Amazon (Nova)

Thesis

Amazon has collapsed its portfolio of high-end Nova models into a single flagship effort — Frontier Model Research (FMR) under Pieter Abbeel — while Nova Premier, Omni, Reel, and Canvas move to keep-the-lights-on status; only Nova 2 Lite/Sonic plus the Nova Forge customization and Nova Act agent tools continue W1. The public trail shows where that redirected engineering and compute is landing: a dense agentic-systems cluster (agent training-data infrastructure, multi-agent security auditing, agentic evals and benchmarks) P4P6P9P12P24, compute-efficiency research on diffusion language models and Trainium silicon P10P20P23E5E21, and a sustained investment in evaluation and synthetic-data artifacts P1P11P16P26. For a data-business operator, Amazon is signaling investment in the infrastructure and evaluation layer around agents and frontier training more than in incremental model launches.

Signal desks

• Hiring: No open-role or requisition evidence is cited in this pack. The only workforce signals are organizational: Amazon's ~80-person San Francisco AGI Lab was closed and its founder David Luan left in Feb 2026, with engineers and compute redirected into Frontier Model Research led by Pieter Abbeel (who joined via the 2024 Covariant deal) W1. Abbeel is named as the leader of the frontier-model research team that builds the base model within AGI W2. No role titles, team-level hiring lists, or locations can be mapped from this pack.
• Forks: No cited evidence in this pack. Every repo metadata record is marked "Fork: no" P1P4P6P8P9P10, and no Amazon fork of an upstream repo appears. The closest data are outward-facing: webie shows 2 external forks of Amazon's own dataset repo P19, and operator-probconserv incorporates modified code from zongyi-li/fourier_neural_operator and amazon-science/probconserv P28 — neither is Amazon forking activity.
• Releases: Diffusion language models ALoDLM-1.7B and ALoDLM-8B shipped on Hugging Face (cc-by-nc-4.0) E1E2P20P21; retrieval models poiss-bge-m3 retriever/reranker E26E28; and first-party finetunes including LAR-27B/LAR-9B E7E10 plus primed-HQwen3 and gpt-oss Eagle long-context variants E48E49E50E54E56E57E59E60. Repo releases include agent-ecc v0.2.2 (transport retry/backlog hardening) E3P2, ammo v1.0.0 E46, chronos-forecasting v2.3.2 E32, uniqsketch v1.6.0/v1.6.1 E34E44, and muss v1.0.0 E55.
• Talking: Amazon Science's public narrative clusters around (a) Trainium silicon co-design — TorchNeuron backend, kernel-centric video generation, the Trainium Frontier competition, and the $110M Build on Trainium credit program E5E21E52E53; (b) safety and formal verification — independent AVERI third-party safety review, Verus, and the Lean Focused Research Organization E9E36E42E51; (c) agentic infrastructure — graph-centric agentic intelligence, SOP-Bench, and why ML research agents don't overfit E19E45E31; (d) architecture efficiency — "priming" models without retraining E15E18; and (e) adjacent application science — controllable speech synthesis for audiobooks and antibody design/characterization E4E24.

Shipping

Concrete artifacts in the pack:

• Models: ALoDLM-1.7B (2.03B params, 486 downloads, 12 likes) and ALoDLM-8B (8.19B params, 1,882 downloads, 62 likes), both cc-by-nc-4.0, initialized from Qwen3 backbones, for text+code with adaptive recurrent diffusion E1E2P20P21P23.
• First-party finetunes: LAR-27B (26.9B params, apache-2.0, 34 downloads) and LAR-9B (8.95B params, apache-2.0, 25 downloads) E7E10; poiss-bge-m3 retriever (MIT) and reranker (apache-2.0) E26E28; primed-HQwen3 Instruct/Reasoner variants E50E54E56E57E59E60; gpt-oss Eagle long-context models E48E49.
• Datasets/benchmarks: evolveQA (4,967 QA records across AWS/Azure/WHO) P11; ListQA list-formatted factual QA eval P16; ATOD (1,000 agentic task-oriented dialogues) P24; SynthAVE (15,906 synthetic multilingual e-commerce products) P26; SpIDER-Bench code-graph retrieval dataset on HF P14; WebIE web information-extraction dataset P19.
• Tools/frameworks: PurpleAudit (multi-agent prompt-injection auditing) P4; agent-ecc (epoch-consistent agent training consumption) P6; LLMZero (agentic RL post-training strategy search) P9; ATTUNE (training-throughput tuning) P12; ammo (vLLM GPU kernel optimizer) E47; whatfits and HypAgent P15P8; data-driven-personas P1.

Research themes

1. Diffusion language models & compute efficiency — ALoDLM's adaptive recurrent diffusion and the dlm-scaling revisit (the 16× diffusion-vs-autoregressive gap shrinks to roughly 3–8× under current protocol) P10P20P23. 2. Agentic training & infrastructure automation — LLMZero uses LLM agents to search RL post-training strategies on verl P9; ATTUNE tunes tokens/s/GPU for Megatron-LM + slime on RayJobs/Kubernetes P12; ammo optimizes vLLM kernels for a specific deployment E47. 3. Agent safety & auditing — PurpleAudit's red/blue co-evolution against indirect prompt injection across LangGraph/CrewAI/AG2 P4; independent third-party safety evaluation via AVERI P7E9. 4. Training-data consistency for agents — agent-ecc seals every training batch under one grader version and re-executes stale labels P6. 5. Evaluation & benchmarks — evolveQA (temporal knowledge), ListQA, ATOD, SynthAVE, SOP-Bench, QUORUM (multi-annotator routing), NLPActiveTesting (active labeling) P11P16P24P26E45E40E38. 6. Retrieval & embeddings — SpIDER spatially-informed dense retrieval over code graphs, poiss-bge-m3 retriever/reranker, QASP query-adaptive vector search P14E26E28P13. 7. Synthetic personas/data — data-driven-personas induces personas from human behavior data and runs Bedrock batch inference P1. 8. Embodied & structured RL — StructRL online structured RL for long-horizon vision-language-action tasks P27; operator-probconserv OOD uncertainty quantification for neural operators P28. 9. Silicon & systems — TorchNeuron PyTorch backend for Trainium, kernel-centric real-time video generation, Graviton5, flat data-center networks P5E5E21E43E33.

Hiring & scaling

No role-level hiring evidence is cited in this pack, which is itself a signal gap for a "what are they hiring for" read W1W2. What the pack does show is a consolidation story: the ~80-person SF AGI Lab closed, founder David Luan departed, and engineering/compute was redirected into a single Frontier Model Research org under Pieter Abbeel W1W2. That implies scaling of one base-model team (plus continued Nova 2 Lite/Sonic, Forge, and Act tooling) rather than a broad portfolio of model teams; a new flagship model is expected at re:Invent later in 2026 W1.

Data-business implications

Connecting only cited evidence to concrete lanes:

• Training-data & eval versioning: agent-ecc is a production-shaped contract for sealing agent training batches under one grader version, re-executing/re-parsing stale labels, and neutralizing undecidable groups P6E3 — a direct template for reproducible eval/training-data pipelines.
• Synthetic data & LLM-as-judge labeling: data-driven-personas generates personas from human behavior data and runs parallel Bedrock inference P1; SynthAVE builds 15,906 synthetic products labeled via a 21-judge LLM arena with adjudication P26.
• Benchmark & dataset assets: evolveQA (temporal knowledge QA, 4,967 rows) P11, ListQA (list-format QA) P16, ATOD (1,000 agentic dialogues) P24, WebIE (web IE from c4) P19, SpIDER-Bench P14.
• Agent security / red-team tooling: PurpleAudit measures attack success rate and utility for prompt injection, with defenses, across LangGraph/CrewAI/AG2 P4; AVERI provides independent safety testing P7E9.
• Training & throughput tooling: ATTUNE raises tokens/s/GPU (6.7–33.9% on held-out models) and gates correctness P12; LLMZero ships end-to-end GRPO/SFT on verl P9; ammo autonomously optimizes vLLM kernels for a specific deployment E47.
• Deployment & inference: ALoDLM ships optimized inference engines P23; TorchNeuron targets reliable PyTorch execution on Trainium via runtime optimizations and agentic evaluation P5E5; kernel-centric video generation targets real-time Trainium inference E21.
• Retrieval / vector search stack: poiss-bge-m3 retriever and reranker E26E28, SpIDER dense retrieval P14, and QASP query-adaptive vector search (ICDE 2027) P13.
• Safety & eval quality: third-party review via AVERI E9P7, discounting correlated LLM judges E41, QUORUM quality-optimized multi-annotator routing E40, and NLPActiveTesting's active label selection E38.
• Product/GTM positioning: Nova models are served pay-per-use through AWS Bedrock W4W5; Amazon frames partnerships as translating discovery into real-world solutions E23 and uses Trainium credits to seed university adoption E53. No revenue figures are claimed here.

Traction highlights

• ALoDLM-8B is the pack's top traction artifact (1,882 downloads, 62 likes); ALoDLM-1.7B trails at 486 downloads/12 likes; the ALoDLM repo shows 52 stars E1E2E18.
• Most new research repos have 0–8 stars, e.g. webie 8 stars/2 forks P19 and operator-probconserv 6 stars/1 fork P28; the older dstc11 intent-induction repo stands out at 49 stars/9 forks P22.
• New finetune releases show very low traction (LAR-27B 34 downloads, LAR-9B 25 downloads) E7E10.
• External discussion is thin: only small HN threads (4 points/2 comments; 3 points/0 comments) on data-center and Lean posts E33E42E43.