☀️ TRENDING AI NEWS 🏢 Apple Agent: Apple is hiring an engineer to lead a new agentic AI assistant that can browse, code and work across apps. 🤖 Meta FAIR Memory: Self-Improving Memory scores 56.0 on Continual Learning Bench versus 44.6 for the best hand-written harness. 🛠️ Perplexity API: The new Orchestrator API routes OpenAI, Anthropic, Google and xAI models through one endpoint with failover. 🚨 Defense AI: A 16-year-old Y Combinator founder is building missiles optimized for 'cost per kill'. |
Everyone spent the week watching Anthropic ring a bell on Nasdaq. Meanwhile, quietly, the biggest names in tech kept shipping the less glamorous stuff: agents, memory and the plumbing that holds them together.
🤓 AI Trivia
What does the 'A2A' in agent gateway talk like MCP and A2A stand for?
The answer is hiding near the bottom of today's newsletter... keep scrolling. 👇

| 🏢 Apple Wants Its Own AI Agent | |

A job listing reveals the plan
Apple is working on an AI agent of its own. A job listing spotted by Bloomberg says the company is looking for an engineer to lead the creation of an agentic AI assistant that can browse the web, work across apps, write code and call services. That puts the iPhone maker in the running against rivals like OpenAI and Anthropic.
Timing is interesting. Just this week, Google gave Gemini on Mac the ability to use your computer, with you choosing which apps it can touch. Apple is clearly not content to leave the agent layer on its own platform to someone else.
The bottom line
A job post is not a product, but it confirms Apple sees agents as core to the next phase of its devices. Expect this to shape how much your iPhone can do on its own.
Get Paid for Data You Already Own
Your business has years of emails, documents, CRM records and call logs. AI researchers need exactly that kind of real-world data, and Mode will pay for it.
You sell a de-identified copy. Your original records and systems remain yours.
No broker fees. Mode buys directly.
You approve what data gets shared.
You never share your passwords.
Sensitive information is removed before delivery.
It takes just a few hours of your team’s time.
Depending on data volume and a full review, companies with several years of records could see 100K–5M.
See if your company qualifies and get a quote.

| 🔬 Meta Lets a Model Write Its Own Memory Prompts | |

Self-improving memory beats hand-tuning
Meta FAIR's Self-Improving Memory setup lets Kimi-K2.6 write its own LOAD and OFFLOAD memory prompts, then distills them into its weights. The optimized prompts score 56.0 on Continual Learning Bench versus 44.6 for the best hand-written harness.
That is a big jump from simply letting the model decide what to remember and what to push out of context. For anyone building long-running AI agents, memory management is one of the stubborn bottlenecks, and this is a clean example of automation beating human prompt craft.
The bottom line
If models can tune their own memory better than people can, hand-written harnesses start to look like a stopgap. Worth watching if you build agents for a living.

| 🛠️ Google Built a Fake Internet to Train Chrome's Agents | |

A sandbox where agents can't go rogue
Google built an alternate-universe copy of the web, a massive sandbox the Chrome team uses to train and test AI browsing agents without the risk of them going rogue on real sites. Think of it as a flight simulator for agents that click, scroll and fill in forms.
It lands the same week Google put AI chats right on the Search homepage, with a collapsible sidebar that lets you continue past AI conversations. The company is pushing agents and chat deeper into products people already use every day.
The bottom line
Testing browsing agents against a safe replica of the web is the responsible way to do this. It also hints at how seriously Google takes the risk of agents misbehaving in the wild.

| 🚨 A 16-Year-Old Is Building AI Missiles | |

Targeting 'cost per kill'
Henrijs Dandzbergs dropped out of high school to join Y Combinator. His startup, Polemos, is building a new breed of missile aimed at making killing as efficient as possible for the US and its military allies, with 'cost per kill' as the metric he is chasing.
It is a stark example of how quickly military AI is moving from big contractors to venture-backed startups run by teenagers. The ethics debate is only getting louder.
The bottom line
Defense tech is now a startup category, complete with accelerator pitches and efficiency metrics. Whether that is progress or a problem depends on who you ask, but it is not going away.

| 🛠️ Perplexity's Orchestrator API Routes Every Major Model | |

One endpoint, built-in failover
Perplexity launched an Orchestrator API that serves OpenAI, Anthropic, Google, xAI and open-weight models through one endpoint, with same-model provider failover and unified billing. A preview perplexity/auto router sends each request to a route you describe.
With outages rattling digital life all week, failover is a very timely feature. If you are weighing gateways more broadly, MarkTechPost also ranked 13 agent gateways for MCP, A2A and LLM traffic, and if you need a quick landing page to showcase whatever you build, our sponsor 60sec.site is an AI website builder that gets you live in seconds.
The bottom line
Model routing and failover are becoming table stakes for production AI. If you build on a single provider, this kind of abstraction is worth a look.

| 🛠️ Tool of the day | |
Gamma - Gamma turns a short prompt or outline into a full slide deck, document or web page. It helps anyone who needs a clean presentation fast, and it has a free plan.
| ⚡ Quick hits | |
Citi and other big banks now say customers are responsible for losses if a third-party AI agent they authorize leaks their credentials or is hacked. Read more →
OpenAI says its models improved chip designs, including its Jalapeño inference chip, and it is partnering with Synopsys on a chip-design model. Read more →
Palo Alto Networks agreed to buy Applied Compute, a platform for training and running open-weight AI models, in a $4.5 billion all-stock deal. Read more →
🌎 Trivia Reveal
The answer is Agent-to-Agent! A2A refers to protocols that let AI agents talk to and coordinate with one another, and agent gateways now govern that traffic alongside model calls and MCP tool access.
💬 Quick Question
Would you trust an Apple-built AI agent to act across your apps, or would you keep it on a short leash? Hit reply and tell me - I read every response!
That's it for today. For more AI news every day, visit dailyinference.com. See you tomorrow!
| 🎁 Share Daily Inference | |
Know someone who loves AI news? Share your link. You get the AI Tools Starter Kit after 1 referral and the AI Insider Briefing after 5.


