Ground Truth
Subscribe
Sign in
Home
Notes
Archive
About
Latest
Top
Discussions
Too Young for Woodstock, Too Old for Gen X
On being born late enough to miss the boom, and early enough to be blamed for it.
Sep 21
•
Jim Karabatsos
3
3
1
What Karate Taught Me About Everything Else
On late starts, and what the body learns that the mind can't
Sep 14
•
Jim Karabatsos
1
Teaching in Prisons: What I Learned
On what you think you'll find, and what's actually there.
Sep 7
•
Jim Karabatsos
2
1
Prompt Injection and Role Confusion
Nine essays on why language alone cannot safely carry authority in an AI system.
Sep 5
•
Jim Karabatsos
1
The Machine Does Not See a Conversation
Behind the friendly chat window is a long, carefully labelled stream of text. That modest engineering fact helps explain both the magic and the risk.
Sep 4
•
Jim Karabatsos
The Labels Holding Up the AI World
A few words such as “user”, “assistant” and “tool” now carry more responsibility than they were ever designed to bear.
Sep 4
•
Jim Karabatsos
When a Web Page Talks Like Your Boss
Prompt injection is not hacker magic. It is what happens when untrusted text is mistaken for an authorised instruction.
Sep 4
•
Jim Karabatsos
The Model Is Not Obeying the Tag. It Is Reading the Room.
A label should establish authority. The evidence suggests that, for a language model, the style of the text can sometimes compete with the label around…
Sep 4
•
Jim Karabatsos
When a Model Trusts the Wrong Working Note
A convincing piece of apparent reasoning can change what a language model does next. That is a problem of provenance, not evidence of consciousness.
Sep 4
•
Jim Karabatsos
A Talking Machine Without a Little Man Inside
“It only predicts the next word” is true, but it is not the dismissal its users think it is. Nor does fluent conversation settle the question of…
Sep 4
•
Jim Karabatsos
Why “Just Ignore the Instructions on the Page” Is Not Security
Good instructions matter. They are not enough when an AI system can read untrusted material and act in the world.
Sep 4
•
Jim Karabatsos
When an AI Agent Treats the Plan as a Suggestion
A plan written into an agent’s context is still only text. If we want commitments, approvals and constraints to hold, they need more support than a…
Sep 4
•
Jim Karabatsos
This site requires JavaScript to run correctly. Please
turn on JavaScript
or unblock scripts