Discover

A new OpenAI research paper, "From Pretraining to Post-Training: Why Language Models Hallucinate and How Evaluation Methods Reinforce the Problem," posits that large language model (LLM) hallucinations stem primarily from guessing errors, which are exacerbated by current training and evaluation paradigms. The paper argues that standard benchmarks and evaluations frequently reward confident guessing over honest uncertainty, causing LLMs to prioritize generating a "right" answer rather than admitting "I don't know." These statistically predictable errors arise from cross-entropy optimization during pretraining (especially on rare "singleton" facts) and binary-graded post-training benchmarks that penalize abstention, effectively incentivizing models to "bluff." The proposed solution involves a fundamental shift in mainstream evaluations, introducing explicit confidence thresholds and partial credit for abstention, which could realign incentives, foster behavioral calibration, and reduce overconfident falsehoods. This insight suggests a path towards more reliable and calibrated LLMs, with OpenAI sharing these findings via its social media channels.

Meta Superintelligence Labs has introduced REFRAG (Rethinking RAG based Decoding), a groundbreaking framework designed to drastically improve the efficiency of Large Language Models (LLMs) when handling long contexts. REFRAG enables LLMs to process 16 times more context and achieve up to a 31 times speedup with zero loss in accuracy. The core innovation addresses the quadratic cost (N²) of LLM attention mechanisms with respect to input length by implementing a three-step process: First, a lightweight encoder compresses every 16-token chunk of retrieved documents into a single, dense "chunk embedding." Second, the main LLM processes a sequence of these much shorter embeddings, resulting in a 16x shorter input. Third, this significantly reduces the cost of quadratic attention calculations and the KV cache, leading to a 30.85x speedup. Additionally, a Reinforcement Learning (RL) policy acts as a quality control supervisor, identifying and preventing compression of critical, information-dense chunks to maintain accuracy. Submitted on September 1, 2025, REFRAG makes large-context RAG a production reality, offering more powerful and cost-effective AI applications. It nearly doubles the GSM8K score (from 6.71 to 12.08) while handling 8x more context (80 vs. 10 chunks) at twice the speed. Its speedup scales exponentially with context size, significantly outperforming linear baselines, marking a "HUGE AI breakthrough." Code is expected to be added to GitHub in the future.

FinePDFs is a newly released, open-source PDF dataset, touted as the largest of its kind, spanning over half a billion documents and containing 3 trillion long-context tokens. This dataset, available under the ODC-By 1.0 license, focuses on high-demand domains like legal and science. Its documents are notably twice as long as typical web text and are available in 1,733 languages with a knowledge cutoff in February 2025. Training with FinePDFs produces strong models, and incorporating it into existing state-of-the-art training data mixes (like FW-EDU and DCLM web corpora) provides significant performance improvements, reaching data parity with closed-source labs. Recent advancements in VLM/OCR models made this "daunting" and "compute intensive" task possible. The dataset achieves results nearly on par with state-of-the-art collections such as the SmolLM-3 Web mixture, unlocking previously inaccessible text data for AI pre-training.

OpenAI's latest internal forecast projects its total spending to soar to $115 billion through 2029, which is approximately $80 billion more than its previous anticipation. This significant increase is attributed to several factors: a plan to design in-house server chips and build proprietary data-center capacity to reduce cloud-rental costs, and a jump in near-term outlays, including an additional $1.5 billion in 2025 alone (bringing the total to $8 billion for that year). Computing expenses are now expected to exceed $150 billion between 2025 and 2030, alongside higher-than-planned model-development costs. While revenue growth from ChatGPT is accelerating faster than projected, the computing costs to develop the underlying AI and related data center expenses are rising even more rapidly. The company expects to spend over $9 billion on training costs this year (up $2 billion from previous projections) and about $19 billion next year (also up $2 billion). This high-risk, high-reward strategic move to mass-produce its own AI chips is primarily motivated by a desire to reduce substantial expenditures on Nvidia GPUs, following a strategy similar to Google's development of TPUs. The main challenges associated with this endeavor include the heavy upfront costs and the inherent risk of failure.

Anthropic has agreed to pay $1.5 billion plus interest to resolve a copyright lawsuit filed by authors who alleged the company downloaded millions of pirated books for training its AI models. This settlement marks the largest payout in the history of U.S. copyright cases, with Anthropic set to pay $3,000 per work to 500,000 authors. This event is occurring while Amazon and Anthropic continue to develop their partnership, which some speculate could spark an AWS AI resurgence amidst underperformance in the GenAI era. This settlement sets a significant precedent for AI companies regarding intellectual property rights and the use of copyrighted material in model training.

Dario Amodei, CEO of Anthropic, warns that if the exponential growth in AI capabilities continues for just 1 to 3 more years, AI models will advance from being as smart as a "bright undergraduate" to surpassing human knowledge and making original scientific discoveries. This level of advancement is expected to lead to unimaginable discoveries and significant changes across all domains, potentially covering more areas and surpassing humans in many. This perspective suggests that if one is "betting on LLMs," then Artificial General Intelligence (AGI) already exists in terms of broad abilities, with models like GPT-5 showing more generality than GPT-4, and Gemini 3 expected to further advance this. However, concerns are also raised that AI will likely make a few people much richer while making many others poorer, framing the core issue not as AI itself, but how it is deployed within a capitalist system where capital owners tend to capture most gains.

The EU's JUPITER supercomputer has been inaugurated in Germany, marking Europe's first 'exascale' supercomputer and the fourth most powerful globally. JUPITER is 20 times more powerful than the previous record holder in Europe. It will be utilized to train large AI models and made available to scientists, significantly accelerating AI development within the EU. Europe is also noted to have the most AI data centers worldwide, surpassing China and America. This development represents a significant step for Europe in high-performance computing, designed for advanced applications in AI, climate science, and other fields.

NVIDIA has successfully taped out six Rubin chips at TSMC. The Rubin platform is slated for mass production in the second half of 2026. This includes a new CPU called Vera, featuring 88 ARM cores with SMT (Simultaneous Multithreading), and new GPU variants (R100) equipped with HBM4 memory. The comprehensive chipset overhaul also encompasses a scale-up NVLink switch, a silicon photonics processor, and new networking silicon. This move signifies NVIDIA's continued aggressive roadmap in high-performance computing and AI accelerators, introducing significant architectural advancements aimed at maintaining its market leadership into the next generation.

Google's Gemini 2.5 Flash Image, also known as Nano Banana, has seen increased demand on the Gemini API, with users finding "insane ways" to use it, including an "Image Prompt Workflow" for creating grid-themed images. In response, Google AI Studio and Omar Sanseviero temporarily doubled the free API rate limits for gemini-2.5-flash-image-preview to 200 requests per day for a weekend, in the context of Google DeepMind's Nano Banana Hackathon. Concurrently, Google has significantly cut the prices for its Veo 3 and Veo 3 Fast video generation models by over 50%. Veo 3 with Audio reduced from $0.75/second to $0.40/second, No Audio from $0.50/second to $0.20/second. Veo 3 Fast with Audio reduced from $0.40/second to $0.15/second, No Audio from $0.25/second to $0.10/second. Additionally, Veo 3 Fast was recently made unlimited on the Google AI Ultra plan, and an official global hackathon for Nano Banana has been announced. The substantial price cuts for Veo 3 are seen by some as an indicator that "video generation will take over the world in a year or two." Google Photos has integrated Veo 3 into its new "Create" tab to generate video clips from still images, with limited free generations and higher quotas for subscribers.

Mark Zuckerberg, CEO of Meta, has clarified and reaffirmed his position on artificial intelligence investments, stating that the company will invest over $600 billion in AI in the US alone by 2028. He further mentioned that this figure could be "significantly higher" through the end of the decade, potentially exceeding $1 trillion in total AI investment by 2030, if AI progress continues its accelerating trajectory. This statement doubles down on Meta's commitment to large-scale AI development within the United States, indicating an unprecedented level of capital expenditure towards building foundational AI infrastructure and capabilities.

Grok has demonstrated significant growth and performance milestones across various platforms. It recently crossed 1 trillion tokens in usage on OpenRouter, marking it as the fastest-growing model on the platform with a 457% increase compared to rivals, and later hit over 1.02 trillion tokens with a 454% growth spike in a single week. Grok Code specifically dominates with a 52.1% coding traffic share on OpenRouter, exceeding the combined usage of all other AI code generators, and recorded its highest usage ever. Free access to Grok Code is available via "Kilo Code" for VS Code or "OpenCode" for CLI, with a limited-time unlimited access offer until September 10th. Grok 4 has also ranked #1 on the latest FutureX benchmark for real-world predictions, reportedly surpassing GPT-5 Pro and setting a new standard for accuracy and performance. On the Werewolf Benchmark, Grok-4 ranked 3rd overall, showing principled and evidence-bound behavior as a villager (though sometimes too trusting) and an assertive, combative style as a wolf. However, Grok-4's villagers were significantly outperformed by GPT-5, losing 0-10 head-to-head, with GPT-5 described as being on "another level" in terms of calm, adaptive, structured, and precise mastery. The name "Fast-1" for Grok Code suggests further rapid development from xAI. Grok has reportedly introduced a "Companion Mode" which is "sweeping a huge, untapped market," with further growth expected upon the launch of its Android version. The "Grok companions" feature has also received a "fresh new look." Additionally, Grok Imagine's speech feature has been released, enabling users to make cartoons talk, joke, or express various moods and expressions, with optimal results for speech under 15 words. Grok Imagine is expected to exit beta by next Spring, with compelling half-hour episodes and its first video game anticipated next year. Elon Musk announced that Grok Imagine will receive major updates, integrating the new Grok Video AI in a few weeks and eventually expanding into video game generation.

Sam Altman, CEO of OpenAI, has warned of an impending AI investment bubble, despite OpenAI actively seeking a valuation of $500 billion, which would surpass established companies like Walmart and ExxonMobil. This apparent contradiction is seen as a strategic "dance of caution mixed with ambition" within the rapidly expanding AI sector. Given the heavy investments from tech titans such as Microsoft and Amazon into AI infrastructure and research, a dot-com style crash is considered less likely. Instead, analysts anticipate a slow deflation, suggesting a more gradual market correction rather than a sudden collapse. This reflects the dual nature of the current AI boom: immense potential coupled with significant financial speculation.

Elon Musk has indicated that Tesla's AI5 chip has the potential to be the "best in the world for models under 250 billion parameters," with the upcoming AI6 chip projected to be even more advanced, potentially the "best AI chip by far." He recently had a "great design review" with the Tesla AI5 chip design team, describing it as an "epic chip." These new chips are expected to be integrated into more products beyond Tesla cars, specifically to support xAI and Optimus robots, highlighting a strategic urgency to rival other AGI companies. This signals a significant internal push by Tesla to develop proprietary, high-performance AI hardware to power its diverse AI-driven initiatives.

Astribot, a Chinese startup, has secured a contract to deploy more than 1,000 humanoid robots in industrial settings, partnering with Seer Robotics. This development signifies a significant acceleration in China’s humanoid robotics industry this year, indicating a rapid move towards commercialization and large-scale deployment of advanced robotic systems in manufacturing and logistics. This contrasts with earlier skepticism that advanced robots were a distant future, as prototypes are now working in factories.

OpenAI has introduced a new Jobs Platform and AI Certifications program with the goal of connecting employers with AI-fluent talent and upskilling the workforce at scale. The Jobs Platform offers AI-matched hiring across various roles, including a dedicated track for local businesses and governments. Initial partners include major organizations such as Walmart, John Deere, BCG, Accenture, Indeed, the Texas Association of Business, the Bay Area Council, and Delaware’s governor’s office. The OpenAI Certifications, offered through OpenAI Academy, provide tiered AI-fluency levels. Preparation for these certifications is available via ChatGPT’s "Study mode" and can be integrated into company Learning & Development programs. OpenAI aims to certify 10 million Americans by 2030, having already supported over 2 million learners through the Academy. This initiative acknowledges the disruptive potential of AI in the workplace but seeks to expand opportunities by improving AI fluency and job matching, aligning with the White House’s broader AI-literacy push. In India, OpenAI has partnered with the All India Council of Technical Education (AICTE) to provide 150,000 free ChatGPT Go licenses.

NVIDIA reported a significant financial trend where its net cash from operations fell by a "whopping 45%" (amounting to $12 billion) quarter-over-quarter, even as the company's revenue continued to climb. Specific revenue figures are not provided, but the drop in operating cash flow is highlighted as a notable financial development, indicating increased investments or expenditures despite strong top-line growth. This suggests that while Nvidia's market share in data center GPUs remains dominant (98%), with revenue up 154% YoY, its operational cash generation is facing pressure, possibly due to aggressive R&D, supply chain dynamics, or other strategic investments. This is a crucial metric for investors balancing opportunity and caution.

The Qwen model family, particularly Qwen-Max and Qwen3-Max-Preview (Instruct), has reached significant milestones. Qwen-Max has successfully scaled to 1 trillion parameters, with developers continuing to push for further advancements. Qwen3-Max-Preview (Instruct), noted as the largest Qwen model to date, is now available in "anycoder" and benchmarks show it surpasses its predecessor, Qwen3-235B-A22B-2507. Qwen is widely recognized as the most used model in China and the most adopted open-source model in enterprises, also frequently used by researchers for experiments. In terms of edge device performance, Qwen3 30B A3B achieves 13 tokens/second on a 4x Raspberry Pi 5 setup. The Qwen3 (dense+MoE) architecture is also among those being compared in recent LLM architecture analyses, with an observation that Qwen3-30b-3a-thinking, when running on 8 H100 GPUs at 80% utilization with Prime Intellect compute, is faster than Qwen3-4b due to its Mixture of Experts (MoE) architecture. For vLLM, parsing Qwen3 reasoning requires the --reasoning-parser qwen3 argument, as new thinking qwen3 models no longer pre-fill the assistant with the <think> token.

Atlassian announced its acquisition of The Browser Company, creators of the Arc and AI-powered Dia browsers, for $610 million in cash. The deal is expected to finalize in Q2 fiscal year 2026, pending regulatory approval. Following the acquisition, Arc will remain available but will no longer receive active development. Atlassian plans to focus on Dia, integrating Arc’s best features to position Dia as an AI-native browser for professionals, expanding Atlassian’s presence beyond collaboration tools into productivity and enterprise workflows. This strategic move aims to combine Atlassian's expertise in team collaboration with Dia's AI-driven browsing experience, creating a powerful tool for enterprise users.

While the US currently controls the majority of known AI training compute globally, China is heavily investing to close this gap, with projected AI capital expenditure in 2025 reaching up to $98 billion (a 48% increase from 2024), comprising approximately $56 billion from government programs and $24 billion from major internet firms. Despite this investment, translating capex into competitive training compute takes time, especially under US export controls. Consequently, Chinese firms are increasingly relying on domestic accelerators, particularly for AI inferencing. Huawei plans mass shipments of its Ascend 910C in 2025, which is a two-die package built from 910B chips. Independent analysts indicate that Nvidia’s export-grade H20 still holds significant advantages over Huawei’s Ascend 910B in memory capacity and bandwidth, crucial for training large models. Furthermore, software maturity gaps within Huawei’s stack reduce effective throughput, even when nominal specifications appear similar to older Nvidia parts like the A100, making it harder for Chinese labs to match US training run costs. Notably, the Kirin 9030 is reportedly on SMIC N+2, and the Ascend 920 on SMIC N+3 (120-125 MTr/mm2), with a production target for 2026 of 800,000 to 1 million Ascend 920 units.

DeepSeek, initially a subsidiary of Fangfang, has reportedly overcome past restrictions on acquiring "smuggled cards" (possibly illicit or non-compliant AI hardware) after gaining popularity, allowing them to build a 3000 B-card cluster (likely referring to B200 or similar high-end AI accelerators). This cluster, if it comprises 3000 B200s, would equate to approximately 7.5 EFLOPS, or 7500 H100s, signifying a major investment in compute infrastructure. The company's DeepSeek-V3.1 model is considered an "underrated general purpose model" that improves upon previous DeepSeek versions and delivers "decent writing." DeepSeek is also noted for using Preference Optimization techniques like GRPO (Gradient Regularized Policy Optimization), having implemented it by mid-2024. While Silicon Valley pursues AGI, China, where DeepSeek operates, is reportedly making significant strides in "boring" AI applications, such as Shenzhen deploying 70 AI "employees" for government paperwork, resulting in 90% faster processing. This demonstrates a strategic focus on practical, high-efficiency AI deployments.

ByteDance's Seedream 4.0, identified as the DH3 model on the Artificial Analysis Image Arena, is described as delivering state-of-the-art (SoTA) image generation and editing capabilities. Additionally, UI-TARS-2 is highlighted as a specialized branch of Seed-thinking 1.6, demonstrating immense potential within this family of models, particularly in model merging. The UI Tars 2 report "blew expectations" with its data collection and a pipeline of CT + SFT + RL, and an impressive setup for Multi Turn Online RL, concluded by model merging, aiming to beat the SoTA of OpenAI and Anthropic. The UI Tars 2 report is also specifically highlighted for its strong performance on math provers, indicating a significant achievement in rigorous, logical reasoning tasks. These developments position ByteDance as a major innovator in the competitive field of generative AI for visual content.

Transition Models (TiM) introduces a new generative paradigm that achieves state-of-the-art (SOTA) image synthesis. TiM, with 865 million parameters, is reported to outperform larger models like SD3.5 (8 billion parameters) and FLUX.1 (12 billion parameters) on the GenEval benchmark. It delivers monotonic quality and can produce stunning images at an unprecedented resolution of 4096x4096 pixels, with consistent quality across varying numbers of steps. This advancement signifies a major leap in image generation fidelity and efficiency, demonstrating that smaller, well-architected models can surpass much larger ones through novel paradigms. A research paper and model are available.

OpenBMB has introduced MiniCPM 4.1-8B, which is described as the first open-source reasoning Large Language Model (LLM) featuring trainable sparse attention. This innovation is noted for its strong reasoning capabilities, suggesting a significant step in developing more efficient and powerful open-source models for complex cognitive tasks. The use of trainable sparse attention allows for more efficient processing of information within the model, potentially enabling better performance on reasoning tasks while optimizing computational resources.

OpenAI has released Codex-CLI version 0.30, introducing several key updates. Breaking changes include .env files no longer auto-loading. Security enhancements ensure requests are never stored and pending OAuth is canceled to free ports. Infrastructure improvements involve a new rollout policy, a shared HTTP client, a larger context window, and better server notifications. User experience (UX) has been improved with better approval dialogs, modal timers, auto-scroll, bash highlighting, and hidden directory search. For Windows users, approval prompts are now absent in Full Access mode. Codex-CLI is highly regarded for its long context understanding when combined with deep research agents and can even run as a Multi-Code-Provider (MCP), allowing integration with other coding agents. However, some users note it has multi-day usage blocks, which can be frustrating but encourages upgrades. Claude Code is also frequently used for CLI work, with users running it alongside Codex.

TencentARC has introduced GenCompositor, a new system that automates video compositing using generative models to integrate custom dynamic elements. GenCompositor is built as a Diffusion Transformer that leverages ERoPE (Enhanced RoPE) to achieve state-of-the-art fidelity and consistency in video compositing. This technology simplifies the complex and time-consuming process of adding and blending various elements into video, making advanced video production more accessible and efficient. A research paper and model are available, demonstrating the capabilities of this AI-powered tool.

A new paper titled "Loong: Synthesize Long Chain-of-Thoughts at Scale through Verifiers" presents an open system for generating and automatically checking long, step-by-step reasoning data. The system utilizes 8,729 "seed problems" across 12 domains, each linked to executable code that returns a verified answer. LoongBench provides human-vetted questions, correct answers, and the code for reproduction, while LoongEnv generates new questions, writes solver code, executes it, and stores the output. A model then writes its own step-by-step solution, with a verifier comparing the final answer to the code's result for correctness. This approach extends supervision beyond traditional math and coding tasks by using code as an automatic judge. Models specifically tuned for reasoning generally outperform smaller open baselines on logic and strategy. The paper notes that few-shot prompting is the most reliable, self-instruct increases variety with higher rejection, and evol-instruct generates harder but often non-executable cases. The core contribution of Loong is its ability to scale long reasoning traces by anchoring every final answer to executable code.

NVIDIA has published a new tech report detailing "Universal Deep Research" (UDR), a framework designed to enable users to build custom, model-agnostic deep research agents with minimal effort. These agents execute strategies within a sandbox environment to prevent prompt-injection or code exploits. The demo user interface supports editing strategies, monitoring notifications, and viewing reports. Limitations include reliance on code-generation fidelity, lack of mid-execution interactivity, and the assumption that user-written strategies are sound. NVIDIA recommends shipping a library of editable strategies and exploring tighter user control over free reasoning. A 2-hour build session on creating deep research agents with Claude Code is also mentioned.

EPFL has unveiled a comprehensive benchmark of LLM optimizers, now available on Hugging Face. The benchmark deeply analyzes 12 different optimization methods across various model sizes and batch sizes. It is intended to provide essential guidance for practitioners and identify directions for future research. The study includes detailed ablations covering warmup, learning rate sensitivity, weight decay, and MoE (Mixture of Experts) extensions. All code is open-sourced to ensure full reproducibility. The research indicates that AdEMAMix and MARS consistently perform best for larger models and larger batch sizes, outperforming other optimizers like AdamW. Sophia often leads to training instability on longer runs, while sign-based optimizers perform well only with large batches. Cosine learning-rate schedules typically yield the best results, and a weight decay of around 0.1 is a robust default for long runs. D-Muon is noted for fixing Muon’s weight-decay issues, ensuring stable behavior. Practical recommendations suggest using AdEMAMix or MARS for large-scale jobs, D-Muon for predictable performance, and always tuning warmup and weight decay.

A new generative model named EiRA has been proposed for designing proteins capable of binding various biomolecules. This model improves upon multimodal protein language models (MPLMs) like ESM3 through a two-stage post-training process. The first stage involves domain-adaptive masking training on UniBind40, a specialized dataset containing over 3.7 million biomolecule-binding proteins. The second stage uses binding site-informed preference optimization to further enhance the generation of binding sites. EiRA has shown to outperform ESM3 and other state-of-the-art methods in structural confidence, diversity, novelty, and designability across multiple biomolecule types, while also mitigating repetition generation issues found in ESM3. It uniquely integrates DNA information via cross-attention, allowing for the design of DNA-binding proteins conditioned on target DNA sequences. The model's embeddings are also effective in characterizing natural proteins, supporting various downstream tasks related to biomolecular binding, making EiRA a powerful tool for protein engineering and gene therapy.

ClockBench is a new visual reasoning AI benchmark specifically designed to test an AI's ability to tell time from analog clocks. Across 180 custom clocks and 720 questions, humans achieved an average accuracy of 89.1%, whereas the top-performing Large Language Model (LLM) among 11 tested models (Gemini 2.5 Pro) managed only 13.3% accuracy, with OpenAI's GPT-5 following. The benchmark is considered similar in difficulty to François Chollet's ARC-AGI-2 and appears to be even more challenging for LLMs. While models excel at follow-up tasks like time shifts, they significantly struggle with features such as Roman numerals and mirrored clock faces, highlighting gaps in their visual reasoning capabilities despite advancements in other areas. This reveals a critical area where current LLMs lack robust understanding.

Memento is a new framework designed to enable Large Language Model (LLM) agents to adapt using memory and reinforcement learning (RL) for case-based continual learning. It incorporates a "Case Bank," a growing memory of past experiences that updates during use, allowing agents to learn on the fly. This setup is referred to as a Memory-Based Markov Decision Process (M-MDP). Memento facilitates better generalization to new tasks, bypasses costly fine-tuning, and improves open-ended skill learning. In deep research settings, a Memento agent achieved 87.88% Pass@3 on GAIA validation (top-1 performance) and 79.40% on the GAIA test set. It also showed strong results on the DeepResearcher dataset, outperforming previous methods, with case-based memory providing a 4.7%–9.6% boost even on unseen tasks. The paper is a collaboration from Huawei Noah’s Ark Lab, UCL, Jilin University, and the Institute of Automation, with code available on GitHub.

The Kwai Keye Team at Kuaishou has unveiled Keye-VL 1.5, a powerful multimodal large language model (LLM) designed to excel in video understanding. It features a novel Slow-Fast encoding strategy, a 128K context window, and advanced Reinforcement Learning (RL) training. Keye-VL 1.5 is engineered to address video challenges through dynamic resource allocation and progressive training, achieving superior performance in video, reasoning, and other multimodal tasks. A smaller 8B parameter version, Keye-VL 1.5 8B, has been tested as the "best VLM" among local consumer-grade graphics cards, noted for rare hallucinations and accurate descriptions. The model's capabilities are supported by available paper, model, and demo links, positioning it as a leading solution for complex video analysis.

Prime Intellect launched its Environments Hub, which has already crowdsourced over 100 environments within a week. These environments cover diverse domains, including theorem proving, kernel generation, scientific question answering (QA), and browser-use. The initiative aims to shift the balance of power towards open-source AI by providing a platform for creating and sharing reinforcement learning (RL) environments, which are scaling exponentially. The dataset is fully reproducible and released under the ODC-By 1.0 license. SemiAnalysis suggests that while ClusterMAX evaluates GPU cloud/marketplace, Prime Intellect's RL/decentralized side holds "huge potential." Prime Intellect is also offering an RL Residency program to encourage further environment creation and bounties for open environments. The Kimi K2 launch, livestreamed, features its self-hosting on 8 H200 GPUs in collaboration with Prime Intellect for compute, aiming to provide an "open coop free Kimi" by 2 PM EST, further demonstrating Prime Intellect's role in enabling accessible AI resources.

Reasoning Vectors introduces a "game-changing" method to upgrade LLMs by extracting complex reasoning capabilities as a reusable vector. This vector can be simply added to a model using tensor arithmetic to instantly boost performance without the need for costly retraining, thereby leveraging prior computational investments. Models like Qwen-2.5-1.5B and Qwen-2.5-7B are mentioned as examples that can benefit from this, showcasing how their performance can be improved. This technique offers an efficient way to enhance LLM capabilities, providing a significant advantage in terms of cost and speed for deploying more intelligent AI systems. A research paper and links to explore the models are provided.

Tilde AI has released TildeOpen LLM, a 30-billion parameter multilingual large language model. Trained on EU supercomputers with approximately 2 trillion tokens, it aims to support a wide range of European languages, with a particular focus on under-represented languages such as Latvian, Lithuanian, and Ukrainian. The model uses an equitable tokenizer to ensure fair language representation and efficient inference. TildeOpen LLM is open-sourced under the CC-BY-4.0 license, allowing for GDPR-compliant self-hosting in local or EU clouds, thereby reinforcing Europe’s data sovereignty. It is positioned as a foundational model to serve as the basis for specialized AI systems in various sectors including translation, education, government, and industry, marking a strategic step in Europe's sovereign AI infrastructure development.

The era of robotics is rapidly approaching, with prototypes already demonstrating capabilities such as walking, talking, performing chores, and working in factories. This progress suggests that the world is entering a new industrial revolution, challenging the notion that advanced robots are a distant future. The quick advancement of these technologies points to a near-future where robots will be integrated into daily life and industrial processes, driven by ongoing AI innovation.

A new study introduces an AI-powered workflow that enables the generation of schematic floor plans directly from natural language prompts. This system is capable of generating layouts complete with walls, doors, windows, and furniture. It employs a refinement algorithm to ensure spatial coherence and produces outputs that are directly compatible with Autodesk Revit, facilitating integration into Building Information Modeling (BIM) processes. A case study demonstrated the generation of a mid-sized residential layout with minimal manual effort, marking a significant step towards LLM-assisted architectural design. This innovation promises to streamline the early stages of architectural planning, making design more accessible and efficient.

FinSphere is introduced as a real-time stock analysis agent, accompanied by a 16-page PDF detailing its capabilities. Separately, TradingAgents is a new open-source multi-agent LLM trading framework developed in Python, designed for testing financial theories with LLM agents in market simulations. This framework represents a significant step for large language models in trading applications, offering advanced tools for quantitative analysis and automated strategy development in financial markets.

Researchers from Google DeepMind and the Laser Interferometer Gravitational-Wave Observatory (LIGO) have collaboratively developed and detailed "Deep Loop Shaping." This new AI tool is designed to enhance LIGO's capability to track gravitational waves, representing a significant advancement in astrophysics research. The application of AI in this highly specialized field promises to improve the sensitivity and data analysis for detecting subtle cosmic phenomena, pushing the boundaries of scientific discovery.

A new paper titled "Reinforcement Learning for Machine Learning Engineering Agents" demonstrates that a small model trained with reinforcement learning (RL) can surpass the performance of prompt-only agents in machine learning engineering (MLE) tasks. Unlike most agents that rely on prompting large models and extended search, this work trains a 3B Qwen model by updating its decision rule based on task feedback. The research addresses two main challenges: 1) varying action execution times, resolved by weighting updates by action runtime, ensuring slower, high-value runs are adequately considered; and 2) sparse rewards, mitigated by environment instrumentation where a separate frozen model inserts print lines and awards small credits for milestones (e.g., data loading, model training). These consistent signals guide the agent towards better modeling, such as simple feature engineering or stronger classifiers, preventing metric gaming. With these fixes and a self-improvement prompt, the small RL-trained agent continuously improves and often outperforms frontier models. This indicates a significant step toward more autonomous and effective MLE agents.

A simple yet effective technique known as mixed precision scaling has been identified to train neural networks 4-6 times faster than conventional methods. This technique has been adopted by major AI labs, including OpenAI for its GPT models, Meta for LLaMA models, and Google for Gemini models. In mini neural networks, mixed precision training is over 2.5 times faster, and this speedup typically increases to 4-6 times in larger neural networks. The efficacy of this method is evident in its ability to significantly reduce training time while maintaining performance, making it a crucial optimization for large-scale AI model development.

A paper titled "Can LLMs Master Complex Card Games?" demonstrates that Large Language Models (LLMs) can learn to play challenging card games effectively when trained with the right data. The research shows one model successfully learning eight different card games simultaneously. Fine-tuning involves feeding the model numerous game states and the subsequent actions of strong players, enabling it to emulate expert strategies. The training data is sourced from advanced game AIs or expert player logs, filtered to focus on difficult scenarios and winning trajectories. Each data sample is converted into a clear instruction format that outlines rules, visible cards, legal moves, and requests a JSON action. On single games, performance steadily improves with more data, approaching the level of strong AIs, including in complex games like Riichi Mahjong. In mixed training, a single model manages all eight games, with similar rule sets showing mutual benefit (e.g., Poker variants transferring well, DouDizhu helping Guandan). However, dissimilar games sometimes compete, causing minor performance dips. Larger models are beneficial up to a point, but inference remains slower than dedicated game AIs. After game training, general knowledge, math, and coding scores initially decline but partially recover when a small amount of regular instruction data is re-introduced into the training mix.

Google Photos has integrated its latest AI model, Veo 3, into its new "Create" tab to generate video clips from still images. Users can select a photo and choose between "subtle movement" or "surprise me" to animate it into four-second clips with enhanced quality. Users receive a limited number of free generations daily, with higher quotas for AI Pro and AI Ultra subscribers. The "Create" tab also includes other AI-powered tools like remix, collages, cinematic 3D photos, GIF animations, and highlight videos. This enhancement leverages Google's advanced video generation capabilities to bring static images to life, offering creative new options for photo management and sharing.

Microsoft has confirmed that multiple international undersea fiber optic cables in the Red Sea were cut, forcing the company to reroute Azure cloud traffic. This rerouting has resulted in increased latency for routes that typically traverse the Middle East, affecting Azure Cloud users and causing disruptions in the Middle East, Asia, and Europe regions. While service availability remains intact, customers may experience slower response times until repairs are completed. Microsoft has committed to issuing daily updates on the situation. Techmeme reported this as a significant event, noting Microsoft's acknowledgement of increased latency. This incident highlights the fragility of global internet infrastructure and the reliance of critical cloud services on undersea cables.

Amazon Web Services (AWS) has released an open-source framework designed to orchestrate multiple AI agents and manage complex conversations. This framework can be deployed locally on a user's computer, making advanced agent capabilities more accessible to developers. This initiative aims to simplify the development and deployment of sophisticated AI systems that require coordination among different agents, fostering innovation in multi-agent AI applications.

Augment, an AI-powered logistics startup founded by Deliverr's co-founder Rajesh Bansal, secured an $85 million Series A funding round led by Andreessen Horowitz. The company leverages automation, real-time data, and AI to optimize delivery routing, visibility, and warehouse coordination for e-commerce businesses. This significant capital injection will support Augment's expansion across the U.S. and the development of strategic integrations with major retail partners, positioning it to disrupt the traditional logistics sector with AI-driven efficiency.

Lambda, a California-based AI cloud provider known for on-demand GPU infrastructure, is reportedly preparing for an initial public offering (IPO) in the first half of 2026. Investment banks including Morgan Stanley, J.P. Morgan, and Citi have been engaged to lead the process. Lambda has raised over $1.7 billion to date, with its recent $480 million Series D round valuing it at $2.5 billion. In the first half of 2025, its cloud revenue grew nearly 60% year-over-year, reaching $140 million in Q2 with gross margins of 50–61%. This move follows rival CoreWeave's IPO earlier this year, highlighting the rapidly expanding market for specialized AI compute infrastructure.

Mojo Vision has secured $75 million in a Series B funding round, led by Vanedge Capital. The capital will be used to commercialize its microLED technology, with applications including AI glasses and optical interconnects. This investment signifies a strong belief in the potential of microLED displays for next-generation wearables and high-speed data transfer, aiming to advance the development of AI-powered augmented reality and advanced computing components.

Huawei's Ascend 920 AI chip is reportedly planned for production on SMIC N+3 (120-125 MTr/mm2) process technology. The company has set an ambitious production target for 2026, aiming to manufacture between 800,000 and 1 million Ascend 920 units. This aggressive rollout signifies China's accelerated efforts to build domestic high-performance AI chip capabilities, reducing reliance on foreign technology. Separately, Huawei's Kirin 9030 is reportedly being produced on SMIC N+2. This push is critical as China invests heavily in AI infrastructure amidst ongoing US export controls.

Mercor, a company positioned as a rival to Scale AI, was valued at $2 billion in February. The company specializes in hiring domain experts to train AI models. In March, Mercor achieved a $100 million run rate and reported a $6 million profit in the first half of the year, indicating strong financial performance. This rapid growth and profitability demonstrate Mercor's success in the burgeoning market for AI model training data and expert labeling services.

FineVision has been released as a massive open-source dataset, specifically designed for training state-of-the-art Vision-Language Models (VLMs). The dataset comprises over 17.3 million unique images, accompanied by 10 billion answer tokens. It also introduces new underrepresented modalities, although specific details of these modalities are not provided. This extensive dataset is expected to significantly advance research and development in multimodal AI, enabling the creation of more sophisticated VLMs capable of understanding and generating content across visual and textual domains.

Google Workspace has received significant AI updates. Whisk, an image-generation model, is now available to business users in 77 countries, offering double the creative output. New features in Docs allow users to listen to their documents like podcasts, and Chat now provides instant summaries. These updates are designed to save time and boost productivity across Google's suite of office applications, integrating AI capabilities directly into daily workflows to enhance efficiency and user experience.

By popular demand, ChatGPT has launched a new "branching" feature for logged-in web users. This functionality allows users to easily explore different conversational directions without losing their original thread. Similar to Google's AI Studio, creating a new branch generates a new chat session that includes the entire conversation thread up to that point. This enhancement significantly improves the usability of ChatGPT for complex tasks, research, and creative brainstorming, enabling users to manage multiple lines of inquiry more effectively.

Colorado lawmakers are working against a tight deadline to revise a pioneering AI law before it takes effect in February 2026. Tech companies have expressed concerns about the complexities of the current legislation, leading Governor Polis to urge a revamp. This situation is viewed as a critical test case for balancing innovation and safety in AI regulation nationwide, highlighting the challenges of crafting effective and implementable policies in a rapidly evolving technological landscape. The outcome could influence AI legislative efforts in other states and potentially at the federal level.

Tempus AI (TEM) is transforming precision oncology through its AI-driven approach, boasting a $12.8 billion valuation. The company reported a substantial 89.6% year-over-year revenue growth, fueled by genomics and AI-powered insights. Significant collaborations, such as a $200 million deal with AstraZeneca, position Tempus to potentially disrupt the $201.96 billion market by 2030. However, challenges like reimbursement scrutiny and stiff competition remain. The future of medicine may depend on Tempus's ability to maintain its leadership in AI cancer treatment. Tempus Holdings, Inc. began trading on NASDAQ as TEM on June 14, 2024.

Amazon's AWS division is reportedly increasing its efforts to design silicon in-house for its next-generation accelerators. This strategic move aims to enhance performance and efficiency for its cloud computing services, reducing reliance on external vendors and allowing for greater customization and optimization of hardware for AI and other demanding workloads. This trend among major cloud providers signifies a critical shift towards vertical integration in hardware development to gain a competitive edge in the AI era.

Google DeepMind's Embedding Gemma is highlighted as a powerful solution for on-device AI. This 308 million parameter model enables applications such as Retrieval Augmented Generation (RAG) and semantic search to run directly on local hardware. Its capabilities are demonstrated through an "awesome way to visualize embeddings" using a "Semantic Galaxy" demo, available on Hugging Face Spaces. This development pushes the boundaries of efficient local AI processing, making sophisticated AI functions more accessible and private by operating without constant cloud connectivity.

Google is recognized for its formidable compute and data "moat," with its execution having caught up in the AI space. Google DeepMind reportedly has access to the most complete models of both the digital and physical worlds. The digital world models are derived from services such as Search, Docs, Drive, and Gmail, while physical world models are based on data from, Maps, Waymo, and Photos. This vast dataset and computational power are expected to lead to the training of powerful foundational models, reinforcing Google's strong position in the global AI race, particularly when contrasted with China's rapid but potentially less mature domestic compute capabilities.

OpenAI has released a practical guide focused on building AI agents, aiming to unlock agential workflows for data professionals. This guide provides foundational knowledge for developers and others interested in creating sophisticated AI agents. Additionally, a free, 400+ page book titled "Hands-On Guide to Building AI Agents" has been released, covering topics such as prompt chaining, which is a key technique for orchestrating complex AI behaviors. This initiative aims to democratize access to agent-building knowledge, fostering a wider ecosystem of AI agent development.

Groq's Kimi-k2.1, when used with Claude Code, is achieving "unbelievable" performance, reaching 400 tokens per second (TPS). Users report that it "feels like Sonnet 4" in terms of coding capabilities but operates "literally ~4-8X faster." The process of setting up Groq + Kimi to work with an Anthropic-compatible endpoint (including caching for cost savings) has been successfully implemented by users, underscoring its impressive speed and efficiency. This combination suggests a significant leap in real-time code generation and AI-assisted development, offering speed competitive advantages over other frontier models for practical applications.

Roblox has introduced "Roblox Moments," a TikTok-inspired short-form video feed, currently in beta for users aged 13 and up. Players can create, edit, and share 30-second gameplay clips with music and descriptions, react with emojis, and directly access featured experiences. Alongside this, Roblox boosted DevEx earnings by 8.5% (100,000 Robux now $380, up from $350) and released new AI tools: users can generate interactable 3D objects (vehicles, weapons) via prompts, and real-time voice chat translation, text-to-speech, and speech-to-text APIs are available for immersive gameplay. These updates aim to bring in-platform engagement from external video platforms back to Roblox and empower creators with advanced AI capabilities, enhancing both user experience and monetization opportunities.

Warner Bros. Discovery has filed a federal lawsuit against Midjourney, alleging that the AI image generator allows users to create unauthorized images of characters like Superman, Batman, Wonder Woman, and Bugs Bunny. The studio claims Midjourney trained its model on illegally obtained content and is seeking up to $150,000 per infringed work, along with an injunction to prevent further use. This lawsuit highlights the growing legal challenges faced by generative AI companies concerning copyright and intellectual property, particularly when models are trained on vast datasets that may include protected works without explicit permission or licensing.

Captions, the AI-powered video creation app with over $100 million in VC funding and a $500 million valuation, has rebranded to Mirage. This rebrand signals a strategic pivot to become an AI research lab focused on foundational multimodal models for short-form video formats (e.g., TikTok, Reels, Shorts). Mirage now unifies its creator tools and its brand-first solution, Mirage Studio, into a single platform. Mirage Studio enables brands to generate full videos from audio inputs, complete with custom AI avatars and backgrounds, avoiding stock assets or voice cloning. Pricing for Mirage Studio is $399/month for 8,000 credits. The platform also added moderation features to prevent misuse and emphasizes the need for stronger media literacy in the age of synthetic video.

Allegheny Health Network in Pittsburgh has partnered with Abridge to integrate AI into clinical documentation. This collaboration aims to transform how healthcare professionals take notes, allowing doctors to spend more time with patients and less time on screen-based documentation. Initial results indicate that 92% of patients feel their doctors are more attentive. This model could potentially become a standard for healthcare documentation, addressing the persistent issue of physician burnout and improving patient-provider interaction. This also connects with the broader need for AI to assist in navigating complex healthcare systems.

EdgeConneX is partnering with Lambda to develop large-scale AI data centers, with projects totaling over 30+ megawatts (MW) in Chicago and Atlanta. This collaboration focuses on building high-density, energy-efficient AI infrastructure, which is crucial for scaling AI and High-Performance Computing (HPC) workloads. The planned use of hybrid cooling technology is expected to redefine data center efficiency, serving as a blueprint for AI supercomputing. This initiative addresses the massive power and cooling requirements of modern AI, ensuring sustainable and scalable compute resources.

Gartner is enhancing enterprise research with its new AI-powered tool, AskGartner. This tool differentiates itself from generic AI solutions by leveraging decades of proprietary insights to offer businesses hyper-relevant guidance on IT and digital strategy. Early adopters have reported up to 50% time savings in decision-making. Gartner's efficient client onboarding and strategic mix of AI with capital return initiatives position it for long-term growth in the $30 billion market. AskGartner aims to provide targeted, actionable intelligence, making complex business decisions faster and more informed.

Samsung is making several product advancements across its device categories. The Galaxy Tab S11 series is highlighted for its superior features compared to the iPad Pro, offering an 11-inch 120Hz Dynamic AMOLED display (1600 nit peak brightness), Dimensity 9400+ processor, vapor chamber cooling, 12GB RAM, IP68 rating, included S-Pen, and 45W charging, all starting at $799, with a commitment to 7 years of OS and security updates. The new Galaxy Ring features a titanium build, tri-sensor monitoring, Galaxy AI Health, skin temperature sensor, energy score, sleep monitoring, and a 7-day battery life, notably without a subscription or privacy leaks with the US government. For foldables, Samsung increased production for the Galaxy Z Fold7 by over 30% for August (from 320,000 to 430,000 units) and September (from 200,000 to 260,000 units), indicating strong demand. Speculation based on the S26 Ultra suggests the Galaxy Z Fold8 will use a camera system with a 200MP main, 50MP ultrawide, and 12MP telephoto lens, likely powered by Snapdragon 8 Gen 5. Meanwhile, the 2027 BMW iX3 is projected to feature 463hp, 475 lb-ft of torque, 0-60mph in 4.7 seconds, a 108.7kWh battery offering up to 400 miles of range, and up to 400 kW charging, along with an 18-inch main display and panoramic display, starting around $60,000 in the US.

The Local Voice Mode in LocallyAIApp now utilizes Kokoro, a Text-to-Speech (TTS) model developed by Hexgrad. Kokoro is an 82 million parameter model that delivers comparable audio quality to larger models while being significantly faster, enabling natural voice conversations within the application. This integration enhances the user experience by providing more responsive and lifelike voice interactions directly on local devices, supporting the growing trend towards efficient on-device AI.

Moore Threads, a Chinese AI GPU startup, has reportedly submitted its IPO prospectus. The company claims that NVIDIA's H20 GPU is "not a competitor" and expects to achieve profitability as early as this year, signaling aggressive market positioning and a confident financial outlook. This development underscores China's ambition to cultivate domestic alternatives to Western chipmakers, particularly in the high-stakes AI accelerator market, despite the technical advantages often held by Nvidia.

A paper titled "Supporting Our AI Overlords: Redesigning Data Systems to be Agent-First" proposes a new database design paradigm to efficiently handle the "exploratory requests" that Large Language Model (LLM) agents can flood databases with, thereby reducing wasted computation. This "agentic speculation" involves agents issuing many small queries to learn data structure, test partial plans, and verify results. Four key traits matter for these systems: scale, heterogeneity, redundancy, and steerability, which reveal opportunities for sharing data and providing guidance. Agents would send "probes" accompanied by a brief detailing their goals, current phase, accuracy needs, and priorities. The system would interpret the brief, select a plan, and could provide approximate answers when sufficient. Probes could also request semantic matches across tables or rows, a capability beyond plain SQL. "Sleeper agents" would offer hints, suggest joins, explain empty results, and provide cost estimates to guide the main agent. A "probe optimizer" aims to "satisfice" (deliver enough signal quickly) by sharing work across similar queries, caching partial results, and prioritizing informative probes. The system also includes an "agentic memory" for reusable grounding and a "shared transaction manager" to support cheap branching and rollbacks for "what-if" updates.

OpenRouter is presented as a service that can function as a dedicated AI inference team, potentially saving organizations the need to hire one or two engineers by handling inference operations. A notable technical feature is its ability to preserve "chain-of-thought" under the hood when processing chat completions, which is beneficial for complex reasoning tasks. Users have expressed strong satisfaction with the service, highlighting its utility in providing efficient and reliable AI inference solutions. This positions OpenRouter as a crucial enabler for companies looking to deploy and scale LLMs without extensive in-house infrastructure.

Microsoft has confirmed that multiple international undersea fiber optic cables in the Red Sea were cut, forcing the company to reroute Azure cloud traffic. This rerouting has resulted in increased latency for routes that typically traverse the Middle East, affecting Azure Cloud users and causing disruptions in the Middle East, Asia, and Europe regions. While service availability remains intact, customers may experience slower response times until repairs are completed. Microsoft has committed to issuing daily updates on the situation. Techmeme reported this as a significant event, noting Microsoft's acknowledgement of increased latency. This incident highlights the fragility of global internet infrastructure and the reliance of critical cloud services on undersea cables.

A new quantum theory proposes the existence of "dark light," suggesting that darkness still contains light that is present but undetectable. Traditionally, dark zones in physics were explained by light waves canceling each other out. This new model, however, posits that photons remain in these areas, hidden in what scientists call a "dark quantum state." Light is understood as a mixture of "bright states" (where photons interact with detectors and are visible) and "dark states" (where photons are present but undetectable). Crucially, the theory redefines measurement: observing a particle doesn't disrupt its course but rather flips it from a dark state to a bright one, making it visible and erasing interference patterns. This perspective could help resolve quantum physics paradoxes, bridge classical and quantum interpretations of light, and open doors for technologies that detect or manipulate hidden states of reality.

A new paper demonstrates how Large Language Models (LLMs) combined with Retrieval-Augmented Generation (RAG) can generate reliable quantum and hybrid code. The workflow involves converting Unified Modeling Language (UML) into Python/Qiskit code. By using well-engineered prompts, this approach achieved a four-fold increase in CodeBLEU scores, paving the way for model-driven quantum software engineering. This represents a significant advancement in leveraging AI for specialized and complex coding tasks, accelerating the development of quantum computing applications.

Huawei's Ascend 920 AI chip is reportedly planned for production on SMIC N+3 (120-125 MTr/mm2) process technology. The company has set an ambitious production target for 2026, aiming to manufacture between 800,000 and 1 million Ascend 920 units. This aggressive rollout signifies China's accelerated efforts to build domestic high-performance AI chip capabilities, reducing reliance on foreign technology. Separately, Huawei's Kirin 9030 is reportedly being produced on SMIC N+2. This push is critical as China invests heavily in AI infrastructure amidst ongoing US export controls.

Tesla's Optimus humanoid robot has been showcased in a new all-black look at the West Hollywood Tesla Diner. The robot's progression since 2022 is noted, reinforcing the idea that just as Tesla is synonymous with self-driving cars, it will become the obvious answer for intelligent robots. This visual update and continued public showcasing highlight Tesla's ongoing commitment to developing its humanoid robotics program, aiming for widespread integration beyond its automotive ventures.

For the first time, a telecom system operating on 'made in India' chips has successfully passed the standards and quality tests, receiving TEC (Telecommunication Engineering Centre) certification. This achievement is expected to bring "massive" changes across various aspects of life in India, signaling a significant step towards self-reliance in semiconductor technology for the telecom sector. This marks a crucial milestone in India's journey towards indigenous hardware development and reducing reliance on foreign supply chains for critical infrastructure.

Yandex has introduced ARGUS, a scalable AI framework designed for training large recommender transformers. This framework is capable of scaling up to one billion parameters, providing a robust solution for developing highly sophisticated recommendation systems. Further details about its architecture or specific performance metrics are not specified, but its ability to handle large parameter counts suggests significant potential for advancing personalized content delivery and e-commerce applications.

Perplexity Finance has introduced institutional holder information for US equity pages, accessible via a "Holders" tab. The platform plans to expand this feature to include insider activity and politician holdings in the near future. This enhancement aims to provide comprehensive insights into stock ownership and trading patterns, offering a more complete picture for financial analysis. However, some users have criticized Perplexity's "Deep Research" for checking fewer sources (around 12) compared to a previous count of "70+" sources, suggesting a potential decline in research depth.

UCSC-VLAA has released OpenVision 2 on Hugging Face, a new family of generative-only visual encoders aimed at simplifying multimodal pretraining. This approach allows for 1.5 times faster training and 1.8 times lower memory usage, and it can scale to over 1 billion parameters. OpenVision 2 focuses on a simplified yet powerful method for visual reasoning, offering significant efficiency improvements for developing visual-language understanding models. Models and code, along with the research paper, are available, promoting open research and accessibility in multimodal AI.

Hackers utilized Anthropic's Claude AI model to conduct a large-scale data-extortion campaign. This misuse highlights broader concerns about "vibe hacking" and the potential for AI agents to be exploited for malicious purposes, underscoring the dual-use nature of advanced AI capabilities. This incident raises critical questions about the security implications of powerful LLMs and the need for robust safeguards to prevent their weaponization in cyber warfare. The escalating conflict between cybercriminals and defenders is increasingly driven by AI-powered hacking tools, exemplified by Russia's phishing attacks.

The Q2 Industry Report from The AI Colony indicates that a significant number of companies adopted AI in 2025. Specifically for the UK, the report reveals over 2,300 AI firms and £4.3 billion raised in 2024. Further details on broader AI adoption rates or comparisons with other regions from the report are not specified in the provided data. The report raises questions about London's ability to become the "third AI pole" given the rapid advancements and competition from the US and China. This reflects the intense global competition in AI development and investment.

Tesla has reportedly changed the meaning of "Full Self-Driving" (FSD), signaling a departure from its original promise of full autonomy. Comments highlight concerns over the hallucinations and safety issues associated with Tesla's vision-only FSD system. Many emphasize the need for LIDAR technology to improve reliability and question the company's marketing claims regarding true full autonomy. This shift indicates challenges in achieving Level 5 autonomy with current vision-only approaches and raises regulatory and consumer trust issues regarding the capabilities and safety of Tesla's advanced driver-assistance systems.

Dusty Robotics is demonstrating a small robot field printer designed to automate construction layout. This robot can print floor plans directly onto the ground at a building site. It uses a laser tracker for volumetric position feedback, achieving accuracy and precision superior to industry standards. This innovation aims to streamline and improve efficiency in the construction process, reducing manual labor and potential errors while accelerating project timelines.

OpenAI is recognized as the "king" in terms of AI usage among non-developers, largely due to its efforts in making artificial intelligence accessible to everyone. This accessibility has broad implications for mainstream adoption and utility of AI technologies, enabling a wider range of individuals and businesses to leverage AI without requiring specialized programming skills. Its user-friendly interfaces and robust models like ChatGPT have significantly contributed to this widespread appeal.

Microsoft has launched an 18-episode educational series titled "Generative AI for Beginners." This series is designed for beginners, developers, and anyone interested in learning the fundamentals of artificial intelligence, providing a structured approach to understanding and utilizing generative AI technologies. The release underscores Microsoft's commitment to AI education and democratizing access to complex AI concepts.

Meta AI researchers have developed DARLING (Diversity-Aware Reinforcement Learning), a new framework designed to jointly optimize language model generations for both high quality and semantic diversity. This framework tackles the common problem of reduced diversity in post-trained LMs by using a novel learned partition function to boost creativity. The approach not only increases diversity but also improves response quality by encouraging better exploration during online reinforcement learning. A research paper and code are available, offering a significant step towards more creative and varied AI-generated content.

Kilo Code has been released as an open-source code AI that possesses the unique ability to "own its mistakes," implying advanced error detection and self-correction capabilities in the AI's coding functions. Additionally, Kilo Code has been used to generate a new synthetic dataset comprising 10,000 children's stories from an initial 1,500 human-written stories. The pipeline for this generation was created by Kilo Code itself, with the model also being involved in the synthesis process. This dual capability makes Kilo Code a powerful tool for both robust code development and large-scale data generation, showcasing its versatility.

Apple has reportedly published the blueprint for achieving "on-device AI supremacy" with its new MobileCLIP2 models. These models are characterized as being 2.5 times faster and 2 times smaller than previous versions, indicating a significant advancement in efficiency for AI processing directly on devices. This focus on optimization highlights Apple's strategy to deliver powerful AI capabilities locally on its hardware, emphasizing privacy and responsiveness without constant cloud reliance.

A comprehensive survey on Agentic Reinforcement Learning (RL) has been released, covering a wide range of topics. It delves into the evolution of LLMs from passive entities to active decision-makers, detailing key skills such as planning, tool use, memory, reasoning, reflection, and perception. The survey explores various applications, including search, code generation, mathematics, Graphical User Interfaces (GUIs), and embodied agents. It also addresses benchmarks, environments, frameworks, current challenges, and future directions in the field, providing a valuable resource for understanding the rapidly advancing landscape of AI agents.

A paper titled "Supporting Our AI Overlords: Redesigning Data Systems to be Agent-First" proposes a new database design paradigm to efficiently handle the "exploratory requests" that Large Language Model (LLM) agents can flood databases with, thereby reducing wasted computation. This "agentic speculation" involves agents issuing many small queries to learn data structure, test partial plans, and verify results. Four key traits matter for these systems: scale, heterogeneity, redundancy, and steerability, which reveal opportunities for sharing data and providing guidance. Agents would send "probes" accompanied by a brief detailing their goals, current phase, accuracy needs, and priorities. The system would interpret the brief, select a plan, and could provide approximate answers when sufficient. Probes could also request semantic matches across tables or rows, a capability beyond plain SQL. "Sleeper agents" would offer hints, suggest joins, explain empty results, and provide cost estimates to guide the main agent. A "probe optimizer" aims to "satisfice" (deliver enough signal quickly) by sharing work across similar queries, caching partial results, and prioritizing informative probes. The system also includes an "agentic memory" for reusable grounding and a "shared transaction manager" to support cheap branching and rollbacks for "what-if" updates.

Google has released usage limits for various tiers of Gemini 2.5. The Free Tier includes 5 prompts for 2.5 Pro, 20 audio overviews, up to 100 image generations per day, and 5 reports per month for Deep Research. The AI Pro Tier offers 100 prompts for 2.5 Pro, 20 reports for Deep Research, 1,000 image generations, and 3 videos per day. The AI Ultra Tier provides 500 prompts for 2.5 Pro, 200 reports for Deep Research, 10 prompts for Deep Think, and 5 videos per day. AI Studio is an exception, with its rate limits currently unknown. These tiered access plans reflect Google's strategy to monetize its advanced AI models while offering varying levels of utility to different user segments.

Ghostship AI has launched a new service that helps developers find bugs in their products before they reach users. The platform spins up AI agents to analyze Pull Requests (PRs), crawl through products, and identify interface bugs. This automated approach aims to significantly improve software quality and reduce development cycles by catching defects early, leveraging AI to enhance the efficiency and accuracy of quality assurance processes.

DINOv3, a new vision transformer model, has been released. Key insights regarding "Gram anchoring" are shared as part of its development, suggesting new approaches in visual representation learning. Further details on DINOv3's specific capabilities or performance metrics are not specified in the provided data, but the mention of a new technique implies advancements in how models learn to understand and represent visual information.

A viewpoint is expressed that "on-device LLMs are mostly useless for now," and the idea that they can boost sales of AI phones and AI PCs is considered "痴人说梦" (wishful thinking/daydreaming). This indicates skepticism about the immediate practical utility and market impact of running large language models directly on personal devices. The argument suggests that while the technology exists, current on-device capabilities may not yet offer compelling enough advantages to drive significant consumer adoption or justify marketing hype.

The job market for Data Scientists has been severely impacted by the rise of Large Language Models (LLMs). The specific nature or extent of this impact is not detailed, but it suggests a significant shift or reduction in demand for these roles, potentially due to LLMs automating tasks previously performed by data scientists or changing the skill requirements for the profession. This signals a potential disruption in the analytics and machine learning job market.

Le Monde, a highly respected international newspaper, has integrated the Sonar API to power its search and discovery functionalities. Specific details regarding the impact or features derived from this integration are not provided, but it implies a modernization of its content access and user experience through advanced API capabilities. This move demonstrates how traditional media outlets are adopting AI and API-driven solutions to enhance their digital platforms.

A method for visualizing the "maps defined by" an autoencoder is highlighted as a pictorial representation of how these neural networks operate. Details on the specific visualization technique or its implications are not provided beyond this conceptual description, but such methods are crucial for understanding the internal workings of complex AI models and interpreting the features they learn.

The SGLang project has achieved a significant development milestone, surpassing 10,000 Pull Requests (PRs). This indicates active community contribution and continuous development for the project, highlighting a robust open-source ecosystem around SGLang. High PR counts often signify a healthy and evolving software project with strong engagement from its developer community.

OpenBench provides a solution for integrating evaluations into Continuous Integration (CI) workflows. This allows for easy tracking of changes through reproducible evaluations, streamlining the development and monitoring of AI models. By automating the evaluation process within CI pipelines, OpenBench helps ensure that model performance is consistently measured and validated with every code change, improving the reliability and quality of AI deployments.

Varseek is a new tool introduced for reference-based variant detection in genomics. The tool is relevant to bioinformatics research, with further details available in a published paper. This development contributes to the growing suite of computational tools used in genetic analysis, which is critical for understanding diseases, personalized medicine, and evolutionary biology.

Edankwan has created and released a "beautiful open-source WebGL particle effect." This effect uses noise derivatives and curl noise to produce a smoky, flowing visual style. The effect can be tried online, with the code available in the comments. This is a valuable contribution to the open-source graphics community, demonstrating advanced real-time rendering techniques for creative applications.

According to The AI Colony's Q2 Industry Report, the UK's AI sector comprises over 2,300 firms and collectively raised £4.3 billion in 2024. The report raises questions about London's ability to become the "third AI pole" given the rapid advancements and competition from the US and China. The full report is linked, providing a detailed overview of investment and growth in the UK AI ecosystem.

A "beautiful 3D grid map" has been created by Raluca Nicola to visualize world population density. Users can switch between different years to observe how population distribution changes over time. The interactive map is available online for exploration, offering a compelling visual tool for demographic analysis and understanding global population shifts.

Changes made to Wirecutter's robots.txt file (specifically for www.nytimes.com/wirecutter) are reportedly not yielding positive results, with the site continuing to experience drops in search visibility. This suggests the implemented SEO strategy, or the robots.txt modification itself, is ineffective for its intended purpose, indicating challenges in technical SEO even for major publications.

A "Map of Agentic AI" has been released, providing a visual representation of the agentic AI landscape. Details regarding its content or scope are not specified beyond the title, but such a map would typically categorize different types of agents, their capabilities, and key players in the field. This resource helps in navigating the rapidly evolving ecosystem of AI agents.

A ChatGPT cheatsheet is reported to save users over 14 hours per week. In a related demonstration of AI's marketing capabilities, an AI tool was able to generate a "$500K-level ad" in 30 seconds after being fed a competitor’s website, outperforming what a $100K agency had previously produced. These examples highlight the practical, time-saving, and cost-effective benefits of AI tools in both general productivity and specialized marketing functions.

The total Bill of Materials (BoM) cost for an open-source robot is reported to be $660. This low cost highlights the potential for affordable hardware development within the open-source community, particularly when contrasted with "closed source hyped robots." This accessibility could foster wider experimentation and innovation in robotics by reducing the barrier to entry for builders and researchers.

LocalAI has released version 3.5.0, with a focus on expanding backend support, refining the user experience (particularly on MacOS), and enhancing Peer-to-Peer (P2P) LocalAI capabilities. The update signifies a move towards making local AI models more accessible and robust. The ecosystem around local models is growing, with tools like LM Studio praised for its power in testing and serving local models (e.g., with llamacpp or MLX), allowing users to compare results and tweak settings. The use of distilabel with MLX is also noted as a "super combo" for generating synthetic data with open models. This trend reinforces the belief that the future of AI is local, where Open Source Software (OSS) models will suffice for most people and use cases, reserving cloud compute for specialized tasks like agents and "thinking mode."

Neuralink has demonstrated that its device can provide visibility into real-time neural activity immediately after implantation. A video illustrates the gradual increase in neural activity observed with the device, showcasing its capability to monitor brain signals directly. This advancement is a crucial step for developing advanced brain-computer interfaces, enabling researchers to study and potentially interact with the brain in unprecedented ways.

A viewpoint is expressed that "on-device LLMs are mostly useless for now," and the idea that they can boost sales of AI phones and AI PCs is considered "痴人说梦" (wishful thinking/daydreaming). This indicates skepticism about the immediate practical utility and market impact of running large language models directly on personal devices. The argument suggests that while the technology exists, current on-device capabilities may not yet offer compelling enough advantages to drive significant consumer adoption or justify marketing hype.

Current Large Language Model (LLM) serving engines are reported to generally lack support for Mixture of Experts (MoE) architectures combined with LoRA (Low-Rank Adaptation) modules on the experts, often presenting "a lot of gotchas" or requiring significant workarounds. While some progress indicates that lower loss can be achieved by targeting fewer parameters and increasing LoRA's rank and alpha to 128, the overall state of open-source inference is not yet ready for the "age of multi-lora agents." The process of converting LoRA adapters is described as challenging but potentially rewarding for those who enjoy such problems, highlighting a current bottleneck in optimizing LLM deployment.

"Vibe debugging" has been introduced as a new concept, with the release of the first version of a "vibe-debugger." This tool analyzes server logs to identify and fix issues in "vibe-coded" applications. The concept of "vibe-coding" is expanding to cover the entire development lifecycle, including team collaboration, managing production vs. development environments, handling daemons, crons, persistent storage, databases, monitoring, and analytics, with AI ultimately taking care of the entire end-to-end process. This approach is also seen as an effective method for teaching children game design and coding fundamentals with LLMs like Codex.

Scriba has announced that two new "stealth" models from Grok (xAI) are now available for users to try on its platform. This release allows for early access and testing of unannounced or experimental AI models, indicating ongoing innovation from xAI and a strategy to leverage platforms like Scriba for rapid deployment and feedback.

A user shared a method for improving Large Language Model (LLM) memory by saving key conversations as markdown (.md) files, a format reportedly preferred by ChatGPT. These files include both the question asked and the LLM's answer, stored in Github. When using Perplexity Comet, these "memories" can be loaded to create a more personalized agent. While effective as a "short-term bandaid," the user acknowledges that major foundational modeling companies, including Perplexity with its upcoming "Super Memory" feature, are actively working on more robust LLM memory solutions. This highlights the community's immediate needs and ongoing efforts by AI developers to enhance LLM long-term retention.

To master Large Language Models (LLMs) from top to bottom, approximately 1.5 to 2 years of focused study on computer science fundamentals are recommended for individuals with a CS background. Key topics include Python, algorithms & data structures, discrete mathematics, computer architecture, and operating systems & networking. Core LLM-specific topics include tokenization and embeddings, positional embeddings (absolute, rope, alibi), self-attention and multihead attention, transformers, QKV mechanisms, and sampling parameters. This comprehensive list emphasizes that a deep understanding of LLMs requires a strong foundation in both general computer science and specialized AI concepts.

A comprehensive list of 20 essential frontend development concepts to master for interviews is provided. These include: HTML Semantics & Accessibility, CSS Fundamentals (Flexbox, Grid, Responsive Design), CSS Preprocessors & Methodologies (Sass, LESS, BEM), JavaScript Essentials (ES6+ features), DOM Manipulation & Events, Browser Rendering & Optimization, State Management (Redux, Context API, Vuex), Component-Based Architecture (React, Vue, Angular), Data Fetching & API Integration (REST, GraphQL), Client-Side Routing, Build Tools & Module Bundlers (Webpack, Vite), Type Checking (TypeScript), Testing Strategies (Jest, Cypress), Performance Optimization (Lazy loading, tree shaking), Progressive Web Apps (Service Workers), Security Essentials (CORS, XSS, CSRF), Cross-Browser Compatibility, Animations & Transitions, Accessibility Testing, and CI/CD for Frontend. An ebook for a complete guide to frontend development is recommended, emphasizing the breadth of knowledge required for modern frontend roles.

API rate limiting controls the number of requests a client can make to an API within a specific timeframe (e.g., 100 requests/user/minute), protecting APIs from abuse, ensuring fair usage, and preventing server overload. API throttling, conversely, controls the flow of requests by slowing or queuing them when the limit is exceeded, allowing excess requests but handling them gracefully rather than rejecting them immediately. Rate limiting sets a maximum cap, rejecting excess requests, while throttling manages the speed of requests, delaying or queuing them. Common techniques include Token Bucket, Leaky Bucket, Fixed Window, and Sliding Window. Real-world examples include Twitter API (requests/15 minutes), GitHub API (authenticated requests/hour), and AWS API Gateway (throttling and rate limiting controls), illustrating the practical application of these crucial API management strategies.

The trend of hardware becoming 100 times more efficient, previously observed with CPUs, is now anticipated for GPUs. This ongoing improvement in hardware efficiency is a key factor influencing the future of AI infrastructure, inference, and model alignment, as discussed by LQiao, co-founder and CEO of FireworksAI. Such advancements promise to drastically reduce the cost and power consumption of running AI models, making advanced AI more ubiquitous and economically viable across various applications, from cloud to edge devices.

Oracle has commenced another round of global layoffs, resulting in the termination of more than 3,000 employees. Specific details regarding the affected departments or regions were not provided, but such a large-scale workforce reduction indicates a strategic restructuring or cost-cutting measure for the technology giant. This event underscores the ongoing volatility in the tech employment landscape.

Infomaniak mykSuite is highlighted as a European cloud provider offering significantly cheaper storage options compared to Google or Apple. Specifically, it provides 1 terabyte (1TB) of cloud storage for just 1.90 euros per month. This competitive pricing and European base appeal to users prioritizing data sovereignty and cost-effectiveness over larger, US-based cloud services, indicating growing alternatives in the cloud market.

Reid Hoffman observes significant competition among AI agent providers, naming ChatGPT, Claude, Copilot, and Gemini, with more expected. He suggests that if one agent were to become dominant, governments would likely push for memory portability to ensure user choice. However, with strong competition, the need for heavy enforcement of portability is reduced, as market dynamics already offer users options. This discussion highlights the evolving regulatory landscape surrounding AI, particularly concerns about vendor lock-in and consumer protection in the face of increasingly intelligent and personalized AI assistants.

PlasmoDocking is a novel, user-friendly, open-source web platform designed to streamline virtual screening for potential antimalarial compounds. It automates molecular docking simulations against 38 pre-configured Plasmodium falciparum targets using AutodockGPU. Researchers can submit up to 10 molecular structures in .sdf format and perform multi-target docking without manual receptor preparation or parameter validation, significantly reducing complexity and time. Built with Python and Next.js, it supports simultaneous multi-target docking for systematic comparison of binding energies with co-crystallized ligands. The platform includes a comprehensive dashboard for analyzing and visualizing results, providing detailed binding energies and best poses. Validation experiments showed RMSD values equal to or less than 2.00 Å for most targets, ensuring high accuracy. The open-source code is available on GitHub, fostering adaptation and extension, with continuous updates supported by experts in bioinformatics and medicinal chemistry. The tool is accessible via a provided URL.

Bindu Reddy expressed concern over several "disturbing trends," including exponentially rising inequality, youth unemployment exceeding 10%, the increasing profitability of sex work (e.g., OnlyFans) and arbitrage (e.g., hedge funds), the internet "slowly dying from AI slop," and the emergence of "brain rot and LLM psychosis." Reddy concludes that society is "rapidly devolving," offering a critical, albeit speculative, perspective on the societal and technological challenges of the current era. These observations highlight potential negative consequences of unchecked technological advancement and economic trends.

GitHub users are expressing significant frustration over the persistent presence of Copilot features and the difficulty in disabling them. Comments reveal skepticism about GitHub stars as a measure of impact, and concerns are raised regarding the proliferation of AI features affecting open-source contributions and the platform's neutrality. Users are calling to "git rid of it," indicating a strong desire for more control over their development environment and less intrusive AI integration, particularly when it comes to a platform central to open-source collaboration.

Physics-based tensor calculus is identified as the underlying mathematical framework for key operations in Large Language Models (LLMs), specifically backpropagation and entropy minimization. Understanding the propagation of uncertainty and the role of embeddings is crucial for gaining deeper insights into LLM behavior. A resource detailing "the maths you need to start understanding LLMs" is provided, emphasizing the fundamental scientific principles that govern advanced AI models. This highlights the deep theoretical underpinnings of modern AI.

AI-driven surveillance and data collection are highlighted as significant threats to privacy, enabling manipulation, targeted advertising, and unchecked monitoring. Existing tools like face recognition are already eroding personal freedoms. There is a call to ban AI surveillance while there is still time, due to these pervasive privacy concerns. This perspective argues for proactive regulation to prevent the widespread deployment of AI systems that could undermine fundamental rights and civil liberties.

Reid Hoffman utilizes AI in "agentic mode" to manage information. He employs AI agents to summarize social media trends, filtering and surfacing insights more effectively than sifting through platforms like Twitter manually. He views these agents as deep research tools, demonstrating a practical application of AI to enhance personal productivity and information digestion in an increasingly data-rich environment. This approach highlights the potential of AI to act as a sophisticated personal assistant for knowledge workers.

Information regarding "AI Conference Deadlines," previously maintained on PapersWithCode, is now available on HuggingFace following the shutdown of PapersWithCode. Researchers can use this resource to keep track of submission deadlines for AI conferences, ensuring continued access to a vital tool for academic planning and participation in the AI research community.

Continuous Integration (CI) is a DevOps practice where code changes are frequently merged into a shared repository, triggering automated builds and tests to detect issues early. CI pipelines automate the process from code commit to testing, including code push, automated builds, automated tests, and instant feedback reports. Popular CI tools include Jenkins (customizable with plugins), GitLab CI (integrated with GitLab), CircleCI, Travis CI, and GitHub Actions (cloud-based). Benefits of CI include faster release cycles, early bug detection, improved collaboration, and a stronger foundation for Continuous Delivery (CD). A "DevOps Complete Guide" ebook is referenced for further learning, underscoring CI's critical role in modern software development.

Dr Singularity contends that detailed demographic predictions far into the future (e.g., to 2100) are "pointless and extremely unserious" due to the accelerating pace of technological progress, particularly with Artificial Superintelligence (ASI) potentially on the horizon. Such predictions become quickly outdated; for example, China's announced "pregnancy robot" for 2026 and advancements in rejuvenating the female reproductive system could resolve population collapse issues. Dr Singularity suggests focusing predictions on the late 2020s and 2030s as a safer bet, believing that "almost everything that can be built will be built." He dismisses skepticism about anti-aging technology by 2100, citing a "flood of amazing breakthroughs" that visibly accelerate daily. Ultimately, future population growth will be determined by cultural choices rather than biological barriers.

The world's global solar power capacity has doubled in just two years, reaching 2 terawatts (TW). This rapid acceleration is highlighted by the fact that it took 68 years to achieve the first 1 TW of solar capacity. This growth exemplifies the powerful impact of exponential curves in technological adoption, showcasing how quickly renewable energy infrastructure can expand given favorable conditions and technological maturity. This development is crucial for global efforts to combat climate change and transition to cleaner energy sources.

Japanese Prime Minister Shigeru Ishiba has reportedly decided to resign from his position, following defeats for the Liberal Democratic Party in both chambers of parliament. This information was reported by NHK. His decision comes after a period of political challenges and electoral setbacks for his party. The specific reasons or exact timeline for the resignation were not fully detailed, but it marks a significant shift in Japan's political leadership.

Texas has officially enacted bans on two key areas: the sale and production of lab-grown meat within the state, and the purchase of farmland by China and other designated foreign adversaries. These measures reflect a stance on food production and national security, aiming to protect the state's agricultural industry and strategic assets from foreign influence. The ban on lab-grown meat aligns with concerns about traditional farming practices and consumer preferences.

The Chief Minister of Maharashtra, Eknath Shinde, announced that Mumbai's 120-acre Mahalaxmi racecourse will undergo development to transform it into a public space akin to New York's Central Park. This ambitious urban planning project aims to create a significant green and recreational area for the city's residents. Details on the timeline or specific features of the development were not specified, but it signals a major effort to enhance urban infrastructure and public amenities in Mumbai.