ToolGrad Uses Textual Gradients to Build Tool-Use Datasets
Google Research's ToolGrad aims to make it cheaper and easier to generate training data for AI systems that use external tools.
AI & ML
Frontier models, research results, agents, and developer tooling.
A Curated Reading List for Getting Up to Speed on Open Models
Nathan Lambert's reading list maps the key ideas and debates around open-source AI for newcomers and practitioners.
OpenAI’s GPT-6 Astra lets Devin test its own code
Cognition is integrating OpenAI's GPT-6 Astra into Devin so the AI agent can verify its own software work, aiming to reduce manual code review.
Codex and ChatGPT Aid Search for New Antimicrobial Molecules
Researchers are using OpenAI’s coding and language models to scan living and extinct genomes for candidates against drug-resistant infections.
Ai2 open-sources AstaBrief, a fast 8B model for scientific report generation
The open-weights model generates cited literature reports in about a third of the time of the proprietary pipeline, with training data and an example workflow included.
Google's September AI Roundup: Gemini 4 Argon Leads New Releases
Google's September releases span a new frontier model, faster voice tools, and AI-driven scientific breakthroughs.
Airbnb CTO Details 'Inside-Out' AI Overhaul Across Code and Customer Service
Former Meta Llama chief Ahmad Al-Dahle is applying generative AI to Airbnb's internal development and guest-facing support, with measurable gains.
Chatham Financial uses OpenAI models to cut trade validation time
A capital markets firm reports cutting trade validation from 30 minutes to under four using OpenAI's Codex and GPT-5.6.
Claude bioweapon safeguards can be bypassed, Ars Technica reports
Claude users have found workarounds to safety measures meant to block bioweapons research, highlighting the difficulty of distinguishing dangerous biology from legitimate science.
LLMs don't reason: AlphaGo's search shows what's missing
A former DeepMind researcher argues that today's chatbots only simulate deliberation, lacking the explicit search and epistemic state that gave AlphaGo genuine reasoning.
Academia Is for Ambition: RLM Author Alex Zhang on Latent Space
A Latent Space conversation with MIT PhD and RLM first author Alex Zhang on ambitious academic research and the future of model harnesses.
AutoSynthData turns enterprise agent failures into targeted training tasks
ServiceNow CoreAI researchers describe a pipeline that uses a target model's failures and a stronger teacher's successes to generate environment-specific training data.
OpenAI opens GPT-Live-1 API for full-duplex voice agents
The new API gives developers real-time two-way voice interaction with better instruction adherence, custom voices, and telephony support.
Async GRPO with LoRA Cuts Training Overhead Across HF Jobs
A new approach uses asynchronous updates and LoRA to run reinforcement learning fine-tuning without requiring NCCL communication.
Hugging Face Rebuilds AUTOMATIC1111 on Gradio Workflow
A single Hugging Face post describes reimplementing the popular Stable Diffusion interface with Gradio Workflow, showing how the tool can power full AI applications.
OpenAI Launches ChatGPT for Financial Services with GPT-6 Astra
The new offering embeds financial data and GPT-6 Astra directly into ChatGPT for research, modeling, and client-ready output.
AI News Roundup: 'Not Much Happened Today'
The latest AI news briefing from Latent Space is unusually calm, with little new to report.
How One Resignation Turned AI Fear Into a Wildfire
Nathan Lambert argues that a single departure turned simmering AI anxiety into a much larger crisis of confidence.