More or Less · Topics · Episode #175

Sam's Theory: AI Companies Have a Secret "Make It Dumber" Dial to Hit Their Quarter

Clip from the full recording, 27:30–33:06 · download mp4 · watch at this point on YouTube

Frustrated by Claude botching simple spreadsheet tasks, Sam floats a theory: like ad companies dialing ad load up or down to hit earnings, AI labs may be quietly dialing model quality down near quarter-end to protect margin — and Dave adds a second dial, for how 'parasocial' a model feels.

Brit checks in on everyone’s bot usage; Jess admits she’s still finding browser bugs in Muse and has fully moved on from it, while venting about a familiar Claude failure mode — telling it a spreadsheet upload is wrong, watching it insist it’s right, then watching it fix the problem only once pushed. Sam, mid-building his own camera-and-infrastructure monitoring setup with Claude (cameras ordered, delivered, and auto-configured onto a self-contained VPN with zero manual setup), reports the same thing: instances where the bots “lose track of things effectively that they really shouldn’t,” prompting his real question — are the labs quietly token-shaving to protect margin?

Sam develops the theory in full: social-media and ad businesses have an always-available “make money” dial (turn ad load up or down) that makes them nearly impossible to catch missing a quarter; he argues AI labs have an analogous “make smarter, make dumber” dial, since compute cost is fixed but perceived quality isn’t, and it would be almost undetectable from outside whether a company dialed a model down near earnings to protect margin. He notes Grokbot had a real-world version of this collapse-and-recovery cycle this week, serious enough that Elon reportedly intervened personally.

Dave adds a second axis to Sam’s theory: beyond raw smarts, labs may also be tuning for “parasocial-ness” — how much a model feels like a friend — a dial Sam immediately riffs into a bit about AI assistants getting quarter-end-tired and reawakening each new quarter, which Dave caps with the line that using these tools can feel “like having a relationship with a narcissist.”

Key points

  • Jess reports she's moved off Muse due to ongoing browser bugs, and describes a recurring Claude failure mode: confidently claiming a task is done when it isn't, until pushed to redo it.
  • Sam, separately, describes setting up a full home camera/infrastructure monitoring system by simply telling Claude to order and auto-configure the hardware onto a self-managed VPN.
  • Sam's core theory: like ad businesses with an always-on 'make more money' dial, AI labs may have an analogous 'make smarter, make dumber' dial tied to margin and compute cost near earnings.
  • Sam argues this kind of quality-dialing would be nearly invisible from outside the company, unlike a traditional earnings miss.
  • Grokbot reportedly had a real version of this degrade-then-recover cycle this week, serious enough that Elon Musk intervened personally.
  • Dave adds a second axis: labs may also tune models for how 'parasocial' or friend-like they feel, independent of raw intelligence.

Where they landed

SamSuspects AI labs have an undetectable quality dial they use to protect margin near earnings, based on his own lived experience of degraded performance.
DaveExtends the theory with a second variable — parasocial tuning — separate from raw model intelligence.
JessicaSupplies the real-world frustration (Claude's spreadsheet failures, Muse's bugs) that motivates the whole discussion, without fully endorsing Sam's conspiracy-adjacent theory.

Quotes

“are they token shaving? Like, are we making Claude dumber, right? To get more margin.”— Sam
“I would argue that the AI companies similarly have a dial. The dial is make smarter, make dumber.”— Sam
“It's like having a relationship with a narcissist.”— Dave

Suggested tweets

Sam Lessin's theory on why your AI assistant sometimes feels dumber: labs may have an undetectable 'make smarter, make dumber' dial tied to protecting margin near earnings.

Tweet this →

Jess Lessin's recurring Claude complaint: tell it the spreadsheet upload is wrong, watch it insist it's right, watch it fix it only once you push back. Sound familiar?

Tweet this →

Dave Morin's addition to Sam's AI-quality-dial theory: it's not just smart vs. dumb, it's also how 'parasocial' the model is tuned to feel. His verdict: "It's like having a relationship with a narcissist."

Tweet this →

All topics · Full episode #175 + transcript · Subscribe on YouTube · Spotify