Recently, I've talked with a few people about artificial intelligence and watched a few presentations. The gap between what some people think are AI capabilities and actual capabilities is a big one.
In the past, on the subject of neural networks, I've touched on recurrent neural networks (RNNs) and convolutional neural networks (CNNs), but the "new" kid on the block are transformers. These were introduced back in 2017 in a paper by Vaswani et al titled "Attention Is All You Need".
- Transformers are a type of neural network based around the ideas of self-attention, positional embedding and multi-headed attention that provides another way of building deep learning tasks like machine translation, text generation and sentiment analysis. Briefly, self-attention allows a model to determine connections or relationships between different things, even if they are far apart from each other in apparent importance and supporting context. The other two elements speed up the processing of the recurrent neural network, like those based on the GPT (generative pretrained transformer) approach. It has added to the ability to process languages that we can now find in newer mobile phones, as one example.
- The newer capabilities can now analyse the sentiment of a text to a certain degree, and create response based on that analysis. Does this mean AI systems have emotions? No, just that the latest models are starting to try and analyse them in documents. Are AI's starting to think for themselves yet? No, but there are those who keep trying to see this in the latest model behaviours. Since the transformer models are looking for connections, they can sometimes make associations they have not been specifically trained for by creating internal models as information is gathered. A little along the lines of the "what happens when I do this" approach. This may form the basis of self-aware creations sometime in the future, but while the current developments are interesting, we are not there yet. As far as I know at least.
- In other AI news Anthropic, which has the Claude AI model, has been sued by three authors. The claim is that Anthropic used the authors' copyrighted works to train their AI model, along with the claim that "Anthropic has built a multibillion-dollar business by stealing hundreds of thousands of copyrighted books". This without permission or paying anything to the authors for their contributions. In one example, Tim Boucher has apparently generated 97 books as of May 2023 using Claude and ChatGPT, all in less than a year. They sell for between US$1.99 (about 68 baht) and $5.99. So someone is making money from the products, just not the authors whose works contributed to the process. There are other similar cases working their way through the legal system and it will be interesting to see how it turns out.
- How safe is that digital wallet? According to a recent paper "In Wallet We Trust: Bypassing The Digital Wallets Payment Security For Free Shopping", digital wallets like Apple Pay, Google Pay and PayPal can be used to conduct transactions using stolen and cancelled cards. To be fair, some of the holes have been patched since the release of the paper, but essentially someone can add a stolen credit card to a digital wallet and even if it is cancelled and replaced, continue to make purchases. The authorisations stay with the wallet so they can be used beyond the time you'd expect.
- I have a friend who is always playing up the chip-making capabilities of China. I smile and nod but in reality, China is still a couple of years behind the USA for logic chips and even more when it comes to memory chips. China lags in the lithographic techniques, as much as five generations if some sources are to believed. The Middle Kingdom will not sit still however, and global manufacturers had better keep moving forward or China will eventually catch up. China is currently beating both Japan and the US in patent filings. China's ultimate goal is to achieve semiconductor self-sufficiency in all aspects of the industry. Back in 2015, the nation set a goal to be 70% self-sufficient in semiconductors by 2025. According to the Information Technology and Innovation Foundation, it will get to 30% by that year but will need an additional $1 trillion of investment to make it all the way.
- When is a security patch not a security patch? When things don't work after applying it. If you haven't guessed already, the latest Microsoft security update, conveniently named CVE-2022-2601, had intended to fix an issue in the GRUB 2 bootloader. This supports the dual boot of Windows and Linux. After the patch users reported the Linux version would no longer boot and received the message: "Verifying shim SBAT data failed: Security Policy Violation. Something has gone seriously wrong: SBAT self-check failed: Security Policy Violation." After disabling Secure Boot things start to work, but that defeats the whole purpose of the protection. Microsoft said it was working with its Linux partners to fix the problem.
James Hein is an IT professional with over 30 years' standing. You can contact him at jclhein@gmail.com.