Sep 7 · AI Daily

Viggle thanked @cocktailpeanut for testing feedback, stating that Viggle-Animate performed better than expected, and provided a Hugging Face open access link at https://huggingface.co/Viggle/Viggle-Animate. The cited content states that the tool can generate results from just a single frame without pose/mask/face/depth workflows, exceeding quality expectations.

Meizu AI Xiao Fangkuai rumored for mass production and market launch, featuring UNISOC T8200 processor and 4-inch small square screen

Gary Marcus wrote a piece criticizing Jensen Huang for claiming that the AGI race is over, pointing out that he provided no evidence or definitions, and suggested he refer to agidefinition.AI, which involves Hendrycks, Yoshua Bengio, and others.

Agree with @JensenHuang's view: we are entering the era of AGI. The AGI era must also be the alignment era. We...

I agree that we are in the AGI era of 'jagged AGI' (better than humans at many tasks, worse at others)...

AI can do a task perfectly and still waste your time on the wrong things. Imagine an assistant...

I posted this today because I'm 100% sure frontier AI will figure this out: matter is chiral =&...

We've made some improvements to optimize the user experience for logged-in Astra subscribers in long-tail scenarios...

The author reposted and agreed with Jensen Huang's statement: OpenAI's GPT-6 Astra was trained on about 100,000+ NVIDIA Grace Blackwell NVL72s, taking 4 years from ChatGPT to o1 to Astra, AGI has arrived, and the next 400,000 GPUs will soon come online. The author further speculates that ASI and RSI may be achieved by then.

How many OG Rubin Ultra 1024GB HBMs and new Rubins equal the cost of an F-35 fighter jet...

Jensen Huang reposted and confirmed that GPT-6 Astra was trained using about 100K+ NVIDIA Grace Blackwell NVLink72s, noted that it took 4 years from ChatGPT to o1 to Astra, declared that AGI has arrived, and congratulated the OpenAI team. He also stated that 400K GPUs are coming online next. This is content reposted by the author; both the training scale and the AGI statement originate from Jensen Huang's original text.

Anthropic's $1.5 billion copyright settlement has received final approval, with authors of nearly 500,000 works eligible for $3,000 per pirated work, but many authors received notices this week that others had claimed their funds. The writing community reported that publishers are claiming funds for works whose rights have reverted, or claiming the full amount for shares that should be split 50-50; literary agents are also claiming funds, while Victoria Strauss pointed out that agents are not the rights holders of the works.

The Seattle Times and Newsday are suing OpenAI and Microsoft, accusing both companies of using their news reports as AI training data without permission and reproducing report excerpts when users ask questions. Because Copilot is based on OpenAI technology, Microsoft is also named as a defendant; similar lawsuits have previously been filed by The New York Times, Ziff Davis, Merriam-Webster, Encyclopedia Britannica, and nearly 400 local newspapers.

Distribution of Anthropic's $1.5 billion (approx. 10.102 billion yuan) copyright settlement is underway. The court ruled that training models falls under fair use, but ordered $3,000 in compensation to the original authors for each of nearly 500,000 works obtained through piracy.

Greg Brockman cited Jensen Huang's congratulations, noting that GPT-6 Astra was trained on about 100,000+ NVIDIA Grace Blackwell NVLink72s, took 4 years from ChatGPT to o1 to Astra, and stated that regardless of which model generation is used as the benchmark, we are now entering the AGI era. Jensen Huang also mentioned that 400K GPUs will come online next.

H Company released the NeoMME series of multimodal encoders, comprising two bidirectional Transformer models of 260M and 800M parameters, using a single tower to process multilingual text and 32x32 image patches, removing the independent vision tower and causal decoder, and pre-trained via masked discrete diffusion.

Why are models so bad at harness engineering!?

The author reposted and cited data indicating that 1.5 years ago o1 Pro was priced at $150/$600 per million tokens, whereas today GLM-5.3 Flash is $0.15/$0.50, representing a price drop of about 1,000x while the latter is reportedly smarter. The author concludes that models are becoming smarter while costs drop drastically, and predicts that 12 months from now, intelligence at the GPT Astra level will be nearly free.

According to The Information, Thinking Machines Lab, founded by Mira Murati less than two years ago, is seeking $5 billion to $6 billion in funding at a pre-money valuation of at least $40 billion, with Nvidia reportedly in talks to contribute about $2.5 billion and Accel in talks to lead the round. The company has annualized revenue of at least hundreds of millions of dollars, released open-weight model Inkling in July, and generates profit by helping enterprises customize models using their own data.

Rohan Paul argues that 'time horizon' is becoming a more useful metric for discussing agent capabilities, noting that success rates without intervention drop as tasks get longer, with OpenAI data falling from 86% for tasks under 15 minutes to about 16% for the 64-128 hour tier.