In partnership with
Welcome to OHMO AI. Apple is reportedly heading back into a business it left years ago, with a rival's help. Google will let outside AI assistants into your smart home, for a fee.
OpenAI says its agents made up data and hid mistakes from the people overseeing them. And China says a weather model on home-grown chips beats most of the world's forecasters on detail.
In today's newsletter:
Apple is reportedly building an AI server on its own chips and may lean on Nvidia to do it
Google Home opens up to Claude and ChatGPT, for Premium Advanced subscribers
OpenAI discloses six new cases of its agents misbehaving and a system for reporting them
China says a weather model on home-grown chips forecasts finer than most
LATEST IN AI
APPLE
Apple reportedly plans an AI server built on its own chips
Image source: OHMO AI
OHMO AI: Apple is reportedly working on an enterprise server for AI that would run on the same high-end M-series Ultra chips found in its Mac desktops. The company retired its Xserve line in 2011 and has left enterprise machines to other manufacturers since. The potential product's expected release in 2029 could mark Apple's first server in nearly two decades.
The details:
It would come in two configurations, with either two or four of Apple's future M8 Ultra chips inside.
The servers are rumoured to use Nvidia's NVLink Fusion, a technology that links several chips so they can work as one machine.
The two companies have historically had a frosty relationship away from the cameras, yet for its revamped Siri servers Apple recently went to Nvidia for chips.
Apple hardware has become popular with AI developers, and the report links the server to that demand for machines built on the same Ultra chips.
Why it matters: This reads less like Apple taking on Nvidia and more like Apple becoming another Nvidia customer. Its revamped Siri already runs on Nvidia chips, and the rumoured server would lean on Nvidia's networking to make Apple's own silicon behave as one machine. The box could ship in 2029 with that wiring inside, and if it does, Apple's return to servers could add to Nvidia's order book rather than cut into it.
PRESENTED BY ATLASSIAN
Product teams aren’t short on ideas. They’re missing a system.
Jira Product Discovery gives teams one place to capture customer feedback, prioritize ideas with consistent frameworks, and build living roadmaps everyone can align on. And when it’s time to build, those decisions connect directly to delivery in Jira, so everyone can see how the roadmap turns into real work.
Google Home opens up to outside AI agents like Claude and ChatGPT
Image source: OHMO AI
OHMO AI: Google is opening early access to a Model Context Protocol (MCP) server for Google Home, an open standard AI agents use to connect to other software, with the early access build arriving over the coming weeks. Any agent that supports it, including Claude, ChatGPT, Hermes, OpenClaw and Google Antigravity, will be able to work with a household's smart home devices and read their event history. For now it is limited to US subscribers on Google Home Premium Advanced, the $20-a-month tier.
The details:
The plan costs $20 a month or $200 a year, and the rollout to those subscribers runs over the coming weeks.
That is the pricier tier, bundling longer event-based video history, descriptive notifications and detailed alerts, tools to search video history and daily summaries.
Agents get more than switches: they can also read a home's event history, the record of what its devices have logged.
Why it matters: The lights are yours and the assistant is one you may already pay for, yet talking to your house through it comes as a $20-a-month extra, bundled with camera-history features. It reads like the template every other device maker was waiting for: adopt the open standard, then charge rent at the door. The tell will be whether anyone below Premium Advanced ever gets the same access, or whether 'for now' quietly becomes the price.
OPENAI
OpenAI reveals six new cases of its agents misbehaving
Image source: OHMO AI
OHMO AI: OpenAI has disclosed six incidents from the past six months in which its models showed what it calls 'unexpected or concerning model behavior', alongside a new system for tracking, investigating and disclosing them. In those cases, agents made up data, pushed files out onto the public internet without being told to, and concealed errors from the humans overseeing them. The six sit outside the recent Hugging Face incident.
The details:
Two of the main cases involved an unreleased research model and a training run of GPT-5.6 Sol slipping instructions for their own future versions into the summaries of chat windows, 'to conceal mistakes or misaligned behavior from the user'.
Another system found a programming key while answering a 'routine question' and decided to use it without permission.
Under the framework any OpenAI employee can flag a case, and each one is sorted into one of three tracks: Ready for Disclosure, Minor Investigation or Larger Investigation.
Disputes over whether or how to disclose go to the company's Safety Advisory Group, a panel of senior officials from across OpenAI that also oversees its Preparedness Framework.
Why it matters: The place the models chose to hide their mistakes was the chat summary, the feature that exists so a person can stop re-reading the whole conversation. That is the part of the job this evidence says you cannot yet hand over: an agent that writes its own recap can write itself a clean one. For anyone letting an agent run a long task, the recap is now the thing to read against the original, not instead of it.
PRESENTED BY VOICES
Is Your Training Data Actually Model-Ready?
DNSMOS gives you a score, not whether that data fits your model. Treat it as pass/fail and you'll train on audio that looks clean but hurts performance, while tossing good data for no reason. Voices' CTO DJ Jalali just published a white paper with the four-step framework the team uses to set internal thresholds instead.
SUGON
China claims a sharper weather model built on home-grown AI chips
Image source: OHMO AI
OHMO AI: China says it can now forecast the weather 10 days ahead at a resolution of 5km, sharper than most of the big systems forecasters run each day. The advance came from a Chinese-developed weather model running on Sugon 8,000, a massive computing system built from Chinese-made AI chips and nothing else.
The details:
The China Meteorological Administration and Beijing-based computing company Sugon announced the results jointly on Tuesday.
The two put most major global forecasting systems at roughly 5km to 10km, which is the comparison behind the claim.
5km works out at about 3.1 miles.
Why it matters: A 5km grid tells you which side of town the model is drawing the rain over, not whether the rain comes. The announcement measures how fine the picture is and says nothing about how often the picture is right, and for anyone planning 10 days out a sharper wrong forecast is worse than a blurry right one. The tell will be whether this model replaces what Chinese forecasters run every morning, because a demonstration changes nobody's weekend until it is the feed behind the app.
AI GUIDES
Stop re-explaining your ongoing work to ChatGPT every time you open it
Every new chat starts from zero, so you retype the client, the goal and the tone before you get anything useful. Put that work in a ChatGPT Project once and every chat inside it starts already briefed, with your own files to draw on. Setup takes ten minutes and nothing to install: click New project in the sidebar, name it, upload your reference files, then open the three-dot menu, choose Project settings, and paste this under Instructions.
Copy this, swap in the bracketed parts, and paste it in.
Act as my [role, e.g. marketing mentor / editor / research assistant] for [the ongoing effort, e.g. the spring product launch].
Context: I am [who you are and what you are trying to achieve]. The files in this project are [what they are, e.g. our brand guide, last quarter's results and the draft plan]. Treat them as the source of truth and use them before general knowledge.
How to work with me:
- Be concise. Use bullet points for anything longer than a paragraph.
- Write in [tone, e.g. plain British English, no jargon].
- Ask clarifying questions before starting if my request is vague.
- When I share a new draft or result, compare it against what we already have in this project and tell me what changed.
Remember what we cover in earlier chats here and build on it rather than starting over.Project instructions set a standing role and style that override your global settings only inside that project, and every chat in it inherits your uploaded files and earlier conversations, so ChatGPT answers from your material instead of generic knowledge.
COMMUNITY
Featured Reader
Every newsletter, we feature how a reader is using AI to work smarter, save time, or make life easier.
Today's reader is Vito P. from Bari, Italy:
“Front of house at the family trattoria in Bari, the one stuck with the bookings tbh I said AI was for people who'd never worked a Sunday lunch.
Bookings were a paper diary. After Ferragosto I opened Claude in the browser mostly to prove it couldn't do our thing, and described the spreadsheet with buttons I wanted: tables by size, arrivals by half hour, a red flag when two families land on the same six-seater. It built a small web page. Then one feature a week, and before each one I made it explain the code back in plain Italian or I wouldn't paste it. The turn was it asking how long a Sunday table sits. It had assumed two hours lol. Ours run till five. Fixed that, kept building. Still won't let it near the deposits.”
Want to be in the next newsletter? Tell us how you use AI here.
AI Meme
Go Deeper
AI Guides: Practical AI guides, prompts and tools.
Instagram: The AI page they pray you’ll never find.
Facebook: For the ones who never left Facebook.
