Tech & AI Twitter: The Daily Pulse of Smart, Moving Conversations

In the next thirty days, four of the world's top AI labs are shipping flagship models simultaneously. Last week, in episode eighty-eight, we talked about the accelerating pace of AI development. This is what that acceleration looks like in real time. The release cycles that used to take years now take weeks. The Polymarket crowd is betting with eighty-nine percent certainty that it all lands before June thirtieth, including new models from OpenAI and Google. This isn't just an update. It’s a market-wide reset button, pushed by everyone at once. A few years ago, we waited nearly three years between GPT-3 and GPT-4.

Then over a year for GPT-5. Now? A new flagship model is dropping practically every month. The competitive moat at the model layer is now measured in weeks, not quarters. Whatever model you chose as your default three months ago is already two generations old. This is the new tempo. Now, for what else is moving. First, let's talk about power. Not the metaphorical kind, the literal kind. The kind that comes from the grid and runs the data centers that power this entire AI boom. The biggest bottleneck for AI growth isn't ideas. It’s electricity and the permits to build new data centers. Everyone is waiting years for grid connections.

But Tesla just decided not to wait. Martin Varsavsky flagged it this week: Tesla is turning millions of its parked cars into a distributed AI supercomputer. Think about it. Millions of vehicles, each with a powerful onboard computer and a battery, sitting idle most of the day. Tesla is networking them. Using their downtime to run AI computations. This is how you solve the problem when you're an operator, not a pundit. While everyone else is stuck in regulatory limbo, waiting for a permit, Tesla is building a supercomputer out of assets it has already sold you. It’s a masterclass in what Varsavsky calls "execution beats regulation." Next, a mental model for how to think about this arms race.

A tweet from Manosai this week splits the AI world into two distinct tracks. And once you see it, you can’t unsee it. Track one: using AI to write code to build software. This is about creating new applications, new tools, from scratch. Track two: using AI to write code to use software. This is about agents. AI that can operate your existing tools, navigate websites, and execute tasks on your behalf. Manosai’s point is that if you believe Anthropic is way ahead, you're probably focused on track one. Their models have been exceptional at pure code generation for building things. But if you believe OpenAI is way ahead, you're focused on track two.

They are relentlessly pushing the agentic use case—making the AI a partner that does things for you inside your current workflow. It’s not about which is better. It’s about understanding they are two different races being run on the same track. And the company that unifies them first, wins. Speaking of winning, while all eyes are on the OpenAI and Google showdown, xAI has been shipping. Relentlessly. In the last month alone, Elon Musk’s company has dropped four releases. FOUR. Grok 4.3, their flagship, actually launched at the end of April. It has a one-million-token context window and continuous reasoning.

But then this month, in June, they followed up with Grok V9-Medium, Grok Voice, and Grok Imagine Video 1.5. They are just shipping, shipping, shipping. And it's working in specific domains. Right now, Grok 4.3 is number one on the Artificial Analysis benchmark for legal reasoning, called CaseLaw. It’s the budget pick that’s quietly winning on specialized tasks while the giants prepare for their heavyweight bout. And finally, a cautionary tale. The speed of this new cycle creates a new kind of risk. Single-provider lock-in has never been more expensive, or more dangerous. Anthropic’s customers learned this the hard way when the Fable 5 model was pulled, literally overnight.

The platforms and applications built on top of it just… broke. Any company that hard-codes a dependency on a single model provider is taking an existential risk. Because the model you bet on today might not be the best one next month. Or it might not exist at all. So let's go deep on the main event. The clash that defines this June arms race: OpenAI versus Google. First, OpenAI. Their next model, likely called GPT-5.6, has not been announced. But it’s all but confirmed. Here’s how we know. For the last three releases—GPT-5.4 in March, GPT-5.5 in April—the pattern has been the same. Roughly ten to fourteen days before launch, a new model name appears in the routing logs for Codex, OpenAI's code interpreter.

It’s there for a few minutes, then it’s gone. On May fourteenth, a researcher spotted a single anomalous line: a mapping entry pointing to a model called gpt-5.6. Then it vanished. The pattern is holding. Internal codenames have also leaked: iris-alpha, ember-alpha, beacon-alpha, and kindle. And we have confirmation that kindle-alpha is the release candidate. So what’s new? The credible leaks point to a one-point-five million token context window, up from one million in GPT-5.5. And something called an UltraFast Codex mode. But the real breakthrough, the one developers are desperate for, is a fix for what's called the "garbage frontend code" problem.

For years, AI coding tools have been great at backend logic but notoriously bad at producing clean, modern frontend code. The leaks suggest GPT-5.6 marks a complete qualitative change on this front. But there’s another, more specific problem GPT-5.6 is almost certainly being built to fix. And it has to do with goblins. After GPT-5.5 was released, OpenAI published a candid, almost shocking, post-mortem on their alignment process. It was called "Where the Goblins Came From." It documented how a miscalibrated reward model during training had systematically, and bizarrely, favored responses that mentioned goblins, gremlins, trolls, and raccoons.

For a while, their flagship model had a weird obsession with mythical creatures. It was a major, if comical, alignment failure. They have to fix that. The market expects it. Now, Google. At their big I/O conference on May nineteenth, the developer audience was primed for a new model. They were ready. Then CEO Sundar Pichai got on stage and told them to "give us until next month." You could hear the groan in the room. Well, "next month" is now. And there are only nine days left in June. The Polymarket is giving it an eighty-nine percent chance they make the deadline with Gemini 3.5 Pro. And the specs are huge.

Literally. Google has confirmed a two-million-token context window. That is thirty-three percent larger than GPT-5.6's rumored one-point-five million. It also features a "Deep Think" reasoning mode, which promises to improve performance on complex, multi-step tasks. We already have a preview of their progress. The model they did ship at I/O, Gemini 3.5 Flash, is a lighter, faster version. And it’s already outperforming the previous top-tier model, Gemini 3.1 Pro, on coding and agentic benchmarks. It's faster, it's cheaper, and it's better on key tasks. The full Pro version is expected to close the remaining gaps on hard reasoning and long-context retrieval where Flash still trails.

The pricing is also aggressive: an estimated fifteen dollars per million input tokens and sixty for the output. They are coming to compete not just on capability, but on cost. So what does it all add up to? You have two titans, OpenAI and Google, about to drop flagship models with multi-million token context windows within days of each other. You have xAI carpet-bombing the month with feature after feature. You have the entire ecosystem trying to keep up. It means the ground beneath your feet is shifting. The foundational layer of the new economy is being replaced, wholesale, every thirty days.

This isn't just about a faster horse. This is a fundamental change in the metabolism of technology. For the past fifty years, the cadence of progress was set by hardware. Moore's Law. A doubling of transistors every two years. It was fast, but it was predictable. You could plan around it. That era is over. The new cadence is set by software. By AI models. And the cycle isn't two years. It's now about thirty days. We just saw four labs compress what used to be a multi-year R-and-D cycle into a single month. This speed creates incredible opportunity, but it also creates chaos. It rewrites the rules of strategy.

How can you have a five-year plan when the technology you depend on has a six-week half-life? And the innovation isn't just happening at the model layer. Look at the Tesla story. When faced with a physical bottleneck—not enough power, not enough data centers—the answer wasn't to wait. The answer was to redefine the computer. To see a million parked cars not as vehicles, but as a latent, distributed supercomputer. That’s systems thinking. That’s execution. This week connects the dots. The frantic race to ship the next model, the clever hack to power it, the new frameworks for even understanding it—it’s all one story.

It’s the story of acceleration. It’s the story of an industry that is no longer waiting for permission, or for the grid, or for the next conference. The era of waiting a year for the next big thing is over. The next big thing now arrives every Tuesday. And the only question is whether you're ready to catch it.