In partnership with
Welcome to OHMO AI. Bill Gates has a list of three things he wants world leaders to have a plan for, and none of them are about better chatbots. Jobs that don't come back. Criminals who need barely any skill. Kids who never learn to make friends.
The rest of today is hardware and memory. OpenAI published the first benchmarks for a chip it built with Broadcom, Apple pushed more AI compute into the machines on your desk, and Claude started carrying what it knows about you between two products. One man is asking for a plan. Everyone else is shipping.
In today's newsletter:
Bill Gates hands world leaders three AI risks and no comfort on jobs
OpenAI's Jalapeño chip claims more AI work per kilowatt than today's best
Apple's M6 and M5 Ultra push AI compute onto the desktop
Claude's memory now carries between chat and Cowork
LATEST IN AI
BILL GATES
Bill Gates hands world leaders three AI risks
Image source: OHMO AI
OHMO AI: Bill Gates has set out three AI risks he wants world leaders to have a plan for. The other two are criminals using AI against targets of any size, and children whose development gets stunted by an AI companion.
The details:
On crime his point is scale. Even criminals with barely any skill could go after individuals, companies and governments.
He says the useful capabilities and the dangerous ones are hard to pull apart.
The kids risk is the personal one. A father of three, he says AI could stunt children's development and replace human relationships.
He uses himself as the example. He worked hard as a child to build social skills and make friends, and doubts he'd have put in the same work with an AI companion around.
Why it matters: The jobs risk lands on people who haven't started a career yet, since the roles he expects to vanish are the entry-level ones young people use to get in. Nobody reaches a senior job without holding a junior one, so the firm that quietly stops taking graduates is also deciding who it has left to promote. That decision gets made in hiring meetings, not in the rooms he's addressing.
PRESENTED BY ARTISAN
Hire Ava, the AI BDR built for enterprise
Ava is the first AI BDR to run outbound end to end, and you decide whether she runs autonomously or on copilot.
She finds leads or ingests accounts from your CRM, enriches them, sends personalized emails on behalf of your reps, follows up, handles replies and books meetings. Website visitor de-anonymization, intent signals, and a parallel dialer come built in.
A small team can manage her centrally for thousands of reps who never log in. Everything syncs two-way with Salesforce and HubSpot.
Ava runs outbound for companies like DoorDash and Grammarly, and one customer deploys her across 1,000+ reps. Ava is SOC 2 Type II audited, SSO and GDPR ready. Ava is how revenue teams grow pipeline without growing headcount.
OPENAI
OpenAI's Jalapeño chip posts its first numbers
Image source: OHMO AI
OHMO AI: The pitch for Jalapeño is what it does with a kilowatt. OpenAI has put numbers on its inference chip, saying it served more tokens per user and more throughput per kilowatt than the leading inference processors on the market now, and returned responses more quickly.
The details:
Same kilowatt, more AI work served out of it.
Per user, the chip also moved more tokens.
Responses came back more quickly too, so the gain isn't only volume.
Why it matters: Power, not chip supply, is the ceiling on how much AI a data centre can serve, so a claim of more work per kilowatt is a claim about how much OpenAI needs to buy from anyone else. Worth watching is whether the numbers hold when someone outside OpenAI runs the same test, because right now the benchmark and the chip belong to the same company.
APPLE
Apple's new chips keep the AI work on your desk
Image source: OHMO AI
OHMO AI: Apple introduced the M6 and the M5 Ultra, two chips built for AI work that never leaves the machine, headed for a new Mac Mini and a new Mac Studio. The M6 has a 12-core GPU, two more cores than the M5, with a Neural Accelerator in every core. Apple says that delivers nearly a 30 percent increase in peak GPU compute for AI compared to the M5.
The details:
The M5 Ultra is Apple's first quad-die chip, fusing two dual-die M5 Max chips.
Apple calls the M5 Ultra its most powerful chip ever.
Its GPU runs up to 80 cores, each with a Neural Accelerator, for up to 4.5x the peak GPU compute for AI compared to the M3 Ultra.
Every AI gain here is measured against Apple's own last chip. The M6 against the M5, the M5 Ultra against the M3 Ultra.
Why it matters: Local compute means the model runs on the desk instead of in someone's rented cloud, so nothing gets metered and nothing leaves the room. Apple is doing the maths against its own previous silicon, which is the comfortable comparison to pick. Worth watching is whether people start running bigger models at home rather than renting them, because nearly 30 percent more peak GPU compute for AI over the M5 is the difference between a model running on your machine and not running at all.
PRESENTED BY ATTIO
Introducing The First Agentic CRM
Get revenue agents, workflows, and automations across every stage of your motion. Access customer data in real time through Attio's web app, MCP, API, and SDK.
Then Ask Attio anything about your business and get instant answers.
It's the CRM that runs the work behind every win.
ANTHROPIC
Claude now remembers you in both places at once
Image source: OHMO AI
OHMO AI: Anthropic has merged the memory behind Claude's chat and Claude Cowork, its automation agent, so what Claude learns in one turns up in the other. A Cowork task running in the cloud now has what you told the chat, and the chat has what Cowork picked up. Sensitive subject matter is left out of memory by default.
The details:
The target is rebriefing. Agents keep needing to be told things they already know, and this removes one round of that.
It works both ways, chat into Cowork and Cowork back into chat.
Anthropic says the default exclusions cover health data, race, ethnicity, religious beliefs, politics and gender identity.
Cowork runs its tasks in the cloud, so the memory travels with work you aren't sitting in front of.
Why it matters: A memory that follows you into an agent is a memory doing work while you're not watching, which is a different thing from a chatbot recalling your dog's name. The default exclusions are the part worth reading, because they're a default rather than a wall. Worth watching is how people find out what Claude has kept, since the moment memory spans two products, deleting something in one of them stops obviously meaning it's gone from the other.
AI GUIDES
Write ChatGPT prompts like you're handing off to a new intern
Ask ChatGPT a plain question and you get a plain, generic answer, then you spend ten minutes fixing it or asking again. Give it the same background, task, and standards you'd give a smart new intern on day one, and you get something close to final on the first try. Costs you ten extra seconds of typing, works in any chatbot, nothing to install.
Copy this, swap in the bracketed parts, and paste it in.
Context: [who you are, what this is for, and any background someone outside your role would need to understand it — e.g. "I'm a program manager sending a weekly status update to a director who has no time for detail"]
Instructions: [the one thing you want done — e.g. "Write a 300-word summary of this week's progress on the [project name] rollout"]
Constraints: [tone, length, format, and what to include or leave out — e.g. "Under 300 words. Plain language, no jargon. End with a one-line note on what's due next week. Do not include numbers that haven't been confirmed."]Context exists because ChatGPT has no memory of your job, your project, or your audience, so without it every answer defaults to generic. Instructions work best when they ask for exactly one output, since a prompt that requests five things at once gets you five half-finished things. Constraints do the editing job upfront, telling it the tone, length, and format you want so you're not rewriting the answer afterward. A good test before you send any prompt: if a stranger outside your job read it, would they know what you want without you explaining further? If not, that's the part to fill in. Treat your first attempt as a draft, see what comes back, and adjust the wording from there.
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 Marek W. from Gdańsk, Poland:
“My grandad went deaf in one ear a couple of years back and he stopped answering the phone. He'd never say he was lonely, he'd say the phone was broken, that was it. So I put an old tablet on the table by his chair that transcribes his calls using AI live in big text and every Sunday me and my mum go through the words it got wrong, mostly names and the neighbour's dog. It's maybe 8 out of 10 on a bad line, worse if the other person mumbles. Last week he was on to his brother in Szczecin for an hour and a half. My mum heard him laughing through the wall and came and stood in the doorway and didn't say anything.”
Want to be in the next newsletter? Tell us how you use AI here.
AI Meme
Go Deeper
AI Tools: A curated list of the AI tools worth knowing.
Instagram: The AI page they pray you’ll never find.
Facebook: For the ones who never left Facebook.
