Saturday Links: Gemini inside, tabular data LLMs, and robotic hand training

January is already halfway through, but the news keeps rolling. Here is this week's crop of interesting AI links:

• Google is making AI in Gmail and Docs free — but raising the price of Workspace. Didn't fancy paying $20 extra to get Goggle Gemini in Google Workspaces? (Let alone didn't want the hoards of pop-ups that come with it?). Google answered the question for you this week by bundling in Gemini access for free and raising the price of all their standard bundles by about 10%. The cynical take in the press has centered on Gemini, maybe not seeing much take up, but no matter the take up, this probably makes sense for Google to do. In one fell swoop, it has upped the bar for what a workspace environment should have. It's also unlikely large percentages of users will make use of AI just yet, so the revenue uplift will probably far outweigh their costs. By the time usage gets widespread, the company will likely have figured out many ways to bring runtime costs down. Lastly, Google will learn even more from legions of users and put pricing pressure on OpenAI and others at the same time. One could argue that this is a bundling strategy that borders on the anti-competitive...
• Robotic exoskeleton can train expert pianists to play faster. Already in use for medical rehabilitation, researchers have shown it's possible to train fast finger movements by using a fine motor-control exoskeleton on the hand. It's hard not to imagine devices like these helping push the edges of what is possible for human dexterity, but one could also imagine that over-use could damage hand function or sensitivity. There must surely be other applications in areas like delicate surgery, art, or bomb disposal.
• Foundation model for tabular data slashes training from hours to seconds. Take this result with a slight grain of salt for now, but there is no doubt an interesting result at the core. Researchers at the University of Freiburg have developed TabPFN, which is a general machine-learning model that is optimized for discovering patterns and correlations in tabular data. Tabular data sets today are often analyzed using a wide range of model types, each of which has to be trained to discover patterns. When using a general model like TabPFN, it may be that patterns can be discovered not by training a model but at inference time. This either means training can be skipped or it means training could be made more efficient with the important correlations already known. The nature paper is here.
• Is software engineering dead in the water? Mark Zuckerberg says mid-level AI engineers might claim coding jobs from professionals at Meta in 2025. Mark Zuckerberg's quote from his Joe Rogan appearance had its viral moment this week. In the quote, he says that larger companies are likely to have AI-based engineers equivalent to mid-level engineers in skill (the video is here). it is obvious that this moment is coming. Whether it is in 2025 and how efficient such systems will really be is much less clear. The other key thing to note is that just because such systems are possible, that doesn't mean they are economical in terms of power usage, supervision, and so on. We're likely to have mixed teams of humans and AI for a long time. As a (human) engineer, though, it seems abundantly clear that the more you understand about business, product, strategy, and big-picture topics such as architecture, the better. The race to become a senior or principle engineer and stay on the payroll is on. It also means that much of the code we have wanted to write but haven't had the budget to might get written as well.
• Anthropic achieves ISO 42001 certification for responsible AI. Anthropic announced this week that it had achieved accredited status for ISO 42001, which is a relatively new set of ISO guidelines on responsible AI. Anthropic is the first of the leading foundation model providers to achieve the certification. There are others who have received the stamp, including Amazon, for its bedrock inference services and Q model/agents. The interesting point about this announcement is what is contained in the recommendation. The standard really focuses on Transparency, Accountability, Data privacy protection, Reliability, and structural / process-related provisions. The standard also makes reference to fairness, explainability, and safety, however. It is really hard to imagine that anything with any confidence could be stated about any of Anthropic's models in this regard. No doubt they have strong processes for evaluation, but LLMs are inherently unpredictable. It's probably better to think of ISO 42001 as indicating that an organization is doing its best to tame the models, but not as a guarantee of safety.

Wishing you a great weekend!