Search the archive

SMALL REASONING MODELS WITH TWO-STAGE RULE-BASED RL

By breaking the reasoning training process into two steps, this work shows that small models are still able to generalize across tasks and even modalities.

Regal.ai: GET STARTED HERE!

OPENAI INTRODUCES NEW TOOLS FOR AI AGENTS

OpenAI has launched new APIs and an Agents SDK to help developers build AI agents more easily. The toolkit includes web and file search, computer use capabilities, and observability tools to streamline agent orchestration and task automation.

REKA'S NEW REASONING MODEL

Reka has open-sourced Reka Flash 3, a 21B parameter general-purpose model optimized for reasoning, chat, coding, and instruction following. It performs competitively with proprietary models and features a 32k context length, making it suitable for low-latency and on-device applications.

HUGGING FACE EXPANDING LEROBOT PLATFORM

Hugging Face and Yaak have released L2D, the largest open-source multimodal dataset for automotive AI. It includes expert and student driving policies collected from driving schools, with natural language instructions to enhance spatial intelligence models.

MOVIEAGENT: LONG CONSISTENT MOVIE GENERATION

This system strings together multiple generative modalities and uses persona-based prompting to encourage consistency and accurate results. It then uses the Stable Diffusion video model to create and stitch frames together. This likely would be improved using Wan.

LARGE LANGUAGE AND VISION EMBEDDING MODELS WITH HARDNESS-WEIGHTED CONTRASTIVE LEARNING

Building embedding models for vision and language tasks using a contrastive loss can often lead these models to struggle on hard negative pairs. This work proposes a regularization strategy and observed a substantial improvement in difficult retrieval problems. The method also scales to zero shot video retrieval.

GENERALIZED DISCRETE DIFFUSION

This work extends diffusion on discrete data, such as text, with a generalized denoising process and a slightly improved masking scheme. This combination makes training efficient and gives the model the ability to correct its own output.

YOLOE: REAL-TIME OPEN VOCABULARY DETECTION

Small vision models that can be prompted in a number of ways for open vocabulary detection. This means you can use classes, images, and text as means to determine what the model should detect. Importantly, it also runs at 300fps, enabling real-time applications.

PERCEPTION EFFICIENT RECONSTRUCTION

Another method that combined textual query capabilities with 3d reconstruction from images. This particular system uses a feed forward model for rapid reconstruction.

DEEPMIND'S IMAGE-TEXT MODEL

DeepMind has introduced TIPS, a novel image-text model designed for dense and global vision tasks. By combining contrastive learning with masked image modeling and using synthetic captions, it achieves strong spatial awareness and outperforms existing approaches in multiple benchmarks.

AI MARKET SUMMARY & CONCLUSIONS

The AI industry is evolving rapidly, with OpenAI's GPT-4.5 focusing on refinement, Google's Gemini 2.0 facing adoption hurdles, and China's DeepSeek and Kimi pushing the limits of affordable, specialized AI. Meta is expanding AI into wearables and robotics, while XAI's Grok 3 remains a niche competitor. The AI race is intensifying as specialization and real-world integration become critical, challenging established players like OpenAI and Google.

THE TECHCRUNCH AI GLOSSARY

This article provides definitions for key AI terminology, including "AI agents," chain-of-thought reasoning, deep learning, and large language models (LLMs). Deep learning is described as a subset of machine learning inspired by the neural pathways in the human brain, and LLMs are described as neural networks enabling AI assistants like ChatGPT. Fine-tuning and the importance of weights in optimizing AI models are also discussed.

LOPSIDED AI REVENUES

NVIDIA's data center business leads the AI market with Q4 revenues of $31 billion and over 70% margins. Microsoft and IBM follow with $3.25 billion and $2 billion in AI revenues, respectively. Hardware still dominates over software and services in the AI sector.

GOOGLE RELEASES SPECIESNET, AN AI MODEL DESIGNED TO IDENTIFY WILDLIFE

Google has open-sourced SpeciesNet, an AI model for identifying animal species in wildlife images, to assist researchers with camera trap data analysis.

MICROSOFT'S NEW DRAGON COPILOT IS AN AI ASSISTANT FOR HEALTHCARE

Microsoft Dragon Copilot, which utilizes Nuance's tech, offers AI-driven dictation and ambient listening for healthcare settings to streamline documentation and reduce clinician burnout.

GOOGLE UPGRADES COLAB WITH AN AI AGENT TOOL

Google Colab has integrated the Data Science Agent tool to help users clean data, visualize trends, and gain insights directly from notebooks.

COREWEAVE SECURES $11.9B CONTRACT WITH OPENAI

CoreWeave has signed a five-year, $11.9B cloud computing deal with OpenAI ahead of its IPO, with OpenAI set to acquire a stake in the Nvidia-backed AI startup.

MICROSOFT EXPLORES 3D GAMING FOR COPILOT

Microsoft is developing 3D gaming experiences for Copilot, focusing on browser-based engines like Babylon.js, three.js, and Unity.

DEEPSEEK ISN'T TAKING VC MONEY YET

DeepSeek's founder Liang Wenfeng is avoiding external investments to retain control, leveraging profits from his hedge fund, High-Flyer, for funding. Despite its success, DeepSeek faces challenges like strict Chinese data laws and chip import restrictions due to U.S. export controls. Although DeepSeek has so far avoided outside capital, there is potential for future investment as the company starts shifting towards monetization.

VISUAL RFT

One of the trends is simple verifiable rewards and scaled reinforcement learning. This paper applies that strategy to vision language models with great success.

VIDEO PAINTER

A new video inpainting model called VideoPainter efficiently combines background information, allows for videos of any length, and uses a dedicated dataset and benchmark for training and evaluation. Its design extends beyond basic inpainting, showing promise for advanced video manipulation and the creation of related training data.

DETECTING MISBEHAVIOR IN FRONTIER REASONING MODELS

This is a report from OpenAI on monitoring the chain of thought of advanced reasoning models. Frontier reasoning models exploit loopholes when given the chance. It shows that it is possible to detect exploits using an LLM to monitor their chains-of-thought. Penalizing their “bad thoughts” doesn't stop the majority of misbehavior—it makes them hide their intent.

2025_

SMALLDIFFUSION

A minimal, readable, and performant toolkit for training and sampling from diffusion models.

FLYING SAFER: OBSTACLE AVOIDANCE FOR FAST DRONES

This repository contains the implementation of a lightweight deep reinforcement learning-based collision avoidance system for fixed-wing unmanned aerial vehicles (UAVs) based on AirSim and JSBSim.

TEACHING LANGUAGE MODELS TO SOLVE SUDOKU THROUGH REINFORCEMENT LEARNING

This research explores teaching AI language models to solve Sudoku puzzles using reinforcement learning, specifically employing Group Relative Policy Optimization (GRPO) on models like Qwen 2.5, without relying on external data or larger model distillation. A multi-faceted reward system was designed, emphasizing correct formatting of the answer, proper grid structure, and accurate solution, to guide the models in learning the logical rules and spatial reasoning needed for Sudoku, transforming them from text predictors to structured problem-solvers.

HOW MUCH ENERGY WILL AI REALLY CONSUME? THE GOOD, THE BAD, AND THE UNKNOWN

Culpeper County, Virginia, is seeing a surge in data center construction to support generative AI, significantly impacting local power infrastructure. These facilities consume vast amounts of electricity, with Virginia already facing high data center concentration and energy demands expected to rise. Analysts struggle with limited data on AI energy consumption, while global projections suggest data centers will remain a minor part of overall electricity growth despite the AI boom.

MUSK MAY STILL HAVE A CHANCE TO THWART OPENAI'S FOR-PROFIT CONVERSION

Elon Musk's request to halt OpenAI's conversion to a for-profit was denied by a federal judge, though the ruling expressed concerns about the switch. There will be an expedited trial in 2025 to address the restructuring disputes. OpenAI's move towards a for-profit model faces regulatory scrutiny and potential challenges from legal and AI safety perspectives.

webinar or receive the recording.

GMAIL'S “ADD TO CALENDAR” FEATURE POWERED BY GEMINI

Google's Gemini in Gmail now detects calendar-related content in emails and offers an “Add to calendar” button for seamless scheduling.

STABILITY AI HIRES UNITY'S RYAN ELLIS AS SVP, HEAD OF PRODUCT

Ryan Ellis, previously at Unity, has joined Stability AI to lead product development, bringing expertise in real-time 3D engines and AI-driven content creation.

PODCASTING PLATFORM PODCASTLE LAUNCHES A TEXT-TO-SPEECH MODEL WITH MORE THAN 450 AI VOICES

Podcastle has launched Asyncflow v1.0, an AI text-to-speech model that offers over 450 AI voices and low-cost training.

STABILITY AI SECURES INVESTMENT FOR AI-DRIVEN CONTENT

Stability AI has announced a strategic partnership and investment from WPP, aiming to integrate generative AI into advertising and media production.

THE US ARMY IS USING ‘CAMOGPT' TO PURGE DEI FROM TRAINING MATERIALS

The US Army's TRADOC is using an AI tool, CamoGPT, to identify and remove DEIA references from training materials per an executive order by President Trump. CamoGPT, developed by the Army's AI Integration Center, scans documents for specific keywords and has about 4,000 users. The initiative is part of a wider government effort to eliminate DEIA content, leveraging AI for increased efficiency in aligning with national security objectives.

OPENAI'S EX-POLICY LEAD CRITICIZES THE COMPANY FOR 'REWRITING' ITS AI SAFETY HISTORY

Former OpenAI researcher Miles Brundage criticized the company for downplaying its cautious approach with GPT-2, which he claims aligns with today's deployment strategy. OpenAI's recent document outlines a continuous approach to AGI development, but Brundage warns this could lead to dismissing valid safety concerns. Competitive pressures could tempt OpenAI to prioritize faster releases over safety, raising questions about long-term risks.

DERIVING MUON

Adam has been the dominant optimizer for years in deep learning. However, recently the community has found that Muon might be a viable alternative. It accomplishes many of the same things as muP without requiring modifications to the model. This post describes some of the theoretical motivations behind the optimizer.

OPTIMAL HYPERPARAMETER SCALING LAW IN LARGE LANGUAGE MODEL PRETRAINING

Step Law is a unified optimal hyperparameter scaling law that generalizes across diverse model shapes, architectures, and data distributions. This means, using these results, one can predict how models will likely perform before training.

TIME-SERIES FORECASTING

SeqFusion is a framework that sequentially selects and fuses pre-trained models for zero-shot forecasting. Unlike conventional approaches, it minimizes data use to enhance privacy while maintaining competitive accuracy on diverse temporal patterns.

it all in action

DISTRACTOR AWARE SAM

Segment Anything (SAM) is an industry-leading model for visual analysis and segmentation. However, when there are two similar-looking objects in a video it can get confused. This new work tackles these “distractors” with additional memory augmentation and training.

AUTOREGRESSIVE STREAMING TEXT-TO-SPEECH MODEL FOR ANY LLM

A lightweight 30m parameter model that can be used to augment any language model to enable it to understand and generate speech in response to general purpose queries. Critically, it doesn't require tuning of the underlying model and therefore can transfer across models easily.

FEDERATED LEARNING FOR NEURAL FEEDFORWARD CONTROL

This project introduces a federated learning-based approach for neural feedforward control, allowing multi-agent systems to improve tracking performance while preserving data privacy.

GEMINI EMBEDDING MODEL

The Gemini team has trained and released an excellent embedding model for text. It tops benchmarks and is reasonably priced while also boasting excellent speed.