#large-language-models

#large-language-models

Boring people don't listen. They tell their own stories, over and over, and never make any attempt to engage in our story, our lives. And so we avoid them. AI, on the other hand, is a superb listener. So much so that people, particularly teens, are turning to chatbots for companionship. But in doing so, do we run the risk of all becoming the same kind of person, wanting the same kinds of friendships, with the same kinds of interactions? In a word, boring.

David Cohen, Chief executive of the Interactive Advertising Bureau (IAB), has climbed onto the dangerous prediction limb, with his 2026 Predictions, oppulently titled Search Takes a Backseat as Brands Battle Inside AI's Black Box. AI seems to be top-of-mind, as four of the predictions (including the top three) focus on artificial intelligence. We will spill those top three below, but leaving out the editorial content which accompanies each of his forecasts:

The story of technology is the story of continual disruption and displacement. New systems and processes send some skills into obsolescence, opening the way for new skills and workflows. Generative AI has triggered the latest "de-skilling." But chatbot technology isn't only transforming jobs and shifting our relationship with information itself. It is also inviting us to relinquish our cognitive independence and bring about a sort of dispossession that is unprecedented.

Cornell Tech faculty made a strong showing at the 2025 Conference on Neural Information Processing Systems (NeurIPS), held Dec. 2-7 in San Diego, presenting 23 research papers at one of the world's premier gatherings for artificial intelligence and machine learning. NeurIPS draws thousands of scholars and industry leaders each year and is widely recognized as a leading forum for breakthroughs in AI, computational neuroscience, statistics, and large-scale modeling.

There's sloppy science, and there's AI slop science. In an ironic twist of fate, beleaguered AI researchers are warning that the field is being choked by a deluge of shoddy academic papers written with large language models, making it harder than ever for high quality work to be discovered and stand out. Part of the problem is that AI research has surged in popularity.

Allie Miller, for example, recently ranked her go-to LLMs for a variety of tasks but noted, "I'm sure it'll change next week." Why? Because one will get faster or come up with enhanced training in a particular area. What won't change, however, is the grounding these LLMs need in high-value enterprise data, which means, of course, that the real trick isn't keeping up with LLM advances, but figuring out how to put memory to use for AI.

Much of the ongoing discourse surrounding AI can largely be divided along two lines of thought. One concerns practical matters: How will large language models (LLMs) affect the job market? How do we stop bad actors from using LLMs to generate misinformation? How do we mitigate risks related to surveillance, cybersecurity, privacy, copyright, and the environment? The other is far more theoretical: Are technological constructs capable of feelings or experiences?

This challenge is sparking innovations in the inference stack. That's where Dynamo comes in. Dynamo is an open-source framework for distributed inference. It manages execution across GPUs and nodes. It breaks inference into phases, like prefill and decode. It also separates memory-bound and compute-bound tasks. Plus, it dynamically manages GPU resources to boost usage and keep latency low. Dynamo allows infrastructure teams to scale inference capacity responsively, handling demand spikes without permanently overprovisioning expensive GPU resources.

I wasn't expecting a conversation about single cells and cognition to explain why a large language model (LLM) feels like a person. But that's exactly what happened when I listened to Michael Levin on the Lex Fridman Podcast. Levin wasn't debating consciousness or speculating about artificial intelligence (AI). He was describing how living systems, from clusters of cells to complex organisms, cooperate and solve problems. The explanation was authoritative and grounded, but the implications push beyond biology.

Welcome to Vibe Coding Video Games with Python. In this book, you will learn how to use artificial intelligence to create mini-games. You will attempt to recreate the look and feel of various classic video games. The intention is not to violate copyright or anything of the sort, but instead to learn the limitations and the power of AI. Instead, you will simply be learning about whether or not you can use AI to help you know how to create video games.

Tim Metz is worried about the "Google Maps-ification" of his mind. Just as many people have come to rely on GPS apps to get around, the 44-year-old content marketer fears that he is becoming dependent on AI. He told me that he uses AI for up to eight hours each day, and he's become particularly fond of Anthropic's Claude. Sometimes, he has as many as six sessions running simultaneously. He consults AI for marriage and parenting advice, and when he goes grocery shopping, he takes photos of the fruits to ask if they are ripe. Recently, he was worried that a large tree near his house might come down, so he uploaded photographs of it and asked the bot for advice. Claude suggested that Metz sleep elsewhere in case the tree fell, so he and his family spent that night at a friend's. Without Claude's input, he said, "I would have never left the house." (The tree never came down, though some branches did.)

When Quentin Farmer was getting his startup Portola off the ground, one of the first hires he made was a sci-fi novelist. The co-founders began building the AI companion company in late 2023 with only a seed of an idea: Their companions would be decidedly non-human. Aliens, in fact, from outer space. But when they asked a large language model to generate a backstory, they got nothing but slop. The model simply couldn't tell a good story.

Fundamentally, they are based on gathering an extraordinary amount of linguistic data (much of it codified on the internet), finding correlations between words (more accurately, sub-words called "tokens"), and then predicting what output should follow given a particular prompt as input. For all the alleged complexity of generative AI, at their core they really are models of language.

Comedians who rely on clever wordplay and writers of witty headlines can rest a little easier, for the moment at least, research on AI suggests. Experts from universities in the UK and Italy have been investigating whether large language models (LLMs) understand puns and found them wanting. The team from Cardiff University, in south Wales, and Ca' Foscari University of Venice concluded that LLMs were able to spot the structure of a pun but did not really get the joke.

"You have to pay them a lot because there's not a lot of these people for the world," Gomez said. "And so there's tons of demand for these people, but there's not enough of those people to do the work the world needs. And it turns out that these models are best at the types of things those people do."

We have the suspicion that the next step for AI models is to start reaching out into the world and affecting the world more broadly,

LeCun founded Meta's Fundamental AI Research lab, known as FAIR, in 2013 and has served as the company's chief AI scientist ever since. He is one of three researchers who won the 2018 Turing Award for pioneering work on deep learning and convolutional neural networks. After leaving Meta, LeCun will remain a professor at New York University, where he has taught since 2003.

That's a crowded market where even her previous firm, 6Sense, offers agents. "I'm not playing in outbound," Kahlow tells TechCrunch. Mindy is intended to handle inbound sales, going all the way to "closing the deal," Kahlow says. This agent is used to augment self-service websites and, Kahlow says, to replace the sales engineer on calls for larger enterprise deals. It can also be the onboarding specialist, setting up new customers.

L.L.M.s are especially good at writing code, in part because code has more structure than prose, and because you can sometimes verify that code is correct. While the rest of the world was mostly just fooling around with A.I. (or swearing it off), I watched as some of the colleagues I most respect retooled their working lives around it. I got the feeling that if I didn't retool, too, I might fall behind.

I think it's long past time I start discussing "artificial intelligence" ("AI") as a failed technology. Specifically, that large language models (LLMs) have repeatedly and consistently failed to demonstrate value to anyone other than their investors and shareholders. The technology is a failure, and I'd like to invite you to join me in treating it as such. I'm not the first one to land here,