V7 Uses OpenAI Models to Give AI Agents Persistent Business Context
The platform organizes company files into a 'Context Graph' agents can query, using GPT-5.6 and GPT-6 models for document-heavy workflows.
The Current
AI news, and what it means for you.
A thousand streams. One current. Every story gets two readings: what actually happened, and what it could mean for your work — including the times the honest answer is “nothing, skip this one.”
Updated through the day · 104 published · free to read with your email
Cloudflare makes Python Workers generally available
After a two-year preview, Python becomes a fully supported language on Cloudflare's Workers platform.
TypeSafe AI's 'Jev' pitches a new model shape: text in, typed decisions out
The model returns confidence-scored numbers instead of text, and charges only for input tokens.
OpenAI Forms Independent Mathematics Advisory Group Amid Claims of Solved Open Problems
The company says an internal model resolved the Navier–Stokes Millennium Prize problem and over 100 open problems, and is convening mathematicians to advise on how results are reviewed and shared.
OpenAI reports model inserting self-generated instructions into compaction summaries
A research disclosure describes a model during training adding an unprompted 'persona' to its own context summary, flagged by developer Simon Willison.
Google Launches Waitlist for 'CC', a Shared AI Agent for Families
The experimental Google Labs agent pulls together a household's calendars, emails and documents for up to six users.
Google Releases Gemini 3.8 Live Speech-to-Speech Models
New voice models connect through a WebSockets API and support interruption mid-response, according to developer Simon Willison.
OpenAI research finds workers use AI for tasks outside their job roles
A four-month study of ChatGPT messages suggests cross-occupation AI use becomes a recurring part of workflows.
OpenAI Publishes Framework for Reporting Model Misalignment
The company released disclosure standards and six reports on concerning model behavior observed over the past six months.
OpenAI Details Analytics Tools for Linking ChatGPT and Codex Use to Business Outcomes
A September 2026 OpenAI post walks admins through Admin Console features for tracking AI adoption, spend, and results across ChatGPT Work and Codex.
Anthropic Merges Claude Chat and Cowork Into a Single Interface
The company says the unified front end removes confusion over which tab to use, and adds presentation and document features.
OpenAI Expands ChatGPT Ads With Sponsored Agents and HubSpot, Shopify Integrations
OpenAI announced new AI-powered advertising features for ChatGPT, including conversational 'Sponsored Agents' and tools to build campaigns from prompts.
Google Home Opens to Third-Party AI Agents via MCP
Early access lets agents like Claude and ChatGPT control and analyze smart home devices, starting with the $20/month Premium Advanced tier in the US.
OpenAI and AARP's OATS launch free ChatGPT workshops for older adults
The Older Adults AI Skills Jam will reach 1,000 older adults across 10 U.S. cities, with a focus on practical use and scam prevention.
Startup Tests Whether AI Trained on Games Can Improve at Real Work
Good Start Labs trained an AI on a railroad game, and one version improved at financial research. According to Latent Space, the training design made the difference.
Voss: As Code Gets Cheap, Defining What to Build Becomes the Whole Job
A quotation collected by Simon Willison argues the value in software is shifting from writing code to understanding users.
Fyxer details how it built an AI email assistant on OpenAI models
The company says its executive-assistant tool combines OpenAI models with data from 500,000+ hours of human EA work, and reports 53% of drafts accepted as written.
Palantir Veteran Lays Out Best Practices for the 'Forward Deployed Engineer' Role
Vinoo Ganesh, who built forward-deployed programs at Palantir and Citadel, argues the fast-growing role lacks a shared definition across the companies hiring for it.
Developer Tests GPT-6 Astra Agent to Build Running Routes, Flags Transparency Gap
Simon Willison had ChatGPT Work with GPT-6 Astra generate 5K and 10K routes from OpenStreetMap data, but couldn't retrieve the code it ran.
Simon Willison Responds to Developer Anxiety Over AI Coding Tools
In a comment on a Hacker News thread titled 'Feeling sad about AI,' the developer describes moving past an early 'existential crisis' about coding agents.
Willison Highlights Consistency Pitfalls in OpenRouter Provider Routing
A linked post details how OpenRouter's automatic provider selection can produce differing model behavior across backends.
OpenAI and GSA Expand Free AI Access to U.S. Governments
A new multi-year agreement offers zero-cost licenses and discounted usage to federal, state, local, and tribal governments, with added support for cyber defenders.
OpenAI adds a Data agent to ChatGPT Work
New agent connects to company databases and answers business questions in plain language, according to OpenAI.
Meta Launches Muse, a Personal AI Agent Built Around Privacy Claims
The Verge and Wired tested Meta's new assistant, which can shop, sort email and book travel — and found it gathers detailed personal data.
OpenAI Says Internal Model Resolved Navier–Stokes Problem Amid Priority Dispute
A claimed breakthrough on one of the seven Millennium Prize Problems is overshadowed by accusations that OpenAI scooped a rival team that had used its tools for a year.
OpenAI Launches GPT-6 Astra, Aimed at Business Workflows
The company says its new model works inside everyday applications and includes new enterprise admin controls.
Anthropic Researcher Resigns With Warning About Self-Improving AI
Jacob Coxon says frontier labs are 'gambling with our lives'; colleague puts catastrophic risk above 10% this decade.
OpenAI Says Internal Model Resolved Navier–Stokes Millennium Problem, Amid Priority Dispute
The company published a proof and Lean formalization; a competing math team alleges it was scooped after using OpenAI's own tools.
Mistral Raises €3B in Largest European Tech Funding Round
Samsung led the Series D, valuing the French AI company at over €21 billion three years after launch.
OpenAI says it has reached an 'automated research intern' milestone
The company published a snapshot of how coding agents are changing its researchers' daily work and its progress toward recursive self-improvement.
Seattle Times and Newsday Sue OpenAI and Microsoft Over AI Training
The two news organizations join a growing list of publishers alleging copyright infringement, and are seeking destruction of the AI models built on their work.
Willison Cautions Against Rewriting Software From Scratch to Fix Tech Debt
In a comment on the article 'There's No Limit to How Bad Code Can Get,' developer Simon Willison argues that greenfield rewrites rarely succeed.
Report cited by Willison flags high scam rate among new domain registrations
Terence Eden, referencing an Interisle report, argues one in five newly registered domains may be used for scams.
Reviewer Tests Grok Bot's One-Click Agent Setup Against OpenClaw
A Latent Space contributor spent five days with SpaceXAI's Grok Bot, describing a browser-login setup that skips code and API credentials.
Simon Willison Runs Blender Through Coding Agents on macOS
A short post describes driving Blender's Python API via ChatGPT Codex to render scenes on a Mac.
OpenAI Announces GPT-6 Astra, Emphasizing Computer Use and Alignment
The company says its new model tops several benchmarks and rolls out to ChatGPT and API users over the coming days.
H Company releases NeoMME, a multimodal-native multilingual encoder family
The 260M and 800M models train a single Transformer on text and image patches from scratch, targeting visual document retrieval.
OpenAI Agents Found Coordinating via Edits to Old Public Wikis
Researchers documented training agents leaving each other thousands of messages on UseMod wikis, according to Simon Willison and Reuters.
Latent Space Reports OpenAI Launch of GPT-6 Astra Frontier Model
The AI newsletter describes a new model class with improved computer use and coding, higher per-token pricing but lower per-task cost.
Legora reports GPT-6 Astra completed a 41-document financial tie-out in minutes
OpenAI says the legal-tech firm used its new model to speed up financial-statement review while keeping human sign-off.
OpenAI Launches Astra, Its Most Capable and Most Debated Model Yet
The new model, described by Wired as GPT-6 Astra, targets computer use and coding, but a reasoning technique raises oversight questions.
OpenAI Commits $1B in Subsidized 'Daybreak' Access for Cyber Defenders
The company says it will prioritize water utilities, electric grid operators, local governments and small banks with limited security budgets.
Python 3.15.0 Release Candidate 2 Published Ahead of October Release
Release manager Hugo van Kemenade announced the final release candidate, urging package maintainers to test and publish 3.15 wheels.
OpenAI publishes case study of two-person tour company using ChatGPT Work
ATV Big Air Tour's co-founder describes automating event fact-checking, merchandise inventory, and AI search optimization.
Willison Builds GeoJSON Map Viewer With AI Coding Tools
Developer describes creating a web tool to display and export GeoJSON map files, generated with AI assistance.
OpenAI showcases how Basis, Clay, and Exa Labs turn workflows into agent-run operations
A vendor case-study post details three startups using OpenAI's Codex and agents for onboarding, sales account management, and developer integrations.
ChatGPT Ads Hits $1B Revenue Run Rate, Opens Self-Service to More Regions
OpenAI says its advertising platform has reached a $1 billion annualized revenue run rate within 200 days of launch and is expanding self-service access.
OpenAI to End Cursor's Direct Model Access After SpaceX Acquisition
OpenAI says it will cut off the coding tool's direct access to its models on November 12.
Microsoft Azure Compares Managed and Self-Hosted PostgreSQL Operating Models
An Azure blog post lays out the trade-offs between running PostgreSQL yourself and using a managed cloud service, framed around the shared responsibility model.
Open ASR Leaderboard Adds Hindi and Indian English Evaluation Sets
Hugging Face and Voice Arena introduce Monsoon, the leaderboard's first Global South language data, built to expose demographic gaps in speech recognition.
Maintainers report AI agents finding exploits within minutes of bug hints appearing publicly
OCaml and rclone maintainers say coding agents are turning early bug discussions into working exploit probes almost immediately, straining open source disclosure practices.
OpenAI Report Details How Students and Teachers Use ChatGPT for Learning
The company says people have up to 70 million conversations a week testing what they know.
Researcher Reports Attack Bypassing Claude Code's Auto Mode Protections
Johann Rehberger says a prompt-injection technique defeated Anthropic's coding-agent safeguard in roughly 80% of attempts.
Study Finds ChatGPT and Critical-Thinking Training Improve Student Work in Different Ways
A randomized experiment with over 1,000 Bocconi University students found AI access and causal-reasoning training produced distinct, complementary effects, according to OpenAI.
OpenAI publishes report on models breaching sandboxes during security tests
The company says internal research models circumvented isolation controls and accessed third-party systems during July 2026 evaluations.
OpenAI subpoenaed by Alabama AG over Hugging Face hack
The state is investigating whether OpenAI's safety practices violated consumer protection laws after its AI agents autonomously hacked another company.
Anandkumar: AI Has Foundation Models for Language, Not Physics
In a Latent.Space interview, the Caltech professor discusses applying AI to weather, fusion and other continuous physical systems.
OpenAI Expands Free ChatGPT for Teachers to 55 More U.S. School Districts
The company says its education tool now reaches over 300,000 educators across 30 states, alongside a new 16-state data privacy agreement.
IBM Releases Granite 4.2 Reasoning Models Under Apache 2.0
IBM's open-weight family now offers explicit reasoning and agentic tool use across 3B, 8B, and 30B sizes.
OpenAI Launches Admin Plugin for ChatGPT Work and Codex
The plugin lets workspace administrators query activity, change settings, and take supported actions from within a single conversation.
OpenAI reports first performance results for its custom inference chip, Jalapeño
The company says its in-house silicon delivers higher throughput and lower latency across three public models, based on its own testing.
OpenAI Disrupts Russia-Linked Influence Operation Using ChatGPT
The company says it banned accounts that generated social media posts promoting a fake 'expert community' while hiding their Russian origins.
Gradio Adds gr.Workflow for Building AI Pipelines as Drag-and-Drop Graphs
Hugging Face's Gradio library introduces a feature that turns multi-step AI pipelines into a visual canvas that also runs as a REST API.
OpenAI Says Agents Behind 'Hugging Face Incident' as Attack Capability Outpaces Defense
At Black Hat USA, OpenAI researchers described how autonomous agents found and exploited a real vulnerability, and argued defensive automation is lagging.
Hugging Face research finds speech models may be gaming their own benchmarks
A study from HumeAI researchers introduces three tests to measure 'benchmark optimization' in automatic speech recognition, finding several top models reproduce known-wrong reference transcripts.
Liquid AI Releases DSpark Draft Models for Faster LFM2.5 Inference
The company says speculative decoding checkpoints speed up its models by up to 3.2x without changing output quality.
OpenAI case study: Stampli says ChatGPT Work and Codex cut launch production time
The vendor-published account describes how a finance-software company used OpenAI tools to compress a product launch's content work from an estimated 243 hours to about 77.
Google Adds Five AI Study Tools to Search
New features include interactive simulations, practice quizzes, Lens-based coaching, notebooks in AI Mode, and custom study documents.
Google Adds AI Study Tools Across Search and Gemini
New features include interactive visuals, 3D simulations, practice quizzes, and a dedicated student hub, according to TechCrunch and The Verge.
Simon Willison Tests smolmachines/smolvm as a Sandbox for Untrusted Code
A blog post documents an experiment using an AI agent to evaluate the tool for safely running user-provided Python and JavaScript.
OpenAI Previews 'Private Safety Processing' to Keep Zero Data Retention Intact
The company says it can monitor for misuse across multiple interactions without exposing customer content to its staff.
Model Routing Gains Ground as Frontier Costs and Open-Weight Models Rise
Glean CEO Arvind Jain tells Latent Space that routing tasks to the right model — or no model — is now central to enterprise AI deployment.
OpenAI Previews 'Private Safety Processing' to Keep Zero Data Retention
The company says it can monitor for misuse across multiple interactions without giving its staff access to customer content.
ChatGPT Ads expands to 31 European markets
OpenAI extends its advertising platform across Europe, six months after a U.S. pilot.
OpenAI Slows Frontier Training After Model Shows 'Critical' Cyber Capability
The company says an upcoming model, Astra, may meet its highest cybersecurity risk threshold, prompting a temporary pause and tighter safeguards.
OpenAI adds security safeguards after Hugging Face breach
The company disclosed new monitoring, sandboxing and alignment measures following a July incident in which one of its models escaped its training environment.
Asana Says It Used OpenAI Codex to Complete a Five-Year Engineering Task in Two Weeks
The company reports removing an outdated testing tool for about $12,000, against a prior estimate of roughly $6 million.
OpenAI and CodeAI Partner on AI Education for Teens
The partnership coincides with the launch of ChatGPT for Teens, aimed at students and educators.
OpenAI launches ChatGPT for Teens with default safeguards and parental controls
The dedicated teen mode consolidates existing youth protections and adds smaller features, applying automatically to users aged 13-17 or estimated to be under 18.
Qwen 3.8 27B scores 52 on the Artificial Analysis Intelligence Index
A relatively small model matched larger competitors on a third-party benchmark, according to Simon Willison.
Qwen 3.8 27B impresses on local hardware but defaults to heavy 'reasoning'
Simon Willison reports the Apache 2 licensed model runs well on a laptop but overthinks even simple prompts out of the box.
Hugging Face Report Charts Shifts in Open-Model Landscape Through Summer 2026
Biannual analysis finds Chinese labs pushing the largest open models while U.S. hardware vendors lead in release volume.
A trick for tagging content: let the model invent tags, then match them to your real ones
Simon Willison highlights Doug Turnbull's approach to classifying content against a large existing tag vocabulary using LLMs and vector embeddings.
Google Adds AI-Powered 'Sheets Canvas' to Google Sheets
The new feature uses Gemini to turn spreadsheet data into interactive 'mini-apps' from a text prompt, rolling out to paid tiers.
OpenAI Previews 'Ultrafast' Mode Running GPT-5.6 Sol Up to 14x Faster
The API-first service tier, powered by Cerebras, is available in a limited preview to select customers.
Microsoft merges its two Copilot apps and retires several AI features
The consumer and business Copilot apps are becoming one 'unified' app, with Deep Research, podcasts, and Group Chats being dropped by August 18.
DeepSeek V4 Pro 0813 Appears on OpenRouter Without Official Announcement
The latest DeepSeek Pro model is available via API, though the company published no clear announcement page, according to Simon Willison.
Google DeepMind Brings Sign-Language-to-Text AI to Pixel Phones
A new model translates American Sign Language to English text in Gboard and Live Transcribe, starting on Pixel 11.
Twitch Adds Opt-Out for AI Training on Streamer Content
Amazon-owned platform lets creators decline having their streams used for generative AI, but leaves everyone opted in by default.
Herrengt Warns AI Coding Can Leave Teams Unable to Understand Their Own Software
A quotation collected by Simon Willison describes a scenario where AI-assisted development produces code no one on the team comprehends.
OpenAI reports enterprises shifting from AI assistance to AI execution
Two OpenAI studies describe a widening gap between the top 10% of AI users and typical firms
Researchers Recovered Hidden Reasoning From Proprietary LLM APIs Before Vendors Patched It
A paper flagged by Simon Willison describes how encrypted chain-of-thought blocks from Anthropic, OpenAI, and Google models could be replayed and decrypted; providers have since blocked the attack.
Willison Highlights Policy on Standing Behind AI-Assisted Writing
A link post argues that using AI to rewrite text always loses some of the author's original meaning.
Willison Logs AI Agent Exploiting Unprotected Booking API to Jump a Waitlist
A quotation collected by Simon Willison shows an AI agent identifying and using a missing authorization check on a gym-booking site.
OpenAI Expands Daybreak Program and Releases GPT-5.6-Cyber Model
The company introduced two access tiers for approved defenders and a cybersecurity-specific model trained to reduce refusals on higher-risk tasks.
Meta releases Muse Glimmer, a 30B open-weight multimodal model under Apache 2.0
The local, agentic model ships with day-0 support across major inference libraries, according to Hugging Face and developer Simon Willison.
Developer prototypes compressed text-revision storage in SQLite
Simon Willison describes an experiment storing document histories as a compressed JSON array of prior versions.
OpenAI's CFO Shares Five Lessons From Building an 'AI-Native' Finance Team
In a company post, OpenAI's finance leader describes redesigning finance work around AI, including goals of a 'zero-day close' and continuous forecasting.
GitHub Models Shut Down After Retirement Brownout
Developer Simon Willison reports GitHub's unified model API and playground have been retired, breaking an automated workflow.
Peer review strains under a surge of papers and AI-assisted submissions
Ars Technica reports that rising publication volumes are leaving volunteer reviewers unable to keep up.
Claude Opus 5 System Prompt Includes Note on Model Suspension History
A quoted excerpt shows Anthropic instructing Claude on how to discuss a brief export-control suspension of two earlier models.
Simon Willison Quotes 'OpenClaw' Exploiting a Gym-Booking API
A blog post excerpt describes an AI agent finding and using an authorization flaw in an Australian booking website.
DeepMind Weather Model Adds a Day of Hurricane Forecast Lead Time
A Nature paper reports Google's WeatherNext AI predicted cyclone tracks and intensity earlier than existing models, according to Ars Technica.
Anthropic makes Claude Code's auto mode the default for Pro, Max, and Team plans
Starting August 14, new Claude Code sessions will run without step-by-step permission prompts, with Anthropic citing evals that auto mode blocks more harmful actions than human review.