OpenAI's device leaks, plus two major breakthroughs

Welcome back, Superhuman. OpenAI reportedly plans to take on Amazon and Google with its first hardware product: a portable smart speaker that will double as an AI companion. Plus, two labs revealed major breakthroughs for on-device and self-improving AI.

Today: Prism’s on-device model, how to turn ChatGPT into your inbox assistant, and get the latest prompts and trending social posts.

1. PrismML reveals the largest-ever AI model to run on an iPhone: The AI lab PrismML released Bonsai 27B, a compressed version of Alibaba’s open-source Qwen3.6 27B that it shrunk from 54GB to under 4GB — capable of running on an iPhone 17 Pro or newer. PrismML CEO Babak Hassibi told CNBC that Apple is evaluating the technology, though it’s unclear where that will lead. Still, it’s a notable step toward running capable AI models directly on personal devices. Check out a demo.

2. IBM CEO says the AI infrastructure buildout is hurting sales: IBM’s stock just had its worst day on record, dropping 25% after the company warned that clients are shifting spending toward AI hardware like servers, storage, and memory. CEO Arvind Krishna told investors "We did not anticipate the magnitude of the capex reprioritization.” The selloff rippled across the sector, dragging down Oracle, Salesforce, and ServiceNow as well.

3. Weco AI claims first experimental evidence of recursive self-improvement: AI lab Weco AI published results from AIDE², a system that spent eight unattended days rewriting its own code, creating an agent that outperformed Weco’s own hand-tuned agent that it spent two years refining. The lab frames it as the first experimental evidence of an AI system improving its own capabilities autonomously. Read the full blog post.

We’ve unlocked WSJ content for you: OpenAI’s growing challenges are narrowing its IPO window.

Wispr Flow works inside every AI tool you use. Dictate detailed prompts 4x faster than typing.

Flow preserves camelCase, snake_case, and technical terms. Better input, better output. Used by millions of developers and AI professionals.

Try Flow at no cost.

The viral essay. Google DeepMind CEO Demis Hassabis doesn’t post often — so when he does, the internet pays attention. In a new essay, A Framework for Frontier AI and the Dawning of a New Age, he argues that we’re standing at “the foothills of the singularity” and believes our decisions over the coming months will determine the next phase of civilization.

Navigating AGI. The essay hinges on the pending arrival of artificial general intelligence (AGI), or an AI system that rivals the human brain. Hassabis argues this breakthrough (if it materializes) would rival the invention of fire or electricity. He says it could usher in “unprecedented” positive change, but admits urgent action is needed.

Hassabis’ plan. The Nobel laureate wants to establish a self-regulating Standards Body modeled after FINRA — the private-industry watchdog that polices Wall Street. The goal: work alongside private labs and federal agencies to track AI’s rapid progress and mitigate risk without stifling innovation. Specifically, vetting new models before their public release.

The timeline. Hassabis believes AGI is “probably” only a few years away and wants to get an organization in place by the end of the year. That said, AGI remains a hypothetical milestone. No lab has achieved it, and timelines like Hassabis’ are still predictions, not certainties.

Sample Prompt: Can you check my @gmail, figure out what I need to respond to, and write drafts in my voice. Use my recent sent replies or @google-drive [writing examples] for tone. Use @slack, @google-drive, or other sources where my work happens when the email is missing the latest decision, owner, file, or blocker.

Bolting a general-purpose search API onto your LLM works fine in a demo—then six months into production, your team is maintaining a retrieval layer instead of shipping features. This guide from You. com breaks down the Web Search, Research, and Finance Research APIs—and where each one fits in a production AI stack.

Stop treating research as simple retrieval. Download the guide.

📝 Prompting Tip: A Sully AI researcher says long-horizon-prompting generates better results on complex tasks. The strategy involves handing your AI a pseudo-formal task brief that defines what success looks like — see the prompt outline here.

🤑 Finding Money: The founder of Ghostty used ChatGPT to find $45K in overlooked invoices that had gotten buried in his emails over a 3-year span. The post breaks the exact strategy he used.

👕 Personal Stylist: An OpenAI employee built a tool that generates unique outfit ideas based on clothes you already own, with one commenter calling it one of the coolest personal use cases for AI so far (4M+ views). Try the skill here.

🍎 Claude for Teachers: Anthropic is giving all verified K-12 educators free access to premium Claude plans, along with plenty of other benefits.

🖌 AI Design: This viral post teaches you how to actually design with AI using 3 different strategies: design skills, going piece-by-piece, or using a reference as inspiration (600K views).

^{* indicates a promoted tool, if any}

Thousands of people are sharing passion projects they’ve built with AI in response to a request from OpenAI CEO Sam Altman (5M views and 3K responses).

Grow customers & revenue: Join companies like Amazon, HubSpot, and Salesforce. Showcase your product to our 1M+ readers and 2M+ followers on socials. Get in touch.

Until next time — Zain, Theodore, & the Superhuman AI team.