← The AI Hype Audit — all 306 verdicts
PARTLY Partly true: the 'Llama 70B on a MacBook for an 11-hour flight' story is a plausible, unverified anecdote; the memory math works on a 64 GB M4 Max, the battery math is the part nobody checked
The claim'A developer reportedly ran Meta's Llama 3.3 70B locally on a MacBook Pro M4 during an 11-hour flight, completing client work entirely offline. Using llama.cpp, it handled a roughly 60,000-token context window while using about 48.6 GiB of memory. Jobs were queued in JSONL, processed one at a time, with results saved automatically and checkpoints every 12 tasks. No cloud. No internet.'
Everything technical in the caption is possible and the details read like a real setup: a Q4-quantized Llama 3.3 70B is about 40 GB on disk, fits in the unified memory of a 64 GB M4 Max, runs under llama.cpp, and a JSONL job queue with periodic checkpoints is how a careful engineer would do batch work offline. The story traces to a Chinese-language post about an engineer processing client cases from a queue.jsonl with checkpoints every 12 tasks, which is where the '48.6 GiB' and '12 tasks' figures come from. What is not established is the headline: 11 hours. Community replies to the original estimated the laptop at roughly three hours and twenty minutes of battery under that load, and nobody has shown the flight, the seat power, or the machine. The page's own word is 'reportedly.' So: a real capability, an unverified anecdote, and a 'no Wi-Fi at 35,000 feet' framing that makes an ordinary local-model workflow sound like a stunt. Note also that a base 'MacBook Pro M4' cannot hold this model; it takes the Max with 64 GB or more.
What holds up
- Reel is a 10-second silent card; the caption is the whole claim and uses 'reportedly'.
- The anecdote circulated as a Chinese-language post (a Taiwanese tech site summarized it): queue.jsonl, checkpoints every 12 tasks, roughly 62.8 MiB per checkpoint; readers estimated about 3h21m of battery at full load.
- Llama 3.3 70B at Q4_K_M is roughly 40.5 GB and runs on an M4 Max with at least 48 to 64 GB of unified memory under llama.cpp; a 60K-token context adds several GB of KV cache, consistent with '48.6 GiB'.
- No primary source (post, video, or repo) from the developer was found; the figure of 11 hours rests on the retelling.
What doesn’t
- '11-hour flight' is the hook and the least supported number; battery under sustained 70B inference is measured in low single-digit hours without seat power.
- 'MacBook Pro M4' undersells the hardware: this needs an M4 Max with 64 GB or more, a $4,000-plus machine.
- An aggregator retelling an anonymous anecdote with 'reportedly' presented as news.
The catch
Running a 70B model offline on a top-spec MacBook is real and useful; the 11-hour flight is a story. If you want the workflow, buy the memory, plug into seat power, and expect a few tokens per second, not a cloud experience.
How to actually do it
- For offline work on a Mac, pick a model that fits comfortably: a 70B at Q4 needs 64 GB of unified memory; 32 GB machines should run a 27B to 32B model instead.
- Use llama.cpp or Ollama with a simple queue file and checkpointing, exactly as the anecdote describes; it is the right pattern for batch jobs without a network.
- Test battery life on the ground first. Sustained inference on Apple silicon drains a MacBook in a few hours; flights without seat power are not the place to find that out.
The technical claims are consistent with real llama.cpp behavior on an M4 Max with 64 GB; the 11-hour flight is an unverified retelling of an anonymous anecdote, and the battery estimate from readers of the original was about a third of that.
- Confidence
- Medium
- Posted by
- AICreatorBase (an AI news aggregator page); reel sent to the self-thread Wed Sep 16, 2026, 7:38 PM
We test hype for free. We build the real thing for a living.
Thirty minutes, no pitch — and you'll leave with something useful either way.
Book a call with Todd or start with the free Business Checkup →The Verdict Weekly
Three verdicts every Friday. Free forever, unsubscribe anytime, no spam — that would be ironic.
© Schreier Group · schreiergroup.com · See a wild AI claim? Drop it here and we'll test it.