Google hasn't announced Gemini Omni. A reddit user just found it anyway. Someone opened their Gemini app, got a pop-up for a model they'd never heard of, and started generating video. What came out has been making rounds on Reddit for the past few days, and the reaction has been "this is scary." The chalkboard video is why.

Most models quit around submission 30 because they stop finding improvement and exit on their own. That's what happened when MiniMax ran a CUDA kernel optimization task against a field of frontier models. Every model except two called it done within the first 30 submissions. M3's best result came on submission 145. After 24 hours. After multiple plateaus where the numbers stopped moving and a reasonable model would have concluded there was nothing left to find. That's the thing MiniMax released yesterday. An AI model with a 1M token context window, native multimodality, and apparently a problem with knowing when to stop.

Who should care If you work with math, science problems, or complex coding tasks and you're looking for something small enough to run locally or cheaply via API, this is worth serious evaluation. The benchmark numbers at 760M active parameters are not normal and the Markovian RSA boost means performance scales with compute budget rather than hitting a fixed ceiling. If you're building agent workflows that need reliable tool calling or multi-step instruction following, look elsewhere for now. The agentic numbers are honest about that gap. Researchers working on test-time compute methods will find the Markovian RSA implementation worth studying regardless of whether they deploy the model itself. The co-design approach — training the model specifically to work with the inference method rather than applying the method after the fact — is an interesting direction that most labs haven't published on at this level of detail. The AMD training story is also worth paying attention to if you care about where the hardware ecosystem goes next. This is the most capable model trained end to end on AMD hardware that anyone has published. That matters beyond just this one release.

OpenAI says one of its internal reasoning models has solved a math problem that has been there on mathematicians' desks since 1946. The problem, first posed by legendary mathematician Paul Erdős, looks almost absurdly simple. Given a set of points on a flat plane, how many pairs can be exactly one unit apart? People have spent nearly 80 years trying to pin down the answer. OpenAI's model didn't just make progress on the problem. According to the company, it disproved a longstanding conjecture that many researchers believed was essentially correct.

Anthropic is adding invisible watermarks to text generated by Claude, and unlike a visible label, the marking is designed to travel with the text when users copy and paste it elsewhere. The move comes as AI-generated content becomes harder to distinguish from human writing and as the European Union begins requiring AI companies to make generated or manipulated content machine identifiable. Anthropic says the marking happens at the model level, meaning it can follow Claude-generated text across different products and surfaces rather than being tied to a particular app. The company also says it may survive some editing. But it raises a question, can AI-generated text actually be made traceable once it leaves the model that created it? And Anthropic's approach suggests the answer may be more complicated than simply adding a hidden signature to every sentence.

Alibaba gave Qwen3.7-Max a kernel optimization task on a hardware platform the model had never encountered before. No documentation or profiling data. No example kernels for the architecture. Just a task description, an existing implementation, and an evaluation script. The model ran for 35 hours. It made 1,158 tool calls. It wrote, compiled, profiled, and rewrote the kernel repeatedly, diagnosing failures, fixing bugs, identifying blocks, and redesigning the architecture multiple times without anyone watching. After 30 hours it was still finding meaningful improvements. The final result was a 10x speedup over the reference implementation. For context: GLM 5.1 ran the same task and reached 7.3x. Kimi K2.6 reached 5x. DeepSeek V4 Pro reached 3.3x. The models that stopped early did so because they issued no tool calls for five consecutive rounds, they concluded they couldn't make further progress and stopped. Qwen3.7-Max didn't stop.

Meta has a new AI model and for the first time in years it is not called Llama. Muse Spark launched yesterday under Meta Superintelligence Labs, a new internal division Meta quietly formed by bringing together researchers from Google DeepMind and other frontier labs. It is natively multimodal, supports multi-agent reasoning, and is available right now at meta.ai. It is also not being released as open weights. That last part is worth sitting with for a second. Meta built one of the most trusted brands in open source AI through Llama. Developers built on it, researchers published with it. Muse Spark continues none of that. No weights, no HuggingFace release, private API preview only. What you get instead is a genuinely capable multimodal model with some benchmark numbers that are hard to ignore and a new reasoning mode called Contemplating that puts it in conversation with Gemini Deep Think and GPT Pro. Whether that trade is worth it depends entirely on what you were using Meta AI for in the first place.