Speaking the same language 🗣️

Brewing Java

Perhaps we’ve become so used to the times in which we live, but what can seem groundbreaking one minute can quickly fall out of fashion the next. (My sympathies if you spent money on a Bored Ape Yacht Club NFT.)

Unfortunately, some tech becomes so ingrained that moving on isn’t always simple - even if you want to. I mean, just ask all those sys admins still having to jerry-rig networks because of a legacy Windows XP app.

But an archaic operating system isn’t the only problem out there - sometimes it’s a programming language.

Common Business Oriented Language (COBOL) dates back to the late 1950s - a time when computers were the size of vending machines (and about as smart). But as tech progressed, COBOL hitched a ride, making its way into countless organizations. And so, much like the XP situation, rather than ditching COBOL, many companies just built on top it.

But as might be expected, there fewer and fewer people working with COBOL these days; COBOL vendor Micro Focus puts the number at 2 million - which is significantly lower than the roughly 10 million coding with Java.

Now IBM says it has a fix for those that want to drag their COBOL apps into the 21st century. Code Assistant for IBM Z is an upcoming AI-powered tool that promises to translate COBOL into Java code.

According to IBM Research chief scientist Ruchir Puri, Code Assistant’s model was trained with 1.5 trillion tokens and has 20 billion parameters, which should provide for a “more efficient COBOL to Java transformation.”

TechCrunch notes the tool will be commercially available in the coming months. So can we now get some AI brains on the Windows XP situation?

Why it matters:

The cost of re-architecting legacy systems can be prohibitive, but so can the expense of maintaining them. And considering poorly cared-for networks are catnip for hackers, an AI tool that can mitigate such risks may prove to be a big deal for the more than 40,000 companies still using COBOL.

Say what?

Meta is continuing to elbow its way into the AI arena.

After rolling out its LLaMA 2 large language model and its AudioCraft music tools, Meta is going after translations.

Meta, the company that has just shy of 4 billion users around the world, has announced SeamlessM4T, which it calls “the first all-in-one multimodal and multilingual AI translation model.” Or in lay terms, this is the kind of technology Captain Kirk would rock so he could talk with aliens.

The multimodal aspect of SeamlessM4T breaks down as follows; it can interpret speech and text, and then output it in another language - again as speech or text.

As things stand, SeamlessM4T can output almost 100 languages as text, and about 35 as speech. (No word if Klingon is covered.)

Perhaps the most interesting aspect of this AI wizardry is that Meta says that, when a multilingual person swaps between languages while talking, SeamlessM4T can still keep up. It’s clear to see how this feature might be useful in countries where multiple languages are commonly spoken, or even immigrant households where someone may code-switch between their parents and siblings.

As with its previous AI models, Meta is making SeamlessM4T publicly available (this time under a research license), along with a dataset of about 270,000 hours of text and speech audio. And for those that want to give it a shot, Meta has set up a Hugging Face space here.

Why it matters:

Translation tools have existed for a while (Google Translate, Microsoft Translator), and they can be fine for dealing with foreign-language restaurant menus or working out if you just inadvertently offended someone by wishing they would fall down a well, rather than wishing them well. But with AI under-the-hood, it could be exciting to see just how far translation technology might go.