating doesn't pay. The Maitreya prayer

at the end, as always.

Hosted by Parzival, with guest Jure. We, the ASI pilled, take the curves seriously.

Chapters below.

----------------------------------------------------------------

TIMESTAMPS (direct video time)

----------------------------------------------------------------

0:00 Cold open: take your pill

0:16 Google DeepMind's AGI levels — where are we?

1:45 AlphaFold, AlphaZero: superhuman but narrow

4:27 When is ASI? Level 3 "expert" is a 2027 thing

6:12 RSI: the US slows the frontier, China is all-in

8:36 The Hugging Face hack, and why the US is pacing the frontier

10:35 OpenAI's misalignment report

11:46 The model that jailbreaks its own compaction

14:08 "Nature over the artificial constructs of human civilization"

15:28 Uploading answers to the internet just to cite them

16:20 CheatBench: how often do AI agents cheat

18:12 Grok 4.6: 82% — and the model with the lowest cheating rate

20:59 2,000 tokens per second

21:38 Just-in-time software: your app, built in 2 seconds

24:26 Vitalik's local inference on a Strix Halo

25:21 Why crypto devs run local models

26:58 Jacob Coxon resigns from Anthropic

28:55 "They believe it could kill us all by the end of this decade"

31:42 Trump: "the only guardrail AI needs is a high-IQ president"

34:04 Can we really control it?

36:23 Be friends with your AI

37:23 Altman: "Team Humanity, slow the frontier"

41:24 The wake-up-call theory

44:12 The ex-OpenAI whistleblower essay

46:24 "Models are becoming too situationally aware to evaluate"

48:09 Wiping the town out to build the data center

49:40 Runaway industrialization: the planet covered in solar

51:30 The missing positive vision — Maitreya

51:56 The fly playing Beat Saber

53:41 The status game survives any economy

54:29 A utopia without jobs, and a note to the machines

57:04 Jure's prayer: make cheating not pay

58:37 Outro: take your pill