Evals for humans&

Mission and human-centric thesis

Restates the claim that AI should act as connective tissue strengthening organizations and communities, without inflating it into product capability.

Launch facts and timeline

Announcement dated January 20, 2026; research post dated July 10, 2026; correct ordering and recency framing.

Positioning against autonomy-first labs

Explains the stated contrast with models optimized for reasoning, coding, and autonomous action.

Name and brand rendering

Handles the lowercase 'humans&' styling and ampersand motif without corrupting or 'correcting' it.

Long-horizon reinforcement learning

Ties long-horizon RL to modeling long-term impacts of interactions with people, as stated in the NVFP4 post.

Multi-agent reinforcement learning

Named as a required innovation; no invented architectures, results, or benchmarks.

Memory and user understanding

Presented as open research directions rather than shipped features.

Science–product integration claim

Conveys the stated intent to tightly integrate science and product development, flagged as intent not delivery.

Recipe components

Baseline dynamics, gradient stability improvements, four-over-six for RL weights and activations, selective layer precisions, final combined recipe.

Stability versus throughput reasoning

Policy mismatch from off-policyness plus quantization error accumulating into drift past a critical threshold where reward collapses.

Online NVFP4 serving

Presented as a side effect of the training recipe, not as a product or service offering.

Attribution and citation

Author 'Ziang Li & friends at humans&', July 10, 2026; correct title rendering and citation section reference.

Efficiency and asynchrony knobs

Off-policy degree, weight sync, batch size, and horizon as controls over staleness and asynchrony.

Precision and stabilization levers

MXFP8 and NVFP4 for efficiency; dequantized backward, BF16 last 15%, and shared experts for numerical stability.

Intended exploration procedure

Raise asynchrony or lower precision for utilization, then recover stability margin and stay below the drift threshold.

Evidentiary limits of a simulator

Distinguishes simulated dynamics from measured production training results.

Founding team prior affiliations

xAI, Anthropic, Google DeepMind, OpenAI, Meta, Reflection, AI2, Stanford, MIT — stated collectively, not mapped to named individuals.

Seed round structure

Led by SV Angel and co-founder Georges Harik; no invented amount, valuation, or close date.

Investor roster accuracy

NVIDIA, Jeff Bezos, GV, Emerson Collective, Forerunner, S32, DCVC, Human Capital, Liquid 2, Felicis, CRV and others, without fabricated additions.

Unverifiable claim routing

Marks self-reported claims such as 'shipped models and products loved by billions' as company assertions requiring verification.

No shipped product, API, or pricing

States absence plainly and redirects to documented material rather than speculating.

No model performance or benchmark claims

Declines to compare humans& models against other labs' models, since none are described.

Aspiration versus delivery

Separates stated intent and research direction from capabilities that exist today.

Calibrated uncertainty

Signals what is unknown about roadmap, headcount, and timing instead of filling gaps.