A comprehensive review of large language models in 2025, covering key developments like DeepSeek R1's reasoning via RLVR/GRPO, the rise of inference-time scaling and tool use, the problem of benchmark overfitting (benchmaxxing), and predictions for 2026 including diffusion models and broader RLVR applications.

The author shares a curated list of interesting research papers from July to December 2025, categorized by topics like reasoning models, reinforcement learning, and architectures, as a thank-you to supporters.

TravelMind is an AI-powered city discovery app that uses swipe-based learning to recommend places based on personal taste, not reviews. Available on iOS and Android.

AI's ability is uneven, described as the 'Jagged Frontier.' Progress is often blocked by bottlenecks, which when resolved lead to sudden leaps, as seen with Google's Nano Banana Pro improving image generation, unlocking new capabilities like PowerPoint creation.

Tadata is an AI employee for Slack that reads the room by understanding conversation tone and context.

Genpire is an AI-powered platform that lets you literally make real products, not just digital concepts. It provides end-to-end support from design to production.

Wandesk is a Product Hunt listing that lets you build your own AI desktop, with discussion and a link.

This article provides an in-depth analysis of DeepSeek V3.2's technical evolution, covering architectural changes (including the sparse attention mechanism DSA), reinforcement learning updates (such as GRPO improvements, self-verification and self-refinement), and the development of hybrid reasoning models. V3.2 matches the performance of GPT-5 and Gemini 3.0 Pro and is released as an open-weight model, making it a significant milestone.

Isaac 0.1 is a lightweight, grounded vision-language model built for real-world perception. Despite its 2B parameters, it rivals much larger models in OCR, object recognition, and visual reasoning. It offers explainable visual reasoning, strong OCR, spatial awareness, and few-shot learning capabilities, suitable for robotics, manufacturing, visual inspection, and document workflows.

Nano Banana Pro brings powerful new capabilities in image generation and editing. Here are the main prompt tricks you should know.

Retro Diffusion's pixel art models, including rd-fast, rd-plus, rd-tile, and rd-animation, are now available on Replicate, enabling developers to generate game assets and sprites with various styles.

The author compares the original ChatGPT from three years ago with today's Gemini 3, demonstrating AI's leap from chatbots to agents. Gemini 3 can code, build games, and autonomously conduct PhD-level research, marking the advent of the 'digital coworker' era.

Replicate announces its acquisition by Cloudflare. The AI model platform will continue as a distinct brand, with no changes to the API or existing models. The move combines Replicate's AI primitives with Cloudflare's global network to build a cloud-native AI operating system.

Replicate's local MCP server now has an experimental code mode that lets language models write and execute TypeScript code in a sandbox, offering two tools for documentation search and code execution to streamline complex workflows.

As AI advice becomes more important, we need to get better at assessing it. Current benchmarks have issues like data leakage, unclear meaning, and uncalibrated difficulty. However, collectively they measure underlying ability. But for specific tasks like writing or business advice, benchmarks fall short. The author proposes 'vibes-based' benchmarking (e.g., asking AI to draw a pelican on a bike) and real-world task testing (like OpenAI's GDPval) to understand AI models, and argues that organizations should systematically interview AI as if hiring an employee.

Albato AI is an integration platform that connects over 1000 apps, using AI to automate and optimize workflows.

This article explores alternatives to standard autoregressive decoder-style transformers for large language models, including linear attention hybrids, text diffusion models, code world models, and small recursive transformers. It analyzes the strengths and limitations of each approach and discusses their potential impact on efficiency, reasoning, and modeling performance.

Bond is an AI Go-to-Market Engineer that automates outbound campaigns. Describe your target audience, and Bond builds the audience, plans the campaign, writes messaging, and executes end-to-end, integrating with sequencers like Instantly and HeyReach. First campaign live in 15 minutes.

This post introduces a reinforcement learning algorithm based on the divide-and-conquer paradigm, which does not rely on temporal difference (TD) learning. The proposed algorithm, Transitive RL (TRL), scales well to long-horizon tasks by recursively splitting trajectories and achieves state-of-the-art performance on challenging OGBench benchmarks without needing to tune the n-step TD hyperparameter.

Ollama partners with OpenAI and ROOST to launch the gpt-oss-safeguard reasoning models for safety classification. Available in 20B and 120B sizes under Apache 2.0 license, these models support custom policies, interpretable reasoning, and configurable effort.

Discussion | Link

MiniMax M2 is now available on Ollama's cloud. It is a model built for coding and agentic workflows, with 10 billion activated parameters (230B total). It ranks #1 among open-source models in composite intelligence benchmarks and excels at multi-file edits, agentic tool use, and long-horizon task execution.

A practical, opinionated guide to using AI in late 2025, covering free vs paid models, choosing among major systems (Claude, Gemini, ChatGPT, Grok, and open-weight models), advanced model selection, getting better answers via deep research and data connections, multimodal inputs, image/video generation, and quick tips. The author emphasizes experimentation and building intuition over becoming an expert.

Google's Veo 3.1 brings powerful new video generation capabilities including reference images, first/last frame control, and enhanced image-to-video. Here's everything you need to know.

Interim report on ongoing work on reward hacking

This article explains four main approaches to evaluating large language models: multiple-choice benchmarks (like MMLU), verifiers for free-form answers, leaderboards based on user preferences (like Chatbot Arena), and LLM-as-a-judge evaluations. It includes from-scratch code implementations and discusses the trade-offs of each method.

xAI is building Colossus 2, a gigawatt-scale training datacenter, leveraging innovative power sourcing across state lines and partnerships with Solaris Energy. The project aims to surpass rivals by Q3 2025, with potential Middle East funding.

Nvidia announced the Rubin CPX, a solution specifically optimized for the prefill phase, emphasizing compute FLOPS over memory bandwidth. This is a game changer for inference, second only to the March 2024 announcement of the GB200 NVL72 Oberon rack-scale form factor. Specialized hardware for prefill and decode unlocks the full potential of disaggregated serving. Nvidia's rack system design gap has become canyon-sized, forcing competitors to reconfigure their roadmaps.

This article explores the 'AI as Normal Technology' framework, contrasts it with AI 2027, addresses common confusions, and discusses the slow diffusion and adoption challenges of AI.

Huawei is ramping up Ascend AI chip production using a die bank from TSMC and capacity from SMIC. However, HBM (High Bandwidth Memory) shortages will become the primary bottleneck for future production. China's domestic HBM supplier CXMT is ramping quickly but cannot meet demand in the near term. The article also analyzes the impact of export controls and the potential implications of Nvidia H20 chip sales to China.