The setup
I’m building a private AI setup at home: an always-on Mac running local models as a personal robotics and agent platform. The planned workload is 20–35B parameter models for daily use, 70B/80B models for experiments, coding agents, browser and computer-control agents, local RAG, and vision work — running concurrently, for the next four to five years. The hardest tasks go to frontier cloud APIs. Everything with credentials stays local.
The hardware shortlist came down to three configurations:
- Mac mini, M5 Pro, 64GB — $3,299
- Mac Studio, M5 Max, 64GB — $3,799
- Mac Studio, M5 Max, 128GB — $5,399
The key specs, verified: the M5 Pro has 307 GB/s of memory bandwidth, the M5 Max has 614 — exactly double. And this is Apple Silicon, so the memory is unified and soldered. Whatever you buy is what you have forever. There is no “add RAM later.”
I ordered the Studio with 64GB. It shipped. Then I started second-guessing: what if a future open model makes the cloud fallback unnecessary? What if 30B-class models turn out to be too weak for browser control, and the real agentic work needs 70B+? Either scenario leaves 64GB as a permanent ceiling. I had a 14-day return window — until about October 9 — and a decision I didn’t trust.
So I did what any of us would do in 2026: I wrote up a clean decision brief and fed the same brief to three different AIs — ChatGPT, Claude, and Muse — to see what they’d say.
The three verdicts
Same facts in, three different answers out:
- Keep the 64GB Studio. The value play: 64GB covers the stated daily workload, 70B experiments are possible at lower quantization, the hardest work goes to the cloud anyway, and $1,600 is real money.
- Return it and get the 128GB. The future-proofing play: unified memory can’t be upgraded later, agentic workloads are memory-hungry (long context means large KV caches), and the return window is the cheapest moment you’ll ever have to change the spec.
- Downgrade to a mid-memory Mac mini and keep the hard work on frontier APIs. The pragmatist play: don’t overbuy local hardware at all — run a modest local box and spend the savings on API credits for the tasks local models can’t handle anyway.
Three AIs, three coherent strategies, zero consensus. [ChatGPT suggests to keep, Muse suggests to upgrade and Claude suggest to downgrade]
Why they disagreed
Here’s the part I found genuinely interesting, and the reason I’m writing this post.
The three AIs did not disagree about any fact. Nobody got the bandwidth numbers wrong. Nobody mispriced anything. They disagreed because my brief under-specified the objective function, and each model quietly filled in a different one:
- Verdict 1 optimizes for value per dollar today.
- Verdict 2 optimizes for regret minimization over five years.
- Verdict 3 optimizes for not overbuilding — it treats my ambition as the risk, not the hardware.
All three are defensible readings of the same brief. The disagreement wasn’t in the reasoning; it was in the unspoken premise about what kind of buyer I am. An AI giving advice is always smuggling in a utility function, and when your prompt doesn’t pin one down, you get whichever one the model defaults to. Ask three models, get three defaults.
This is worth internalizing beyond Mac shopping. When AI advisors disagree, the move isn’t to average their answers or pick the one that flatters you — it’s to notice which question each one thought you asked, and then decide which question is actually yours.
How I decided
Two realizations settled it for me.
First: “wait a year and upgrade later” is not a real option. The return window is the only cheap moment to change the memory spec. After it closes, upgrading means selling used (roughly $800–1,000 in depreciation) and paying full price again. Waiting doesn’t preserve optionality; it destroys it. The choice was never “now vs. later” — it was 128GB vs. living with 64GB for the life of the machine.
Second: the memory question was really an ambition question. If my honest future is one 27B model and API calls for everything else, 64GB is plenty and even oversized. If I’m actually going to build the concurrent multi-model setup — a 120B-class planner, a fast worker, a dedicated GUI grounder for computer control, all resident at once — 64GB doesn’t fit at all. (Concrete detail: gpt-oss-120b at 4-bit quantization is ~64GB by itself. It does not fit on a 64GB machine, period.) The $1,600 wasn’t buying gigabytes; it was buying a commitment to the project.
So the experiment I’m running is not a hardware benchmark. It’s a weekend of actually building, to find out which user I am.
The uncomfortable footnote
Make of that what you will. You could call it sycophancy — the AI just agreeing with whoever argued last. You could call it updating on new information, which is what reasoning is supposed to do. I genuinely can’t tell which it was, and I think that ambiguity is the most important thing in this post. An AI that refines its answer as you sharpen the question is useful. An AI that just drifts toward your latest worry is a mirror. The two look identical from the inside of the conversation.
The defense I have is the same one I’d offer for any advisor, human or otherwise: I didn’t accept any verdict on authority. I used the disagreement to find the real question — which turned out to be about my own ambition, not the hardware — and answered that myself.
Where it stands
At the time of writing, the plan is: spend a weekend building on the 64GB Studio, run the actual concurrent workload, watch the memory pressure, and decide before the return window closes. If the machine feels effortless, the 128GB justifies itself. If I find myself thinking “one model and APIs would’ve been fine,” I’ll keep the 64GB and spend the $1,600 on OpenRouter credits instead.
Either way, the three AIs did their job. Not by agreeing — by disagreeing in a way that showed me what I actually had to decide.