Mira Murati's New AI Doesn't Wait for You to Finish Talking — and That's the Point

Mira Murati's New AI Doesn't Wait for You to Finish Talking — and That's the Point

Mira Murati's Thinking Machines Lab, conspicuously quiet since she left OpenAI, just stepped into the light with its first model: a 276-billion-parameter system built for live interaction. The strange part isn't the parameter count — it's the architecture. The model processes voice, video, and text in 200-millisecond chunks, reacting in a continuous streaming loop rather than waiting for you to finish your thought. A second background model handles slower reasoning and tool work, so the live model never has to pause.

That changes the feel of the whole product category. Most AI tools still behave like extremely fast email: you ask, they answer, you wait. Thinking Machines is aiming at something closer to a collaborator in the room — one that can catch a visual change, count reps, translate speech as it happens, and jump in at the right moment instead of politely waiting its turn. As Murati put it, "the way we work with AI matters as much as how smart it is."

The competitive angle is real. OpenAI, Anthropic, and Google have spent the last year racing to make their models smarter. Murati is making hers harder to ignore — and the product decision is the message: the next wave of AI won't just be better at answers, it'll be better at presence.

Gobble's Take: The AI that wins may be the one that stops acting like software and starts acting like a sharp colleague who knows exactly when to interrupt.

Sources: Pascal Bornet Substack · AM Data Lakehouse Substack

Anthropic Forms Project Glasswing to Test Unreleased Mythos Model — Wall Street Banks Also Testing Internally

Anthropic is letting a select group of tech firms access an unreleased AI model called Mythos through a new initiative named Project Glasswing. Amazon, Apple, Microsoft, Cisco, and other organizations are participating. The companies will use Mythos to hunt for flaws in their own products and share findings with industry peers. Anthropic said it does not have plans yet to release Mythos to the general public, and will use Project Glasswing's findings to inform guardrails for the technology.

Wall Street banks are also starting to test Mythos internally as Trump administration officials encourage them to use it to detect vulnerabilities. JPMorgan Chase & Co. was the only bank named as part of an initiative to test the model, but other major financial institutions have also gained access or expect to in the coming days.

Mythos has already uncovered a 27-year-old bug in critical internet software and a 16-year-old vulnerability in video game code that automated testing tools had scanned five million times without detecting.

Gobble's Take: Anthropic is sharing Mythos with a limited set of companies and financial institutions before any public release, with Trump administration officials urging banks to use it to find vulnerabilities.

Source: Monique Malcolm Hay Substack

Google Is Trying to Make Android Feel Like It Has a Brain, Not Just Apps

At its Android Show event — one week before Google I/O — the company unveiled Gemini Intelligence, a cross-device AI platform designed to carry out agentic tasks inside apps using on-screen context. The companion hardware push is just as telling: a new line of AI-native "Googlebook" laptops, built with Dell, HP, Lenovo, Acer, and Asus, shipping this fall, featuring a "Magic Pointer" AI cursor shown in a live demo. The laptops will run Android phone apps and files, blending ChromeOS, Android, Google Play, and Gemini into a single surface.

The important shift isn't the laptop. It's the plumbing. Google is weaving AI directly into Android rather than bolting it on as a separate chatbot button. Smaller tools round out the picture: a filler-word-stripping dictation feature called Rambler, a Create My Widget tool, and Gemini auto-browse in Chrome, running on-device. Individually, these are conveniences. Together, they're a platform play — AI embedded so deeply into the OS that it becomes invisible infrastructure. Apple is still promising a Siri revival. Google is trying to make the whole operating system do the asking for you.

If it works, your phone stops being a place you search and becomes a place that acts. That sounds useful until you realize convenience is exactly how platforms get under your skin — and then under your thumb.

Gobble's Take: The real AI race is no longer about who has the smartest chatbot — it's about whose OS starts finishing your sentences before you open your mouth.

Source: Pascal Bornet Substack

Cerebras Priced at $185, Opened at $350, and Briefly Hit a $106 Billion Market Cap — on Day One

Cerebras Systems hit the Nasdaq on May 14 under the ticker CBRS and immediately rewrote the usual IPO script. The company had set its price at $185 per share — already $25 above an upwardly revised target range, which had itself been raised from the initial range just days earlier. It raised $5.55 billion selling 30 million shares. Then the stock opened at roughly $350, an 89% pop, with shares briefly surging above $370 in the first minutes. The order book had closed approximately 20 times oversubscribed.

The company behind the pop is stranger than the number. Cerebras builds wafer-scale chips — processors the size of a dinner plate — on the premise that giant AI models shouldn't have to be split across dozens of separate chips constantly shuttling data between each other. Founder Andrew Feldman, a serial entrepreneur who previously sold a microserver startup to AMD, spent a decade pursuing what much of the chip industry dismissed as a sophisticated science project. Eighteen months ago, Cerebras couldn't complete a routine IPO because U.S. regulators were concerned that AI chips in its Abu Dhabi supercomputers could — through a chain of corporate ownership — end up serving models running in Beijing.

That regulatory drama is now priced in, or at least priced past. At the IPO price of $185 per share, on a fully diluted basis, Cerebras was worth approximately $56 billion — though the fully diluted market cap at open crossed roughly $106 billion once shares began trading. What Wall Street is actually betting on is simpler than the architecture: that AI demand is now running into physical bottlenecks, and that a chip designed to keep models in one place rather than scattered across a rack might be worth the absurdity of its size.

Gobble's Take: When a chip company 20x oversubscribes its IPO, it usually means investors have decided that AI's next constraint isn't intelligence — it's plumbing, and someone just cornered the pipe.

Source: Long Yield Substack

DeepSeek and Moonshot Are Turning China's AI Race Into a War of Efficiency, Not Brute Force

DeepSeek — reportedly a team of around 200 people — is in talks for a funding round that would value it at up to $50 billion, potentially raising between $3 billion and $7 billion, led by China's state-backed semiconductor investment fund. Meanwhile, Moonshot AI, the company behind the Kimi chatbot, recently closed a $2 billion round at a $20 billion valuation. These are not scrappy underdog numbers. They are frontier valuations for teams built around constraint rather than scale.

The hardware pressure driving this is real. Restrictions on Nvidia's top-end chips — the H100 and H200 are largely off-limits — forced Chinese AI labs to find workarounds: stockpiling earlier chips, building around domestic alternatives like Huawei's Ascend processors, and engineering models that extract more from less compute. The result, per reporting on the sector, is models trained and run at a fraction of the cost of Western counterparts — sometimes one-tenth to one-hundredth the expense. China's latest five-year plan, covering 2026–2030, mentions AI more than 50 times and includes an "AI+" action plan to embed intelligence across the broader economy.

The takeaway for anyone watching the global AI market: the frontier is no longer only where the biggest American labs are spending the most money. It's also where the most constrained teams are learning to squeeze performance out of whatever silicon they can actually access. Scarcity, it turns out, is a surprisingly good engineering teacher.

Gobble's Take: The next AI breakout may come from the team that literally cannot afford to waste a single GPU cycle — which means the chip embargo might be backfiring in slow motion.

Source: José Luis Chavez Calva Substack

Quick Hits

• Cursor lands inside Microsoft Teams: The AI coding assistant announced on May 11 that users can now @mention it directly in any Teams channel to delegate tasks or pull information — moving it out of the code editor and into where engineering teams actually coordinate. AM Data Lakehouse Substack
• Anthropic doubled Claude Code rate limits in under a month: A SpaceX partnership added 300 megawatts of new compute — equivalent to more than 220,000 Nvidia GPUs — in under a month, using the Colossus One facility originally built for xAI's Grok training workloads; users who'd been hitting output limits during peak hours got relief with no price increase. AM Data Lakehouse Substack

In Case You Missed It

Yesterday's top stories:

• Liquid Helium, One Core, Nine-Point-Two Gigahertz: A New CPU World Record
• Amazon Kills 13 Kindle Models on May 20th — Owners Are Already Fighting Back
• AI Is Eating the DRAM Supply — and PC Gamers Are the Ones Going Hungry
• Greg Brockman Is Back at OpenAI — and Now He's Running the Whole Product Operation