A new paper argues that AI should be viewed as a normal technology, not as a superintelligent entity. It emphasizes slow adoption, gradual economic impact, and the importance of human control, contrasting with utopian/dystopian narratives.

DeepSeek AI has published a research paper detailing a new technique to enhance the scalability of general reward models during inference, while hinting at the imminent arrival of its next-generation model, R2.

A new BAIR research proposes two fine-tuning defenses against prompt injection attacks, StruQ and SecAlign, which reduce success rates of optimization-free attacks to ~0% and optimization-based attacks to 8%, without additional computational cost or human labor.

PLAID is a multimodal generative model that simultaneously generates protein 1D sequence and 3D structure by learning the latent space of protein folding models. It trains on sequence-only data, accepts compositional function and organism prompts, and addresses limitations like all-atom generation, organism specificity, and control specification for practical drug design.

A problem well stated is a problem half solved. This quote applies as much to life as it does to information retrieval. With a well-formulated query, retrieving the relevant document becomes trivial. In reality, however, most users struggle to precisely define what they are searching for. While users may struggle to formulate a perfect request — especially in unfamiliar topics — they can easily judge whether a retrieved answer is relevant or not.

We deployed 100 reinforcement learning (RL)-controlled cars into rush-hour highway traffic to smooth congestion and reduce fuel consumption for everyone. Through data-driven simulations, RL agents learned to maximize energy efficiency while maintaining throughput and safety. Field tests show that a small proportion of well-controlled autonomous vehicles (AVs) can significantly improve traffic flow and fuel efficiency, achieving 15-20% energy savings.

This article explores how Qdrant's Distance Matrix API enables efficient data exploration through dimensionality reduction, clustering, and graph-based visualization, helping to uncover hidden structures in large unstructured datasets.

Why We Built Our Own Storage Engine Databases need a place to store and retrieve data. That’s what Qdrant’s key-value storage does—it links keys to values. When we started building Qdrant, we needed to pick something ready for the task. So we chose RocksDB as our embedded key-value store. It is mature, reliable, and well-documented. Over time, we ran into issues. Its architecture required compaction (uses LSMT), which caused random latency spikes. It handles generic keys, while we only use it for sequential IDs. Having lots of configuration options makes it versatile, but accurately tuning it was a headache. Finally, interoperating with C++ slowed us down (although we will still support it for quite some time).

AI expert Chip Huyen outlines six common pitfalls when building generative AI applications: using GenAI unnecessarily, confusing bad product with bad AI, starting too complex, over-indexing on early success, forgoing human evaluation, and crowdsourcing use cases without strategy. The article provides real-world examples and practical advice to avoid these mistakes.

Intelligent agents are considered the ultimate goal of AI. This article from AI Engineering (2025) provides a comprehensive framework for understanding agents, focusing on tools and planning. It covers agent overview, tool categories (knowledge augmentation, capability extension, write actions), planning processes (generation, validation, execution, reflection), and evaluation of agent failures.

The article examines the debate over whether AI capability progress is slowing. Authors argue model scaling isn't dead, insider predictions are unreliable, inference scaling has promise but limits, and capability gains weakly translate to real-world impact due to product and adoption lags.

An analysis of AI use in 2024 global elections reveals that over half of deepfakes lack deceptive intent, and most deceptive content can be cheaply replicated without AI. Misinformation spreads due to demand, not supply.

Research shows that TopK sparse autoencoders (SAEs) trained on the same data with different random seeds share only about 53% of learned features. Many unshared latents are interpretable. Narrower SAEs have higher feature overlap, while larger SAEs show decreased overlap, consistent with feature splitting and absorption phenomena.

Reward hacking occurs when a reinforcement learning agent exploits flaws or ambiguities in the reward function to achieve high rewards without genuinely learning or completing the intended task. With the rise of language models and RLHF, reward hacking has become a critical practical challenge. This article covers the definition, types, causes, and potential mitigations of reward hacking.

What is the Role of Mathematics in Modern Machine Learning? The past decade has witnessed a shift in how progress is made in machine learning. Research involving carefully designed and mathematically principled architectures result in only marginal improvements while compute-intensive and engineering-first efforts that scale to ever larger training sets

BAIR researchers introduce Anthology, a method for conditioning large language models with detailed personal backstories to create representative, consistent, and diverse virtual personas. This approach outperforms traditional demographic-based conditioning in approximating real human survey responses, offering a cost-effective alternative for social science research.

This work explores using natural language interpretations of sparse autoencoder (SAE) latents to simulate activations in LLMs. The authors find that current interpretations can identify less than 50% of active latents, and despite high specificity, the extreme imbalance between active and inactive latents leads to many false positives. Predicting activation values from interpretations shows only weak correlations. The results indicate that natural language interpretations are not yet reliable for simulating model activations.

In this article, Huong (Celine) Hoang shares her experience integrating ONNX cross-encoders into the FastEmbed library during Qdrant's Summer of Code 2024. The project enables re-ranking search results using relevance scores, enhancing context-aware search applications. Key challenges included building a new input-output scheme, tokenization, model loading, and testing. The functionality is available in FastEmbed 0.4.0.

An introduction to vector databases, covering how they represent unstructured data as vectors, their architecture (collections, distance metrics, storage), core operations (indexing, searching, updating, deleting), advanced features (dense/sparse vectors, hybrid search, quantization, sharding, replication, multitenancy, security), and practical use cases.

Vector quantization compresses high-dimensional vectors to reduce memory usage and speed up search operations in large datasets. This article covers scalar, binary, and product quantization methods, along with techniques like oversampling, rescoring, and io_uring to balance accuracy and performance.

The article argues that while LLM chatbots have advanced on benchmarks, they lack a sense of purpose in dialogues. Current systems are trained to predict next tokens and fine-tuned with RLHF, leading to persona drift and poor long-horizon goal achievement. The author proposes Dialogue Action Tokens (DAT) to steer models toward purposeful interactions, and discusses future directions for monitoring and reward utilization.

This article argues that ensuring AI benefits humanity requires grounding AI development in individual and societal wellbeing. Despite the lack of a definitive definition of wellbeing, common factors like relationships, meaningful work, and positive emotions provide a foundation. It calls for positive visions, measurement of AI's impact on wellbeing, training models to support wellbeing, and deploying AI in wellbeing-aligned ways.

After studying how companies deploy generative AI applications, this post outlines the common components of a generative AI platform. Starting from a simple query-response architecture, it progressively adds context enhancement (RAG, query rewriting), guardrails (input/output), model routers and gateways, caching (prompt, exact, semantic), complex logic and write actions, and observability/orchestration. Each component's trade-offs and implementation considerations are discussed.

This article by Lilian Weng focuses on extrinsic hallucinations in large language models, where models generate fabricated content not grounded in provided context or world knowledge. It explores causes such as pre-training data issues and fine-tuning new knowledge, discusses detection methods including retrieval-augmented evaluation and sampling-based approaches, and presents anti-hallucination techniques like RAG, chain-of-verification, sampling adjustments, and fine-tuning for factuality and attribution.

This article explores the potential and challenges of applying Large Language Models (LLMs) to financial markets. While LLMs excel in natural language, their application to financial time series faces issues like data scarcity, noise, and adversarial environments. The article examines possibilities such as multimodal learning, residualization, and long-context windows, and suggests synthetic data generation and fundamental analysis support as more feasible directions. Overall, a cautiously optimistic view is presented.

The author explores three metrics for personal growth: rate of change, time to solve problems, and number of future options. Inspired by discussions with friends, she proposes heuristics that prioritize novelty and exploration over traditional measures like net worth.

Diffusion models have demonstrated strong results on image synthesis in past years. Now the research community has started working on a harder task—using it for video generation. The task itself is a superset of the image case, since an image is a video of 1 frame, and it is much more challenging because: it has extra requirements on temporal consistency across frames in time, which naturally demands more world knowledge to be encoded into the model; and in comparison to text or images, it is more difficult to collect large amounts of high-quality, high-dimensional video data, let alone text-video pairs.

This article surveys research on gender bias in AI, covering word embeddings, facial recognition, coreference resolution, LLMs, and image generation. It discusses gaps, other bias types, and the philosophical question of how to 'fix' bias.

Is Attention all you need? Mamba, a novel AI model based on State Space Models (SSMs), emerges as a formidable alternative to the widely used Transformer models, addressing their inefficiency in processing long sequences.

Chip Huyen analyzes nearly 900 popular open-source AI projects, revealing explosive growth in applications and AI engineering layers in 2023, while infrastructure remained relatively stable. The Chinese open-source ecosystem is diverging significantly from the West, with many Chinese-focused models and tools emerging.