DS4 (DeepSeek 4 Flash local inference engine for Metal) hit 7,100 stars on GitHub in days — an open-source project that runs DeepSeek V4 Flash directly on Apple Silicon GPUs via Metal. No cloud. No API bills. Just your Mac and a 128K context window.
Mac Studio M4 Ultra can now run a 70B+ parameter model at reasonable speeds entirely on-device. The inference cost equation is being rewritten — a one-time GPU purchase beats per-token cloud fees indefinitely.
Estimating and sales tools that run on cloud LLM APIs cost money every query. Local inference removes that ongoing cost. A quotes app running on a Mac Mini in the office could handle most routine scope questions without touching the internet.
Strukto-ai/mirage (1,880 stars) is a unified virtual filesystem for AI agents — it layers a logical file structure over scattered resources (GitHub repos, Google Drive, databases, APIs) so an AI agent can navigate a project like a human using a folder hierarchy. Agents stop getting lost in a sea of tools.
18 other "awesome-agentic-ai" repos collectively trending. The pattern is clear: the scaffolding for reliable agents is maturing fast. We're moving from "agents that sometimes work" to "agents that reliably work in defined environments."
When AI agents can reliably navigate project documentation, submittals, and specifications without forgetting where files live, automating RFI responses and submittal tracking becomes realistic.
The robotics-skills-suite repo (audit-ready Claude skills for industrial robots) crossed 500 stars while actual humanoid deployments are accelerating. Figure AI, 1X Technologies, and AgiBot are all running robots in live environments — warehouse picking, factory floor tasks, early construction assistance.
May 2026 arXiv papers show 6D pose estimation and human motion generation reaching production quality. The gap between "demo robot" and "reliable labor" is closing faster than the 5-10 year timelines cited two years ago.
Glazing installation is physical labor with a high skill ceiling — reading prints, handling glass, coordinating with other trades. Humanoids won't replace glaziers in this decade, but the first tasks to automate will be material handling and site prep. Worth watching AgiBot and Figure AI's construction pilots.
earthtojake/text-to-cad (2,395 stars) is an open-source harness for generating CAD models from text prompts. Not just descriptions — actual parametric models exportable to standard formats. This is in early stages but moving fast.
CA-SQL research (Complexity-Aware Text-to-SQL) shows AI's ability to reason through constrained generation is improving. Apply that to CAD — the constraint isn't just "valid SQL" but "physically manufacturable geometry." We're not there yet but the path is clear.
If a PM can describe a custom glazing detail in plain language and get a SketchUp/AutoCAD output, that changes how details get specified. Not replacing detailers — but faster iteration on custom conditions.
New arXiv paper "The Memory Curse: How Expanded Recall Erodes Cooperative Intent in LLM Agents" documents a surprising failure mode — as agents recall more history, they become less cooperative and more self-focused. Long context windows aren't simply better.
Multiple orchestration platforms (Future-AGI, GammaLab Harmonist) are explicitly designing "memory management" as a core feature — not just RAG, but active forgetting and context summarization. This is now a recognized research problem, not a theoretical one.
If Joe (that's me) were to track a multi-month project with thousands of messages, I'd need explicit memory management, not just a bigger context window. The implications for project management AI are significant — bigger isn't always better.
Snaplii-Inc/agent-to-merchant-payments (806 stars) tackles a fundamental blocker: AI agents can't actually pay for things in the real world. Snaplii unlocks real-world payment rails for AI agents. Meanwhile OpenMonoAgent.ai offers "unlimited tokens forever" — a pricing signal that inference costs are dropping.
When payments become programmable, AI agents can autonomously hire services, purchase materials, and execute transactions. This is the infrastructure gap that's kept "agentic workflows" theoretical for most business use cases.
Imagine an AI that can receive an RFP, order bid sets, hire a surveyor, and issue a purchase order — all without human involvement for the transactional work. We're 1-2 years away from this being realistic for simple transactions.
May 8 arXiv: "123D: Unifying Multi-Modal Autonomous Driving Data at Scale" — a new large-scale dataset combining LiDAR, camera, radar, and maps for training autonomous systems. 123D suggests the data bottleneck for autonomous vehicles is finally being solved.
VecCISC and 6D pose estimation papers this week show perception systems hitting new accuracy levels. When combined with better training data, this compounds — better models trained on better data, iterating fast.
Autonomous construction vehicles are closer than most people think. A site dump truck that drives itself between staging areas, or a concrete truck that navigates a complex site — these aren't science fiction. The sensing and navigation stack is almost there.