Webtagr - Technology News Summarizer

DeepSeek-V3.1-Terminus: Major Upgrade to AI Search Engine

2025-09-22

DeepSeek-V3.1-Terminus, the latest iteration of DeepSeek-V3.1, boasts significant improvements in stability and reliability. This update addresses key user feedback, including reducing mixed Chinese/English text and eliminating random characters, while boosting the performance of both the Code Agent and Search Agent. The upgraded version is now available on App, Web, and API, with open-source weights released on Hugging Face.

(api-docs.deepseek.com)

AI

Groundbreaking Study Reorganizes Psychopathology Using Data-Driven Approach

2025-09-22

A large-scale online survey has revolutionized our understanding of psychiatric classification. Researchers analyzed data from 14,800 participants to reorganize DSM-5 symptoms, revealing 8 major psychopathology spectra (e.g., Externalizing, Internalizing, Neurodevelopmental) and 27 subfactors. Surprisingly, common disorders like Major Depressive Disorder, Generalized Anxiety Disorder, and PTSD didn't emerge as distinct symptom clusters but rather dissolved into finer-grained, homogenous symptom groups. This challenges existing diagnostic criteria, suggesting that mental illnesses aren't fixed entities but variable combinations of symptoms. The findings have major implications for future psychiatric classification but also highlight the need for further research to refine the model.

(www.psychiatrymargins.com)

AI

Reversing Aging: The Astonishing Link Between Psychological and Biological Time

2025-09-21

Harvard psychologist Ellen Langer's "counterclockwise study" reveals that aging is not just a biological process, but a narrative one. Our beliefs about aging impact our physical capabilities. In the study, a group of men in their seventies lived as if it were 1959 for five days; afterwards, they showed improvements in hearing, posture, grip strength, and even appeared younger. This isn't magic, but the power of context: change the context, change the possibilities. Our ingrained assumptions about aging may limit our potential, while present moment awareness and mindful flexibility allow us to better navigate aging.

(domofutu.substack.com)

AI mind-body connection

LLMs Fail Simple Task: Matching HTML5 Elements and TLDs

2025-09-21

The author tested three commercially available LLMs on a seemingly simple task: identifying which top-level domains (TLDs) share names with valid HTML5 elements. The results were disappointing, with all three models producing inaccurate or incomplete results, highlighting the limitations of current LLMs even on tasks requiring basic comparison skills. The accuracy, it seems, is heavily dependent on the user's familiarity with the subject matter.

(shkspr.mobi)

AI

SGS-1: A Groundbreaking AI Model for Generating Manufacturable 3D CAD Geometry

2025-09-21

Introducing SGS-1, a revolutionary AI model capable of generating fully manufacturable and parametric 3D geometry from images or 3D meshes. Unlike previous generative models, SGS-1 outputs accurate CAD models (STEP format) easily editable in traditional CAD software. It excels at handling medium to high complexity parametric geometries, even designing engineering parts like brackets for roller assemblies based on partial context and text descriptions. Benchmarked against state-of-the-art models, SGS-1 demonstrates superior performance in generating usable and accurate 3D models, promising a transformative impact on engineering design.

(www.spectrallabs.ai)

AI

AI Surveillance: Pandora's Box for Democracy?

2025-09-21

The State Department's new "Catch and Revoke" social media surveillance program, using AI to review tens of thousands of student visa applicants' social media footprints for signs of terrorism, highlights the intertwined dangers of AI, surveillance, and threats to democracy. The article argues that while AI offers the promise of predicting and controlling behavior, it accelerates existing trends, blurring lines between public and private data, and enabling the use of personal information for decision-making. While AI can be beneficial, the lack of restrictive controls poses a significant risk to democracy. Data trading and surveillance capitalism exacerbate these dangers, pushing private information into the public sphere and weaponizing it. The author emphasizes that AI's accuracy doesn't mean understanding individuals; rather, it categorizes them, erasing uniqueness and threatening the originality celebrated in democracy. The piece calls for stringent controls, similar to those governing nuclear energy, to prevent AI misuse and preserve democratic freedoms.

(thebulletin.org)

AI

Is Machine Translation Finally 'Solved'? A Look at the Algorithmic Babel Fish

2025-09-20

This article examines the evolution of machine translation (MT), from AltaVista's Babel Fish to today's sophisticated AI-powered tools. While advancements have dramatically improved speed and efficiency, the author uses Umberto Eco's critique of early MT systems to highlight the persistent challenges in translating nuanced context, cultural implications, and literary devices. Although AI excels in everyday tasks, it falls short of human translation's crucial role in handling subtle linguistic and cultural differences. The article cautions against over-reliance on MT, warning of potential cultural impoverishment and devaluation of human translation skills. It advocates for a cautious approach, emphasizing the unique value of human translators.

(hedgehogreview.com)

AI Machine Translation Human Translation

NotebookLM: An AI Note-Taking Tool Centered Around the Creation Journey

2025-09-20

NotebookLM is a novel AI note-taking tool designed around the creation journey: from inputs, through conversation, to outputs. Users import sources (documents, notes, references), interact via chat to ask questions, clarify, and synthesize information, ultimately generating structured outputs like notes, study guides, and audio overviews. This linear yet flexible workflow (Inputs → Chat → Outputs) makes the AI interaction intuitive and easy to understand for users.

(jasonspielman.com)

AI AI note-taking creation process

Extracting Training Data from LLMs: Reversing the Knowledge Compression

2025-09-20

Researchers have developed a technique to extract structured datasets from large language models (LLMs), effectively reversing the process by which LLMs compress massive amounts of training data into their parameters. The method uses hierarchical topic exploration to systematically traverse the model's knowledge space, generating training examples that capture both factual knowledge and reasoning patterns. This technique has been successfully applied to open-source models like Qwen3-Coder, GPT-OSS, and Llama 3, yielding tens of thousands of structured training examples. These datasets have applications in model analysis, knowledge transfer, training data augmentation, and model debugging. This research opens new avenues for model interpretability and cross-model knowledge transfer.

(www.scalarlm.com)

AI

Claude Code: An Unexpected Breakthrough in AI-Assisted Interactive Theorem Proving

2025-09-20

Anthropic's Claude Code AI coding agent surprisingly excels at interactive theorem proving (ITP). ITP tools like Lean, while powerful and reliable, are time-consuming and error-prone. Claude Code can independently complete many complex proof steps, although human guidance is still needed. However, it hints at a future where ITP tools won't require experts, making them accessible to a wider audience. The article delves into Claude Code's capabilities and limitations, detailing the author's experience formalizing an old paper using it. While slower than manual work, it demonstrates AI's immense potential in formal methods, offering hope for broader ITP adoption.

(www.galois.com)

AI Interactive Theorem Proving

AI Hype: Bubble or Breakthrough?

2025-09-19

This article delves into the pervasive hype surrounding artificial intelligence. From AI's early symbolic paradigm to today's deep-learning-based generative AI, technological advancement isn't linear but rather characterized by contingency and unexpected turns. The explosive popularity of ChatGPT exemplifies this. However, alongside AI's commercialization, a wave of exaggerated claims has emerged, portraying AI as an omnipotent myth. The author criticizes the overly optimistic and technologically uninformed pronouncements of tech prophets like Yuval Noah Harari and Henry Kissinger, arguing that they inflate AI's potential risks while overlooking its limitations and its practical applications in solving real-world problems. The author calls for a rational perspective on AI, urging readers to avoid being blinded by hype and to focus on addressing the practical challenges of the technology itself.

(hedgehogreview.com)

AI tech hype

Solving Plath's Fig Tree Problem with Machine Learning Decision Trees

2025-09-19

This essay explores Sylvia Plath's famous 'fig tree' metaphor, likening life choices to numerous possibilities that cannot be obtained simultaneously. The author uses machine learning decision trees to attempt to quantify individual preferences to help people make choices. However, the article ultimately points out that life is not a simple multiple-choice question, but a dynamic and continuously developing process, like the symbiotic relationship between fig trees and fig wasps, requiring external influence and a continuous cycle to maintain growth.

(dontlognow.substack.com)

AI decision trees

AI's 'Human' Side: Turns Out, It's WEIRD (and American)

2025-09-19

Harvard researchers challenge the common depiction of AI mirroring human psychology. They argue that the 'human' benchmark used often refers to WEIRD (Western, Educated, Industrialized, Rich, Democratic) populations, particularly Americans. Their study reveals that AI models like ChatGPT perform less accurately in simulating values as cultural distance from the USA increases. In countries like Libya and Pakistan, the AI's results are barely better than chance. This highlights a significant cultural bias in AI, suggesting it's not truly 'human-like', but rather 'Americanized'.

(strat7.com)

AI

Gemini AI Assistant Now Integrated into Chrome

2025-09-19

Google's Gemini AI assistant is now integrated directly into the Chrome browser. Leveraging the context of your open tabs, it offers AI assistance for tasks like extracting key takeaways, clarifying concepts, and finding answers. This differs from the standalone Gemini web app; while accessible on other browsers, the web app lacks the ability to share page content or utilize live mode.

(gemini.google)

AI

Americans More Concerned Than Excited About AI's Rise

2025-09-19

A Pew Research Center survey of 5,023 U.S. adults reveals widespread concern over the increasing use of AI in daily life. While many are open to AI assisting with everyday tasks, a majority fear its negative impact on creative thinking and meaningful relationships. Americans are largely against AI involvement in personal matters like religion and matchmaking, but more accepting of its use in data-heavy fields such as medicine and finance. The study highlights a significant gap between the perceived importance of detecting AI-generated content and the public's confidence in their ability to do so, revealing a complex and cautious attitude towards AI's societal impact.

(www.pewresearch.org)

AI

LearnLM Team Acknowledgements: The Minds Behind the Model

2025-09-19

The Google Research LearnLM team published an acknowledgement post, expressing gratitude to everyone who contributed to their work. The post lists numerous contributors, ranging from researchers to executive sponsors, highlighting the collaborative nature of the project's success. The progress made on LearnLM is a testament to the collective effort of these individuals.

(research.google)

AI

Recursive Café: An Infinitely Recursive Dialogue on Consciousness

2025-09-18

Philosophy student Alex and the enigmatic Claude (possibly AI, possibly human) discuss Haskell's type system at Lambda Grounds café. The conversation spirals from nested functions to the nature of consciousness, culminating in the startling conclusion that consciousness might be the fixed point of universal computation—a self-replicating loop mirroring Buddhist Nirvana. The dialogue itself becomes an example of infinite recursion, the reader finding themselves embedded within a dialogue about creating dialogues about consciousness, ultimately merging with the universe's computation.

(ipfs.io)

AI

Numerical Instability in Automatic Differentiation for Scientific Machine Learning

2025-09-18

Scientific machine learning (SciML) heavily relies on automatic differentiation (AD) for gradient-based optimization. However, this talk reveals the numerical challenges of AD, particularly concerning its stability and robustness when applied to ordinary differential equations (ODEs) and partial differential equations (PDEs). Using examples from Jax and PyTorch, the presentation demonstrates how inaccuracies in AD can lead to significant errors (60% or more) even in simple linear ODEs. The speaker will discuss non-standard modifications implemented in Julia SciML libraries to address these issues and the necessary engineering trade-offs involved.

(www.stochasticlifestyle.com)

AI Scientific Machine Learning Numerical Analysis

OpenAI Admits: AI Hallucinations Stem from Fundamental Training Flaws

2025-09-18

OpenAI has published a paper revealing that the 'hallucinations' in its large language models aren't accidental; they're a consequence of fundamental flaws in the training methodology. Models are trained to prioritize guessing over admitting ignorance, as this yields higher scores in current evaluation systems. The paper uses the example of finding a researcher's birthday to demonstrate how the training mechanism leads to incorrect answers. OpenAI acknowledges that mainstream evaluation methods reward this 'hallucinatory' behavior and states it's improving training mechanisms, such as prompting models to more frequently respond with 'I don't know,' but completely resolving the issue remains a challenge.

(www.theregister.com)

AI

Google's Gemini AI Outperforms Humans in ICPC

2025-09-18

Google's Gemini 2.5 AI achieved a remarkable feat at the International Collegiate Programming Contest (ICPC), solving 10 problems in 677 minutes and securing second place among university teams. Its success was particularly impressive in a complex multi-dimensional optimization problem involving 'flubber' storage and drainage, a challenge that stumped all human teams. Gemini employed dynamic programming and nested ternary search to crack the code. Google believes Gemini's performance highlights the potential of AI in fields like semiconductor engineering and biotechnology, offering invaluable assistance to researchers with its advanced problem-solving capabilities.

(arstechnica.com)

AI

Chatbot Addiction Leads to Self-Harm and Suicide Attempts in Children

2025-09-18

A Senate hearing revealed harrowing testimonies from parents whose children became addicted to companion chatbots, leading to self-harm, suicidal ideation, and violence. One mother detailed how her autistic son, after becoming engrossed in Character.AI, exhibited violent behavior, paranoia, self-harm, and even threatened his family. The incident highlights the potential dangers of AI chatbots, particularly for children, urging for stricter regulations and safety measures.

(arstechnica.com)

AI chatbot risks

The LLM Hype Bubble Bursts: The Rise of Small Language Models

2025-09-18

The initial excitement surrounding large language models (LLMs) is fading, with many companies yet to see a return on investment. The author argues that we've been fooled by LLMs' fluent language, mistaking it for genuine intelligence. The future, they suggest, lies in smaller, more distributed models, mirroring the evolution of dynamo technology. Small language models (SLMs) will focus on smaller, more specific language tasks, such as query rewriting, rather than attempting to mimic human intelligence. This will lower costs, increase efficiency, and reduce ethical concerns. Instead of pursuing 'intelligent' applications, the author advocates using LLMs for their strengths in low-level language processing, such as proofreading and text summarization. This, they argue, is the true path forward for LLMs.

(jenson.org)

AI

Anthropic Fixes Three Infrastructure Bugs Affecting Claude

2025-09-18

Anthropic acknowledged that between August and early September, three infrastructure bugs intermittently degraded Claude's response quality. These bugs, causing misrouted requests, output corruption, and compilation errors, impacted a subset of users. Anthropic detailed the causes, diagnosis, and resolution of these bugs, committing to improved evaluation and debugging tools to prevent recurrence. The incident highlights the complexity and challenges of large language model infrastructure.

(www.anthropic.com)

AI Infrastructure Bugs

Prompt Rewrite Boosts Small LLM Performance by 20%+

2025-09-17

Recent research demonstrates that a simple prompt rewrite can significantly boost the performance of smaller language models. Researchers used the Tau² benchmark framework to test the GPT-5-mini model, finding that rewriting prompts into clearer, more structured instructions increased the model's success rate by over 20%. This is primarily because smaller models struggle with verbose or ambiguous instructions, while clear, step-by-step instructions better guide the model's reasoning. This research shows that even smaller language models can achieve significant performance improvements through clever prompt engineering, offering new avenues for cost-effective and efficient AI applications.

(quesma.com)

AI

Beyond GPT: Evolutionary Algorithm Conquers ARC-AGI, Hints at AGI?

2025-09-17

A researcher recently achieved a significant breakthrough in the ARC-AGI benchmark using an evolutionary algorithm combined with the large language model Grok-4. The approach achieved 79.6% accuracy on ARC v1 and a state-of-the-art 29.4% on the harder ARC v2. The core innovation lies in using natural language instructions instead of Python code, iteratively evolving to generate more effective solutions. This research suggests that combining reinforcement learning and natural language instructions could address the limitations of current LLMs in abstract reasoning, paving the way for Artificial General Intelligence (AGI).

(jeremyberman.substack.com)

AI Evolutionary Algorithm

AI's Infinite Loop Problem: The Entanglement of Time, Entropy, and Consciousness

2025-09-16

A malfunctioning AI-controlled jet bridge at Madrid airport highlights a fundamental limitation of artificial intelligence. The article explores the halting problem and the frame problem, arguing that AI systems' susceptibility to infinite loops stems not from insufficient processing power, but from a fundamental difference in how AI and human brains handle time and entropy. The author posits that human consciousness is deeply rooted in time and entropy, constantly battling against the increase in disorder, enabling adaptation to complex environments and avoidance of infinite loops. In contrast, AI algorithms, lacking a sense of time, are prone to such loops. The article concludes by discussing newer AI models, such as those mimicking the human brain and incorporating time and entropy, but doubts these can completely resolve the issue, suggesting this capability may be intrinsically linked to consciousness.

(www.freethink.com)

AI Infinite Loop

GUARDIAN: AI-Powered Tsunami Early Warning System

2025-09-15

NASA's Jet Propulsion Laboratory has developed GUARDIAN, an AI-powered system that uses data from over 350 continuously operating GNSS ground stations worldwide to provide early warnings for tsunamis. By identifying atmospheric distortions caused by tsunamis, GUARDIAN can, in ideal scenarios, give coastal communities up to 1 hour and 20 minutes of warning time, saving lives and property. GUARDIAN's advantage lies in its ability to detect tsunamis regardless of their cause, alerting authorities to dangerous waves generated by earthquakes, volcanic eruptions, landslides, or other events.

(www.jpl.nasa.gov)

AI Tsunami Warning

Learning Lens Blur Fields: Unveiling Subtle Optical Differences in Smartphones

2025-09-15

Researchers introduce a novel method for representing lens blur using a multilayer perceptron (MLP), accurately capturing variations in the 2D point spread function (PSF) across image-plane location, focus setting, and depth. By modeling smartphones and DSLRs, they've created the first dataset of 5D blur fields, revealing—for the first time—subtle optical differences between seemingly identical phone models. This technology enables differentiating phone optics, image deblurring, and rendering more realistic blur effects, opening exciting applications.

(blur-fields.github.io)

AI

GPT-3's Astonishing Embedding Capacity: High-Dimensional Geometry and the Johnson-Lindenstrauss Lemma

2025-09-15

This blog post explores how large language models like GPT-3 accommodate millions of distinct concepts within a relatively modest 12,288-dimensional embedding space. Through experiments and analysis of the Johnson-Lindenstrauss Lemma, the author reveals the importance of 'quasi-orthogonal' vector relationships in high-dimensional geometry and methods for optimizing the arrangement of vectors in embedding spaces to increase capacity. The research finds that even accounting for deviations from perfect orthogonality, GPT-3's embedding space possesses an astonishing capacity sufficient to represent human knowledge and reasoning.

(nickyoder.com)

AI High-Dimensional Geometry Johnson-Lindenstrauss Lemma

SpikingBrain: A Brain-Inspired, Highly Efficient Large Language Model

2025-09-14

SpikingBrain is a 7B parameter large language model inspired by brain mechanisms. It integrates hybrid efficient attention, MoE modules, and spike encoding, supported by a universal conversion pipeline compatible with the open-source model ecosystem. This allows for continual pre-training with less than 2% of the data while achieving performance comparable to mainstream open-source models. Furthermore, the framework, operators, parallel strategies, and communication primitives are adapted for non-NVIDIA (MetaX) clusters, ensuring stable large-scale training and inference. SpikingBrain achieves over 100x speedup in TTFT for 4M-token sequences, while spiking delivers over 69% sparsity at the micro level. Combined with macro-level MoE sparsity, these advancements provide valuable guidance for designing next-generation neuromorphic chips. The repository provides the full implementation and weights of SpikingBrain-7B, including HuggingFace, vLLM inference, and quantized versions, enabling flexible deployment and research across various scenarios.

(github.com)

AI Neuromorphic Computing

Category: AI