Today in Tech & AI: The Smartest Twitter Threads You Missed

A new AI model can now hold a conversation in two-hundred millisecond ‘micro-turns,’ effectively listening and speaking at the same time. Last week on episode forty-eight, the entire conversation was dominated by Anthropic's power plays and its refusal to share its cyber-AI with governments. This week, the board has been completely reset by a company called Thinking Machines, proving that the most important breakthroughs aren't always about who has the biggest model, but who has the fastest connection to the real world. The age of turn-based AI is over. The age of the duet has begun. That’s the big one, but the ground is shifting everywhere.

While Thinking Machines redefines real-time interaction, OpenAI is making a move that feels like a direct response to last week's drama. They are now actively negotiating with the European Commission to give EU authorities access to a specialized, cyber-focused version of GPT-5.5. This isn't the public model. This is the one designed to identify and exploit software vulnerabilities. It’s a hacking tool. And OpenAI is offering to hand over the keys to help defend critical infrastructure. The contrast couldn't be sharper. Anthropic, as we discussed, is keeping its own version, Mythos, under lock and key, citing trust issues.

OpenAI’s executive George Osborne put it bluntly, saying, “Frontier AI companies... shouldn’t be the sole arbiters of who has cybersecurity in this world.” It's a bold play for geopolitical relevance, painting Anthropic as isolationist while positioning OpenAI as a responsible global partner. The EU, for its part, seems to be welcoming the transparency. Then, there's Elon Musk. Just when you think the cast of characters is set, he launches a new AI software firm with the completely straight-faced, not-at-all-a-joke name… "Macrohard." Yes, Macrohard. Beyond the troll, he's framing the entire AI endeavor as an existential battle for humanity's future.

He’s back on his crusade for transparency, regulation, and safety, arguing that unchecked commercialization is putting civilization at risk. This isn't a product launch; it's a political campaign. Musk is drawing a line in the sand, creating an ideological pole for regulators and the public to rally around. He’s building a narrative that casts the rapid, closed-door development at other labs as reckless. Whether Macrohard produces a single line of code is almost secondary. Its primary product is a story. A story about who you should trust with the future. And finally, under the hood of all this, a change that could accelerate everything.

Sakana AI and NVIDIA just dropped a new paper on a technology called TwELL. It’s a new type of sparse CUDA kernel, which is a very technical way of saying they found a way to make GPUs work smarter, not just harder. Large language models are full of empty space—what’s called activation sparsity. Up to ninety-nine percent of the numbers in certain calculations can be zero. But GPUs have always been bad at skipping the zeroes; they just plow through everything, wasting time and energy. TwELL fixes that. It cleverly packs the important numbers together inside the GPU's own workflow, achieving a twenty-point-five percent speedup in inference and a nearly twenty-two percent speedup in training.

This isn't a small tweak. It's a fundamental efficiency gain that makes every single model from every single company cheaper to run and faster to train. It’s a rising tide that lifts all boats, and it just arrived. So let's go back to Thinking Machines. Because what they announced is more than just a faster chatbot. It represents a fundamental break with the past. For the entire history of AI interaction, there's been one constant. The turn. You speak, you wait. The AI processes, it speaks, it waits. Even the most advanced models we have today, like GPT-Realtime-2 or Gemini 3.1-Flash, are just making that wait time shorter.

They’re perfecting the art of the turn. VentureBeat put it perfectly: “The basic interaction mode across text, imagery, audio, and video remains the same: the human user provides an input, waits, and the AI model provides an output.” This is the paradigm that TML-Interaction-Small just shattered. This isn't about shortening the wait. It's about eliminating it. The model operates in what they call two-hundred-millisecond "micro-turns." This is a full-duplex architecture. It processes your voice, your face, and the text you’re typing, all at the same time, while it is also generating its own voice, images, and text.

It can interrupt you. You can interrupt it. It can nod along while you’re speaking. It can finish your sentence. It can notice you’re looking away from the screen and ask if you're distracted. This isn't a conversation with a machine. This is a collaboration. The two-hundred-millisecond cycle time is not an arbitrary number. That's the approximate threshold for fluid, natural human conversation. Below that, things feel seamless. Above that, you start to notice the lag. Thinking Machines built their model to live on our side of that line. How did they do it? They threw out the old playbook. Instead of taking audio, converting it to text, feeding it to a model, getting a text response, and converting that back to audio—a chain that creates latency at every step—they built a system with encoder-free early fusion.

That means it’s taking in the raw audio signals and raw image patches directly. It co-trains all the modalities from scratch, inside one transformer. The AI is listening to the sound of your voice, not just the words. It's seeing the pixels of your face, not just a tag that says "human." This is what gives it the speed and the nuance to react in real time. The architecture itself is also a work of art. They separated the system into two parts. There's an "interaction model," which is the TML-Interaction-Small 276B model. Its only job is to manage the immediate, real-time dialogue. It maintains presence.

It handles the back-and-forth. It’s the fast-twitch muscle of the system. Then, there's a "background model." This can be any other powerful model, maybe one that’s much larger and slower. When you ask a complex question that requires deep reasoning, a web search, or tool use, the interaction model doesn't stop talking to you. It keeps the conversation going while it asynchronously passes the hard problem to the background model. When the answer is ready, the background model streams the results back, and the interaction model weaves it seamlessly into the conversation. Think about what that unlocks.

You could be on a video call with two other people, and the AI is translating for everyone in real-time, in their own voices, without anyone ever pausing. You could be a pilot in a simulator, and the AI is your co-pilot, calling out instrument readings and responding to your commands instantly. You could be a developer coding, and the AI is your pair programmer, not just suggesting code blocks, but pointing out a typo in your variable name as you are typing it. This is the difference between a tool and a partner. A tool waits for your command. A partner anticipates your needs. This is why the raw parameter count—276 billion—is the least important number here.

It's a big model, yes, but it’s a Mixture-of-Experts model with only twelve billion active parameters at any given time. It's designed for efficiency, not just brute force. The real story is the architecture that allows it to use those parameters to create a continuous, fluid presence. It’s a shift from thinking about AI as a giant brain in a box to thinking about it as a distributed nervous system, with reflexes and reason operating on different loops. So while OpenAI is battling for political alliances in Brussels and Elon Musk is battling for the narrative with names like "Macrohard," Thinking Machines just quietly changed the physics of the game.

They're not just building a better version of what we have. They're building the next thing entirely. This technology isn't publicly available yet, and that's the big caveat. But the paper is out. The demo videos are circulating on Twitter. And the core idea—that the turn is a bug, not a feature—is now loose in the world. What connects all of this week's news is a pivot. For the past few years, the AI race has been a simple contest of scale. Who can build the biggest model? Who can secure the most GPUs? Who can raise the most capital? It was a race of brute force. This week signals the beginning of a new race.

A race of architecture. The question is no longer just "how big?" but "how smart?" How do you design a system? Thinking Machines is rewriting the architecture of interaction. Sakana AI and NVIDIA are rewriting the architecture of computation on the chip itself. OpenAI is attempting to rewrite the geopolitical architecture of who gets access to this power. And Elon Musk is trying to rewrite the ideological architecture that governs it all. The era of simply scaling up is giving way to an era of clever design. The most important breakthroughs are no longer happening just in the size of the models, but in the elegance of the systems that connect them to our world.

The interface is becoming the innovation.