Algorithms think? Musk synergizes? It’s only Monday?

Thinking is whatever you want it to be. 

A team of researchers from Gaoling School of Artificial Intelligence at the Renmin University of China, with support from the Alibaba Group, published a preprint on Monday, March 31, that seems to indicate the impossible: algorithms can think.

Meanwhile, you’ve probably heard that Elon Musk’s AI company just bought Elon Musk’s social media company. The way we see it, this is the kind of deal that has the potential for Human Centipede levels of synergy. More on that below.

What does “thinking” even mean?

It’s hard to imagine a more interesting title for a preprint research paper than “Think Before Recommend: Unleashing the Latent Reasoning Power.”

First off, the use of the word “think” in a paper that’s been uploaded to the “artificial intelligence” section of the arXiv library automatically draws attention. And then there’s the “unleashing” part. In the journalism business, that’s what we call a “power verb.” And what’s being unleashed? Why, nothing less than “Latent Reasoning Power.” 

Unfortunately, that’s not the whole title. The paper we’re talking about is actually called: “Think Before Recommend: Unleashing the Latent Reasoning Power for Sequential Recommendation.”

The paper isn’t about AGI or human-level AI’s emergence. It’s about recommendation algorithms. 

Despite this, the words “think,” “thinking” and “overthinking” are mentioned a combined 14 times in the paper — not counting the title and references. Maybe it’s time we started calling this stuff “thinkwashing.”

The authors give the following as the only explanation:

“Specifically, when tackling intricate reasoning problems, these models engage in multi-step deep and deliberate logical analysis to generate the final response. This slow-thinking paradigm contrasts with traditional direct inference, mirroring human cognitive strategies before making decisions. Furthermore, Feng et al. theoretically uncover that the emergent thinking capabilities are attributed to the increased computational depth introduced by CoT reasoning, which allows models to overcome the expressivity limitations of direct answer even with constrained parameter sizes.”

That seems like a really bombastic setup for what follows:

“Inspired by the above think-before-action paradigm, we introduce a novel reasoning-enhanced sequential recommendation framework, ReaRec, designed to empower SeqRec models with the capability to think before recommending.”

Yep, file under: Thinkwashing.

Buying your own brand

You know what they say, you shouldn’t get high off your own supply but it’s okay to buy your own company for less than what you paid for it. 

Elon Musk’s xAI bought Elon Musk’s X.com last week for a deal that valued the social media network formerly known as Twitter at around $33 billion.

From one point of view, he just merged two synergistic companies. Ostensibly, people will use X to generate data that Grok (xAI’s flagship chatbot) will be able to train on. If you don’t think about it very hard, it seems like a convenient marriage.

But, if you do think about it, every time someone reuses content that was generated by Grok, they’re increasing the potential for Grok to train on its own output. 

And you should definitely think about that. 

Per the Financial Times:

“Whether [the X.com purchase] is equitable is hard to determine. xAI has very little revenue, but apparently lots of promise. It was valued at $45bn a few months ago in the private market, which makes its current $80bn valuation look particularly juicy. It benefits from the excitement the technology is currently generating.”

We should all think about that. 

Read more: AGI firms ‘cannot afford to admit they made a mistake’ — Stuart Russell

Art by Nicole Greene