NO/FOMO

Independent AI signal, once a day

The AI briefing worth opening.

ISSUE DATE2025-11-06ENGLISH EDITION
This issue
—
All time
—

Hacker News

6 stories
01

Kimi K2 Thinking, a SOTA open-source trillion-parameter reasoning model

Moonshot AI has reportedly introduced Kimi K2 Thinking, an advanced open-source reasoning model that sets a new state-of-the-art (SOTA) benchmark in its category. Distinguished by its massive scale, Kimi K2 Thinking reportedly boasts a trillion parameters, positioning it among the most formidable large language models developed to date. This development underscores a significant stride in AI research, particularly in the domain of complex reasoning capabilities, which are crucial for addressing intricate problems across various domains. The open-source nature of Kimi K2 Thinking is particularly noteworthy, as it promises to democratize access to cutting-edge AI technology, fostering collaboration and accelerating further innovations within the AI community. The model's focus on advanced reasoning suggests potential applications across diverse fields requiring sophisticated analytical and problem-solving skills, from scientific discovery and engineering to intricate logical tasks and general artificial intelligence. Its release is expected to provide researchers and developers with a powerful, accessible tool to push the boundaries of artificial intelligence and explore new avenues for AI applications.

Kimi K2 ThinkingLarge Language ModelState-of-the-Art AIOpen-source AIAI ReasoningArtificial IntelligenceLarge Language ModelDeep Learning
02

LLMs Encode How Difficult Problems Are

A recent study titled "LLMs Encode How Difficult Problems Are" from ArXiv explores the intriguing hypothesis that Large Language Models (LLMs) inherently encode or represent the difficulty of various computational and cognitive problems within their internal structures. This research delves into how these advanced AI systems, beyond simply providing solutions, might possess an intrinsic understanding or a learned proxy for the inherent complexity of tasks presented to them. The paper likely investigates the mechanisms through which this 'difficulty encoding' manifests, potentially analyzing internal model states, activation patterns, or predictive confidence levels in relation to problem complexity. Such a capability suggests a deeper level of meta-cognition in LLMs, implying they can not only process information but also assess the challenge associated with that processing. The findings could have significant implications for advancing our understanding of AI intelligence, leading to the development of more sophisticated and adaptive AI agents capable of self-assessing task difficulty, optimizing resource allocation, and potentially offering insights into why certain problems are harder than others. This research contributes to the broader field of AI interpretability and cognitive science, shedding light on the nuanced capabilities of modern LLMs.

Large Language ModelsProblem DifficultyAI CognitionModel RepresentationMachine Learning ResearchLarge Language ModelArtificial IntelligenceMachine Learning
03

Show HN: TabPFN-2.5 – SOTA foundation model for tabular data

Prior Labs has announced the release of TabPFN-2.5, an advanced tabular foundation model that sets a new state-of-the-art for tabular data analysis. This latest iteration significantly expands capabilities, now processing datasets with up to 50,000 samples and 2,000 features, marking a fivefold increase over its predecessor, TabPFN v2. TabPFN-2.5 operates as a pretrained transformer, leveraging in-context learning derived from training on over a hundred million synthetic datasets. It delivers state-of-the-art predictions efficiently in a single forward pass, eliminating the need for hyperparameter tuning across both classification and regression tasks. The model also boasts native support for missing values, various feature types including categorical and text, and enhanced robustness against outliers and uninformative features, consistently outperforming traditionally tuned tree-based models.

TabPFN-2.5tabular datafoundation modelstate-of-the-artmachine learningMachine LearningDeep LearningArtificial Intelligence
04

Mathematical exploration and discovery at scale

The article, titled "Mathematical exploration and discovery at scale," examines the transformative potential of applying advanced computational techniques, particularly within the realms of artificial intelligence and machine learning, to fundamentally alter how mathematical research is conducted. It posits that by employing these scalable technologies, the process of mathematical exploration – encompassing conjecture generation, hypothesis testing, and the discovery of new theorems – can be dramatically amplified. This paradigm shift aims to expand the capacity for tackling intricate mathematical problems, enabling researchers to explore vast solution spaces and identify novel patterns and structures that might elude traditional human-centric methods. The discussion highlights a future where computational tools serve as powerful accelerators, not just for verifying existing proofs, but actively contributing to the generation of new mathematical knowledge, thereby pushing the frontiers of discovery and innovation in mathematics.

Mathematical DiscoveryComputational MathematicsAI in MathematicsAutomated ReasoningLarge-Scale ComputationArtificial IntelligenceMachine LearningOthers
05

Show HN: qqqa – A fast, stateless LLM-powered assistant for your shell

qqqa is an open-source, LLM-powered command-line assistant designed to streamline shell interactions by eliminating the need to switch to a browser or ChatGPT for common tasks. Developed with a focus on the Unix philosophy, qqqa offers a stateless approach, differentiating itself from many stateful coding agents. The project includes two distinct binaries: 'qq' for "quick questions," functioning as a read-only tool to recall forgotten commands, and 'qa' for "quick agent" actions, which proposes shell commands and requires user approval before execution. This design emphasizes safety and user control. While compatible with any OpenAI-compatible API, the developer highlights optimal performance with Groq and gpt-oss-20b, noting its near-instant response times for up to 1,000 tokens per second. The project aims to provide a fast and efficient solution for developers seeking an integrated AI assistant within their shell environment.

LLM-powered assistantShell utilityOpen-sourceCommand-line interfaceAI agentArtificial IntelligenceLarge Language ModelAI Agent
06

AI Slop vs. OSS Security

The discussion "AI Slop vs. OSS Security" addresses the growing concerns regarding the integration of AI-generated content and code, often termed "AI Slop," into Open Source Software (OSS) projects and its potential impact on security. As generative AI tools become more prevalent, developers might be tempted to incorporate AI-assisted or entirely AI-generated contributions without thorough vetting. This practice poses significant risks, including the introduction of subtle vulnerabilities, backdoors, or simply low-quality, buggy code that could compromise the integrity and stability of critical OSS components. The article likely explores the tension between accelerating development cycles with AI and maintaining rigorous security standards in the open-source ecosystem. It underscores the challenges in auditing AI-generated code, identifying potential malicious insertions, and ensuring the overall trustworthiness of software supply chains reliant on OSS. The narrative emphasizes the urgent need for robust verification mechanisms, developer education, and advanced security tools to mitigate these emerging threats and safeguard the future of open-source security against the influx of potentially problematic AI-derived contributions.

AI-generated codeOpen Source SoftwareSoftware SecuritySupply Chain SecurityCode QualityArtificial IntelligenceGenerative AIMachine Learning

GitHub

4 stories
01

微舆：多智能体舆情分析系统

"Weiyu" (BettaFish) is an innovative, multi-agent public opinion analysis system designed to empower users to break through information silos, accurately reconstruct the true landscape of public sentiment, predict future trends, and support informed decision-making. Users simply articulate their analysis requirements, and the system's intelligent agents autonomously process and analyze vast amounts of data from over 30 mainstream domestic and international social media platforms, including millions of user comments. Its core technical advantages include AI-driven 24/7 all-domain monitoring across diverse social media, a powerful hybrid analysis engine integrating five specialized agents with fine-tuned and statistical models for deep, accurate, and multi-dimensional results, and advanced multimodal capabilities to parse short video content and structured information cards. Furthermore, it features a unique Agent "Forum" collaboration mechanism, fostering higher-quality collective intelligence through moderated debates. The system also supports seamless integration of public and private data, all within a lightweight and highly extensible Python-based framework, making it a versatile data analysis engine beyond just public opinion.

Multi-Agent SystemPublic Opinion AnalysisAI-driven MonitoringSocial Media AnalyticsSentiment AnalysisArtificial IntelligenceLarge Language ModelAI Agent
02

Skyvern

Skyvern is an open-source platform designed to automate browser-based workflows using a combination of Large Language Models (LLMs) and computer vision. It provides a simple API for automating manual tasks across a vast number of websites, overcoming the fragility of traditional, XPath-dependent browser automation methods. Instead of relying on static code, Skyvern leverages Vision LLMs and a "swarm of agents" approach to dynamically understand and interact with web pages. This enables it to operate effectively on unfamiliar websites, resist layout changes, and apply a single workflow across diverse platforms. Key features include SOTA performance on WebBench for RPA-adjacent "WRITE" tasks, support for complex LLM reasoning, and functionalities like form filling, data extraction, file downloading, and various authentication methods including 2FA and password manager integrations. It also supports multiple LLMs and integrates with workflow tools like Zapier and Make.com, making it a versatile solution for automating real-world scenarios such as invoice downloading, job applications, materials procurement, and retrieving insurance quotes.

Large Language ModelComputer VisionBrowser AutomationAI AgentRobotic Process AutomationArtificial IntelligenceLarge Language ModelComputer Vision
03

NocoBase

NocoBase is an advanced, AI-powered no-code platform distinguished by its profound extensibility and data model-driven architecture, enabling organizations to swiftly adapt and significantly reduce development costs. Unlike traditional form/table-driven systems, NocoBase decouples UI from data structures, offering unparalleled flexibility and supporting diverse data sources. It seamlessly integrates AI capabilities into business systems, allowing for the definition of AI employees for varied roles and fostering secure AI-human collaboration within workflows. The platform boasts an intuitive "what you see is what you get" interface, making complex system development accessible to non-programmers. Furthermore, NocoBase features a microkernel, plugin-based architecture where all functionalities are extensible, supporting custom plugins for pages, blocks, actions, APIs, and data sources. This design ensures infinite scalability and customization for any business case, eliminating the need for years of development or millions in wasted resources.

No-code platformAI-powered platformExtensible architectureData model-drivenPlugin systemArtificial IntelligenceAI AgentOthers
04

LocalAI: The Free, Open Source OpenAI Alternative

LocalAI is an open-source, free alternative to OpenAI, offering a drop-in replacement REST API compatible with OpenAI's specifications for local AI inferencing. It enables users to run Large Language Models (LLMs), generate images, and process audio locally or on-premise, even with consumer-grade hardware, without necessarily requiring a dedicated GPU. The project supports a wide array of model families and various backends, including llama.cpp, vLLM, transformers, whisper.cpp, and stablediffusion.cpp, with extensive GPU acceleration options across NVIDIA, AMD, Intel, and Apple Metal architectures. Key features include P2P inferencing, object detection, a reranker API, embeddings generation, constrained grammars, and an integrated WebUI. LocalAI is also part of a larger "Local Stack Family," integrating with LocalAGI for agent management and LocalRecall for persistent memory, making it a versatile platform for building sophisticated local AI applications.

Local AIOpenAI AlternativeLarge Language ModelGenerative AIAI AgentArtificial IntelligenceLarge Language ModelGenerative AI

huggingface

6 stories
01

LiveTradeBench: Seeking Real-World Alpha with Large Language Models

Large language models (LLMs) achieve strong performance across benchmarks--from knowledge quizzes and math reasoning to web-agent tasks--but these tests occur in static settings, lacking real dynamics and uncertainty. Consequently, they evaluate isolated reasoning or problem-solving rather than decision-making under uncertainty. To address this, we introduce LiveTradeBench, a live trading environment for evaluating LLM agents in realistic and evolving markets. LiveTradeBench follows three design principles: (i) Live data streaming of market prices and news, eliminating dependence on offline backtesting and preventing information leakage while capturing real-time uncertainty; (ii) a portfolio-management abstraction that extends control from single-asset actions to multi-asset allocation, integrating risk management and cross-asset reasoning; and (iii) multi-market evaluation across structurally distinct environments--U.S. stocks and Polymarket prediction markets--differing in volatility, liquidity, and information flow. At each step, an agent observes prices, news, and its portfolio, then outputs percentage allocations that balance risk and return. Using LiveTradeBench, we run 50-day live evaluations of 21 LLMs across families. Results show that (1) high LMArena scores do not imply superior trading outcomes; (2) models display distinct portfolio styles reflecting risk appetite and reasoning dynamics; and (3) some LLMs effectively leverage live signals to adapt decisions. These findings expose a gap between static evaluation and real-world competence, motivating benchmarks that test sequential decision making and consistency under live uncertainty.

Large Language ModelsLive TradingAI AgentsPortfolio ManagementFinancial MarketsLarge Language ModelAI AgentArtificial Intelligence
02

UniAVGen: Unified Audio and Video Generation with Asymmetric Cross-Modal Interactions

Due to the lack of effective cross-modal modeling, existing open-source audio-video generation methods often exhibit compromised lip synchronization and insufficient semantic consistency. To mitigate these drawbacks, we propose UniAVGen, a unified framework for joint audio and video generation. UniAVGen is anchored in a dual-branch joint synthesis architecture, incorporating two parallel Diffusion Transformers (DiTs) to build a cohesive cross-modal latent space. At its heart lies an Asymmetric Cross-Modal Interaction mechanism, which enables bidirectional, temporally aligned cross-attention, thus ensuring precise spatiotemporal synchronization and semantic consistency. Furthermore, this cross-modal interaction is augmented by a Face-Aware Modulation module, which dynamically prioritizes salient regions in the interaction process. To enhance generative fidelity during inference, we additionally introduce Modality-Aware Classifier-Free Guidance, a novel strategy that explicitly amplifies cross-modal correlation signals. Notably, UniAVGen's robust joint synthesis design enables seamless unification of pivotal audio-video tasks within a single model, such as joint audio-video generation and continuation, video-to-audio dubbing, and audio-driven video synthesis. Comprehensive experiments validate that, with far fewer training samples (1.3M vs. 30.1M), UniAVGen delivers overall advantages in audio-video synchronization, timbre consistency, and emotion consistency.

Audio-Video GenerationCross-Modal InteractionsDiffusion TransformersLip SynchronizationSemantic ConsistencyMultimodalGenerative AIComputer Vision
03

Diffusion Language Models are Super Data Learners

Under strictly controlled pre-training settings, we observe a Crossover: when unique data is limited, diffusion language models (DLMs) consistently surpass autoregressive (AR) models by training for more epochs. The crossover shifts later with more or higher-quality data, earlier with larger models, and persists across dense and sparse architectures. We attribute the gains to three compounding factors: (1) any-order modeling, (2) super-dense compute from iterative bidirectional denoising, and (3) built-in Monte Carlo augmentation; input or parameter noise improves AR under data constraint but cannot close the gap. At scale, a 1.7B DLM trained with a ~1.5T-token compute budget on 10B unique Python tokens overtakes an AR coder trained with strictly matched settings. In addition, a 1B-parameter DLM achieves > 56% accuracy on HellaSwag and > 33% on MMLU using only 1B tokens, without any special tricks, just by repeating standard pre-training data. We also show that rising validation cross-entropy does not imply degraded downstream performance in this regime.

Diffusion Language ModelsSuper Data LearnersAutoregressive ModelsPre-trainingData EfficiencyNatural Language ProcessingDeep LearningLarge Language Model
04

MME-CC: A Challenging Multi-Modal Evaluation Benchmark of Cognitive Capacity

As reasoning models scale rapidly, the essential role of multimodality in human cognition has come into sharp relief, driving a growing need to probe vision-centric cognitive behaviors. Yet, existing multimodal benchmarks either overemphasize textual reasoning or fall short of systematically capturing vision-centric cognitive behaviors, leaving the cognitive capacity of MLLMs insufficiently assessed. To address this limitation, we introduce MME-CC (Multi-Modal Evaluation benchmark of Cognitive Capacity), a vision-grounded benchmark that organizes 11 representative reasoning tasks into three fundamental categories of visual information: spatial, geometric, and knowledge-based reasoning, and provides fine-grained analyses of MLLMs' cognitive capacity across these dimensions. Based on MME-CC, we conduct extensive experiments over 16 representative MLLMs. Our study reveals that closed-source models currently lead overall (e.g., 42.66 for Gemini-2.5-Pro vs. 30.45 for GLM-4.5V), while spatial and geometric reasoning remain broadly weak (less than or equal to 30%). We further identify common error patterns, including orientation mistakes, fragile cross-view identity persistence, and poor adherence to counterfactual instructions, and observe that Chain-of-Thought typically follows a three-stage process (extract -> reason -> verify) with heavy reliance on visual extraction. We hope this work catalyzes a shift toward treating the cognitive capacity of MLLMs as central to both evaluation and model design.

Multi-Modal EvaluationCognitive CapacityVision-Centric ReasoningMLLMsEvaluation BenchmarkMultimodalLarge Language ModelComputer Vision
05

LEGO-Eval: Towards Fine-Grained Evaluation on Synthesizing 3D Embodied Environments with Tool Augmentation

Despite recent progress in using Large Language Models (LLMs) for automatically generating 3D scenes, generated scenes often lack realistic spatial layouts and object attributes found in real-world environments. As this problem stems from insufficiently detailed, coarse-grained instructions, advancing 3D scene synthesis guided by more detailed, fine-grained instructions that reflect real-world environments becomes crucial. Without such realistic scenes, training embodied agents in unrealistic environments can lead them to learn priors that diverge significantly from real-world physics and semantics, degrading their performance when deployed. Thus, verifying the alignment between the fine-grained instruction and the generated scene is essential for effective learning. However, current evaluation methods, such as CLIPScore and vision-language models (VLMs), often fail to reliably assess such alignment. This shortcoming arises primarily from their shallow understanding of 3D scenes, which often leads to improperly grounded scene components. To address this, we introduce LEGO-Eval, an evaluation framework equipped with diverse tools designed to explicitly ground scene components, enabling more accurate alignment assessments. We also present LEGO-Bench, a benchmark of detailed instructions that specify complex layouts and attributes of real-world environments. Experiments demonstrate that LEGO-Eval outperforms VLM-as-a-judge by 0.41 F1 score in assessing scene-instruction alignment. Benchmarking with LEGO-Bench reveals significant limitations in current generation methods. Across all evaluated approaches, success rates reached at most 10% in generating scenes that fully align with fine-grained instructions.

3D Scene SynthesisFine-Grained EvaluationTool AugmentationEmbodied EnvironmentsLarge Language ModelsLarge Language ModelGenerative AIAI Agent
06

The Sequential Edge: Inverse-Entropy Voting Beats Parallel Self-Consistency at Matched Compute

We revisit test-time scaling for language model reasoning and ask a fundamental question: at equal token budget and compute, is it better to run multiple independent chains in parallel, or to run fewer chains that iteratively refine through sequential steps? Through comprehensive evaluation across 5 state-of-the-art open source models and 3 challenging reasoning benchmarks, we find that sequential scaling where chains explicitly build upon previous attempts consistently outperforms the dominant parallel self-consistency paradigm in 95.6% of configurations with gains in accuracy upto 46.7%. Further, we introduce inverse-entropy weighted voting, a novel training-free method to further boost the accuracy of sequential scaling. By weighing answers in proportion to the inverse entropy of their reasoning chains, we increase our success rate over parallel majority and establish it as the optimal test-time scaling strategy. Our findings fundamentally challenge the parallel reasoning orthodoxy that has dominated test-time scaling since Wang et al.'s self-consistency decoding (Wang et al., 2022), positioning sequential refinement as the robust default for modern LLM reasoning and necessitating a paradigm shift in how we approach inference-time optimization.

Sequential ScalingInverse-Entropy VotingLanguage Model ReasoningSelf-ConsistencyTest-Time ScalingNatural Language ProcessingLarge Language ModelMachine Learning