I Trained a Small Network to Compress Data, and It Ended Up Drawing a Pentagon.
A from-scratch NumPy rebuild of Anthropic toy superposition uses hand-derived gradients to show a tiny network compressing data into a pentagon.
Coding with AI
AI coding assistants and agents, testing and debugging with AI, and practical setups for developers.
TypeSafe Jev-Driven GraphRAG: A System One Method for Scalable Knowledge Graphs
TypeSafe Jev GraphRAG separates fast calibrated graph decisions from LLM reasoning and generation for scalable, auditable knowledge graphs.
llm-typesafe version 0.1a0 released
llm-typesafe v0.1a0 is an early-alpha plugin that connects TypeSafe AI's Jev model to Simon Willison's llm CLI. The article covers installation via one command, setting the API key, and running Jev's …
NVIDIA Unveils SoL-Pi: Self-Driving Research Cycles That Reduce Coding Agent Token Usage by as Much as 49%
NVIDIA, NTU, and MIT researchers released SoL-Pi, an MIT-licensed extension for the open-source Pi coding agent that cuts recorded token traffic by 44.7–49% on EdgeBench and lowers API costs by roughl…
Accelerating physics computations with Python
After a year of robotics research, the team released an open-source Python library wrapping the MuJoCo physics engine. It offers clean APIs for building simulation environments, performance-focused de…
Music Player Machine
Jukebox is a neural network that generates music as raw audio across many genres and artist styles, even producing basic singing. Unlike MIDI-based systems, it creates rich, recording-like sound. Its …
OpenAI's Codex Model
OpenAI has announced an improved version of Codex, an AI model that translates plain English into working code, now available through a private beta via the OpenAI API. The upgrade promises faster pro…
Tackling mathematical word problems
Researchers have built a specialized AI system that solves grade school math word problems at nearly double the accuracy of a fine-tuned GPT-3 model, achieving 55% versus children's 60% on the same te…
Presenting embeddings for text and code
OpenAI has launched embeddings, a new endpoint available through the OpenAI API. It helps developers handle natural language and code-related tasks such as semantic search, clustering, topic modeling,…
Cracking selected formal mathematical olympiad problems
A neural theorem prover built for the Lean proof assistant has learned to solve demanding high-school olympiad math problems, including AMC12 and AIME contest problems and two adapted IMO problems. Tr…
Building next-gen applications with the OpenAI Codex engine
OpenAI's Codex engine is currently powering 70 different applications that span a wide range of use cases. All of these applications are built through the OpenAI API, demonstrating Codex's versatility…
Upgrades to the fine-tuning API and a broader custom models program
OpenAI announces upgrades to its fine-tuning API, offering developers enhanced control over the model training process. Alongside these improvements, the company is expanding its custom models program…
Collaborative API Alliance with Stack Overflow
Stack Overflow and OpenAI have announced an API partnership pairing the web's leading technical knowledge platform with the most widely used large language models. The collaboration grounds OpenAI's m…
Refined Approaches for Training Consistency Models
Consistency models are a new family of generative models that can produce high-quality samples in a single step. They combine the speed of GANs with more stable, non-adversarial training, offering a m…
The API Now Supports Structured Outputs
Structured Outputs guarantees that API model responses adhere exactly to developer-supplied JSON Schemas, eliminating unreliable parsing, markdown wrapping, and invented fields. This guide explains ho…
Introducing SWE-bench Verified
OpenAI released SWE-bench Verified, a human-verified subset of the original SWE-bench benchmark. Each task has been manually checked for correctness and clarity, removing flawed or ambiguous items. Th…
Programming with OpenAI o1
Cognition CEO Scott Wu explains why OpenAI's o1 model marks a meaningful shift in AI coding: instead of instantly generating answers, o1 reasons through problems step by step, weighing trade-offs like…
Solving quantum physics problems using OpenAI o1
Quantum physicist Mario Krenn of the Max Planck Institute uses OpenAI's o1 reasoning model as a genuine thinking partner for deep scientific problems. The article explains how o1's extended reasoning …
Bringing vision capabilities to the fine-tuning API
OpenAI has extended fine-tuning for GPT-4o to support both images and text, allowing developers to train the model on custom visual data for tasks like product recognition, defect inspection, and docu…
Unveiling the Realtime API
This article explains how speech-to-speech models replace traditional multi-stage voice pipelines (ASR, LLM, TTS) with a single model that processes raw audio directly, dramatically reducing latency a…