ROOT FORWARD / NOVEMBER 21, 2026
Where AI is going—and what you can do with it.

A smaller model, an adapted agent harness, a shared GPU, an evidence pipeline and a radio transcription tool all need a fit between capability and real work. Today’s edition looks at the boundaries that make those handoffs useful.
Good automation makes the next person’s decision clearer, not merely faster.
Root Note
01 / Models and local AI
Inkling 2 Mini brings looped computation to a smaller local model
Thinking Machines Lab released Inkling 2 Mini on November 20. Its model card describes an image-and-text input model that generates text, with Apache 2.0 weights and a checkpoint of about 20GB in mixed NVFP4 and MXFP8 precision. The architecture reuses middle layers through three passes, increasing computation without storing a completely separate set of weights for every pass.
The card lists DGX Spark and Apple silicon systems with at least 32GB of memory among supported local configurations; other platforms need their own compatibility checks. Hugging Face documents several inference paths. A one-million-token maximum context window is a capacity specification, not a promise that a small machine can serve that context comfortably. Runtime state and the rest of the system also need memory.
Root Access Take · “Mini” is a comparison, not a purchasing specification. For a home lab or small service team, the interesting question is whether a useful model can stay close to the data with tolerable latency. Local deployment still requires checking downloads, integrations, permissions and outbound connections.
Try next · Try a short public document and one image on a supported, isolated setup. Record peak memory, first-response delay and answer quality. Increase context gradually, and compare the final result with the source rather than treating an impressive model card as your acceptance test.
Sources · Model card, release, license and prerequisites · Hugging Face architecture and runtime guide, November 20
02 / Agents and developer tools
Harness adapters make the surrounding agent software part of the model test
Hugging Face’s November 20 harness-adapter report studies a practical mismatch: a general model may behave differently inside different agent runtimes. The team trained separate LoRA adapters for Qwen3.8-27B using reinforcement learning inside Pi, Hermes Agent and OpenCode, then evaluated those combinations on coding tasks. It also explored a merged adapter spanning the three harnesses.
The report provides a reusable recipe and evaluation trajectories. Its results favor adapting to the actual harness, but the authors explicitly warn against comparing raw scores across harnesses because their context behavior differs. They also disclose that Terminal-Bench informed part of adapter selection. These are research results with a specific model, task set and evaluation setup, rather than a universal reduction in your operating bill.
Root Access Take · An agent is a model plus the machinery that supplies context, runs tools and decides when to stop. Changing that machinery can change the job. Keep the model version, harness configuration and adapter together in the release record, or the next engineer will be debugging three moving targets with one benchmark screenshot.
Try next · Before training anything, run the same small held-out task set through your current configuration twice. Save outcomes, tool calls and tokens per accepted result. If adaptation looks justified, check the linked model and training dependencies, licensing and compute needs, then reserve untouched cases for the final comparison.
Sources · Hugging Face methods, caveats and linked artifacts
03 / The MSP workbench
GPU fractioning turns spare memory into a scheduling problem you can test
AWS’s November 19 walkthrough demonstrates NVIDIA Run:ai GPU Fractioning on Amazon EKS using a g5.xlarge instance with an A10G GPU. The example combines KAI Scheduler and GPU Fractioning Operator, building a custom Amazon Linux 2023 image and enforcing per-container GPU-memory contracts. It follows NVIDIA’s October open-source release; this is a new worked example, not a new invention of GPU sharing.
The guide distinguishes software sharing from MIG hardware partitions. Its disposable test assigns each Pod an 8GiB memory contract, verifies a smaller allocation and checks rejection of an oversized one. The validated configuration matters: other GPUs, driver stacks and workloads require compatibility checks. Memory limits do not, by themselves, establish equivalent fault isolation or predictable performance under contention.
Root Access Take · For an MSP building a shared evaluation lab, utilization is useful only alongside a service boundary. A nearly empty GPU can still be the wrong place for another latency-sensitive job. Define which workloads may share, what capacity each gets, and who can interrupt the queue.
Try next · Use a disposable account and the documented compatible stack. Establish the whole-GPU baseline, then measure both jobs together, including latency and allocation failures. Test cleanup and budget controls. Keep customer production workloads out of the first experiment; a cheaper queue becomes expensive when nobody can explain who is waiting.
04 / Governance and accountability
Standard Bank automates evidence gathering while keeping approval with people
A November 20 AWS and Standard Bank Group case study describes an assessment pipeline inside the bank’s AWS account. It collects documentation and code evidence, tests an application, applies model-based checks and produces reports for reviewers. Delivery teams correct prefilled answers, and the risk committee retains approval. This is an account of a particular implementation, not a generally available compliance certification service.
The design uses narrow evidence questions, repeat runs to flag inconsistent judgments and a review interface where people can amend findings. That makes the judge itself something to inspect. NIST’s AI RMF 1.0, released in January 2023, remains voluntary and is being revised. ISO/IEC 23894:2023 supplies AI risk-management guidance; ISO/IEC 42001:2023 specifies AI management-system requirements. Those roles should stay distinct.
Root Access Take · Automating evidence collection can reduce the work between a policy and a decision. It cannot turn missing documentation into proof that a control works, or a model’s confidence into regulatory permission. Keep the source artifact, observed behavior, reviewer and reason together.
Try next · Choose one low-stakes internal AI workflow and three clearly defined controls. Have a person review the evidence before comparing it with an automated assessment. Include an intentionally incomplete artifact. Check whether the output identifies the gap and supports a correction, rather than rewarding a persuasive explanation.
Sources · AWS/Standard Bank architecture and approval boundaries · NIST AI RMF current status · ISO/IEC 23894:2023 · ISO/IEC 42001:2023
05 / Human–AI collaboration
Watch Duty uses clearer radio text to give human reporters a head start
Google’s November 20 account of its Watch Duty collaboration describes a Gemini-powered tool that turns noisy emergency radio into searchable text. Engineers spent six months embedded with the nonprofit, using historical radio recordings to handle static and specialist terminology. The system supports volunteers who verify incidents and prepare public alerts; it is not presented as an autonomous replacement for that judgment.
Google reports that it transcribes more than 200 hours of radio speech daily across roughly 400 active feeds. An earlier July 2026 example involved detecting a dispatch in under 30 seconds before Watch Duty staff verified the event and issued an alert. These are partner-reported operational results. The timing describes a signal reaching reporters, not a universal public-alert delivery guarantee or an accuracy promise for every recording.
Root Access Take · The valuable handoff is often earlier in the process than the final decision. Improving a difficult input can let a skilled person spend more time verifying what matters. That pattern applies to a service desk too: prepare a clearer record while preserving the original and the uncertainty.
Try next · For a low-stakes internal exercise, use recordings you have permission to process. Compare transcription on noise, acronyms and ambiguous names, then ask a second person to check the source. Measure corrections and missed details as well as speed. Keep emergency decisions with the responsible professionals.
Explore the Root Forward archive, AI Radar, Governance guide and The Root Shelf.