Here’s what’s on our plate today:

• 🧪 MiniMax's M3 reveals why frontier AI is no longer one race.

• 📰 Who benefits from the Anthropic crackdown, China's green AI hurdles, and Google's AI mistakes, horror fiction for fact?

• 🧠 Brain Snack: assemble the open-weight pieces, spend on the one frontier that's still hard.

• 🗳️ Poll: What does a fragmenting frontier mean for the AI race?

Let’s dive in. No floaties needed…

The Laboratory

TL;DR

• The bundle comes apart: MiniMax's M3, released June 1, 2026, combines frontier coding, a 1M-token context window, and native multimodality in one downloadable open-weight model, a combination that until now lived only inside closed labs.

• Different labs, different axes: DeepSeek chased long context, Moonshot chased multimodality, Alibaba chased efficiency. Each capability advanced separately, proving they were never fused to begin with.

• Cheap, not smart: Long context turned out to be an engineering and cost problem, not a leap in intelligence. As the computational burden fell, the capability spread fast.

• Reasoning is the holdout: Coding, context, and multimodality all yielded to better engineering, and so, eventually, did the abstract puzzles of ARC-AGI-2. The next layer has not. On ARC-AGI-3, the interactive reasoning test released this March, no leading model clears 1%, the clearest frontier still standing.

MiniMax’s M3 reveals why frontier AI is no longer one race

For the past couple of years, frontier AI models have differentiated themselves from others by ticking all the boxes of a checklist that required them to write software, understand images, process enormous amounts of information, hold long conversations, and generally perform a wider range of tasks than their competitors. These abilities appeared together so consistently that it was easy to assume they were all manifestations of the same underlying thing. A frontier model was simply a model that had more of whatever made AI powerful.

That assumption has shaped how investors think about the industry, how governments think about technological leadership, and how the public imagines the race between companies like OpenAI, Google, Anthropic, and a growing collection of challengers from China. This meant that, for most, the race to develop the most capable AI model was straightforward, since the goal was to build the smartest model.

However, a model released earlier this month is forcing a rethink of that assumption. And it does this by showing that many advanced AI capabilities can be developed separately and stitched together afterward.

On June 1, 2026, Chinese AI company MiniMax introduced M3, a model that drew attention for combining several capabilities traditionally associated with the industry's most advanced systems. It can write code, work with extremely large amounts of information at once, and understand images and video alongside text. More importantly, MiniMax released it as an open-weight model, meaning developers can download and run it themselves rather than access it exclusively through a company's servers.

However, the announcement may sound like another routine milestone in the relentless competition to build better AI; a closer look raises a more interesting question. If combining coding, long context, and multimodality is itself considered an achievement, then those capabilities must have been separable to begin with: three distinct problems that happened to be bundled together inside the same expensive models. That is precisely what the open-weight ecosystem has spent the past year demonstrating.

The distinction may sound subtle, but it points toward a larger shift in how AI progress is unfolding around the world.

AI is starting to come apart

The easiest way to see what is changing in AI is to look at what different labs have been building. Over the past year, DeepSeek has become known for handling massive amounts of information while maintaining strong performance. Moonshot focused on models that can work across text, images, and video while carrying out complex tasks. Alibaba focused on efficiency, finding ways to deliver strong performance without increasing costs. Rather than pursuing the same goal, each company picked a different problem and pushed it as far as it could.

That is what makes MiniMax's M3 interesting. The company did not invent coding, long memory, or multimodal AI. Others had already demonstrated each capability separately. What M3 accomplished was to bring them together into a single downloadable model. That this counts as an achievement at all suggests these capabilities were separable from the start, distinct challenges that happened to coexist inside the industry's largest and most expensive models.

Long context shows how barriers fall

The clearest example is long context, the ability to work with enormous amounts of information at once. For years, it looked like a frontier capability available only to the richest companies, largely because handling that much data demanded computational resources few could afford. Yet the more researchers worked on the problem, the more it came to resemble an engineering challenge rather than a mystery of intelligence. Rather than making models fundamentally smarter, new techniques helped them search through information more efficiently, reducing the computational cost of working with large documents and conversations. As that burden fell, context windows (the amount of text a model can take in at once) expanded rapidly, and a capability once reserved for frontier models became far more widely available.

Seen through that lens, M3's significance is less about a breakthrough and more about integration. It suggests that several capabilities once viewed as signs of frontier AI are becoming modular components that can be improved independently and then assembled into a single system. What looked like one frontier is increasingly revealing itself as a collection of smaller ones.

And once the bundle comes apart, an uncomfortable possibility follows: many of these capabilities may have been expensive rather than genuinely difficult, appearing hard only because nobody had yet found a cheaper way to achieve them.

A similar pattern can be found across much of the industry's recent progress. Improved coding performance resulted from advances in training methods and reinforcement learning, in which models learn through trial and error guided by feedback. Multimodality benefited from training pipelines designed to work with text, images, and video from the outset. Each advance removed a specific constraint, and each removed constraint made another frontier capability more widely accessible.

Viewed from that perspective, the progress of open-weight models begins to look less like a dramatic chase after a distant leader and more like a gradual process of dismantling barriers. Capabilities that once seemed inseparable from the largest and most expensive models increasingly look like individual engineering problems that can be solved, shared, and incorporated elsewhere.

The frontier that still resists

That process naturally draws attention to the remaining barriers, and abstract reasoning has been the most stubborn of them. When ARC-AGI-2 launched in 2025 to test a model's ability to identify patterns, infer rules, and handle unfamiliar situations, no leading model solved more than about 3% of its tasks, far below what the same systems managed on their easier predecessors. For a while, it looked like the one capability engineering could not simply buy. Then it followed the others: over the next year, frontier models climbed to roughly 75%-85% on ARC-AGI-2. The barrier fell, only later than the rest.

The frontier moved rather than vanished. ARC-AGI-3, released in March 2026, trades static puzzles for interactive problems a person can pick up in seconds, and it has reset nearly every model to near zero, with the best scoring well under 1%. So reasoning is not one wall either. The static, abstract version gave way to better training and architectures, as did the rest. The fluid kind, entering an unfamiliar situation and working out its rules in real time, remains one of the clearest gaps between today's AI systems and human intelligence, which brings us back to MiniMax. M3 matters because it bundles several capabilities that the industry increasingly recognizes as distinct achievements. What appeared to be a single frontier now looks more like a collection of frontiers, each falling on its own timetable and for its own reasons. The capabilities that have become cheaper, more efficient, and easier to reproduce are steadily being folded back into unified systems. The capabilities that remain difficult to achieve stand out more clearly with each passing year, highlighting the questions the industry has yet to answer.

Brain Snack (for Builders)

Wednesday Poll

Quick Bits, No Fluff

• Who benefits from the Anthropic crackdown: As the Trump administration cracks down on Anthropic, rivals like OpenAI and Google stand to gain ground, reshaping the competitive map of frontier AI.

• China's green AI hurdles: China's push to power AI projects with green energy is hitting real obstacles, experts say, complicating its bid to lead on both AI and clean power.

• Google's AI mistakes horror fiction for fact: Google's AI Overview presented SCP horror fiction as real, the latest embarrassing reminder that AI summaries still can't tell stories from facts.

Meme Of The Day

The Toolkit

• Dust: No-code platform for building custom AI agents that connect to your company's tools and data, so teams can automate workflows without engineering.

• Krea: Real-time AI image and video generator with a creative-first interface, great for designers who want to actually steer the output instead of fighting prompts.

• Lavender: AI sales email coach that scores your drafts in real time and suggests rewrites to lift reply rates. 

Rate This Edition