Presented by
Good morning, AI enthusiasts. Remember that scene in Her where the AI knew your whole life because it had been listening all along? Two Harvard dropouts just made that real.
Mira's new AI glasses capture every conversation, building a personalized AI from the whole thing. Whether the world is ready for always-on recording in glasses form is about to be tested.
In today's recap:
MIRA launches AI glasses that capture everything
GPT-5.5 cracks a 12-hour cyberattack in 10 minutes
Build an infographic with ChatGPT Image 2.0
AI outperforms doctors, detects cancer years early
4 new AI tools, news, and more
PRESENTED BY VIKTOR
Last week Viktor wrote a brief, built a landing page, and opened a pull request.
Last week, Viktor wrote a campaign brief, built a landing page, opened a pull request, generated a board-ready PDF from live Stripe data, and sent a follow-up email to a churned customer. All from Slack. Same colleague that also pulls your reports and monitors your dashboards. 5,700+ teams. 3,000+ integrations.
AI HARDWARE
MIRA launches AI glasses that never forget
Mira
Recaply: Mira, the startup co-founded by Harvard dropouts AnhPhu Nguyen and Caine Ardayfio, just opened orders for AI glasses that capture every conversation and build a personalized AI from the wearer's entire life context.
Key details:
Mira's glasses use an always-on microphone to listen and transcribe conversations in real time, with an AI layer surfacing relevant information to the wearer, giving what co-founder Ardayfio describes as "infinite memory" without requiring any active prompting.
The company shipped its first product in January 2026, founded after Nguyen and Ardayfio went viral for demonstrating facial recognition on Meta's Ray-Ban glasses to identify strangers in public, which led to their departure from Harvard.
Mira positions itself directly against Meta's smart glasses, arguing Meta won't build a truly always-on recording device due to privacy reputational risk, leaving the market open to a startup willing to move first.
Orders are now open for the latest Mira model, backed by General Catalyst, with the founders describing it as "the most personalized form of AI ever."
Why it matters: The wearable AI category has been waiting for a product that goes beyond voice commands. Mira's bet is that continuous capture, not on-demand queries, is the real unlock for personal AI. The privacy questions are real and the founders have already lived through viral controversy. But for builders who want an AI that actually knows their life, the product solves something no phone assistant can, which is memory without friction. Whether the market is ready for always-on recording in glasses form is the only question that matters now.
OPENAI
GPT-5.5 cracks cyberattack in 10 minutes
AISI
Recaply: OpenAI's GPT-5.5 just cleared a multi-step cyberattack simulation in 10 minutes and 22 seconds, a task that took a human expert 12 hours, costing just $1.73 in API usage.
Key details:
The task was a custom Rust virtual machine reverse-engineering challenge, requiring GPT-5.5 to recover the VM's instruction set, disassemble its bytecode, solve a cryptographic password check, and submit the flag, with no human assistance and only Bash and Python tools in a Kali Linux container.
GPT-5.5 achieved a 71.4% pass rate on AISI's Expert-level cybersecurity tasks, compared to 68.6% for Anthropic's Claude Mythos Preview, 52.4% for GPT-5.4, and 48.6% for Anthropic's Opus 4.7.
GPT-5.5 is only the second frontier model to solve AISI's corporate network attack simulation end-to-end; Anthropic's Claude Mythos Preview was the first, completing a task estimated at 20 human hours.
AISI published the evaluation on April 30, 2026, using a basic ReAct agent scaffold; the human expert used professional tools including Binary Ninja, gdb, Python, and Z3 to solve the same challenge.
Why it matters: People have debated whether AI cyberattack tests reflect real threats or just lab games. AISI's evaluation was built with real cybersecurity firms. It's not a controlled toy problem. GPT-5.5 broke open a custom virtual machine, decoded its hidden logic, and cracked a password, with no help, in 10 minutes. The $1.73 cost is what makes this story different. When a skilled cyberattack can be run for less than two dollars, the math for defenders changes fast. And it won't get cheaper slowly.
GUIDES
Build a shareable infographic in 10 minutes using ChatGPT Image 2.0
Recaply: In this tutorial, you will learn how to build a finished, shareable infographic using ChatGPT Image 2.0, skipping design tools entirely and going from idea to export-ready visual in under 10 minutes.
Step-by-step:
Open ChatGPT (Pro required for Image 2.0) and pick a topic with 3-7 data points or steps, for example "AI vs human: key differences in 5 areas" or "5 practical uses of GPT-5.5 this week." The tighter the scope, the cleaner the output.
Write a structured prompt that includes four things: the layout type (vertical list, side-by-side comparison, or numbered steps), a color palette ("dark navy background, white text, orange accents"), a font style ("bold headers, clean sans-serif body"), and the exact text strings you want displayed, word for word.
Set the aspect ratio in your prompt before generating: 1:1 for social posts, 4:5 for Instagram feed, 16:9 for slide decks, or 9:16 for stories. Add "flat vector style, minimal infographic" to stop the model from generating photorealistic output.
Evaluate the first version: check that the headline is the largest element, body text is legible, and sections are clearly separated. Then iterate with precise follow-up prompts: "make the headline 30% larger," "change background to dark navy," "add a small icon left of each bullet," "left-align all text."
Right-click the finished image in ChatGPT and save as PNG. Drop it directly into a social post, newsletter, or slide. Save your structured prompt as a template and swap in new data points each time you need a fresh visual.
Pro tip: Add "no gradients, solid fills only" to your prompt if the infographic will appear on different backgrounds. Solid fills look clean in both light and dark contexts and screenshot better on mobile.
AI RESEARCH
AI outperforms doctors, detects cancer early
Mayo Clinic
Recaply: Harvard researchers just published a landmark Science study finding AI outperforms emergency physicians at diagnosis, while Mayo Clinic separately validated an AI that detects pancreatic cancer up to 3 years before clinical diagnosis on routine CT scans, both published the same week.
Key details:
OpenAI's o1 preview was tested on 76 Boston ER cases, evaluated by two physicians who didn't know whether assessments came from AI or expert doctors, matching or exceeding expert performance across all three triage stages, with particular strength when the least information is available.
In challenging cases from Massachusetts General Hospital published in The New England Journal of Medicine, AI achieved near-optimal performance; Harvard senior co-author Arjun Manrai said the results "shocked a lot of folks," noting AI's performance on complex diagnostic benchmarks dating back to 1959.
Mayo's REDMOD model identified 73% of prediagnostic pancreatic cancers at a median of 16 months before clinical diagnosis, nearly double specialists' detection rate without AI, and nearly 3x more early cancers in scans taken more than 2 years before diagnosis.
Both studies were published April 29-30, 2026, in Science and Gut respectively, with Mayo advancing the work into prospective clinical testing through its AI-PACED study across multiple institutions.
Why it matters: AI in healthcare has been a long time coming. These two studies are different from the hype. They're peer-reviewed, published in top journals, and run by institutions with no product to sell. The pancreatic cancer story hits hardest. More than 85% of patients are diagnosed too late, with a 5-year survival rate below 15%. REDMOD runs on CT scans that are already being taken. It doesn't need new machines or new workflows. It just needs to run in the background. Two major institutions publishing results like this on the same day means something has changed.
TOOLS
Trending AI Tools
🗂️ Mesa - A versioned, POSIX-compatible filesystem for AI agents
🔬 Silico - GoodfireAI's platform for building AI models
☁️ Manus Cloud Compute - Manus's dedicated cloud machine that runs bots, scripts, and automations 24/7
🕶️ MIRA - AI glasses with an always-on microphone that capture every conversation and build a personalized AI from your life context
NEWS
What Matters in AI Right Now?
Elon Musk confirmed under oath in a California federal court that xAI used distillation techniques on OpenAI models to train Grok, saying it was a "general practice" among AI companies.
Biohub announced the Virtual Biology Initiative, a five-year effort to create the open data foundation needed to build predictive models of the human cell, with a $500M commitment from the Zuckerberg-backed nonprofit.
Anthropic launched Claude Security in public beta for Claude Enterprise customers, an AI-powered tool that scans codebases, validates findings, and suggests patches that users can review and approve.
MIT genetic engineer Kevin Esvelt shared transcripts with The New York Times in which ChatGPT explained how to spread biological payloads via weather balloon, Gemini ranked pathogens by agricultural damage potential, and Claude produced a recipe for a novel toxin.
Standard Intelligence raised $75M from Sequoia and Spark Capital, with angels including Andrej Karpathy, Stanley Druckenmiller, and Milan Kovac. The 6-person San Francisco team is building FDM-1, a computer-use model series targeting superhuman performance on general computer tasks through video pretraining.
Manus introduced Cloud Computer, a dedicated cloud machine that runs bots, scripts, and software 24/7 without requiring users to manage servers or keep their laptop on.
Alibaba's Qwen team released Qwen-Scope, an open interpretability toolkit built with Sparse Autoencoders trained on Qwen3 and Qwen3.5 models, covering 7 models and 14 SAE configurations.
Microsoft introduced a Legal Agent in Word built to follow the structured workflows lawyers use while keeping them fully in control, with every clause and redline tracked.
Cursor launched Security Review, a cloud agent feature that automatically analyzes codebases for security vulnerabilities as part of its Cloud Agents infrastructure.
X announced a rebuilt, AI-powered advertising platform using AI to enhance campaigns with better results, more relevant placements, and precise targeting.
🧡 Enjoyed this issue?
🤝 Recommend our newsletter or leave a feedback.
How'd you like today's newsletter?
Cheers, Jason
