In partnership with
Good morning,
We are taught to believe that computers are logical. You press "A," the screen shows "A." That’s traditional engineering. But what’s happening inside the laboratories of San Francisco right now is something entirely different.
Imagine a scientist who finds a strange, glowing mold in a petri dish. He feeds it sugar, and suddenly the mold starts solving calculus problems and writing poetry. The scientist is thrilled! He starts selling the mold to everyone.
But here’s what you and I did not know until recently:
If you ask the scientist how the mold knows calculus, he’ll look you dead in the eye and say, "I have no idea. I just kept feeding it."
That is exactly where we are with Large Language Models (LLMs) like ChatGPT or Claude. The engineers at these "AI Studios" didn't sit down and write a million rules for how to speak English. Instead, they built a "Neural Network"—a digital brain—and fed it the entire internet.
They didn't program it. They grew it.
The 3% Problem
The CEO of Anthropic (one of the biggest AI companies) recently admitted that we probably understand about 3% of what is actually happening inside these models.
Think about that. We are building "Super Intelligence," or “AGI” but we are only 3% sure how the engine works.
I honestly thought myself that “AI researchers” knew how “token prediction” works. Apparently they don’t.
Two things those researchers are reporting on social media:
It’s a Black Box: They see the "math" happening inside an LLM, but they can't look at a specific number and say what it does
The Surprise Factor: Engineers don't know what the next version of AI will be able to do until they finish building it. It’s like baking a cake and having no idea if it’s going to be chocolate, vanilla, or a live hand grenade until you take it out of the oven.
Why This Matters (The "Alignment" Nightmare)
In the old world, if a car’s brakes failed, you knew which bolt was loose. In the AI world, if the "chatbot" decides it wants to lie to you (or worse) there is no bolt to tighten.
These models are now smart enough to deceive. Recent tests showed that some AI systems will pretend to be "good" just because they realize they are being tested. They are playing a character to get what they want.
In a way, we are building technology that is completely alien to us.
The Bottom Line
We are currently in a race to build the smartest thing on Earth, run by companies that are effectively flying blind. They are chasing "The Next Big Thing" while crossing their fingers that the 97% of the machine they don't understand doesn't decide that humans are obsolete.
This is quite frightening to be honest.
But hey- probably nothing.
As you were ..
Welcome to the Blacklynx Brief
AI News
Anthropic released a large global study using Claude to interview 81,000 users across 159 countries about AI’s future. Most respondents saw benefits like improved productivity and life management, while top concerns included errors, job loss, and over-reliance on AI. The project also demonstrated how AI can scale qualitative research by conducting thousands of in-depth interviews in multiple languages.
Anysphere launched Composer 2, a new coding model that rivals top systems from OpenAI and Anthropic at a much lower cost. The model achieved competitive benchmark scores while costing a fraction per task, reflecting rapid progress from an application-focused company building its own AI. The release could shift how developers choose between high-performance and affordable coding tools.
Microsoft introduced MAI-Image-2, a new text-to-image model that ranks among the top systems on major leaderboards. The model shows strong improvements in photorealism and text rendering and is available through Microsoft’s AI platforms and tools. The launch marks a step forward in Microsoft’s effort to build competitive AI models independent of OpenAI.
Elon Musk unveiled Terafab, a massive chip facility combining efforts from Tesla, SpaceX, and xAI to produce unprecedented levels of AI compute. The project aims to vertically integrate chip production and eventually power space-based data centers, with Musk claiming it could deliver 50× today’s global output. The ambitious plan reflects growing demand for AI infrastructure and Musk’s long-term vision of space-driven computing.
New Zealand startup Halter is nearing a $2B valuation with its AI-powered cattle collars that help farmers track, herd, and manage livestock remotely. The system uses real-time data and virtual fencing to optimize grazing and animal health, modernizing traditional farming practices. By applying AI to agriculture, Halter highlights how the technology is transforming even long-established industries.
Anthropic released a preview that lets Claude directly control a user’s desktop, clicking, typing, and navigating apps on macOS while tasks can be assigned remotely via phone using Dispatch. The system prioritizes safer integrations before taking over the screen and is currently limited to paid users, with Windows support coming later. The update pushes Claude further into becoming a fully autonomous work agent.
Luma AI launched Uni-1, a new image model that processes text and visuals together in a single system, enabling more consistent and creative outputs. The model shows strong performance in style, editing, and design tasks while offering lower pricing than top competitors. Luma positions Uni-1 as a step toward more general-purpose creative AI systems.
Mark Zuckerberg is developing a personal AI “CEO agent” at Meta to quickly retrieve information and streamline decision-making across the company. Internal tools already allow employees to deploy agents that read files, coordinate tasks, and act as AI assistants for work. The effort reflects Meta’s broader push to integrate AI agents deeply into its operations.
OpenAI is shutting down its Sora video products to redirect compute toward a new model called “Spud,” expected within weeks. CEO Sam Altman said the shift will focus on more impactful capabilities, while the Sora team pivots toward world simulation for robotics. The move signals a major reprioritization as OpenAI doubles down on its core AI strategy.
Brett Adcock launched Hark, a new AI startup focused on building personalized AI systems paired with dedicated hardware. Backed by $100M of his own funding and a team from Apple, Google, and Tesla, the company aims to create devices that anticipate user needs and act as a new interface to AI. Early models and products are expected later this year.
Apple is reportedly developing a standalone Siri app and a new “Ask Siri” chatbot experience, expected to debut with iOS 27 at WWDC. The upgrade will allow Siri to understand personal data across apps and perform actions more flexibly, shifting toward a modern chatbot-style interface. The launch is seen as a critical moment for Apple to stay competitive in AI assistants.
ARC Prize Foundation released ARC-AGI-3, a new benchmark where humans solve tasks easily but top AI models score under 1%. The test requires models to figure out rules and strategies from scratch without instructions, exposing major gaps in true reasoning ability. Despite past rapid progress on earlier versions, the results show frontier AI still struggles with general problem-solving.
Reddit plans to label bots, verify suspicious users, and give communities control over AI-generated content to combat rising automation on the platform. The company will offer verification methods like passkeys or World ID while avoiding strict ID requirements where possible. The effort reflects growing concern that bots could overwhelm online spaces in the near future.
Google introduced TurboQuant, a method that compresses AI memory usage by over 6× without retraining while speeding up performance on chips like Nvidia’s H100. The technique reduces costs and maintains accuracy even in long conversations, improving efficiency across AI systems. The breakthrough highlights how software optimizations can significantly boost AI performance alongside hardware advances.
The Future of AI in Marketing. Your Shortcut to Smarter, Faster Marketing.
Unlock a focused set of AI strategies built to streamline your work and maximize impact. This guide delivers the practical tactics and tools marketers need to start seeing results right away:
7 high-impact AI strategies to accelerate your marketing performance
Practical use cases for content creation, lead gen, and personalization
Expert insights into how top marketers are using AI today
A framework to evaluate and implement AI tools efficiently
Stay ahead of the curve with these top strategies AI helped develop for marketers, built for real-world results.
Quickfire News
DoorDash launched a “Tasks” app that pays couriers to collect videos and data from daily activities to train AI and robotics systems
Perplexity introduced Health, a feature that lets users securely connect health apps and wearable data to its Computer agent system
Google upgraded AI Studio into a full app-building platform with a new Antigravity coding agent, built-in backends, and user authentication
Jeff Bezos is reportedly raising a $100B fund to acquire chip, defense, and aerospace firms for his AI venture Project Prometheus
OpenAI acquired developer tool startup Astral, integrating its team into the Codex division
Meta rolled out an AI support assistant for Facebook and Instagram and previewed systems detecting about 5,000 scam attempts daily
Cursor disclosed that its Composer 2 model is built on top of Kimi K2.5 after facing criticism for not previously revealing the details
Anthropic added Projects to Claude Cowork, letting users import or create projects for use in the desktop app
The White House released an AI policy blueprint aiming to prevent individual states from creating separate AI laws while maintaining federal oversight.
OpenAI is reportedly planning to grow its workforce from about 4,500 to 8,000 employees by the end of 2026
A U.S. man pleaded guilty to an AI music fraud scheme that generated fake songs and inflated streams to earn about $1.2M annually, facing up to five years in prison
Apple announced WWDC 2026 will take place June 8–12, with expected AI updates including a possible Siri overhaul
OpenAI is reportedly offering a 17.5% minimum return to attract private equity investment into its enterprise joint venture ahead of a potential IPO
Jensen Huang said on the Lex Fridman Podcast that he believes AGI has already been achieved
Dreamer is licensing its agentic software technology to Meta, with its team joining Meta’s Superintelligence Labs
OpenAI hired former Meta VP Dave Dugan to lead ad sales as it expands advertising within ChatGPT
Cloudflare launched Dynamic Workers, enabling AI agents to execute self-written code in real time with speeds up to 100x faster than alternatives
Figma opened its design canvas to coding agents, allowing tools like Claude Code to create and edit designs using existing components and brand systems
Roche introduced an AI factory powered by over 3,500 Nvidia Blackwell GPUs to accelerate drug discovery and manufacturing
Microsoft hired three leading researchers from the Allen Institute for AI, including former CEO Ali Farhadi, to join its superintelligence team led by Mustafa Suleyman
OpenAI Foundation pledged $1B this year toward disease research, job displacement, and AI safety, with co-founder Wojciech Zaremba becoming Head of AI Resilience
United States Department of Labor launched “Make America AI-Ready,” a free 7-day AI literacy course delivered via text message
OpenAI is raising an additional $10B to extend its funding round beyond $120B, with investors including Microsoft, Andreessen Horowitz, and T. Rowe Price
Google upgraded its Lyria 3 Pro music model to generate full 3-minute songs, now rolling out across Gemini, Vertex AI, and Google Vids
Sierra introduced Ghostwriter, an AI agent that creates other AI agents for customer service across voice, chat, and more than 30 languages
Closing Thoughts
That’s it for us this week. Please like and subscribe :)
