Open inference compilation, proof certificates as deliverables, personal agent twins: today's openings all sit on turning heavy work into reusable assets.

OpenAI is expanding ChatGPT Ads globally, with a $1B annualized revenue run-rate and free plus low-cost options. The direct implication: the top vendor is itself lowering the entry barrier, so application-layer pricing and features get squeezed from both sides. The window for telling an "I can call a model too" story is narrowing.

For engineering teams deploying inference on non-NVIDIA accelerators. They previously had to hand-write and maintain a kernel compilation pipeline per xPU; DeepJIT takes kernel code and performs just-in-time compilation at runtime, producing directly callable compiled results. This is a direct product of the trend where model vendors open-source the compilation layer of their inference stacks, and the wedge sits in inference deployment services for domestic or in-house accelerators. The supported hardware list and delivery form still need verification, so it stays on watch.

Mathematics researchers and formal verification engineers checking proofs in partial differential equations and fluid mechanics open this public code repository so the Lean proof assistant can ingest certificates shipped alongside paper results and check each derivation step, yielding a machine-checkable pass or fail. The trend is that frontier model teams now ship formal proof certificates as public research deliverables rather than paper appendices; the wedge is in domains where being wrong is expensive — mathematics, cryptography, chip verification.

Public evidence is not yet sufficient to confirm concrete workflow value for these three, so they are uniformly on watch. The bar is the same: validate sustained use in a real workflow before deciding whether the opportunity merits investment. mythos-5-incident-transcript appears incident-record related by name, but without a verifiable delivery form, no further inference is drawn.

1mil lets founders input concept text during early ideation; the AI scores ideas against predefined criteria and returns a structured scorecard, with feasibility still left to the founder — the wedge is a real pain-point library for a vertical, not generic algorithmic scoring. 3dviz-pro-max turns natural-language descriptions into runnable Three.js or Blender code and scene assets, a fit for vertical knowledge bases in 3D e-commerce display or data dashboards. agent-me distills personal knowledge, memories and decisions into an open-source, inspectable agent twin; personal memory is becoming the retention layer of the agent stack, but skip the generic twin and go where compliance matters.

Vercel Sandbox now offers 64 GB of storage, Copilot is available in the AI SDK harness layer, Amazon SageMaker Inference added prefix-aware routing, HyperPod supports model caching to cut cold starts, Bedrock Knowledge Base added video and image search via Marengo 3.0, and Amazon Quick is generally available on desktop. All of it pushes the cost of "just running it" down. The cheaper the infrastructure, the more differentiation can only come from understanding a specific scenario.