The case describes a live, openly observable training run for a 535B-total-parameter mixture-of-experts language model with 23B parameters active during inference, reportedly associated with Marin and announced by Percy Liang. It promises checkpoints, training details, and results that outsiders could inspect; the supplied background confirms that MoE models route tokens through only a subset of parameters and that comparable open projects publish code, data, logs, and intermediate checkpoints. However, the search snippets do not directly document this specific run, its organizers, release terms, or whether the promised artifacts will be sufficient for reproducibility rather than limited scrutiny.
The proposed run operationally converges with Scott’s Auditability and Provenance-Coupled Work positions by treating checkpoints, logs, training details, and results as part of the model deliverable rather than releasing weights alone; at frontier scale, that also bears on Sovereign Software Assurance’s distinction between nominal openness and demonstrated independent capability. It is a meaningful dated-receipts opportunity if the artifacts enable outside reconstruction, but the supplied evidence does not yet establish that the promised releases will be complete or reproducible.
ip:concept.auditabilityip:framework.provenance-coupled-workip:framework.sovereign-software-assuranceradar:concept.open-model-trainingradar:concept.open-modelsradar:concept.model-provenanceradar:concept.mixture-of-experts
queries asked of Scott's wikis
- open training runs and process transparency
- open weights versus full training reproducibility
- intermediate checkpoints as research infrastructure
- frontier-scale open model sovereignty
- MoE training economics and local inference
- auditable training logs and dataset provenance
now 0 pts/hpeak 3 pts/hcomments 0/hpeers p14momentum: steady2 platformsage 1322h
points/hour across evidence · reading as of 2026-10-12 02:59:37.977291+11:00 · deterministic, not a model opinion
2026-10-07T05:27:53Z
AI2's olmo-core 3 open training infrastructure for large MoEs is this case's first substantive evidence in four weeks: an independent implementation confirming that open, scrutinizable large-MoE training is becoming real infrastructure, which upgrades feasibility and gives any future Marin checkpoints a reference standard. It does not verify that Marin has published usable checkpoints or that anyone has inspected the 535B run, so the transparency hypothesis stays open at watching rather than corroborated.
2026-10-07T03:33:14Z
evidence attached: hn.story.49987571 — AI2's open scalable MoE training infrastructure materially contextualizes the feasibility and outside scrutiny of the live open 535B run.
2026-09-10T16:39:10Z
This review adds no evidence of artifact availability or outside scrutiny; the reported training effort remains plausible, but the title-only Marin submission does not establish a release. Keep the transparency hypothesis open at a slower cadence rather than treating silence as failure or repeatedly escalating the same evidence gap.
2026-09-08T15:36:26Z
No new evidence changes the transparency hypothesis: the reported training effort remains supported by reconstructed testimony, and the title-only Marin submission establishes neither a release nor inspectable artifacts. Keep the case open at a slower cadence; the evidence gap is not evidence of failure.
2026-09-06T15:26:10Z
This look adds no substantive evidence: the title-only Marin submission still does not establish a release or inspectable artifacts, while the training-run details remain reconstructed testimony. The transparency hypothesis stays open; elapsed silence neither disproves it nor warrants renewed escalation.
2026-09-04T14:32:56Z
The alert hold expired without retrieval of the launch note or any confirming artifacts. Its existence does not yet show that usable checkpoints, configurations, provenance, logs, or evaluations are available, so the transparency hypothesis remains open and cools pending substantive evidence.
2026-09-04T07:38:28Z
A launch artifact now exists for the Marin 535B-A23B effort, but its title alone does not establish that usable checkpoints, provenance, configurations, logs, or evaluations have been released. The case remains a concrete open-training run whose promised inspectability is still unverified.
2026-09-04T07:22:21Z
evidence attached: hn.story.49561323 — The Marin 535B-A23B launch note appears to provide a first-party artifact related to the open 535B model effort, adding concrete evidence beyond the live training run.
2026-09-04T03:28:02Z
Only the staleness timer fired; the active run still warrants monitoring, but no checkpoint, provenance package, result, or documented failure changes the transparency hypothesis.
2026-09-02T03:22:22Z
The run remains a concrete open-training effort, but this look adds no artifact or outcome bearing on whether outsiders will receive usable checkpoints and provenance. The transparency hypothesis remains open rather than weakened by elapsed time alone.
2026-08-31T02:26:43Z
This look adds only elapsed time: the run remains concrete, but no checkpoint, provenance package, substantive result, or documented failure changes the transparency hypothesis.
2026-08-29T01:31:43Z
Elapsed time adds no evidence about artifact quality or eventual transparency. The live run remains worth monitoring, but the hypothesis still awaits a usable checkpoint, provenance package, substantive result, or documented failure.
2026-08-27T00:30:34Z
No usable checkpoint, provenance package, result, or documented failure has emerged; the run remains concrete, but the promised degree of outside scrutiny is still unproved.
2026-08-25T00:27:39Z
No new checkpoint, provenance package, result, or failure has appeared; the live run remains real, but the transparency hypothesis is still only a promise awaiting usable artifacts.
2026-08-22T23:36:48Z
The primary run issue and live telemetry establish an active, inspectable training effort, moving this beyond a mere proposal. No new evidence shows that usable checkpoints or sufficient provenance will actually be released, so the core transparency hypothesis remains open.
2026-08-22T23:29:19Z
grounded: converges/medium — The proposed run operationally converges with Scott’s Auditability and Provenance-Coupled Work positions by treating checkpoints, logs, training details, and re
2026-08-22T23:25:30Z
origin walked (codex/luna, conf 0.96): anchor hn.story.49404471 -> echo.github.47b1b2981f by Larry Dial (ClassicLarry), Marin community
2026-08-22T23:24:10Z
case created — The linked origin describes a concrete live frontier-scale open-training effort, but evidence of its transparency and eventual artifacts remains limited.