X and Visa partner for socmed payments, Starlink txt-ing coming for iPhones, and OpenAI complains about Chinese IP theft

Twitter-X seems set to launch the payments function its owner and PayPal co-founder Elon Musk talked about some time ago, later this year and in partnership with Visa. How exactly it'll work remains to be seen but it could be similar to China's WeChat Pay, with a digital wallet, and United States only, at least to start with.

We’re excited to partner with @XMoney on the launch of X Money Account.

Visa Direct will make it possible for US X Money Account users to fund and transfer money in real-time with their debit card. https://t.co/h1t0ofX9NE

— Visa (@Visa) January 28, 2025

The real time payments functionality will no doubt be of interest to some, but even with Visa's backing, success isn't guaranteed for X. Ask Goldman Sachs, which is now trying to get out of its Apple Card partnership as it turned into a loss maker for the bank. Apple Card uses the MasterCard payments network.

We'll see how the XMoney card fares, if people trust a stagnating social network with their money, and are willing to overlook Musk's political leanings, that penchant for doing "Roman salutes" in public and openly supporting Germany's far-right AfD party. TSLA investors are reportedly not happy with the vehicles being called "swasticars"; neither is Poland over Musk suggesting that Germans should just forget about their past.

One NZ Apple users might get Starlink Satellite TXT service

Speaking of Musk, his SpaceX company has been working with Apple to figure out why the Direct to Cell or Direct to Device texting service doesn't work. One NZ had to launch Starlink-based Satellite TXT recently without iPhone support as a result, which given Apple's sizeable market share locally took some wind out of the sails for the telco.

Bloomberg's Apple guru Mark Gurman reported that US telco T-Mobile has notified users that they're in the Starlink D2D beta, with support for the technology being added to the iOS operating system version 18.3 that was released recently.

Yours truly's iPhone 16 Pro Max has been updated to iOS 18.3, but no satellite setting is to be found on the device. Ah well. All will be revealed at some point, but meanwhile, where does that leave Apple's GlobalStar satellite investment?

OpenAI called out as hypocrites

Tech stocks are see-sawing thanks to DeepSeek releasing its V3 and R1 artificial large language models (LLMs) which, despite being censored, seem pretty good and efficient overall. The big thing about that AI drop was DeepSeek suggesting the models were trained at a relatively low cost, and without using a huge amount of expensive, state-of-the-art graphics cards like Western competitors do.

Instead, DeepSeek claimed it used just 2048 Nvidia H800 GPUs to train the big V3 model, costing around US$5.6 million. It could be true, or China has managed to obtain on the black market some 50,000 Nvidia H100 cards, which are way more powerful (and expensive), and they were used to train the DeepSeek LLMs. 

That could just be scurrilous rumours, because Microsoft-backed OpenAI now says it has evidence that DeepSeek used "distillation" to train its models cheaply. As in, DeepSeek used OpenAI's application programming interfaces (APIs) to train its own models, which might indeed be a quick and cost-efficient way of transferring knowledge from existing AIs.

You get the human input added to the models with distillation for example, something which is otherwise slow and cumbersome, and perhaps save the US$2 an hour OpenAI reportedly paid Kenyan workers to filter out horrendous stuff from ChatGPT responses.

It only took the smallest amount of time before OpenAI and its chief executive Sam Altman were being derided as utter hypocrites for complaining about DeepSeek's alleged intellectual property theft. OpenAI is being sued left, right and centre for copyright infringement, accused of snarfing data to train its models without permission or compensation. 

As the above plays out, Alibaba which like Amazon runs a big cloud computing division, has released the Qwen2.5-Max which it reckons beats DeepSeek and a bundle of US LLMs. It's only just been released, but a quick trial run suggests Qwen2.5 is good, able to generate images and video, has soon-to-come web search and more. Goodness knows what it was trained on though.

The whole circus highlights fundamental problems with AI which in turn suggests that investors really do need to be cautious. More AI surprises coming up, no doubt, but in the meantime 新年快樂!