As Tina walks onto the stage in front of hundreds of people she is beaming. She's collecting her British Sign Language (BSL) certificate which is the culmination of a journey that began with tragedy. Learning BSL has helped me say words that I cannot speak, she says. In 2018, while returning from a holiday, Tina fell down a flight of stairs and was in a coma for six weeks. The accident caused a traumatic brain injury that dramatically changed her life, leaving her struggling to speak.

How cost-effectively can strong reasoning abilities be achieved in language models? Driven by this fundamental question, we present Tina, a family of tiny reasoning models achieved with high cost-efficiency. Notably, Tina demonstrates that substantial reasoning performance can be developed using only minimal resources, by applying parameter-efficient updates during reinforcement learning (RL), using low-rank adaptation (LoRA), to an already tiny 1.5B parameter base model. This minimalist approach produces models that achieve reasoning performance which is competitive with, and sometimes surpasses, SOTA RL reasoning models built upon the same base model. Crucially, this is achieved at a tiny fraction of the computational post-training cost employed by existing SOTA models. In fact, the best Tina model achieves a >20\% reasoning performance increase and 43.33\% Pass@1 accuracy on AIME24, at only \$9 USD post-training and evaluation cost (i.e., an estimated 260x cost reduction). Our work reveals the surprising effectiveness of efficient RL reasoning via LoRA. We validate this across multiple open-source reasoning datasets and various ablation settings starting with a single, fixed set of hyperparameters. Furthermore, we hypothesize that this effectiveness and efficiency stem from LoRA rapidly adapting the model to the structural format of reasoning rewarded by RL, while largely preserving the base model's underlying knowledge. In service of accessibility and open research, we fully open-source all code, training logs, and model weights \& checkpoints.

Generative artificial intelligence (GenAI) has made significant progress in understanding world knowledge and generating content from human languages across various modalities, like text-to-text large language models, text-to-image stable diffusion, and text-to-video Sora. While in this paper, we investigate the capability of GenAI for text-to-model generation, to see whether GenAI can comprehend hyper-level knowledge embedded within AI itself parameters. Specifically, we study a practical scenario termed train-once-for-all personalization, aiming to generate personalized models for diverse end-users and tasks using text prompts. Inspired by the recent emergence of neural network diffusion, we present Tina, a text-conditioned neural network diffusion for train-once-for-all personalization. Tina leverages a diffusion transformer model conditioned on task descriptions embedded using a CLIP model. Despite the astronomical number of potential personalized tasks (e.g., $1.73\times10^{13}$), by our design, Tina demonstrates remarkable in-distribution and out-of-distribution generalization even trained on small datasets ($\sim 1000$). We further verify whether and how \Tina understands world knowledge by analyzing its capabilities under zero-shot/few-shot image prompts, different numbers of personalized classes, prompts of natural language descriptions, and predicting unseen entities.

Feeld Notes is a column about a middle-aged woman who suddenly realizes she wants to have sex again--and the beguiling app she uses to do it. The first man I had sex with in the decade since my divorce was not so much a man as, well, a boy. He was 29 years old, with a lean torso, olive-brown skin, and dark hair and eyes. He was more than 20 years younger than me. His name was Enrique, and like many of us on the app where we met, he looked different in his photographs than he did in real life.

The scale and pace of scientific endeavors to develop a vaccine for Coronavirus is unprecedented. There are already 25 different candidate vaccines in clinical trials around the world according to the World Health Organization (WHO) as of July 2020. Oxford University and AstraZeneca have recently announced promising early-stage results, claiming a vaccine might even be available later this year. AI is playing two important roles in the quest for an effective vaccine: firstly, by analyzing and understanding viral protein structures to guide the elements of a vaccine; and secondly, by helping researchers find relevant research papers that are being published at an accelerating rate. Around the world, organizations have created AI tools, shared data sets and research results, and then shared them freely with the global scientific community to help them find the papers relevant to their specific research, to review the breadth of recent findings, and uncover insights.

If a sci-fi film ever names a planet Tina, blame Janelle Shane. Her hobby is training neural networks on data sets to create amusing names, be it rescue kittens ("Mag Jeggles" and "Snox Boops"), Pokémon ("Tortabool"), and -- perhaps most famously -- paint colours ("Sudden Pine" and "Turdly"). In the latest round, the neural network was trained on 700 planets from Star Wars. That's worth doing, she says, because "Kepler-452b" isn't catchy enough to shout out to "the ship's engineer during a raging ion storm", and there's thousands of exoplanets that need better names. After feeding in the data supplied by a fan, her neural network spat out better suggestions, such as "Bartan", "Vantos" and "Nananon".

From MOOCs (Massive open online courses) to the use of iPads in schools, there's no doubt that technology is changing the way that we learn. Chatbots could have a similar effect, too -- by offering a means by which children can better interact with the subjects they're studying. That was the conclusion reached by UK-based tech company rehabstudio. During a recent hackathon event, they came up with the idea of creating an "edubot" that would enable kids to ask questions to a tyrannosaurus rex (called Tina) using Facebook Messenger. The finished product was a collaboration between rehabstudio (providing the tech) and National Geographic Kids (providing the data.)

'Tina the Turner' cheese robot helps keep the wheels of cheddar production in motion British dairy farmers have become the first in the world to employ a robot to turn their award-winning cheddar cheese. Nicknamed Tina the Turner, the robot is a modified Swiss machine, used to turn Comte. Say'cheese': Cheesemakers Richard and Tom Calver (pictured) of Somerset have employed the cheese cheddar handling robot to automate the vital turning process, which helps the cheese to mature No comments have so far been submitted. Why not be the first to send us your thoughts, or debate this issue live on our message boards. By posting your comment you agree to our house rules.