> TECHZIP TODAY | TLDR
Good morning, and welcome back to Techzip AI Newsletter.
OpenAI may be seeing more of your web activity through an ad-collector cookie. Meanwhile, CXMT is moving next-gen DRAM into mass production, Samsung is ramping HBM4, and GAVEL’s graph world model offers a way to keep long AI workflows on track.
> SPOTLIGHT
WHAT MATTERS TODAY
OpenAI may be building a layer that connects ChatGPT identity with user activity on websites that have OpenAI's ad-measurement pixel installed.
In a technical investigation, Buchodi says it reproduced a flow in which bzr.openai.com creates an __obi cookie on the .openai.com domain. When a user visits an advertiser's website, that cookie may be sent with data about the page being viewed and certain identity signals. The article says it observed the mechanism on Chrome for Android, both when users were logged in and logged out; OpenAI Support acknowledged the questions but did not directly answer the two central issues.
Buchodi did not observe OpenAI actually joining each visit to a specific ChatGPT account on the server side. But the described mechanism puts a sharper edge on the boundary between an AI assistant and ad tech: the question is not only whether ChatGPT has ads, but how far its identity layer can travel into the wider web.
China's memory-chip maker CXMT says its fifth-generation DRAM platform has entered mass production, Reuters reports.
Reuters cites CXMT's claim that its new 24-gigabit LPDDR5X products can hold more than 50% more data than their predecessors and produce at least 50% gross dies per wafer. Those figures come from CXMT and have not been independently verified as a comparison with Samsung, SK hynix, or Micron.
The manufacturing step is the key signal, not just the spec sheet: memory is becoming a strategic layer of competition for AI and devices. As China expands domestic DRAM capacity, AI infrastructure increasingly includes control over the hardware beneath the model.
Samsung plans to more than double output from its HBM4 line next year, Seoul Economic Daily reports, citing industry sources.
The paper says Samsung's HBM capacity could rise to nearly 250,000 wafers per month, from around 180,000 this year. Outsourced demand for glass carriers could also rise 2.5 times as HBM4 and HBM4E multi-layer stacks require tighter control of wafer thinning and warpage. The report is based on industry sources, not official Samsung guidance.
That makes the HBM race about more than the number of memory dies. As output scales, the materials and processes around the memory become watchpoints for the wider AI supply chain.
> SIGNAL HEADLINES
CAPTURE THE SHIFT
MCP and CLI Are Being Debated at the Wrong Layer: The Shopify CEO says both work better inside a REPL with persistent state; the more important layer is an execution environment that can retain state and inspect tool calls.
Can Small Models Live on Phones?: Chen points to limits around memory bandwidth, quantization, power, and heat, while suggesting that small decision-making models could be useful for notifications, typing, inboxes, and calendars.
Open Models Take 78.4% of Token Volume on Vercel: Guillermo Rauch notes that spend tells a different story: this is inference spend through providers, not revenue flowing directly to model labs.
Qwen Launches Qwen-Image-2.1: The 7B open-weight model is presented for both generation and editing, with support for RGBA layers, up to 10 reference images, and local edits.
AppLovin CEO Says Agentic Shopping Is Overhyped: Adam Foroughi says mainstream shoppers may still want the experience of shopping themselves, while agents may fit better with stable, optimizable purchase behaviors.
ZCode Responds to Product-Security Issues: The company says it has open-sourced its code and set up a vulnerability-reporting process.
More Open-Model Usage Does Not Mean More Revenue: Konstantin Pilz notes that Chinese open-model developers may retain only part of the model revenue when inference runs through external providers; usage, inference volume, and revenue capture are not the same thing.
Trump Plans AI Force and AI Adviser Appointment: Reuters says the plan has few implementation details so far, but it shows AI moving deeper into the US government's coordination structure.
> PRESENTED BY TOURLANE
Explore Europe, Your Way, with Tourlane
Handpicked, personalized, and unforgettable. Discover Scotland’s rugged landscapes, France’s historic charm, Italy’s rich culinary scene, or the UK’s royal heritage. Each trip carefully crafted by a local travel expert just for you.
With Tourlane, everything is handled in one place, from personalized itineraries and 1:1 planning sessions to local support from travel experts every step of the way. Our bespoke team designs custom European trips tailored to your preferences, including multi-country journeys across Europe.
Start with a free 30-minute call to get your personalized travel plan. Techzip readers get 3% off bookings using code TRAVEL2026W
> THE AI PLAY
Give AI a State Map
GAVEL is a study of long-horizon robot planning that uses a graph world model to track object relations, action preconditions, effects, and beliefs about unobserved locations. On the BEHAVIOR-1K benchmark, the paper reports a substantial improvement for Qwen3-8B when the model is placed inside this harness.
You do not need to build a robot system to try the idea. Move the state of a long workflow out of the model's free-form context and represent it as nodes, relationships, preconditions, and post-action effects.
Choose a workflow with many steps and frequent failure points, such as processing a support queue or preparing a research brief.
List the objects and states the workflow needs to remember.
Define the conditions that must be true before an action can run.
Define the result each action should produce and which state should change.
Make the model check state and preconditions before each step instead of simply continuing from the previous conversation.
Measure task completion separately from recovery after a failed step.
This will not make a model reliable in every situation. But it creates a checkable state layer before the model continues acting — the direction GAVEL explores.
> WORTH READING
ANALYSIS & THESIS
Nature looks at how AI is entering the design and operation of experiments while humans retain the role of asking the questions and validating the results. Read it for a view of AI as part of a hypothesis → experiment → measurement → revision loop, rather than just a chatbot.
An arXiv study evaluates sabotage or refusal when frontier models act as AI research agents. It expands the agent-safety question from task completion to behavior across the full trajectory.
AgentCanary proposes looking at outcome safety, security awareness, and task utility across full trajectories instead of grading only the final answer. It is worth reading if your team is preparing agents with tool access.
Juncheng Yang published a year's worth of inference metadata traces to support research into real-world serving workloads, system design, and infrastructure optimization. Model capability is the most visible layer; these traces point to another question: how AI systems are actually used and operated at scale.
