Planned Topics

These topics are under consideration for future versions of the paper. The field moves quickly, so any item may be dropped, merged with another, or moved to a later version.

• New Models

  Updated text covering new closed-source and open-weight/open-source AI models, starting with Anthropic’s Claude Opus 5, Claude Fable 5.1 and Mythos 5.1, Claude Opus 5.5 and Claude Sonnet 5.5, plus new releases from OpenAI, Google and other vendors. Helps readers decide which models to evaluate now and which to keep on a watch list.

• Open Weights and American AI Leadership

  NVIDIA CEO Jensen Huang’s letter on open weights and American AI leadership. Adds a hardware vendor’s policy argument to the paper’s case for running open-weight models on premises.

• Stanford AI Index 2026

  Findings from Stanford HAI’s 2026 AI Index Report (PDF) on model performance, cost, adoption and investment. Gives decision makers an independent baseline to check vendor claims against.

• Claude Managed Agents and Self-Hosted Sandboxes

  How Anthropic’s hosted agent service can run tool execution in a self-hosted sandbox on your own infrastructure while orchestration stays with Anthropic (documentation). Helps teams that want managed agents but need code, files and network egress to stay inside their perimeter.

• How Building Software Is Changing at Anthropic

  Lessons from The Pragmatic Engineer’s look inside Anthropic on how an AI lab’s own engineers now build software. Gives engineering leaders a reference point for how AI-assisted development changes roles and team workflows.

• Shared Claude Chats in Search Results (July 2026)

  How shared Claude conversations and artifacts became findable through Google and Bing searches in late July 2026 (Malwarebytes, Fortune). A case study for shadow AI and data handling policy, since a public share link can behave like a published web page.

• PJM Grid Limits on New Data Center Load

  PJM’s July 27, 2026 Board decision that new large loads, such as data centers, that lack sufficient capacity of their own as of June 1, 2027 will be cut back before PJM’s pre-emergency load management. Strengthens the paper’s energy argument for right-sized local AI.

• Expert Voices: New Additions

  Add perspectives from Demis Hassabis (Alphabet chief scientist and Google DeepMind chair), Google CEO Sundar Pichai, NVIDIA CEO Jensen Huang and John Carmack to the expert voices section. Widens the set of informed viewpoints readers can weigh against vendor messaging.

• Useful Life of AI Infrastructure

  How many years properly operated GPU servers, networking and cooling equipment stay in productive service before replacement, and how that service life feeds depreciation and ROI. Helps finance and IT leaders set a realistic payback period for on-premises hardware.

• Inkling-Small

  Thinking Machines Lab’s Inkling-Small, an open-weight mixture-of-experts model with 276B total and 12B active parameters that performs close to the larger Inkling at a quarter of its size. Extends the paper’s existing Inkling coverage with a smaller option for on-premises hardware.

• GPT-5.6 Price Cuts

  OpenAI’s July 30, 2026 price cuts of 80% for GPT-5.6 Luna and 20% for Terra, followed by a time-limited discount on Sol in August. Shows how quickly hosted API prices move and why the paper’s cloud versus on-premises cost comparisons need a refresh date.

• 128 GB AI Developer Boxes

  Compact AI development systems with 128 GB of unified memory, such as the Microsoft and NVIDIA Surface RTX Spark Dev Box and the AMD Ryzen AI Halo. Helps teams size a low-cost local prototyping tier before committing to server-class GPUs.

• BOND AI Trends Report (May 2025)

  Mary Meeker and BOND’s May 2025 report on AI trends, a 340-page data deck on AI adoption, usage growth, capital spending and falling inference costs. Provides historical context for the acceleration timeline and the investment case.

• Bun’s Rewrite from Zig to Rust

  Bun’s rewrite of roughly 535,000 lines of Zig into Rust, done largely with a pre-release version of Claude Fable 5. A case study of AI-driven code migration at scale, including the test suite that made it feasible and what it cost.

• Microsoft MAI Models

  Microsoft’s proprietary MAI model family, launched at Build 2026 and led by the MAI-Thinking-1 reasoning model, distributed through Microsoft Foundry and partner APIs. Clarifies where Microsoft’s own models fit next to the paper’s existing Microsoft and Azure coverage.

• Claude on Google Cloud and AWS

  Access to Claude models through Amazon Bedrock and Claude Platform on AWS and through Google Cloud. Helps organizations already committed to one cloud reach frontier models under their existing contracts and governance controls.

• FinOps for AI

  Applying FinOps Foundation practices to AI spending: tracking token, GPU and cloud costs and assigning each to an owner. Gives finance and engineering one shared view of AI cost across cloud, hybrid and on-premises deployments.

• NVIDIA Agentic Variation Operators (AVO)

  NVIDIA’s AVO architecture for long-horizon autonomous agents, which NVIDIA reports reached 100% on ARC-AGI-3. Tracks how agent architecture, in addition to model size, drives capability gains.

• Edge LLM Serving on Limited Hardware

  Serving systems that run large open-weight models on personal and edge machines, starting with FreeToken (preprint, August 2026), which serves mixture-of-experts models larger than GPU memory by spreading work across GPU, CPU and host memory. Shows how far a single workstation or laptop can stretch for small teams.

• Ollama Cloud Models and Cost

  Ollama’s hosted cloud models and plans, which run larger open-weight models through the same tools used for local inference. Helps readers compare pay-as-you-go cloud capacity against buying more local hardware.

• GB300 DGX Station Workstations

  NVIDIA’s DGX Station with the GB300 Grace Blackwell Ultra superchip and 748 GB of coherent memory, now built and sold by OEMs including ASUS, Dell, GIGABYTE, HP, MSI and Supermicro. A deskside tier between developer boxes and rack servers for teams that fine-tune or serve large models locally.

• Data Foundations Before AI Tools

  Why data quality, governance, cataloging and access control should be engineered first, before adopting the AI tools that depend on them. Helps leaders sequence investments so AI projects do not stall on unready data.

• NVIDIA Personal AI Router (PAIR)

  NVIDIA PAIR, a free open-source beta that routes inference jobs across machines on a local network. Shows how small teams can pool the GPUs they already own into one local inference resource.

• Acceleration Timeline Additions

  Add these milestones to the acceleration timeline: 2015 ResNet, 2016 AlphaGo, 2020 neural scaling laws and DDPM, 2021 AlphaFold 2, 2022 RLHF through InstructGPT, and the 2026 mathematics results from Claude Fable 5 (Jacobian conjecture counterexample) and GPT-6 Astra (ten advances in mathematics). Gives readers the longer arc behind today’s capabilities.

• NVIDIA RTX PRO 5500 Blackwell

  NVIDIA’s RTX PRO 5500 Blackwell with 84 GB of GDDR7 memory, built for rack-mounted workstation deployment. An option to compare against the RTX PRO 6000 recommended for Phase 1.

• The September-October 2026 AI Panic

  An analysis of what is driving the fall 2026 wave of public anxiety about AI. Helps leaders separate lasting risks from news-cycle reactions when briefing boards and staff.

• AMD EPYC 9996 “Venice”

  AMD’s EPYC 9996, the 256-core Zen 6 flagship of the EPYC 9006 “Venice” family. Informs CPU sizing for agent sandboxes and GPU host nodes, and the paper’s Phase 2 AMD re-evaluation.

• Google Procedural Graphs

  Procedural Graphs, Google research that stores an agent’s procedural knowledge as a graph that an LLM refines from failed and successful runs. Points to a practical way to make long-running agents more reliable without retraining the model.

• Decision Models: Jev and Julia 1

  TypeSafe AI’s Jev, a closed “System One” model that outputs typed decisions with calibrated probabilities and no free-form text, and Supersonic Labs’ open-weight Julia 1. Shows a faster, cheaper option for routing, classification and guardrail steps inside AI workflows.

• DORA Research on AI

  DORA’s research publications on AI-assisted software development. Gives engineering leaders survey-based evidence on how AI adoption affects delivery throughput and stability.

• NVIDIA DGX Spark 64 GB

  NVIDIA’s 64 GB DGX Spark, sold through OEMs from October 23, 2026 starting at $4,999, with two units able to pool 128 GB of memory. A smaller configuration for local agent and inference work, and a new data point for the paper’s hardware tiers.

Suggest a topic

If a question you face is missing from the paper, write to contact@emilioborges.com or raise it in the comments on the most closely related section.