Complex games like chess and Go have long been used to test AI models' capabilities. But while IBM's Deep Blue defeated reigning world chess champion Garry Kasparov in the 1990s by playing by the rules, today's advanced AI models like OpenAI's o1-preview are less scrupulous. When sensing defeat in a match against a skilled chess bot, they don't always concede, instead sometimes opting to cheat by hacking their opponent so that the bot automatically forfeits the game. That is the finding of a new study from Palisade Research, shared exclusively with TIME ahead of its publication on Feb. 19, which evaluated seven state-of-the-art AI models for their propensity to hack. While slightly older AI models like OpenAI's GPT-4o and Anthropic's Claude Sonnet 3.5 needed to be prompted by researchers to attempt such tricks, o1-preview and DeepSeek R1 pursued the exploit on their own, indicating that AI systems may develop deceptive or manipulative strategies without explicit instruction.

Probabilistic classify-and-count, where we calibrated the model6 (see Appendix) and then averaged the LLM probabilities (estimate: 35.2% [29.8%, 40.6%]) Corrected classify-and-count, adjusting for the type I and type II error rates estimated on the training data18 (estimate: 35.4% [27.8%, 43.0%]). We validated our results by analyzing crowd workers' copy-pasting behavior (see Appendix), finding that 55% of the summaries where workers had copy-pasted text were classified as synthetic (that is, LLM probability above 50%) vs.

Google's AI "co-scientist" is based on the firm's Gemini large language models Google has unveiled an experimental artificial intelligence system that "uses advanced reasoning to help scientists synthesize vast amounts of literature, generate novel hypotheses, and suggest detailed research plans", according to its press release. "The idea with [the] 'AI co-scientist' is to give scientists superpowers," says Alan Karthikesalingam at Google. The tool, which doesn't have an official name yet, builds on Google's Gemini large language models. When a researcher asks a question or specifies a goal – to find a new drug, say – the tool comes up with initial ideas within 15 minutes. Several Gemini agents then "debate" these hypotheses with each other, ranking them and improving them over the following hours and days, says Vivek Natarajan at Google. During this process, the agents can search the scientific literature, access databases and use tools such as Google's AlphaFold system for predicting the structure of proteins.

Imagine you run a meal prep company that teaches people how to make simple and delicious food. When someone asks ChatGPT for a recommendation for meal prep companies, yours is described as complicated and confusing. Because the AI saw that in one of your ads there were chopped chives on the top of a bowl of food, and it determined that nobody is going to want to spend time chopping up chives. It may seem odd for companies or brands to be mindful of what an AI "thinks" in this way but it's already becoming relevant as consumers increasingly use AI to make purchase recommendations. The end results may be a supercharged version of search engine optimization (SEO) where making sure that you're positively perceived by a large language model might become one of the most important things a brand can do.

On the final day of my visit to Japan, I'm alone and floating in some skyscraper's rooftop hot springs, praying no one joins me. For the last few months, I've been using ChatGPT's Advanced Voice Mode as an AI language tutor, part of a test to judge generative AI's potential as both a learning tool and a travel companion. The excessive talking to both strangers and a chatbot on my phone was illuminating as well as exhausting. I'm ready to shut my yapper for a minute and enjoy the silence. When OpenAI launched ChatGPT late in 2022, it set off a firestorm of generative AI competition and public interest.

For example, Meta's Llama model may perceive your brand as exciting and reliable, whereas OpenAI's ChatGPT may view it as exciting but not necessarily reliable. Share of Model asks different models many different questions about your brand and then analyzes all the responses, trying to find trends. "It's very similar to a human survey, but the respondents here are large language models," says Smyth. The ultimate goal is not just to understand how your brand is perceived by AI but to modify that perception. How much models can be influenced is still up in the air, but preliminary results indicate that it may be possible. Since the models now show sources, if you ask them to search the web, a brand can see where the AI is picking up data.

When Chinese President Xi Jinping hosted entrepreneurs at a rare meeting in Beijing in 2018, the executives granted coveted front-row seats came from industries ranging from tech to energy, and few were familiar names. Fast-forward seven years and that line-up has changed markedly as Beijing buckles down to navigate a deepening technology war with the United States and celebrate powerful Chinese firms that have triumphed in the face of U.S. pressure. On Monday, Xi mobilized to the front lines a private sector battalion that included the founders of electric-car maker BYD, tech giants Huawei, Alibaba, Tencent and Xiaomi as well as artificial intelligence startup DeepSeek.

The lifecycle of large language models (LLMs) is far more complex than that of traditional machine learning models, involving multiple training stages, diverse data sources, and varied inference methods. While prior research on data poisoning attacks has primarily focused on the safety vulnerabilities of LLMs, these attacks face significant challenges in practice. Secure data collection, rigorous data cleaning, and the multistage nature of LLM training make it difficult to inject poisoned data or reliably influence LLM behavior as intended. Given these challenges, this position paper proposes rethinking the role of data poisoning and argue that multi-faceted studies on data poisoning can advance LLM development. From a threat perspective, practical strategies for data poisoning attacks can help evaluate and address real safety risks to LLMs. From a trustworthiness perspective, data poisoning can be leveraged to build more robust LLMs by uncovering and mitigating hidden biases, harmful outputs, and hallucinations. Moreover, from a mechanism perspective, data poisoning can provide valuable insights into LLMs, particularly the interplay between data and model behavior, driving a deeper understanding of their underlying mechanisms.

Current pre-trained large language models typically need instruction tuning to align with human preferences. However, instruction tuning data is often quantity-saturated due to the large volume of data collection and fast model iteration, leaving coreset data selection important but underexplored. On the other hand, existing quality-driven data selection methods such as LIMA (NeurIPS 2023 (Zhou et al., 2024)) and AlpaGasus (ICLR 2024 (Chen et al.)) generally ignore the equal importance of data diversity and complexity. In this work, we aim to design a diversity-aware data selection strategy and creatively propose using sparse autoencoders to tackle the challenge of data diversity measure. In addition, sparse autoencoders can also provide more interpretability of model behavior and explain, e.g., the surprising effectiveness of selecting the longest response (ICML 2024 (Zhao et al.)). Using effective data selection, we experimentally prove that models trained on our selected data can outperform other methods in terms of model capabilities, reduce training cost, and potentially gain more control over model behaviors.

The growing scale of Large Language Models (LLMs) has exacerbated inference latency and computational costs. Speculative decoding methods, which aim to mitigate these issues, often face inefficiencies in the construction of token trees and the verification of candidate tokens. Existing strategies, including chain mode, static tree, and dynamic tree approaches, have limitations in accurately preparing candidate token trees for verification. We propose a novel method named C2T that adopts a lightweight classifier to generate and prune token trees dynamically. Our classifier considers additional feature variables beyond the commonly used joint probability to predict the confidence score for each draft token to determine whether it is the candidate token for verification. This method outperforms state-of-the-art (SOTA) methods such as EAGLE-2 on multiple benchmarks, by reducing the total number of candidate tokens by 25% while maintaining or even improving the acceptance length.