EleutherAI

• Stated work across '70 languages' and the Polyglot initiative for non-English language modelling signals some willingness to widen the data aperture beyond Anglophone corpora.
• Open-weight release model is in principle compatible with community re-use, which can lower barriers for communities to inspect what their data trained.

• No reference to Indigenous data sovereignty or the CARE Principles for Indigenous Data Governance anywhere in the public-facing copy.
• No mention of consultation with Indigenous communities or stewards of oral/non-textual knowledge.
• Large web-scraped pretraining corpora (the Pile lineage) are extractive by default; no acknowledgment of consent, provenance, or benefit-sharing for community-originated text.
• Multilingual breadth is framed quantitatively ('70 languages') rather than relationally — counting languages is not the same as honouring the knowledge systems they carry.

Language breadth is a genuine but thin signal. The complete absence of data-sovereignty framing, consent, or relational accountability for community knowledge keeps this near the floor; quantity of languages is not stewardship.

• Research on 'contamination effects on evaluations' and how 'properties of models emerge and evolve' during training shows awareness that models carry the history of their training process.
• The org's own origin as a grassroots Discord collective building open alternatives to proprietary labs is itself a historically-situated counter-position to corporate AI.

• No acknowledgment of colonial or extractive data legacies underlying the web corpora it trains on.
• No discussion of the geopolitical economy of compute — GPU access, donated/sponsored cloud, or the labour behind data cleaning and annotation.
• No transparency about regulatory or jurisdictional constraints shaping what it can release.
• Narrative treats open research as a self-evident good without situating it in the history of who has historically been able to 'open' and 'extract' knowledge.

There is technical-historical awareness of training dynamics, but no political or material history of data, compute, or labour. Methodological history without geopolitical history earns a low-mid score.

• Polyglot and the 70-language scope are the strongest concrete signal of multilingual ambition among comparable small labs.
• Framing positions itself explicitly 'beyond English-centric approaches,' naming Anglophone default as a problem to be addressed.
• Open release means non-Western researchers can in principle adapt models rather than only consume them.

• No evidence of consultation with cultural scholars, linguists from the relevant language communities, or native-speaker evaluators in the public copy.
• No discussion of preserving culturally specific reasoning patterns — multilingual coverage is presented as tokenisation/coverage, not as epistemic plurality.
• Western categorical logic (benchmarks, evals, 'capabilities') remains the unmarked universal frame against which all this is measured.
• '70 languages' risks being a coverage metric rather than a depth commitment; no per-language quality or community-ownership detail surfaced.

Multilingual work is real and above peers, lifting this above the indigenous-knowledge floor, but it stays at coverage-level inclusion. Languages are counted, not consulted; reasoning pluralism is absent.

• Open-weight releases of many LLMs are the single strongest verification affordance available — third parties can independently inspect, replicate, and audit.
• Active interpretability programme (Eliciting Latent Knowledge) and explicit work on evaluation contamination directly address the limits of evidence.
• 'Deep Ignorance' project on filtering pretraining data is an unusually concrete, falsifiable safety intervention published openly.
• GitHub repositories, academic publications (Google Scholar presence), and a 'Summer of Open Science' initiative all support replication.

• Public copy references rigorous evaluation but does not surface independent third-party audits of training data or bias in the materials read.
• Known-limitation disclosures are implied through the research agenda rather than stated as explicit model limitation statements on the landing pages.
• Evaluation-contamination awareness is named but the remediation status is not quantified in the visible text.

Open weights plus open code plus published interpretability and contamination research make this the lab's strongest lens by a wide margin. Held below 9 only because explicit independent-audit and formal limitation disclosures are not surfaced in the visible copy.

• EleutherAI's history includes generative/creative model work (its lineage touches open text and image generation), leaving an implicit door to affective and aesthetic dimensions.

• The public framing is entirely instrumental — capabilities, evaluation, alignment, safety — with no language of feeling, ambiguity, or poetic uncertainty.
• No acknowledgment of affective or intuitive dimensions of model behaviour or of the research itself.
• No recognition of emotional labour in community moderation or volunteer research.
• Attention is framed around efficiency, rigour, and capability, not other modes of attending.

The register is uniformly technical-instrumental. No space is made for the affective, ambiguous, or aesthetic, which keeps this near the floor despite the org's creative-model heritage.

• Explicit engagement with 'responsible scaling' and alignment/governance positions the lab as actively modelling risk trajectories of more capable systems.
• ELK and mesaoptimization work directly concern the futures of agentic/advanced systems and what could go wrong as 'models get smarter.'
• 'Deep Ignorance' (filtering pretraining data for protective measures in open-weight models) is forward-looking safety engineering for the open-release future.

• No environmental cost or compute-energy disclosure in the visible text — a notable omission for an org training many LLMs.
• No engagement with labour-displacement risk from the systems it helps make broadly available.
• Governance is named but there is no description of democratic or participatory deliberation over agentic systems — governance reads as technical, not civic.
• Whose futures are being optimised (and who bears the downside of open release) is not interrogated.

Strong technical futures-modelling around alignment and scaling, but the civic, environmental, and labour dimensions of 'whose future' are absent. A genuine mid-score: serious about capability futures, silent on distributive ones.

• Low-barrier open community model (Discord, GitHub, open publications) genuinely lowers participation thresholds versus closed corporate labs.
• Non-English / Polyglot work gestures toward Global South and minority-language developers as potential beneficiaries and contributors.
• Open weights let under-resourced researchers build without gatekept API access or licensing.

• No evidence of structured participatory design with Global South developers — openness is passive (anyone may join) rather than active (deliberate inclusion).
• No mention of disability-community accessibility of either the models or the research channels.
• No labour-representative engagement and no compensated feedback channels — participation is volunteer/unpaid by default, which selects for the already-resourced.
• Discord-centric organising can itself exclude those without the time, language, or connectivity to keep up.

Open-by-default participation is real and better than gatekept labs, but unpaid, self-selecting, English-Discord-mediated inclusion is structurally thin. No compensation, accessibility, or deliberate Global South co-design lifts it only to low-mid.

• The 'Deep Ignorance' project is itself a near-trickster move: deliberately making a model NOT know certain things inverts the field's reflexive 'more capability is better' creed.
• The whole org is a structural inversion of the closed-lab consensus — open weights as a refusal of the proprietary default carries an implicit 'the official story is wrong' energy.
• Studying evaluation contamination is a willingness to name that the field's own scoreboards may be rigged — auditing the audit.

• The public copy is earnest and solemn; no irony, satire, or self-mockery is deployed as a disciplined instrument.
• The lab does not visibly turn its scepticism on its own seriousness — openness is treated as self-evidently good rather than tested by its opposite (e.g. open release as a dual-use hazard).
• No space is held where the lab's own narrative is allowed to be contradicted; 'Deep Ignorance' hints at the dual-use tension but the copy does not name it as a contradiction.
• Counter-consensus posture is sincere conviction, not the disciplined inversion that trickster knowledge prizes.

There is real structural inversion in the work (deliberate ignorance, contamination audits, anti-proprietary stance), which earns a genuine mid-score, but the lab never turns the blade on its own openness-as-virtue, and the register is solemn rather than ironic. Latent trickster, undeployed.