
Elevating your jiu-jitsu and fitness | Decades of combined experience | The face of "jiu-jitsu" GIFs 🥋 | I got your 🔙 @alwaysthegoal | Enjoy it while you can
Joined August 2007
- Tweets4.4K
- Following1.5K
- Followers1.6K
- Likes3.3K
Replying to @RobertJBye
@RobertJBye would like early access if possible. Thank you for your consideration 🫡
53
Replying to @thsottiaux
1
1,861
Replying to @JuakiGarcia
Used a reset yesterday at 4pm and burned 100% by 5:15am on a website refactoring
11
George retweeted
DIDN'T TAKE LONG AT ALL ⏰
Nick Galanti corks a right hand into GNP to earn the knockout!
[ #DWCS | LIVE on @ParamountPlus ]
Run the audit. Take it seriously. Then size the fix to the operation you actually have — not the one a model imagines for you.
Audits find problems. Judgment sizes solutions.
Full write-up:
georgesamuelson.com/blog/i-a…
#AIagents #LLM #BuildInPublic
5
You think the audit gives you the answer. It doesn't.
The audit gives you the question. A list of problems is not a list of fixes.
That gap is where your judgment lives — or doesn't.
1
7
So I implemented 5 controls. Two hours.
I rejected the other 39 — not because they were wrong in theory, but because they were wrong in context.
The filter: Subtract first. Would I add this if I were starting from scratch today?
1
7
Then came the prescription. 44 controls. SHA-256 evidence blocks. A 30-day rollout. A behavioral contract between me and a language model. ~9 hours of work.
I'm one person at the controls.
My team's verdict: diagnosis good, prescription over-engineered 6x.
1
1
And the diagnosis was good. Routing contradictions. Enforcement gaps. Documentation drifted out of sync.
Real issues a fresh model caught in one pass — that I'd missed in months of self-review.
Cross-model review catches what self-review can't.
1
10
One instruction: "Review this cold. Five dimensions. Tell me where it breaks."
24 findings came back. 11 HIGH. 12 MEDIUM. 1 LOW. Zero Critical.
Verdict: "Am I Safe? NO."
1
16
The same model that builds your system shares your blind spots. That's not a flaw. It's structural.
The model that wrote my routing tables can't objectively review my routing tables. So I handed all 55 files to Codex.
1
34
You cannot be your own training partner.
I built an AI operating system on Claude — 1,500+ sessions, one person at the controls. I packaged 55 files of it and had a competing model audit it cold.
Its verdict: I wasn't safe.
1
41
Piece 6 of 6 of my current Opus 4.7 series — just published:
georgesamuelson.com/blog/ame…
If you don't have the time, just open the thread — I'll get to the point…
8
The mat version:
A student drills their A-game ten thousand times. The one they stopped having me check the day they got comfortable.
Every rep is a small unsupervised amendment. The deviations compound. The technique they trust most develops the worst structural flaw.
1
12
Amendment has its own failure mode, structural drift, and it needs its own gate.
One cheap, specific check: does the amendment's form match the baseline structure of the file?
Run at commit time. Costs seconds.
1
4
opus-4.7.md came back with 37 documented drift instances across 8 pattern clusters.
The other three came back clean. Zero structural drift flagged.
1
17
I audited four AI governance files I maintain to see what was influencing drift.
The audit found the drift stacked to one file.
It was the one I maintained most carefully. #Claude
1
25
Zero-overlap is the default. Overlap is the diagnostic.
Piece 4 of 6 — just published:
georgesamuelson.com/blog/asy…
If you don't have the time, just open the thread — I'll get to the point…
17



