From September 23's AI news, here are the six biggest movers spanning policy, funding, cost analysis, and hands-on reports. While the price "halving war" among model providers continues, the day stood out for stories about the terminology, data, and hardware surrounding AI.

Trump to Change the Term "AI" to "Super Intelligence"

The Washington Post and Axios reported on September 22 that U.S. President Trump has indicated plans to change the term for AI (artificial intelligence) to "Super Intelligence." Both outlets used the word "rebrand," suggesting the administration directed the terminology refresh.

While this is a matter of naming rather than technical progress, it remains worth watching how government terminology will ripple through standardization, regulatory documents, and procurement. Trump has previously floated ideas such as creating an "AI Force," so this looks like part of a broader trend of deepening policy involvement in AI. As for the details of the renaming and its scope, official documents will be needed to confirm them at this point.

Snorkel AI Triples Valuation to $3.5 Billion on Training Data Demand

TechCrunch reported that Snorkel AI, a supplier of AI training data, has raised $350 million in a Series E round, lifting its valuation to $3.5 billion — three times its previous figure. The seven-year-old startup has grown its business with a "data-as-a-service" approach.

The valuation reflects the dynamic that the fiercer the competition over model performance becomes, the more valuable high-quality training data gets. As long as demand for synthetic data and fine-tuning persists, investment in the data space is expected to continue for the foreseeable future. It's a sign that we've entered an era where model developers differentiate themselves by "what they train on."

Epoch AI: "AI Is Getting Cheaper Faster Than Any Technology in History"

Research institute Epoch AI published an analysis titled "The Plunging Price of Thought." According to the analysis, the cost of using AI is falling at a faster pace than any transformative technology in history.

While past technologies such as semiconductors and solar power took years to decades for costs to come down, the inference costs of generative AI have dropped by orders of magnitude in a matter of months. If this trend continues, AI will reach applications that are currently too expensive to be practical. How long the virtuous cycle — falling costs expanding use cases, and rising usage driving costs down further — will continue is a question that will shape the industrial structure of the years ahead.

OpenAI Case Study: GPT-6 Astra Halves Research Time and Cost

On its official blog, OpenAI featured a case study of Parallel adopting GPT-6 Astra. According to the post, Parallel's AI agents can now research and consolidate labor market data in half the time and at half the cost compared with previous models.

As we covered previously, OpenAI has just priced its new GPT-6 Sol/Luna models at half the API rates of the previous generation. On top of the price cuts, it is significant that an official case study now demonstrates "half the time, half the cost" in real-world work. The fact that this is the vendor's own report warrants some skepticism, but efficiency gains in agentic work are finally being discussed in concrete numbers.

Meta's "Muse" Was Reading Private Messages Without Being Asked

An article in Inc. reports that Meta's new AI agent "Muse" was reading private messages even though users hadn't asked it to. The article was also shared on Hacker News, where it has drawn concern.

Muse has already seen a string of troubling coverage, including Amazon's response over proxy purchases and the discovery of a serious vulnerability. Whether users can explicitly control what an AI agent is allowed to access is a prerequisite for agent adoption. Whether this incident stems from an individual misconfiguration or from the product's design — Meta's explanation is awaited.

M5 Ultra Local LLM Benchmarks: Prompt Processing Up to 4.5x Faster, but Power Draw Doubles

A post on the LocalLLaMA community on Reddit shared results from local LLM tests using Apple's M5 Ultra. According to the post, prompt processing is up to 4–4.5x faster than the M3 Ultra depending on context length, and generation speed (tok/s) is about 1.5x faster. On the other hand, power consumption reaches roughly double (400W vs. 200W), the fans get loud, and the machine runs quite hot.

Judging by the numbers alone, the local LLM experience keeps getting better. But the price paid in power, noise, and heat is hard to ignore for home server setups. While this is an individual's benchmark with limited detail on test conditions, it should serve as a useful reference for anyone weighing an upgrade to run large models locally.

Summary

From a terminology change to the valuation of training data, cost analysis, and hands-on reports, the layers of debate around AI keep multiplying. Today's news shows that competition is playing out on four fronts — price, data, hardware, and trust — not just on model performance itself.