ROOT FORWARD / NOVEMBER 6, 2026
Where AI is going—and what you can do with it.
A hosted model arrives ahead of its weights. A search connector returns a finished answer with citations. An insurer gives each software agent a handoff, disability advocates ask who gets to shape a driverless ride, and a satellite sends back a poem. Each story turns on what crosses the boundary and who can verify it.
Every handoff needs evidence the next person can use.
Root Note
01 / MODELS / MULTIMODAL
Inkling 2 Small arrives on Tinker ahead of its weights
Thinking Machines Lab released Inkling 2 Small on November 5 through Tinker. The model card describes a mixture-of-experts model with 594 billion total parameters, 17 billion active, text and image inputs, text output, and a context limit of up to one million tokens. Inference is available through the Tinker Playground, with fine-tuning and inference access through Tinker. The developer says it intends to release weights under Apache 2.0 the following week, after external testing.
The hosted release lets teams try tasks such as inspecting a screenshot, explaining a code change, or working through an interface with visual feedback. Those are useful pilot cases. The launch demonstrations and benchmark tables show what the provider measured in its setup. They leave open how the system performs with your data, review standards, and budget.
Root Access Take · Active parameters describe the fraction of the model involved in a token’s computation. Deployment still has to account for the weights and runtime overhead. A fast hosted trial and a viable self-hosted rollout are two separate purchasing and engineering decisions.
Try next · Start with synthetic screenshots and a small set of known failures. Track whether the model identifies the right control or state and can cite what it saw. Confirm distribution terms and hardware requirements before planning local deployment.
Sources · Thinking Machines Lab launch and availability · Model card and current terms
02 / INTEGRATIONS / PUBLIC WEB
AWS Web Answer gives agents a synthesis with citations
AWS made Web Answer generally available for Amazon Bedrock AgentCore on November 5. A Gateway administrator can add the AWS-managed connector as a target. Connected agents can discover and call it over MCP. It returns a synthesized answer and citations from the public web. The AWS walkthrough uses a Gateway in US East (N. Virginia) with IAM inbound authentication, an authorized caller, and permission for the Gateway’s service role to use the connector.
The useful distinction is where the synthesis happens. With Web Search, a team can get search results and decide how to interpret them. With Web Answer, the connector does more of that work. That can simplify a support assistant that needs a public vendor answer, at the cost of giving up some control over how sources are selected and combined.
Root Access Take · Make the citation part of the acceptance test. A link can be genuine and still support a different product version, plan, or region. A polished answer that sends an engineer to the wrong runbook consumes the time you were trying to save.
Try next · Try questions with known vendor answers and deliberately similar documentation. Track whether each citation supports the specific claim, then record latency and tool-call charges before expanding access.
Sources · AWS launch and region · AWS setup walkthrough · AWS connector and language details
03 / MSP PRACTICE / DELIVERY
Verisk makes each software agent hand off something reviewable
Verisk and AWS published a November 5 case study about agents working at different stages of Verisk’s software delivery process. The examples start with Jira tickets. A requirements agent proposes missing acceptance criteria and flags ambiguity. A separate technical grooming stage uses the approved ticket and context from repositories and API specifications to propose an implementation plan. The workflow then proceeds through developer, review, and quality roles.
The case study describes human checkpoints before stages advance. A product owner reviews proposed requirements, developers validate the plan and remain responsible for committed code, and review feedback passes through the normal repository workflow. The important change is that decisions can return to the existing record rather than living in a chat transcript the next reviewer never sees.
Root Access Take · Teams can start with a single handoff that already has an owner. A polished specification is useful only when the person approving it can see what changed and can send it back. That approval needs to mean something before the next system acts on it.
Try next · For one recurring service change, ask an agent to identify gaps in a ticket using an approved checklist. Save the suggestions separately, require a named reviewer, and measure missing criteria, rework, and reviewer minutes.
Sources · Verisk and AWS delivery case study
04 / GOVERNANCE / ACCESSIBILITY
Disability advocates put the whole driverless trip on the agenda
The American Association of People with Disabilities announced its first Autonomous Vehicle Summit for November 5 in Washington, DC, with its We Will Ride initiative. The association said the program would bring disability advocates, accessibility specialists, vehicle companies, and government officials together. Its agenda included safety, policy, community engagement, and the staff and practices needed to support accessibility.
This is a convening by an advocacy organization. It is useful as an example of how people who depend on a system can help define what counts as success. An autonomous vehicle might navigate the street correctly while its booking, boarding, or support process leaves a rider stranded. A test that stops at the vehicle’s driving performance misses the other systems that make a trip possible.
Root Access Take · Bring users into the definition of the service early enough to change the design. For a technical team, that means treating accessibility as a property of the entire workflow. A route can be technically successful while the person it was designed to serve never makes it out of the pickup zone.
Try next · For a service pilot, walk through booking, pickup, boarding, disruption, and human escalation with the people who will use it. Give each step a way to report a failure and a person responsible for resolving it.
Sources · AAPD summit announcement and agenda
05 / HARDWARE / INFERENCE
Google’s orbital TPU sent back a poem
Google reported on October 30 that Project Suncatcher had returned its first AI response from an orbital prototype. The satellite launched October 1. Google says the team sent a prompt during a ground-station pass, ran an onboard Gemma 4 model, and received a haiku on the downlink near Svalbard. The company also reports testing further prompts and plans to study operation under thermal stress and radiation exposure.
A short poem is a surprisingly good commissioning payload. It exercises the uplink, the inference hardware, the model, and the downlink without making anyone interpret a large data dump. This older item earns its place as an example of how to test a whole system with a small, unmistakable result. It is one provider-reported prototype milestone, with sustained performance, long-term reliability, and economics still to demonstrate.
Root Access Take · Start with a test that can tell you which boundary worked. If the first mission is to prove the signal can make the round trip, a simple input and a checkable output beat a complex benchmark whose failure could mean almost anything.
Try next · Borrow the idea for a remote edge device. Send a small known task across the real connection, time each hop, and check what the device reports when it loses the link.
Sources · Google Research prototype milestone
Explore the Root Forward archive, AI Radar, Governance guide and The Root Shelf.