June 2, 2026
AI launch or messy breakup post?
MAI-Thinking-1
Microsoft’s new AI flex sparks hype, side-eyes, and a full-on website rage fest
TLDR: Microsoft unveiled MAI-Thinking-1 as a serious new in-house AI push, showing off test scores and talking up its future plans. But commenters were more animated about the drama: a hated website design, hints that Microsoft is drifting from OpenAI, and doubts about whether the company’s favorite test results prove much.
Microsoft’s new MAI-Thinking-1 was supposed to be a big “look what we built” moment: benchmark charts, bold claims about strong science and coding performance, and a recruiting pitch for brilliant people to come build the next wave of artificial intelligence. But the real show was in the comment section, where readers instantly turned the launch into a spicy mix of hype, suspicion, and pure annoyance.
The loudest reaction? Not even the model itself — it was the website. One commenter blasted the page for “absolutely disgusting scroll jacking,” complaining that even accessibility mode didn’t save the experience. Ouch. That set the tone for a thread where people seemed half-impressed, half-ready to throw tomatoes. Others saw bigger drama: with Microsoft now showing off its own model, some readers declared the long-rumored split from OpenAI is finally getting real. That turned the launch into less of a product update and more of a celebrity breakup watch.
Then came the shady eyebrow-raise over Microsoft’s claim that the model used “clean” and licensed training data, with one commenter basically yelling, “Shots fired?” at rivals accused of feeding their systems mystery internet sludge. Meanwhile, skeptics weren’t buying every score on the chart, with one grumbling that a favored coding test was “junk” and demanding tougher proof. And in classic launch-day comedy, another user noticed the menu promised seven modes, showed five models, and only let people use four — the kind of messy math that comment sections live for.
Key Points
- •The article reports post-training evaluation results for MAI-Thinking-1 on public STEM and agentic coding benchmarks.
- •It states that comparison numbers for other models were taken from official model cards.
- •Benchmark scores are reported as percentages unless otherwise noted, and unavailable values are marked with dashes.
- •The article references a section labeled Table 2 for pre-training metrics.
- •MAI says it is expanding its compute capacity, working with product teams to reach billions of users, and hiring for development of next-generation models.