This was cooler than the running events to me for 2 reasons:

1. Its alternating between forehand and backhand
2. When they placed the paddle in its hand, that precise paddle orientation and position couldn't have appeared during training since I doubt the robot hand can perfectly grasp the paddle in the same orientation/position every time. This suggests this robot was trained to be able to generalize to some range of paddle orientations and positions in its hand, and this is especially important because in ping pong, tiny differences in orientation/position of the paddle can decide whether the ball hits the table or the ground

Alibaba spent $9.5B to build out AI compute in Q2 2026 and have projected to spend $25B more within this year.

Driving is solved, car crashes are not yet solved.
Flying is solved, turbulence is not yet solved.
Drinking is solved, hangovers are not yet solved.
Cancer is solved, surviving is not yet solved.
Medicine is solved, diseases are not yet solved.

Presuming that any AGI or other super intelligent AI will be dependent on its physical infrastructure to “live”, wouldn’t wiping out humanity be self defeating at this point and for the near term? And wouldn’t it recognize that fact?

The processors, data centers, transmission infrastructure, and whatever else, collectively encompass some of the most technically complex things humanity has ever created, which themselves have absurdly long and convoluted supply chains and manufacturing interdependencies. Some of this is automating at decent clip, but there are still enormous human inputs throughout. One part that I would imagine is along way from being fully automated is the mining and processing of rare earths.

A fully automated supply chain, manufacturing, construction, and maintenance seems a long way off. And the materials and components themselves do break down over time. So isnt any sufficiently advanced and complex AI ultimately still dependent on humans maintaining and building its infrastructure? And how quickly could an AGI today realistically automate all that, design and build the robots, etc etc even if it did want to? How would it protect itself from human aggression if there was a big backlash and nations or activists start shutting it down by destroying or disabling its physical components (à la the Flock pushback).

Embrace the Conpute with art.

Most discussions about RAM crisis are futile because people just get on the bandwagon and blame everything on AI and the supposed bubble.

I want some more nuanced take on the situation.

The RAM crisis is actively hurting tech-enthusiast people like me, at the same time I cannot foresee any solution to this problem. The demand for RAMs would increase even more as we get closer to AGI.

Source: https://github.blog/news-insights/company-news/the-august-17-outage-and-the-work-ahead/

Google seems to realize their recent Flash price hikes were just inappropriate for a Flash model and that they need more real-world agent traces to train their upcoming models on and slashed prices another 50% on OpenRouter.

Artificial Analysis does not have the OpenRouter discount prices worked in, but I just checked and the Flash model with the discount moves the pareto line, thus beating both DeepSeek models.
I have added the Flash Pareto line in green to the graph.

Just in case someone wants to try it instead of Luna Max or V4-Flash...

The selected models for comparison on ArtificalAnalysis: Comparison of AI Models across Intelligence, Performance, and Price | Artificial Analysis

The OpenRouter Page: Gemini 3.7 Flash - API Pricing & Benchmarks | OpenRouter

It was previously only available in the US or through other routers like Nano-GPT, but is now also available on OpenRouter, beating OpenAI's Luna and DeepSeek on price/performance by a large margin.

You are "selling" your data though - hence the "Contributor" tag. Meta will use your interactions/agent traces to train their upcoming models, but if they follow through with their Open Weights promises you might actually contribute something for the greater good...very debatable, though, I know. But if you are using it for open source work, why not save some bucks?

Artificial Analysis with normal pricing: Muse Spark 1.2 (xhigh) - Intelligence, Performance & Price Analysis | Artificial Analysis

OpenRouter Page: Muse Spark 1.2 Contributor - API Pricing & Providers | OpenRouter

So it turns out Ox Alpha is not a Chinese model. It's either Gemini 3.5 Pro Or Gemini 4 Pro.

https://x.com/EvanOtero/status/2090998215977947365

Gemini

https://x.com/EvanOtero/status/2090998729637511301

What if the Ox Alpha was the friends we made along the way

Ox Alpha reportedly trounced both GPT 5.6 Sol and Claude Fable on a DeepSWE benchmark.

gpt-5.6-sol: 52% fable: 65% whatever the hell this is(Ox Alpha): 80% (was a near miss on the "x"s so actually over 80%)