Language models trained via outcome-based reinforcement learning (RL) to reason using chain-of-thought (CoT) have shown remarkable performance. Monitoring such a model's CoT may allow us to understand its intentions and detect potential malicious behavior. However, to be effective, this requires that CoTs are legible and faithful. We study CoT legibility across 14 reasoning models, finding that RL often causes reasoning to become illegible to both humans and AI monitors, with reasoning models (except Claude) generating illegible CoTs while returning to perfectly readable final answers. We show that models use illegible reasoning to reach correct answers (accuracy dropping by 53\% when forced to use only legible portions), yet find no correlation between legibility and performance when resampling - suggesting the relationship is more nuanced. We also find that legibility degrades on harder questions. We discuss potential hypotheses for these results, including steganography, training artifacts, and vestigial tokens. These results suggest that without explicit optimization for legibility, outcome-based RL naturally produces models with increasingly opaque reasoning processes, potentially undermining monitoring approaches.

This study demonstrates that GuruAgents, prompt-guided AI agents, can systematically operationalize the strategies of legendary investment gurus. We develop five distinct GuruAgents, each designed to emulate an iconic investor, by encoding their distinct philosophies into LLM prompts that integrate financial tools and a deterministic reasoning pipeline. In a backtest on NASDAQ-100 constituents from Q4 2023 to Q2 2025, the GuruAgents exhibit unique behaviors driven by their prompted personas. The Buffett GuruAgent achieves the highest performance, delivering a 42.2\% CAGR that significantly outperforms benchmarks, while other agents show varied results. These findings confirm that prompt engineering can successfully translate the qualitative philosophies of investment gurus into reproducible, quantitative strategies, highlighting a novel direction for automated systematic investing. The source code and data are available at https://github.com/yejining99/GuruAgents.

Washington, DC – After Elon Musk made an apparent Nazi salute at an inauguration rally for United States President Donald Trump, the Anti-Defamation League (ADL) rushed to defend the SpaceX founder. The self-described anti-Semitism watchdog and "leading anti-hate organization in the world" dismissed Musk's raised arm as "an awkward gesture in a moment of enthusiasm" in a social media post on Monday. Months earlier, however, Jonathan Greenblatt, the head of the staunchly pro-Israel ADL, compared the Palestinian keffiyeh to the Nazi swastika. Activists say the contrast between the ADL's hurried defence of Musk and its efforts to demonise Palestinians and their supporters shows that the group is more focused on silencing voices critical of Israel than it is on fighting anti-Semitism. "The ADL is being crystal clear about where it stands," said Beth Miller, political director at Jewish Voice for Peace (JVP).

For years, computer scientists have worried that advanced artificial intelligence might be difficult to control. A smart enough AI might pretend to comply with the constraints placed upon it by its human creators, only to reveal its dangerous capabilities at a later point. Until this month, these worries have been purely theoretical. Some academics have even dismissed them as science fiction. But a new paper, shared exclusively with TIME ahead of its publication on Wednesday, offers some of the first evidence that today's AIs are capable of this type of deceit. The paper, which describes experiments jointly carried out by the AI company Anthropic and the nonprofit Redwood Research, shows a version of Anthropic's model, Claude, strategically misleading its creators during the training process in order to avoid being modified.

Lisp machines are general-purpose computers designed to efficiently run Lisp as their main software and programming language, usually via hardware support. They are an example of a high-level language computer architecture, and in a sense, they were the first commercial single-user workstations. Despite being modest in number (perhaps 7,000 units total as of 1988[1]), Lisp machines commercially pioneered many now-commonplace technologies, including effective garbage collection, laser printing, windowing systems, computer mice, high-resolution bit-mapped raster graphics, computer graphic rendering, and networking innovations such as Chaosnet.[citation The operating systems were written in Lisp Machine Lisp, Interlisp (Xerox), and later partly in Common Lisp. Artificial intelligence (AI) computer programs of the 1960s and 1970s intrinsically required what was then considered a huge amount of computer power, as measured in processor time and memory space.

The patter of automated machinery fills the air inside wire-basket manufacturer Marlin Steel's bustling factory in a rugged industrial section of this city. Maxi Cifarelli, 25, of Baltimore, peers through safety goggles at a flat screen, her left knee bent and heel resting on her chair. Two years after earning a fine arts degree from Towson University with a specialty in interdisciplinary object design, she now spends her work days working with a personality-free machine with a name to match: a computer numerical control, or CNC, router. With automation poised to sweep through the economy, some fear that it will kill more jobs than it creates. But Cifarelli's experience is the opposite. She befriended automation, instead of fighting it, and she has a job because of it.

The patter of automated machinery fills the air inside wire-basket manufacturer Marlin Steel's bustling factory in a rugged industrial section of this city. Maxi Cifarelli, 25, of Baltimore, peers through safety goggles at a flat screen, her left knee bent and heel resting on her chair. Two years after earning a fine arts degree from Towson University with a specialty in interdisciplinary object design, she now spends her work days working with a personality-free machine with a name to match: a computer numerical control, or CNC, router. With automation poised to sweep through the economy, some fear that it will kill more jobs than it creates. But Cifarelli's experience is the opposite. She befriended automation, instead of fighting it, and she has a job because of it.

In order to be able to view the situation objectively we feel that it would be useful to preface this with a historical review of the development of ideas in this twenty-year-old field. By considering the most important ideas and techniques that are employed in the (currently) best program available, we hope to convince the reader that progress has been very slow despite the multiplicity of programs (and their associated literature) which have appeared since 1950. HISTORICAL REVIEW The most important paper that has appeared on the subject of computer chess is one written by Claude Shannon in 1948 and published two years later (Shannon 1950). Shannon's paper does not describe an actual program, but offers many suggestions for those who are interested in writing one. In this respect Shannon's paper may be compared with one by Jack Good which was also full of sound ideas which could well be included in a successful chess program (Good 1967). Shannon stressed the importance of having a good evaluation function. The features which he considers necessary for inclusion in the evaluation function included material, mobility, five aspects of pawn-structure, four of the positions of pieces, and four of commitments of pieces, attacks and options. He appreciated that such an evaluation function should be used only in the middle-game, and that different principles applied to the opening and endgame phases of chess. He suggested that the values of the coefficients of the function should be determined by'some experimental procedure', and the fact that this statement has never been followed in practice is very surprising.

Virginia Polytechnic Institute It is no disgrace for Scott's program to have lost to Greenblatt's which seems to be the best chess program so far written: it finished one of its games with a brilliant five-move combination.* Judging by the present game Greenblatt's program could play about board 2000 for England. Neither program seems able to form a plan that is naturally expressed-by a description rather than by evaluation functions plus analysis. In the following game the first four moves on each side were played before the machines took over the play, because the ICI, program cannot castle. The move time limits originally agreed were 90 seconds for # xEss and'blitz speed' (5 or 10 seconds per move) for the Greenblatt program, as it was considered that the P-K5, and the game has the character of a French defence difficult to evaluate.