charles leifer | Cave of Forgotten Dreams
//charlesleifer.com
/blog
/code
/resume
Cave of Forgotten Dreams
June 02, 2026 10:08
ai
thoughts
1 comments
Who is he that hideth counsel without knowledge? therefore have I uttered that<br>I understood not; things too wonderful for me, which I knew not.
"Hold on a sec-". My new boss' now-familiar MacOS desktop appeared in the video<br>call, browser with Claude open, dominating the screen. I watched as he<br>copied the transcript of our call up to that point (he records transcripts of<br>every call in order to feed the text into AI), and began a new chat with the<br>prompt: "Say where Charlie's right, and where he's wrong. Say where I'm right<br>and where I'm wrong." He pasted the transcript and hit enter. I consulted my<br>avatar in the lower-right. We both waited in silence while Claude thought. We<br>were 45 minutes into a call about product roadmap and a possible customer<br>announcement before this interruption. Soon the cursor started skipping along as words began filling the<br>screen. Then, we read aloud through the findings one-by-one. Claude had helpfully given us a bulleted list to work through,<br>an even number of findings for each of us. I felt called-upon to gallantly agree with Claude's softly (oh-so-softly) couched<br>criticisms of my viewpoint, while conceding everywhere Claude expressed subtle<br>(oh-so-subtle) approval of my boss. The call ended shortly afterwards, somewhat<br>awkwardly for both of us. I had just experienced the most baffling<br>mixture of radical transparency and impossible opacity.
Have I missed the mark, or, like true archer, do I strike my quarry? Or am I<br>prophet of lies, a babbler from door to door?<br>- Aeschylus (Agamemnon)
At the beginning of his tenure, I initially answered my new boss' questions,<br>posed via Slack, with as much candor and detail as I regularly applied to<br>technical matters with my colleagues. After a couple of these, I noticed that<br>my responses were never engaged with beyond an emoji reaction. Instead I<br>observed my responses being fed directly into his AI assistant, and the<br>AI's responses returned to me verbatim - often just linked, but sometimes with the invitation: "Check this out". There was<br>no attempt at subterfuge, my boss has always been refreshingly open, but these<br>were forms of interaction I never had imagined until I was in them. It is an<br>unsettling, helpless feeling to see "Charlie says ..." in someone else's chat<br>prompt. It is even more unsettling to be named by the AI in it's responses, as<br>if I'm not there.
I adapted by becoming an occasional ventriloquist, front-running my own<br>comments, using AI to avoid a gaffe that his AI would jump on which might<br>discredit me, and to avoid coming off as an AI-naysayer. Mentally, I began<br>modeling his model, and the one-sentence summary that would likely emerge as<br>the punchline. I was stuck in an absurdist game of telephone, but felt I had to<br>play or accept the consequences of his AI turning against me. On his side, I<br>can only imagine that the challenges of running a new company required him to<br>have his attention on so many things that he reached for the quick AI summary.<br>Yet, I was literally living out the recursive feedback loops I wrote about<br>regarding agentic coding patterns, only in real life interactions with another<br>person. Meanwhile, I was fielding random questions like, "could AI agents write<br>us an API over the weekends?", being sent links to gstack<br>(which famously shat out a prodigious log of code once), and generally acting<br>as both buffer and technical reviewer for whatever the AI hype crowd was<br>talking about in a given week, so my response could be fed back into Claude. I<br>felt rather overwhelmed.
This disruption of my consciousness did not go unnoticed by me, but I was<br>wholly unable to pull myself out of it. It felt urgent, and I wanted to be rid<br>of the intrusion. I sensed weakness in the models and I decided to lawyer up<br>and go on the offensive. My plan was simple: litigate, convict, and imprison<br>these models in the court of my mind. (Of course it didn't present itself to<br>me in those terms, but looking back that's the color of it.)
Since a case needs evidence, I became a forensic analyst poring over output<br>produced by Claude. My first experiment back in January predates this madness,<br>but the germ of a strategy was there. I would overlook the sycophancy, the<br>reassuring tone, the confidently-stated errors. The case I committed to was<br>that correctness is not interchangeable with coherence, and that coherence is<br>far more important for code. As of Opus 4.5, it was undeniable that Claude<br>could produce islands of reasonably correct code. But how would it manage the<br>big picture: coherence and unity among the members? Through a series of more<br>ambitious tasks on my own open-source projects, whose undercurrents of intent I<br>understood deeply, I began building a case.
I'm not above it.
I had accepted a free six-month trial of Claude Max, so I had access to the<br>best frontier models...