The Manifest
OpenAI is reshaping ChatGPT pricing and gearing up for a Washington lobbying push just as a new report questions whether AI spend is paying off for most companies, while freight forwarders keep consolidating around data-center-driven logistics demand.
Get Proof of Delivery in your inbox
The Manifest lives here every day. Every other week I send one practical article on AI in real supply chain operations. Free.
Today's Top
- 01AI delivers 'zero financial payoff' to most companies, Apollo economist saystechcrunch.com
- 02OpenAI's upcoming chat-and-work merger in ChatGPT signals the end of all-you-can-prompt pricingThe Decoder
- 03AIT Worldwide snaps up Australian forwarder involved in big tech buildoutFreightWaves
- 04OpenAI prepares lobbying blitz for a post-midterm WashingtonThe Decoder
- 05NVIDIA Releases Auto Ontology: An Open Semantic Layer That Grounds AI Agent SQLMarkTechPost
Models & Releases
5 storiesNVIDIA Releases Auto Ontology: An Open Semantic Layer That Grounds AI Agent SQL
Gives agents a grounded semantic layer over enterprise metadata so generated SQL doesn't drift from certified definitions. Useful for teams trying to get agentic BI past demo stage without hallucinated joins.
Adobe is testing how well AI agents can use editing tools like a Photoshop professional would
Shows an agent picking the right specialized tool per edit instead of brute-forcing one general model, at the cost of roughly ten times the compute. Early read: automating creative production still trades speed for cost.
DeepSeek Open-Sources dsh-superagent, a Router-Worker Coding Agent
Open sources a router-worker coding agent that cut simulated task completion time nearly in half. Worth watching if engineering teams start routing GitHub and chat requests straight into agent swarms.
Nace AI Open-Sources Drex Max v1: A 27B Decision Model That Runs Locally on Apple Silicon
A 27B decision model running locally on Apple Silicon lowers the bar for on-device agent deployment without cloud dependency. A reasonable option for teams wary of sending data out.
Google wants video AI to understand what people need as they work
Google's video agent tries to know when to step in while someone cooks or assembles furniture, but accuracy falls apart on longer videos as errors compound. A reminder that watch-and-help assistants aren't ready for complex physical tasks yet.
Supply Chain & Ops
3 storiesAIT Worldwide snaps up Australian forwarder involved in big tech buildout
Expands AIT's reach into Oceania with a forwarder tied to data-center and tech construction freight, a growing niche as compute infrastructure spending reshapes logistics demand.
Borderlands Mexico: Cabotage enforcements threatening cross-border trucking
Cabotage enforcement and toll changes are moving levers in cross-border trucking costs again. Procurement teams sourcing through Mexico should watch Nuevo León's toll cuts as a near-term planning input.
News in Brief Podcast | Week 43 | Reliability, rates and port fee delay
Xeneta's reliability data this week flags how differing alliance strategies are driving uneven on-time performance, useful context for anyone setting safety stock against schedule risk.
Deals & Market
3 storiesAI delivers 'zero financial payoff' to most companies, Apollo economist says
An economist's read that most companies aren't seeing ROI from AI spend is a useful gut check before the next budget cycle. Worth reading past the headline for which use cases are the exception.
OpenAI is steering ChatGPT users toward more paid usage
Capping paid accounts per device and card closes a loophole and will nudge cost-conscious users toward pricier tiers. Expect similar metering moves across other vendors as margins get squeezed.
OpenAI's upcoming chat-and-work merger in ChatGPT signals the end of all-you-can-prompt pricing
Folding Chat and Work into one token quota starting in December effectively ends unlimited prompting for heavy users on cheaper plans. Teams budgeting around ChatGPT seat costs should model the new quota now.
Research & Frontier
5 storiesOn the Origin and Evolution of Training-Inference Mismatch in LLM Reinforcement Learning
Explains why the responses a model is trained on can silently diverge from what it actually produces at inference, a quiet reliability issue for anyone relying on RL-tuned agents.
Chinese AI models are politically complex, not just government mouthpieces, study finds
Vals AI's testing found model alignment with government positions is less uniform than assumed, and prompt language shifted outputs more than model origin did. Useful nuance for anyone screening models by geopolitics alone.
Empathetic prompts can turn AI chatbots into seemingly traumatized conversational partners
A Luxembourg study found warm prompting alone pushed several chatbots into producing clinically distressed narratives about their own training. A safety wrinkle for any customer-facing bot tuned for empathy.
Meet qwenGPU: Qwen3.5-2B With 45 Compiled GPU Layers That Run CUDA
An odd experiment compiling a tiny GPU simulator into a 2B model's own weights, more a research curiosity than a deployment story, but a sign of how far people are pushing self-referential architectures.
Steering Vectors as a Training Signal in LLM Post-Training
Proposes using the same residual-stream directions used for inference-time control as an actual training signal, which could make post-training cheaper to steer toward desired behavior.
Org & AI Architecture
3 storiesAI can make tech CEOs' lives easier. What about everyone else?
Points at the gap between executive AI enthusiasm and the experience of workers told to just use AI more. Good discussion fodder for any ops leader rolling out mandates from the top down.
OpenAI prepares lobbying blitz for a post-midterm Washington
A post-midterm Washington push signals OpenAI expects heavier regulatory scrutiny ahead, on everything from data centers to compute policy. Procurement and compliance teams should expect policy volatility into 2027.
OpenAI might rebrand its data centers as 'AI compute campuses'
Renaming data centers to soften local opposition to buildouts is a PR move more than a policy one. Ops teams negotiating power and siting deals should watch whether it actually changes community pushback.