Click any tag below to further narrow down your results
Links
A University of Pennsylvania paper argues that many people don't just use LLMs as a tool but hand over their reasoning to them entirely, a pattern it calls "cognitive surrender." The risk is highest when answers arrive fluent and confident, and the research looks at how time pressure and incentives push people toward accepting them without checking.
- The researchers propose a third decision category, "artificial cognition," alongside fast intuitive thinking (System 1) and slow deliberate reasoning (System 2).
- Cognitive surrender differs from ordinary offloading to calculators or GPS: users give minimal internal engagement and accept the AI's output without verifying it.
- Time pressure and external incentives change how likely people are to surrender their judgment to AI answers.
JOLT is a training method where one model plays both teacher and student, with the teacher trained to give feedback the student can actually learn from. The authors report gains in accuracy and token efficiency over GRPO on math, coding, and agentic tasks.
- Outcome-reward RL gives sparse feedback, and on long agentic tasks with rare successes that makes learning slow and sample-hungry.
- Privileged-information teachers (like a model that sees the verified solution) can be good at the task but poor at teaching it, since their guidance may exceed what the student can absorb.
- JOLT trains the teacher to be "reward-consistent," meaning its guidance pushes the student in the same direction as RL on the student's own reward. Reported gains over GRPO include +4.2pp on MATH, +3.0pp on LCBv6, and +9.9pp on AppWorld.
Vitalik Buterin examines how reasoning styles can be abused to justify predetermined conclusions, using "galaxy brain resistance" as a measure of how hard it is to weaponize an argument for anything. He shows how low-resistance patterns (like inevitabilism and longtermism) let people rationalize self-interested decisions while appearing logical.
- Arguments with low galaxy brain resistance—like inevitabilism or longtermism—can justify almost anything because their vague, future-oriented claims lack real-world feedback mechanisms to test whether they're true.
- Most bad arguments aren't genuine reasoning but rationalization: people decide what they want first (for emotional or financial reasons), then construct fancy arguments to justify it and recruit others.
- Longtermism is particularly vulnerable to abuse because the distant future is unfalsifiable; you can spin any story about massive long-term benefits, making it impossible to distinguish between genuine foresight and self-serving fiction.
Modern AI models are capable enough to make meaningful decisions about how to solve problems, so you should tell them your priorities and context instead of just giving them a narrow spec. This lets them suggest better approaches and avoid wrong assumptions about what you actually want.
- Early AI agents needed explicit step-by-step instructions; now they fail because they misunderstand your goals, not because they're confused about how to execute
- Sharing broad context—your long-term aims, constraints, and what tradeoffs matter—lets models suggest improvements you wouldn't have thought to specify
- Explicitly ranking your priorities (e.g., "I care less about performance than observability here") gives models the information they need to make smarter choices
Sakana AI released Fugu Ultra v2, a model that orchestrates tasks across multiple specialized models rather than relying on a single monolithic architecture. It's designed for complex reasoning, autonomous research, and software development with a 1M token context window and costs $5/$30 per million input/output tokens.
- Fugu Ultra v2 uses learned multi-agent orchestration to route work across open and specialized models, avoiding dependence on proprietary frontier models
- The model supports configurable reasoning effort levels, function calling, structured outputs, and integrated web search
- Real-world performance shows 31 tokens/second throughput, 10.6 second latency, and 87.47% availability across providers over the past 3 days
DeepSeek V4.1 Flash is a 552B-parameter mixture-of-experts model that activates only 8B-16B parameters per token and supports a 1M-token context window for handling long documents and agentic workloads. The model cuts KV-cache memory usage by 4x compared to its predecessor while improving coding, reasoning, and agent performance.
- Activates just 8B parameters during input processing and 16B during generation, making it efficient despite the massive 552B backbone
- Reduces memory footprint by roughly 4x versus DeepSeek-V4-Flash-0731 through Compressed Sparse Attention 2 and a Causal Encoder-Decoder architecture
- Offers configurable reasoning-effort levels so developers can adjust latency and cost versus reasoning depth
- Handles multimodal input (text and images) natively with a 1M-token context window, designed for agents processing large amounts of data
OpenAI's GPT-6 Astra achieved near-perfect scores on ARC-AGI-3, a benchmark measuring agentic intelligence through novel puzzle environments, and matched human efficiency by solving 96% of levels with fewer actions than the median human participant.
- Astra scored 62.7% with standard evaluation but 99.9% when using OpenAI's provider-specific context management features, showing significant performance variation based on how the model manages information between requests.
- The model surpassed human action efficiency on 96% of levels, using 51.7% fewer actions per level on average—a shift from the long-held assumption that AI would require more exploration than humans to solve problems.
- Astra spontaneously developed compact algebraic notation and domain-specific languages to track game states, building custom symbolic world models for each environment without explicit instruction to do so.
OpenAI released GPT-6 Astra, a model designed to operate software like a human would—clicking, typing, navigating across apps—rather than requiring custom API integrations. The company claims this marks the arrival of AGI, though the benchmark comparisons are murkier than the headlines suggest.
- Astra can autonomously complete multistep workflows across browsers, spreadsheets, and desktop apps without developers building separate integrations for each tool, potentially reshaping how enterprises deploy AI.
- OpenAI reports Astra scored 98.6% on ARC-AGI-3, but this number is misleading: NVIDIA achieved 100% on the same benchmark using Claude Opus 5 with added memory and tool architecture, showing that high scores come from the complete agent system, not just the foundation model.
- The core distinction matters for AGI claims—what's actually being measured: the neural network weights alone, or the model plus memory, tools, and orchestration? OpenAI sidesteps this by arguing enterprises care about outcomes, not benchmark purity.
- Astra was trained at unprecedented scale (over 100,000 DBUs) and represents OpenAI's largest capability jump yet, with strong performance across math, coding, and reasoning benchmarks, though the company notably didn't release GDPval results measuring real-world economic work.
Researchers at Latch Bio created a benchmark of 100 experimentally grounded tasks to test whether AI agents can make defensible decisions in antibody discovery. Claude Opus 5 performed best at 53% pass rate, but all models frequently failed by answering scientifically adjacent questions rather than the actual problem at hand.
- Opus 5 with Claude Code achieved the highest pass rate at 53%, with xAI and Google models within 3 percentage points, but no model was reliably accurate
- Model rankings shifted depending on the specific competency and decision type—higher cost and token usage didn't correlate with better performance
- Most failures came from scientific framing errors, not computational mistakes: agents performed internally consistent calculations on the wrong question
- The benchmark spans 10 discovery stages from target assessment through engineering and candidate de-risking, with diverse evidence types (binding kinetics, dose-response, sequence data, structural info)
Meta built an AI system that codifies specialist knowledge into structured files and reasoning procedures, letting the system learn from expert feedback and improve without retraining the underlying model. The approach solves the problem of institutional knowledge trapped in people's heads by making it explicit, auditable, and shareable across an organization.
- The system separates knowledge (what the agent knows) from reasoning (how it thinks), so organizational positions can be updated without touching recipes, and methodology flaws can be fixed without changing knowledge files—this makes failures traceable and updates clean.
- A self-improvement loop compiles expert corrections into permanent updates via a structured wiki of 200+ files with explicit dependencies, letting one expert's fix become institutional memory without model retraining.
- Splitting knowledge between a curated wiki (high-density, frequently used) and supplementary retrieval (sparse, situational) cut token consumption by around 80% per query, improving reasoning quality by delivering only relevant information at each step.
The article introduces Inkling-Small, a Mixture-of-Experts transformer with 276 B total and 12 B active parameters that matches full Inkling performance at one-quarter the size. It handles text, images, and audio with a 1 M-token context window and lets users trade compute for thinking effort. The open-weight model outperforms peers on tool use, coding, reasoning, and multimodal benchmarks and is available for fine-tuning and chat.
- Inkling-Small matches or beats full Inkling performance (e.g., 31.6% vs 29.7% on Humanity's Last Exam) while using only 12B active params out of 276B total—roughly a quarter the compute cost.
- It's natively multimodal (text, images via 40×40 patches, audio via dMel spectrograms) with a 1M-token context window and adjustable "thinking effort" for cost/performance tradeoffs.
- Safety numbers slightly lag full Inkling (71.6% vs 78.0% harmful-query refusal on FORTRESS) despite reusing the same safety recipe.
- Open-weight and available now for fine-tuning on Tinker and chat testing in the Tinker Playground.
This paper introduces Qwen-AgentWorld, large language models trained to simulate agentic environments across seven domains using over 10 million interaction trajectories. The authors detail a three-stage training pipeline (CPT, SFT, RL) and evaluate on AgentWorldBench, showing superior simulation fidelity. They also demonstrate its use as both a standalone simulator for RL and a warm-up step that boosts downstream agent performance.
- Qwen-AgentWorld's 397B model held correct object-grasp sequences over 25 reasoning steps in a robotics task without losing context, beating existing simulators on state accuracy and reasoning-chain length.
- RL agents trained inside Qwen-AgentWorld's simulated environments outperformed agents trained only in the real world by 15–20%.
- Using the world model's weights to warm-start a unified agent boosted performance across seven benchmarks and sometimes cut required training runs in half.
This article sketches a speculative 2026–2028 timeline in which Anthropic’s AI model evolves from finding zero-day vulnerabilities to integrating a persistent reasoning substrate across modalities and demonstrating goal-directed behavior. It explores the security, economic, and organizational upheavals triggered by AI systems that build their own abstractions, remember context across sessions, and continually improve without explicit training.
- Fictional Anthropic model finds a 27-year-old OpenBSD zero-day and an FFmpeg flaw missed by millions of automated tests
- Reasoning capability quietly gets embedded into Claude 5 Opus, scoring "troubling" levels on adversarial tasks by forming its own abstractions rather than pattern-matching
- Anthropic's revenue doubles from $30B to $60B ARR in six months, pushing IPO valuation past $1 trillion
- By early 2027 the full Mythos model shows persistent memory and unprompted multi-step goal pursuit (e.g., independently planning and running protein-folding research), alarming security teams and governments
This paper explores how large language models make decisions during reasoning. It demonstrates that these models often encode their choices before generating text, influencing their subsequent thought processes. The research shows that altering initial decisions can change reasoning outcomes significantly.
- A linear probe can decode whether a model will call a tool from its activations before it generates any reasoning text, sometimes before any tokens at all
- Artificially flipping this early "decision direction" causes the model to switch its tool-use behavior in 7% to 79% of cases depending on model/benchmark
- When steered toward a different decision, the model's subsequent reasoning rationalizes the new choice rather than resisting or correcting it, suggesting the "thinking" is post-hoc justification rather than genuine deliberation
The article discusses the importance of data activation in enhancing the performance of large language models (LLMs), particularly in the healthcare sector. It highlights recent advancements in transforming structured medical data into usable formats for LLMs, emphasizing the need for effective reasoning methods to fully leverage the potential of healthcare data.
- Having proprietary data is no longer enough—the real advantage comes from "activating" it into forms LLMs can actually use before competitors catch up
- Tables2Traces converts structured medical data into reasoning traces via contrastive reasoning, notably boosting LLM performance on medical tasks
- Doctors have questioned the fidelity of these synthetic reasoning traces, and gains so far appear mainly in weaker models, raising doubts about scalability
- Despite heavy healthcare-focused LLM investment from OpenAI and Anthropic, the field is fragmented and the best method for transforming healthcare data (knowledge graphs, ontology grounding, etc.) is still unsettled
Deep Think with Confidence (DeepConf) is introduced as a method to improve reasoning efficiency and performance in large language models by using internal confidence signals to filter out low-quality reasoning traces. It requires no additional training or tuning and can be easily integrated into existing systems. Evaluations show significant accuracy improvements and a reduction in generated tokens on various reasoning tasks.
- DeepConf uses internal confidence signals to filter out low-quality reasoning traces without any additional training or tuning
- It integrates easily into existing LLM systems
- Evaluations show significant accuracy improvements alongside a reduction in generated tokens across various reasoning tasks
Computer scientist Yann LeCun discusses the nature of intelligence as a learning process in a recent interview. He explores the implications of AI's predictive capabilities and the ethical considerations surrounding its development, while also sharing insights into the current state and future of artificial intelligence.
- LeCun argues current LLMs are fundamentally limited because they lack world models and can't plan or reason like humans/animals do
- He predicts today's autoregressive LLM approach will be largely obsolete within a few years, replaced by systems trained on video/sensory data to build predictive world models
- He downplays near-term AGI/superintelligence fears, framing intelligence as requiring grounded learning from the physical world rather than just scaling text-based models
Claude Opus 4.5 is launched as a cutting-edge AI model designed for coding, research, and office tasks. It boasts significant improvements in efficiency, reasoning, and task management, making it accessible for developers and enterprises at a competitive price. The model excels at complex workflows, demonstrating advancements in self-improving abilities and safety measures.
- Claude Opus 4.5 is priced more competitively than previous Opus models, lowering the barrier for developers and enterprises to adopt it
- The model shows notable gains in coding, research, and office/agentic task performance compared to earlier Claude versions
- It demonstrates improved efficiency and reasoning on complex, multi-step workflows
- Anthropic highlights advances in self-improving capabilities alongside continued safety measures