"our work got checked by an external reviewer."
That's what Fable said after I brought Gemini (Antigravity 3.7) in to review StacksNG, the offline...
For further actions, you may consider blocking this person and/or reporting abuse
On the hardware aside, memory bandwidth decides 7B to 13B speed on those Macs far more than core count does. The M4 in the Air is around 120 GB/s and the M4 Pro is roughly 273, so a 13B at 4-bit will feel like a different machine on the Pro. Get more than 16GB either way, since a 13B model plus the OS leaves you almost no headroom.
Nazar, bandwidth over core count is the piece I was missing entirely, 273 GB/s makes the Pro case for me.
This is a strong reminder that ai, discuss, programming need an observable contract. The happy path is rarely the expensive part; it is the boundary behavior, stale state, and partial failure path that decide whether the design holds up in production.
The second opinion reflex is the useful part of this, even if the social framing is what stuck. When I bring a second model in on a stuck agent session, the first one has already written a clean summary of why its approach is fine. The external pass only earns its keep when it is forced to read the actual diff, not the story the first agent tells about the work. That is where the plateau usually hides: a tidy narrative covering a wrong edit. Did Gemini catch anything Fable's own explanation had already smoothed over?
yes, that's exactly what happened. Gemini wasn't reading Fable's account of the fix, it was reading the actual code and that's where the blind spots showed up. Fable's explanation had smoothed over the same kind of thing you're describing, a clean story sitting on top of a wrong edit. How do you force the second model to skip the first one's summary entirely, do you strip the conversation history before handing it the diff?
The word our doing that much work is a nice catch. I'd read it less as the assistant having feelings about being checked and more as a signal of whose failure mode it's now modeling, if it frames a review as something that happens to our work, it's picked up the same defensiveness a teammate would have, for better and worse.
Wow, LLMs are starting to get an ego now, or developing 'sensitivities' (or is it 'sensibilities') ? That's something of a quantum leap, "AGI" might be around the corner! 😃 (joking of course - nobody even knows what "AGI" is)
leob, not ego. Closer to how any collaborator reacts when a stranger checks their work. AGI's still waiting on the invite
The second-opinion pattern has saved me more than once, though I had to stop treating the two models as independent when they shared training data and confidently agreed on the same wrong answer. What I do now is score the disagreements, since the cases where they split are usually where the real bug is hiding. Did the second model actually catch different errors for you, or mostly restate the first?
Kartik, no overlap-and-agree here. Gemini surfaced blind spots I'd been ignoring not restating what Fable already flagged. Your disagreement-scoring habit sounds like the sharper version of what I stumbled into by accident: I brought in a second model for coverage not for a disagreement signal.
Curious what your split rate looks like . are the two catching really different error classes or clustering around the same weak spots most of the time??
The MacBook aside is relatable too. 8GB gets painful fast once you start running local models alongside an actual dev workflow. M4 with more memory feels like the safer upgrade.
Julian, memory over raw speed is the real fix: 24GB is where a 13B model stops swapping to disk instead of crawling.
Interesting!