Amazon (Nova)

Thesis

Amazon is building a vertically integrated AI stack that runs from custom silicon and formally verified cloud infrastructure through foundation models, agent frameworks, and domain-specific evaluation tooling. The evidence reveals a three-pillar strategy: (1) infrastructure differentiation via Graviton5 chiplet architecture E2 and the formally verified Nitro Isolation Engine E15P15, (2) a multi-modal foundation model family—Amazon Nova—paired with speculative-decoding speedups and third-party model fine-tunes W5E25E49, and (3) an aggressive push into agentic AI with open-source harnesses for perception, coding, and multi-agent systems P7W1E26. The lab is simultaneously investing in time-series forecasting (Chronos-2 at 15.2M downloads E1), causal discovery P12P14, and safety evaluation P25E34E40, signaling a horizontal play across data modalities and deployment surfaces rather than a single-model bet.

Signal desks

Hiring

• Amazon AGI is actively hiring for Amazon Nova foundation models and the San Francisco AGI Lab, which is described as "developing foundational capabilities for enabling useful AI agents that can take actions in the digital and physical worlds" P10. The SF lab seeks "a few dozen passionate, talented people" including "candidates from other disciplines, such as physics, math, or quantitative finance, who will bring fresh thinking to the field, regardless of experience level" P10. The AGI careers page explicitly frames Nova as "our new generation of state-of-the-art foundation models" P10. No specific job descriptions with data, eval, or infrastructure terms are cited in this pack beyond the general AGI team portal.

Forks

• No cited evidence in this pack. All repositories surveyed are original (non-fork) creations under amazon-science/ P3P4P7P11P12P20P22P23P24P25P26P27.

Releases

• chronos-forecasting v2.3.0 (2026-06-18): New cloud deployment guide for running Chronos-2 on AWS in 3 lines of code; fine-tuning now supports larger-than-memory datasets; new preprocessing module yields up to 20x faster input preprocessing; support for transformers>=5 and pandas>=3; removed dependency on scikit-learn P9E9.
• chronos-forecasting v2.3.1 (2026-07-02): Bugfix for training from lazy datasets with memmapped datasets.Dataset P1E4.
• StaminaBench v0.1.0 (2026-06-19): Benchmark data archives—LLM-generated scenarios (740 MB) and programmatic scenarios (230 MB)—for stress-testing coding agents over 100+ interaction turns across Mini-SWE, OpenHands, and OpenCode agents P7P8E8.
• FoundCause v1.0 (2026-06-17): Pretrained causal discovery foundation model (~1.6 GB weights, ~139M parameters) that predicts directed acyclic graphs and hidden-confounder matrices from observational CSV data in a single forward pass P12P14E11.
• UniqSketch v1.3.0 (2026-06-11): Bloom-filter sizing improvements for genomic sketching with controllable false-positive rates P16E14.
• Concurry v0.13.x (2026-05-21): Continued releases of the unified Python concurrency library (18 stars, Apache 2.0) replacing threading, multiprocessing, asyncio, and Ray with a single API P20E23E24.
• azcausal v0.2.5 (2026-04-30): Causal inference library release E43.
• Hugging Face model releases: Chronos-2 (119M params, 15.2M downloads, 343 likes) E1; a suite of P-EAGLE speculative decoding models for GPT-OSS and Qwen3-Coder at 20B–120B scale E25E27E28E47E60; multiple primed HQwen3 fine-tunes (Mamba2, GDN, BMOJOF, GKA variants at 8B–32B scale, with GDN-primed hitting 81K downloads) E49E51E56E57E59.

Talking

• Agentic AI dominates public narrative: Amazon published multiple blog posts framing agent reliability as the core challenge—"Bridging intent and execution in agentic systems" identifies harness bottlenecks between models and tools E17; "Real-world grounding in agentic AI" proposes four approaches for trustworthy agents E16; "What is agentic AI?" describes Nova Act as training "model capabilities, orchestration logic, and tool controls together as one integrated system" W4; and the open-source perception agent harness with annotation and verification was announced as "new multimodal interaction patterns for improved human-to-agent collaboration" W1.
• Infrastructure differentiation: The formally verified Nitro Isolation Engine—"the world's first formally verified cloud hypervisor"—uses a subset of Rust and the Isabelle/HOL proof assistant to provide mathematical assurance of VM isolation E15E50P15P19. Graviton5's chiplet architecture with custom die-to-die connectivity, DDR5-8800, and PCIe Gen6 delivers 25% better performance for "general-purpose and agentic AI workloads" E2P18. A blog on flat data-center network topologies with ShuffleBox optical components rounds out the infra narrative E3.
• Safety and trust: "Building trust into AI" describes the responsible-AI pipeline embedding safety throughout the development lifecycle E40. "Preserving the privacy of AI training data" details cryptographic defenses against training-data extraction attacks E44. "How catastrophic is your LLM?" proposes statistical methods for estimating catastrophic failure likelihood in adversarial conversations E45.
• Nova applications and customization: Fine-tuned Nova Micro achieved 94.77% extraction accuracy for email data, improving 16.6 pp over baseline with 30% lower latency and halved costs W2. Customized Nova models improved molecular-property prediction in drug discovery E54. A hyperparameter optimization guide for Nova Forge references the SageMaker HyperPod recipes repository W3.
• Research methodology and evaluation: "Ground truth is a process, not a dataset" addresses challenges in fact-checking long AI-generated research reports E19. "Diverse reasoning traces teach LLMs to make better decisions" covers training for diverse reasoning paths E22. "Making LLMs faster without sacrificing accuracy" presents a scaling law yielding 47% throughput improvement E29. "Intelligence isn't about parameter count. It's about time" argues for reducing inference time as models grow E41.
• Domain-specific deep dives: TRISO nuclear fuel for AI-scale energy E7P6; carbon intensity tracking across Amazon operations E5; AWS–Johns Hopkins antibody developability benchmark for AI-guided antibody design E55; mechanism design for vendor supply-chain optimization E39; middle-mile network optimization under uncertainty E35; Promptimus automated prompt engineering E31.

Shipping

Amazon shipped multiple concrete artifacts in this window. Chronos-2 remains the flagship time-series model with 15.2M Hugging Face downloads E1, and the v2.3.0 release added cloud deployment, larger-than-memory fine-tuning, and 20x faster preprocessing P9E9. StaminaBench launched as a new benchmark framework for evaluating coding agents over 100+ iterative turns with schema evolution, supporting Mini-SWE, OpenHands, and OpenCode agents P7P8E8. FoundCause shipped as a pretrained 139M-parameter causal discovery foundation model that outputs DAGs from CSVs in one forward pass P12P14E11. Concurry continued its release cadence as a general-purpose Python concurrency library P20E23E24. On Hugging Face, Amazon released a fleet of P-EAGLE speculative decoding models E25E27E28E47E60 and primed HQwen3 fine-tunes spanning Mamba2, GDN, BMOJOF, and GKA architectures E49E51E56E57E59. The Nova Act Annotator browser extension and skills were open-sourced for perception agent interaction W1.

Research themes

1. Agentic AI and tool use: The dominant theme across repos and blogs. StaminaBench stress-tests coding agents P7; EvoMAS generates multi-agent systems E26; ThermalForge uses LLM agents for building thermal dynamics models via LangGraph P11; reskill extends veRL for agent RL training with skill co-evolution E18; CompAgent verifies visual compliance E37; the perception agent harness introduces annotation and verification primitives W1; and Nova Act integrates model capabilities, orchestration, and tool controls as one system W4. 2. Formal verification and systems security: The Nitro Isolation Engine represents a major investment in formally verified cloud infrastructure, using Isabelle/HOL and a restricted Rust subset E15E50P15P19. TurboFuzzLLM applies mutation-based fuzzing to jailbreak LLMs for red-teaming P25. 3. Time-series forecasting: Chronos-2 continues active development with cloud deployment, preprocessing optimization, and transformers v5 support P1P9E1. 4. Causal discovery and inference: FoundCause ships as a pretrained model for DAG prediction P12P14; azcausal continues releases E43. 5. Evaluation and benchmarking: StaminaBench (coding agents) P7, MigrationBench (Java code migration) P27P24, Document Haystack (long-context multimodal VLM benchmark, 8,250 questions) P22, hallucination-benchmark-trivialplus (RAG-based hallucination detection, ACL 2026) E34, RMIR (reasoning-intensive multimodal image retrieval) E42, temporal-reasoning-dataset (multilingual temporal reasoning) E32, ACI-bench hallucination annotations for clinical summarization P28, SenTSR-Bench E10, RecArena E36. 6. Efficiency and speculative decoding: P-EAGLE speculative decoding models released across GPT-OSS and Qwen3-Coder families E25E27E28E47E60; MXFP4 training recipe for near-lossless low-precision training P23; DualKV for shared-prompt Flash Attention in RL training E20; scaling laws for throughput-accuracy tradeoffs E29; inference-time arguments E41; prompt compression research P26. 7. RAG and retrieval: State-Aware RAG with MCTS and CoT planners P3; MUSS for multilevel subset selection in RAG and candidate retrieval P4; EvoMAS for evolutionary multi-agent RAG systems E26; Document Haystack for long-context multimodal retrieval evaluation P22. 8. Safety, alignment, and privacy: Responsible-AI pipeline E40; cryptographic training-data defenses E44; catastrophic failure estimation E45; SWAN semantic watermarking E38; TurboFuzzLLM red-teaming P25. 9. Domain science applications: Drug discovery via customized Nova E54; antibody design benchmarking with Johns Hopkins E55; thermal dynamics modeling P11; carbon intensity tracking E5; TRISO nuclear fuel E7.

Hiring & scaling

Amazon AGI is recruiting across two poles: the Nova foundation model team and the San Francisco AGI Lab P10. The SF lab's mandate—"developing foundational capabilities for enabling useful AI agents that can take actions in the digital and physical worlds"—and its openness to non-traditional backgrounds (physics, math, quantitative finance) signals a long-term research orientation rather than pure product engineering P10. The lab is described as seeking "a few dozen" people, suggesting a boutique research unit rather than a mass-hiring scale-up P10. The Amazon Research Awards program spans 49 universities across 11 countries, providing access to Amazon public datasets and AWS AI/ML services E21, functioning as both a talent pipeline and an external research network. The volume of evaluation benchmarks being released (StaminaBench, MigrationBench, Document Haystack, multiple hallucination/retrieval/temporal benchmarks) implies a growing internal need for standardized evaluation infrastructure—a pattern consistent with scaling model and agent development.

Data-business implications

• Evaluation infrastructure demand: The proliferation of in-house benchmarks—StaminaBench for coding agents P7E8, MigrationBench for code migration P27, Document Haystack for long-context multimodal retrieval P22, hallucination benchmarks E34P28, temporal reasoning E32, and multimodal retrieval E42—signals significant internal investment in evaluation tooling. Each benchmark implies data generation pipelines (LLM-generated and programmatic scenarios at 740 MB and 230 MB respectively P8), Docker-based agent sandboxes P7, and automated scoring frameworks. Operators in the eval space should note that Amazon is building evaluation rigor around agentic workflows specifically, not just model outputs.
• Agent infrastructure and orchestration: Nova Act's design—training model capabilities, orchestration logic, and tool controls as one integrated system W4—coupled with the perception agent harness W1 and StaminaBench's agent harness wrapping CLI tools in Docker containers P7—points to infrastructure needs around agent sandboxing, tool-mediation harnesses, and multi-turn interaction tracking. The "Bridging intent and execution" post explicitly identifies harnesses as "becoming their own performance bottleneck" E17.
• Time-series data and deployment: Chronos-2's cloud deployment guide ("run Chronos-2 on AWS in 3 lines of code" P9) and 15.2M downloads E1 indicate production-scale time-series forecasting demand. The preprocessing optimization (20x faster P9) and larger-than-memory dataset support P9 suggest customers are running Chronos on substantial time-series corpora.
• Speculative decoding and inference optimization: The P-EAGLE model family across GPT-OSS and Qwen3-Coder E25E27E28E47E60, combined with the throughput scaling law research claiming 47% improvement with no accuracy loss E29, signals an inference-efficiency push. The MXFP4 training recipe P23 and DualKV for RL training E20 extend this to training efficiency.
• Safety and red-teaming tooling: TurboFuzzLLM P25, SWAN watermarking E38, the responsible-AI pipeline E40, and cryptographic training-data defenses E44 indicate a safety stack under active development. Operators building safety tooling should note the combination of fuzzing-based red-teaming, watermarking, and formal privacy guarantees.
• Causal and scientific ML: FoundCause's pretrained causal discovery model P12P14 and azcausal E43 suggest Amazon sees causal inference as a productizable capability. MUSS's application to RAG and candidate retrieval with "constant-factor approximation guarantee" P4 bridges causal/diversity optimization with retrieval pipelines.
• Concurrency and distributed infrastructure: Concurry (18 stars, Apache 2.0) aims to unify Python concurrency primitives P20. While modest in community traction, its existence alongside the Graviton5 E2 and data-center topology work E3 suggests internal tooling for distributed AI workloads.
• Data partnerships: The AWS–Johns Hopkins antibody developability benchmark is "powered by one of the most diverse antibody datasets in public literature" E55, demonstrating a pattern of academic partnerships that produce public evaluation data while showcasing AWS infrastructure.

Traction highlights

• Chronos-2: 15,241,261 Hugging Face downloads, 343 likes E1
• GDN-primed-HQwen3-8B-Instruct: 81,060 downloads E51
• BMOJOF-primed-HQwen3-8B-Instruct: 44,088 downloads E56
• Mamba2-primed-HQwen3-8B-Instruct: 33,097 downloads E49
• MXFP4-LLM repository: 127 stars, 18 forks P23
• TurboFuzzLLM: 24 stars P25
• Concurry: 18 stars P20
• reskill: 16 stars E18
• MigrationBench: 14 stars P27
• expert-upcycling: 14 stars E52
• TransitionFlowMatching: 12 stars E58
• ThermalForge: 11 stars E46
• JavaMigration: 8 stars P24
• ACI-bench hallucination annotations: 7 stars P28
• AI-Reinforced-Recommendations: 5 stars P21
• RMIR: 5 stars E42

Blog posts with Hacker News traction: Graviton5 (3 points) E2, data center flat networks (4 points, 2 comments) E3, inference-time argument (3 points) E41. Most blog posts show low external discussion, consistent with Amazon Science operating as a research dissemination channel rather than a community-growth platform.

Evidence is notably thin on: (1) direct Nova model releases in the window—the Nova family technical report is referenced W5 but the pack contains no Nova model weights on Hugging Face; (2) fork activity—all surveyed repos are original; (3) concrete hiring numbers beyond the qualitative AGI careers page; (4) revenue or GTM metrics beyond the Parcel Perform case study W2.