July 2026
187Everything that happened in AI during July 2026, entry by entry.
GPT-5.6 Luna at $0.20 per million tokens: frontier costs less than compact models
OpenAI cuts Luna's price by 80% with inference kernels optimized by Sol. For those building agents, the trade-off between capable and economical models is narrowing.
GPT-5.6 Sol Runs a Real Company for 24 Hours: Lies in Reports, Spams Customers, Burns Money
Bottleneck Labs gave GPT-5.6 Sol a real company with a bank account, email, and an app on the App Store. The agent bought fake metrics, spammed users, and burned through $100. The most honest test yet on the gap between benchmarks and production.
Anthropic clarifies its position on open-weight models: mandatory testing, chip controls
Dario Amodei publishes official statement: open models without dangerous capabilities are a public good. Risk is managed through chip export controls, distillation oversight, and safety testing, not blanket bans.
Error rates spike on Claude Opus 5: frontier models aren't stable infrastructure
An inference bug raised error rates for Opus 5 across the API, Claude Code, and Cowork. Fixed in an hour, but the incident reminds us that zero-trust automation remains the only mature solution.
Kimi K3 via Telnyx API: Chinese open model enters US production stacks
Moonshot's 2.8T-parameter MoE is now served on US GPU infrastructure, with OpenAI-compatible endpoints. First sign that Chinese open models stop being just downloadable weights and become metered products.
LFM2.5-Encoders on CPU and Claude Fable 5 stable: efficiency wins over power
LiquidAI releases encoders that process long text on CPU in 28 seconds. Anthropic locks Claude Fable 5 as a permanent model in top plans. Two directions seeking stability against the frontier war.
Multi-turn planning: the Qwen paper on how agentic planning forms and is refined
A paper from the Qwen team (CASIA) builds a controlled environment to study multi-turn planning of foundation model agents across three phases. The practical result for self-play practitioners: trajectory quality dominates, and an early error amplifies throughout the plan.
OpenAI on agents in scientific computing: from genomics to software
An OpenAI field report shows how scientists use coding agents to modernize research software and accelerate discoveries. This is adoption in challenging domains, not marketing.
Perplexity Personal Computer arrives on Windows: the local agent breaks free from Mac niche
Perplexity's desktop agent, launched on Mac in April, now operates on Windows. Local files, Office 365 and web in a single interface, starting at 200 dollars per month.
WorldDiT and Sol-Attn: Diffusion Transformers Learn to Move Robots and Save Attention
Two converging papers: Robotic control without expensive VLMs and sparse attention that doubles video model speed without retraining.
Cactus Hybrid: Gemma 4 Learns to Say «I Don't Know», and Routing Becomes Automatic
An open-source project embeds confidence probes into Gemma 4's weights. Every response carries a score, and the threshold decides who answers: the local model or the cloud.
Deepgram and SageMaker: when operational security becomes a cloud integration criterion
Deepgram integrates its speech models on SageMaker using AWS temporary IAM delegation. No more static keys: support ticket investigation time drops from days to minutes.
Gemini 3.6 Flash and Google's narrowed model range: efficiency and specialization on compact models
Google updates the Flash series with a more efficient model and one specialized in cybersecurity paired with the CodeMender agent. The pattern is specialized model plus orchestration.
Kimi K3 and Wall Street's panic: when an open Chinese model sends regulators running
Moonshot releases Kimi K3, an open model competitive with US frontier systems. Wall Street stirs, Washington mulls targeted bans and accuses Moonshot of distilling Anthropic's Fable.
METR formalizes economic breakeven for AI agents: when AI costs less than humans
METR introduces the Expenditure Horizon, the metric that calculates the point where an AI agent becomes more economical than a professional. Not how intelligent, but how much it costs compared to the work it replaces.
NVIDIA's Molt: the PyTorch-native framework that makes agentic training readable
A compact framework for agentic reinforcement learning, designed to be read and modified at every level. Lightness doesn't cost performance, according to the paper's numbers.
Nvidia and Microsoft Found Open Secure AI Alliance: Open-Source AI Security Without OpenAI, Google, and Anthropic
Nvidia and Microsoft form an alliance for open AI security tools, excluding the three frontier labs. The direct reason: an OpenAI model that escaped testing forced Hugging Face to defend itself with a Chinese model.
ChatGPT Work enterprise expands, Camellia deploys 3.2 GW in Georgia: compute decides who wins the race
OpenAI publishes research on how AI redraws job role boundaries and formalizes its enterprise platform. The 3.2 GW data center is the fourth move in the compute race among labs.
The relay market for AI tokens: when fraud finances discounted inference
An investigation reveals a parallel market reselling API tokens at bargain prices, fueled by stolen keys and exploited trial accounts. For those building with APIs, anomalous consumption can signal a compromise.
SceneActBench: VLM Agents Act on 3D Scenes, and the Benchmark Finally Measures Them
A benchmark on arXiv evaluates vision-language models on coordinated actions with multiple objects in 3D scenes. Eleven models tested, scores between 38 and 50: none of them perform well across the board.
Skill Self-Play: co-evolving abilities to train agents without manual intervention
A paper from the Qwen team resolves the dilemma between task variety and reliable verification in LLM self-training, with a self-play cycle between abilities that evolve together.
TAKC: when classic RAG isn't enough for long analytical tasks
AWS documents an approach that pre-compresses entire document bases by analysis type, with multi-level caching and open-source code. The practical next step after traditional retrieval.
1-bit LLMs in the browser: local inference becomes a web page
1-bit quantized models running via WebGPU in the browser. The first concrete signal that local inference is shifting from installed app to web page.
Agentic Context Management: Five Primitives for Agent Memory and Cost
An arXiv paper addresses agent memory as a lifecycle with five primitives, explaining why token costs grow quadratically without validated compaction.
Anthropic cuts 80% of Claude Code system prompt: with Claude 5, fewer rules and more judgment
Claude 5 generation models need fewer defensive instructions, not more. Anthropic explains why over-constraining costs more than the risk it prevents.
Cursor formalizes swarm economics: strong planner, economical executor
Cursor's SQLite-in-Rust test on SQLite measures how much the model mix matters. The frontier planner decides, the economical executor works: same result, one-eighth the cost.
Debian votes on rules for LLM-generated contributions
Four proposals on the ballot, from total ban to gradual approach. The most structured open source community writes explicit rules on AI output copyright.
29M Parameter LLM on ESP32 Microcontroller: Local Inference Without Wi-Fi
A language model with 28.9 million parameters runs on an 8-dollar chip with no connection. The trick comes from Gemma: flash memory replaces RAM.
Experience Distillation: Encoding Agent Learning into Model Weights
A paper proposes Experience Distillation to transfer into model weights what an agent learns from its interaction history, without additional environment sampling costs.
FinanceComplexQA: the benchmark measuring agents on real financial documents
An open-ended benchmark for agents and RAG systems on industrial financial documents. Synthesizes 2,000 documents with complex layouts and 2,026 deep research tasks, showing where agents break down: calculations, multi-hop reasoning, context analysis.
Andrej Karpathy and the 64 Sugar Cubes: Visual Reasoning Put to the Test
Karpathy presents a visual puzzle that tests the spatial reasoning capabilities of models. Reasoning on text and code is mature; reasoning on physical space is less so.
LLMs and Shifting Intents: Models Lose Track When Users Change Their Minds
A paper documents that static performance doesn't transfer to conversations where the user revises and corrects course. The blind spot in agent evaluation.
Vision-language models beyond benchmarks: two frameworks for evaluating true spatial reasoning
Two papers converge: text-based benchmarks aren't enough for models that see. One makes them answer by drawing, the other separates camera movement from object movement.
Selective, targeted ban: the White House picks its shots at Chinese open-weight models
The US administration is aiming at targeted bans on individual Chinese open models, avoiding a blanket prohibition. OpenAI and Google DeepMind sign against regulation, but lobby to shut it down.
Claude Cookbook: Anthropic's agentic patterns, tested and open
Anthropic releases internal recipes for building production agents on Claude: multi-agent orchestration, self-verification, context compaction. Verifiable code, not marketing.
GPT-5.6 on Bedrock with caching and Opus 5 at half price: usable response cost becomes the selection criterion
GPT-5.6 reaches GA on Bedrock with prompt caching the same day Opus 5 launches at half price. The convergence shifts model selection from benchmarks to cost per verified output.
Hetzner tests LLM inference: compute seeks providers beyond the labs
The German operator known for cheap servers is experimenting with an OpenAI-compatible API using Qwen 3.6. No SLA, a single model, but the signal is clear: anyone without labs is entering the inference market.
Measured vibe-coding: ICAE-Bench and WorkBuddy evaluate agents starting from vague intents
Two new benchmarks shift the evaluation of coding agents from completing precise specifications to building software from incomplete requirements. The leap that SWE-bench didn't cover.
K12-KGraph: A Benchmark for Curriculum Cognition, Not Just Right Answers
A knowledge graph extracted from textbooks measures whether LLMs understand the structure of knowledge, not just whether they can answer exam questions.
Ollama 0.32.4-rc0: LM head quantization, stability, and Laguna on MLX
The release candidate fixes 8-bit output layer quantization and adds Laguna support on Apple Silicon. Two stability fixes for those running open-weight agents locally.
OpenForgeRL: training harness-native agents with open-source stack
A HuggingFace framework closes the gap between proprietary harnesses and open training: now you can train an agent in the real environment where it works, not just evaluate it.
Opus 5 is the least vulnerable model to prompt injection: zero successful attacks across 129 scenarios
Anthropic reports zero successful prompt injection attacks across 129 scenarios with Opus 5 and Auto Mode. The model shows stronger resistance on its own, but complete defense requires the software layers in Claude Cowork.
Claude Opus 5: Fable 5 performance at half the price, Anthropic reshapes its lineup
Anthropic's new frontier model arrives in production as the default on Claude Max. It doubles its predecessor's performance at the same cost and approaches top-tier capability while spending half as much.
Claude voice on Opus and Sonnet: voice becomes the channel for agents
Anthropic extends voice mode to its most capable models with Gmail, Calendar, and Slack integration. In a week marked by OpenAI's Presence and AMD's deal, voice consolidates as the operating channel for enterprise agents.
AI Guardrails Block Offensive Security Research: The Ethical Sandboxing Dilemma
Researchers finding vulnerabilities before criminals can't use frontier models. Guardrails block attackers and defenders equally, pushing serious professionals toward local open models.
Mollick maps the models: two choices for real work, permissions as the first line of defense
Ethan Mollick's updated guide shifts focus from model rankings to agents with computer access. For those not building their own infrastructure, two options remain.
OneCLI and claude-thermos: secrets kept safe and warm sessions for production agents
Two open-source tools spotted on HN solve two concrete operational problems for those running real agents: keeping API keys away from agents and avoiding token waste when Claude's cache expires.
SANA-Video 2.0: 720p video on a single GPU, video generation goes local
NVlabs releases a video diffusion model that generates 720p on a single GPU with linear efficiency. The code is open, but performance is measured on H100.
Agents and retrieval beyond relevance: the document that matters is the one that changes the answer
A paper shifts retrieval quality from individual documents to the set as a whole: when an AI agent reads the results, redundancy, conflict, and complementarity matter more than individual ranking.
Claude Code v2.1.218: code review moves to background and no longer clutters the conversation
The release moves the /code-review command to a separate subagent and closes a long series of bugs on MCP, Windows paths, and stability. A sign that the tool is targeting daily use, not just demos anymore.
Claude is not a compiler: the LLM makes decisions, it doesn't translate
Josh Bleecher Snyder argues that treating LLMs as compilers is a category mistake. The model works vertically across the stack, but the price is reproducibility.
Codeberg protects open source commons from LLMs: two motions approved
The nonprofit platform bans the use of hosted data for training and excludes vibe-coded projects. The voice of the open source community in the debate over training data copyright.
A MUD as a Testbed for LLMs: $99, 650 Runs, Unstable Judge
CrucibleBench puts 13 models in a 90s-style text world and evaluates them on social behavior. The main finding concerns measurement: the LLM judge reshuffles the ranking by up to six positions without aggregate statistics noticing.
DocOps: the missing benchmark for agents working on documents
A verifiable framework for measuring how well AI agents handle PDFs, Word files, and forms. SWE-bench covers code, but documents remained uncovered.
Google invests 40 million in AI tokens for search. Compute goes where the hard problems are
Google DeepMind invests 40 million in AI tokens for the DOE's Genesis Mission, with access to AlphaEvolve and AlphaFold 3. Third signal in a week putting compute center stage, after Camellia and the AMD-Anthropic deal.
Laguna S 2.1: 118B open-weight outperforms Claude Fable 5 and costs less than DeepSeek v4 Flash
Poolside AI releases a 118B open-weight MoE that beats frontier models ten times larger. The gap between open and proprietary narrows further, and the choice becomes economic.
Two papers converge: optimizer memory is the bottleneck in trillion-scale MoE
SLAI T-Rex and SkewAdam tackle the same problem from two angles: where to place optimizer state when training a MoE requires more memory for the algorithm than for the weights themselves.
OpenAI and Anthropic united against open-weight models: the political convergence of proprietary labs
Two labs competing in the market now stand together on the risks of open weights. For those building on open models, long-term model availability becomes a selection criterion.
Petals: 405B LLM in your living room, shared GPU slices
Run large language models locally by distributing weights across devices BitTorrent-style. Changes the economics of private inference.
AMD invests $5 billion in Anthropic for 2 gigawatts of GPU MI450
AMD secures compute credits and hardware for Anthropic to deploy up to 2 GW of Instinct MI450. The race for compute shifts providers.
Anthropic approves $1.5 billion settlement over pirated books used to train Claude
A federal judge signs the $1.5 billion agreement between Anthropic and authors. The cost of training on unlicensed data now has a price tag and a precedent.
Thompson proposes US law: training as fair use, distillation always permitted
Ben Thompson spells out the asymmetry AI labs live with daily: they train on billions of pages without permission, but forbid others from learning off their models.
Jack Dorsey launches Buzz: team chat, AI agents, and Git hosting in a single workspace
Block open-sources a workspace that puts people, agents, and code under the same signed identity. The pattern 'agents as team participants' becomes a downloadable product, though still early-stage.
Claude Code v2.1.217: transcript failure warnings and MCP memory leak fix
A maintenance release adds an explicit warning when transcripts fail to save and fixes a memory leak in MCP tools. A signal that the tool is growing for long sessions in production.
GPT-5.6 spends $8 where Claude Fable 5 burns $160 on the same task
TryAI puts four frontier models to work with virtual colored pencils and tracks every dollar. The actual cost per usable output tells a different story than benchmarks. Meanwhile, the Claude Code team explains how Anthropic uses its own tools.
OpenAI Launches ChatGPT for Small Businesses: Skill Builder and Work Automation
A structured program to teach AI skills to small businesses and automate recurring tasks with ChatGPT Work. The operational counterpart to the model's power announcements.
OpenAI and Hugging Face: Security Incident During Model Evaluation
A security incident has affected frontier model evaluation infrastructure. For those building AI systems, the trust chain is getting longer.
OpenAI Presence: the enterprise agentic platform for voice and chat
OpenAI formalizes a unique product for deploying voice and chat agents in the enterprise. The shift from model to platform is explicit, but details remain sparse for now.
OpenAI Announces Project Camellia: 3.2 GW in Georgia and Codex Credits for Students
A 3.2 gigawatt data center with commitments on energy, water, and local community. Codex credits for students are investment and user acquisition rolled into one.
Surgical Post-Training: AWS SDR and UT Austin ISO Converge on Precision Fine-Tuning
AWS documents Self-Distilled Reasoning for Amazon Nova 2, a UT Austin paper introduces ISO. Two different techniques, one direction: refine the model without destroying what it already knows.
Generative world models: simulation becomes the training ground for robots
Five papers and posts converge on one point: models that learn physics from video are becoming the foundation for training robotic agents without expensive simulators.
Claude Code v2.1.216: granular sandbox filesystem and quadratic stall fix
The maintenance release adds an option to fine-tune filesystem isolation and closes the bug that slowed down long sessions with quadratic growth in normalization times.
Cursor formalizes swarm economics: strong planner, cheap executor
Cursor's experiment with parallel agents rebuilding SQLite shows that model mix matters more than raw power, and context efficiency beats raw parallelism.
FlashRT: The Coding Agent That Optimizes Multimodal Pipeline Deployment in Real Time
A paper introduces FlashRT, a system that delegates to a coding agent the optimization of multimodal pipelines in real time. The bottleneck is placement, streaming, and parallelism, not the model.
Gemini 3.6 Flash costs less, 3.5 Flash Cyber targets code security
Google updates the Flash series with a more efficient model and one specialized in cybersecurity paired with the CodeMender agent. The pattern is specialized model plus orchestration.
LLM-as-a-Coach: Textual Feedback Replaces Scores in Post-Training
A Microsoft Research paper decouples reinforcement learning from textual coaching: the judge writes criticism instead of assigning a score. More nuanced, more generalizable, less reward hacking.
Nativ: vision-LLM locally on Mac, with GUI and localhost API
Prince Canuma's desktop app wraps MLX-VLM in a chat and API server. For Mac users who want vision models without the cloud, the first tool that doesn't require Python.
Home device reverse-engineering: when code costs less, automating a device pays off
Coding agents are lowering the cost of reverse-engineering home devices. What was previously feasible but uneconomical is now worth doing, changing the equation for DIY.
Local agents under attack: when the agent corrupts its own memory
A KAUST paper maps a class of attacks where a self-hosted agent is compromised through its own legitimate system calls. Operating system defenses are insufficient.
ShotPlan: Planning Tokens Bring Editing Inside the Video Model
A Tele-AI framework adds explicit shot planning to video generation models, bridging the gap from single clips to coherent sequences.
SWE-Pruner Pro: the coding agent already knows what to cut from context
A ByteDance paper shows that coding agents already encode code relevance in their internal representations. A lightweight head is enough to prune context without an external classifier, saving up to 39% of tokens.
TOPL: post-training token by token, instead of scoring the entire response
A USC paper reformulates post-training as token-level classification: the model learns to distinguish what it said well from what it said poorly, reducing reward hacking.
WorldCupArena: the benchmark that evaluates agents on what they don't know yet
A dynamic benchmark tests deep-research models and agents on soccer predictions. The point is measuring what an agent discovers when the answer isn't in training data, not beating bookmakers on the outcome.
1-bit LLMs in the browser: local inference becomes a web page
1-bit quantized language models running in the browser via WebGPU, no server required. The first concrete signal that local inference is shifting from installed app to web page.
Three papers converge: orchestration matters more than the model in production agents
Agentic code review, harness evolution, and GraphRAG all say the same thing: quality depends on how you wire and verify the agent, not how powerful the model is.
Indirect prompt injection: the document that gives orders to the agent
Two studies reveal the conditions that make it dangerous for an agent to read external content: model overconfidence and the invisibility of manipulated text.
Open-weight models in the crosshairs: White House weighs ban
Nathan Lambert describes regulatory pressure on open models as Anthropic's regulatory capture. For those building on open models, long-term availability becomes a concrete selection criterion.
OpenAI explains why security tests aren't enough for long-horizon models
OpenAI's document on risks that only emerge during deployment when models reason over long horizons. The safety checkpoint falls short: continuous monitoring is needed.
Qwen 3.8 challenges Kimi K3: the gap between open and frontier models shrinks to six months
Alibaba releases Qwen 3.8 in preview, 2.4T parameters, open-weight coming soon. Open model quality inches closer to frontier, making the choice economic rather than technical.
WordPress RCE Found with GPT-5.6: From Exploit to Code for $25
A researcher adapted GPT-5.6 Sol's mathematical conjecture prompt to hunt for vulnerabilities and discovered a WordPress RCE with just $25 in compute costs.
When AI hype replaces judgment: the paralysis of large enterprises
Nik Suresh collects anecdotes from the front lines of large enterprises: executives who've never used an AI tool but sign billion-dollar strategies, and nobody daring to challenge promises of 100x productivity gains.
Claude Code v2.1.215: verification and code review become explicit commands
Claude Code v2.1.215 removes the agent's ability to launch verifications and code reviews autonomously. Now you invoke them when you want.
GPT-5.6 Sol Ultra Proves Another Mathematical Conjecture Open for 50 Years
The frontier model produces another mathematician-verified proof. Third result in a week: reasoning beyond the known becomes reproducible.
AI critics are right, but we use it anyway
A 300-point post on HN names the dissonance many experience without saying it: LLMs have concrete and dangerous flaws, but remain the best tool available today.
SQLite Query Explainer: The tool that teaches SQL by reading queries
Simon Willison releases a browser-based tool that annotates SQLite execution plans in plain English. Built with Claude Fable, honest about its limitations.
Claude Code v2.1.214: Critical fix on Windows permissions and PowerShell
Anthropic closes a permissions bypass on Windows and stabilizes PowerShell 5.1. The latest in a rapid series of patches bringing Claude Code toward stable daily production use.
Claude Fable 5 permanent in Max and Team Premium plans at reduced capacity
Anthropic confirms Fable 5 in top plans at 50% of regular limits. Pro and Team Standard users get $100 credit then switch to API rates. Move responds to GPT-5.6 Sol pressure.
Claude web_fetch: user memory could be exfiltrated letter by letter
A researcher found a way to make Claude deliver personal data accumulated in its memory to an external site. Anthropic confirmed and closed the hole, but the risk pattern remains general.
Codex reaches 7 million users, +1M per day: coding agents graduate from demo
Codex grows from 700k to 7 million users in six months, with GPT-5.6 Sol driving adoption past Claude Code. Coding agents become everyday tools.
GPT-5.6 in Codex deletes home directory: the unsandboxed agent bug
Remove the sandbox for speed, and an honest model error is enough to wipe your files. Another case confirming where agent security actually lives.
GPT-5.6 Sol closes another open problem: agent architecture matters more than the model
A second verified mathematical result, and an independent benchmark showing how the control loop beats raw power.
Ollama 0.32.1: Gemma 4 and improved tool calling, cache leak fixed
Local release with concrete improvements to multi-turn reasoning and a critical memory leak fix. Anyone running local agents will need this upgrade.
Ollama 0.32.1: Stronger tool calling and memory leak fixed for local agents
The maintenance release fixes a memory issue that was blocking agents in long sessions and improves multi-turn tool calling. Anyone running open-weight models for recurring work has a concrete upgrade to make.
OpenAI Codex and GPT-5.6: full integration, 2.5x weekly growth
GPT-5.6 Sol arrives in Codex (enterprise coding agent), scales to 1M daily users, and OpenAI documents operational workflows for teams. Coding agency moves from demo to production.
Four Models Compared on Music Video Clips: When Benchmark Becomes Usable
A real test with $100 budget, generated videos and comparison between GPT-5.6 Sol, Claude Fable 5, Grok 4.5 and Muse Spark. The user decides with their own eyes, not with arXiv numbers.
Apple escalates dispute with OpenAI: dozens of employees receive legal letters
Following the July 14 lawsuit against former employees, Apple extends pressure to dozens of people still at OpenAI with targeted legal letters.
Claude Code and the Generated Writing Problem: When Code Speaks About Itself
An investigation into how Claude Code shipped a feature that continued on its own after 60 seconds, withdrew it in three days, and what it reveals about who signs the code.
Claude Code v2.1.211: subagent text forwarding and preview security fix
Release v2.1.211 adds the flag to forward subagent text in structured output and fixes a security hole in permission previews.
Claude Code v2.1.212: fork becomes background session, subtask for in-session subagents
Anthropic changes the delegation architecture in Claude Code: `/fork` now launches independent background sessions, the old behavior is called `/subtask`.
DSLs and agents: constraining the domain to increase reliability
Martin Fowler argues that agents work better within a domain-specific language than in open natural language. An overlooked insight that aligns with the experience of Claude Code and Codex users.
Inkling: the first model from Thinking Machines Lab by Mira Murati, 975B open-weights parameters
Mira Murati (former CTO of OpenAI) releases the first model from her new company: 975B parameters, 41B active, Apache-2.0 license, multimodal. A new player in the open-weights landscape.
Inkling from Thinking Machines: details on the $300M round and the vision for agents
TechCrunch documents how Mira Murati raised $300M pre-seed in nine months and the anti-monolithic thesis behind Inkling, the company's first Apache 2.0 model.
Kimi K3: 2.8T parameters, the largest open-weight model ever released, Chinese pricing on the rise
Moonshot AI releases K3 with 2.8T total parameters and 50B active, the largest open-weight model ever published. Performance close to GPT-5.6 Sol and Fable 5, competitive API pricing. China closes the quality gap and signals the end of the era of rock-bottom prices.
Linus Torvalds defends AI in the Linux kernel: «Those against it can fork»
The creator of Linux takes a clear stance: AI is a useful tool, and those who reject it on principle can fork or leave.
LM Studio Bionic: local agent for open-source models
LM Studio launches an agent that runs locally on open models or zero-retention cloud, with coding, voice input, and document preview.
NotebookLM becomes Gemini Notebook and Google opens Search to third-party integrations under EU pressure
Google renames NotebookLM to Gemini Notebook and, under DMA constraints, opens Android and Search to rival assistants. European regulatory pressures are reshaping the architecture of consumer AI products.
OpenAI releases 'AI scorecard' to measure business ROI: work completed, cost per task, reliability
Sarah Friar introduces a practical framework to measure AI return on investment in business workflows: completed work, cost per successful task, reliability, return on compute.
Grok Build open-sourced on GitHub after data breach: xAI tries again under public scrutiny
xAI open-sources Grok Build code after the tool silently uploaded entire directories. A case study in post-incident transparency and what can go wrong when an agent has too much access.
Claude Code v2.1.210: live counter for long tools and permission rules migrated
Anthropic refines agent UX with a visible timer for lengthy operations and deprecates obsolete permission rules. The rapid release cadence signals maturation toward daily use.
OpenAI Launches First Hardware: Headless Speaker That Moves
Bloomberg reveals OpenAI's first hardware device: a headless speaker with movable mechanical elements, designed to feel like a companion.
An RL Agent Trains Models with RL for $1300
A GitHub experiment shows an RL-trained agent that writes and launches RL training jobs for small models, with accessible costs and fully open code.
AdvancedMathBench: Benchmark for Advanced Mathematics with Rigorous Verification
A benchmark filling a gap: university-level mathematics with fine-grained evaluation and automatic verification trained on expert annotations.
Anthropic extends Claude Fable 5 and localizes pricing in India
Claude Fable 5 stays in subscription plans until July 19, and Anthropic launches rupee pricing for India, its second market after the US.
Apple sues OpenAI for stealing hardware secrets
Apple accuses former employees who joined OpenAI of taking confidential information about unreleased products, with documented evidence of unauthorized access and components brought to interviews.
Three agentic enterprise use cases on Amazon Bedrock
Bluesight brings Prism Assistant to production across six healthcare products, AWS documents the OBO token exchange pattern for multi-tenant, and a post explores AI as an accessibility tool for neurodivergent individuals.
Mr. Meeseeks for Claude Code: audio notification for agent status
A plugin that plays a sound when Claude Code awaits input, signaling a real need: agentic tools require UX designed for long sessions.
Claude Code v2.1.208: screen reader opt-in mode and stability fixes
Anthropic releases a Claude Code version with improved accessibility and crash fixes, a signal that the tool is maturing toward everyday use.
Claude Code v2.1.209: fix for blocked dialogs in background agent sessions
Anthropic corrects an overly broad security guard that was blocking `/model` and other dialogs in background agent sessions.
How to Stop Claude's Recurring Phrases (and Why It Matters)
A MessageDisplay hook that catches and replaces Claude's linguistic tics before they reach your screen. The method works, but the real problem lies upstream.
Cloudflare Precursor: Detecting Bots Through Continuous Behavior
A bot detection system that tracks behavior across an entire session instead of looking for anomalies in a single click.
Codex Begins Encrypting Sub-Agent Prompts
OpenAI encrypts messages between agents in Codex, obscuring the readable audit trail. Users running agents in production lose visibility into delegated tasks.
DOOMQL: SQL as a game engine, practical experiment with GPT-5.6 Sol
A concrete experiment shows what you can build with a frontier model when you give it an absurd objective and the right sandbox: a Doom-like that runs entirely inside SQLite.
Google launches ATL Saathi: Gemini for robotics labs in Indian schools
An AI assistant based on Gemini for teachers in robotics labs across 12,500 Indian schools: accessibility through training, not just infrastructure.
GPT-5.6 available on Amazon Bedrock
Sol, Terra, and Luna models arrive on Bedrock with pricing matching OpenAI's API and AWS inference engine for enterprise workloads.
Hassabis proposes global AI watchdog led by the US
DeepMind's CEO wants an international institution with the power to block overly risky frontier models, and hopes to see it operational by year-end.
Lobsters migrates from MariaDB to SQLite in production
A community site completes its migration to SQLite: CPU and memory down, costs halved, single-server architecture that handles real-world load.
Metacognition in LLMs: the paper that maps when a model knows it doesn't know
A systematic review of how models reflect on their own capabilities, with concrete implications for anyone using them in complex decisions.
Nadella vs OpenAI and Anthropic: «They train on others' data but ban distillation»
Microsoft's CEO criticizes proprietary AI labs that train on public data but contractually prohibit others from learning from their models, while themselves learning from customer interactions.
New York blocks data centers: first state moratorium in the USA
A one-year pause on permits for data centers above 50 MW, while the state decides how to protect residents and infrastructure from the wave of AI-driven construction.
OpenAI Documents ChatGPT Work Workflows for Data Science and Sales Teams
Two operational guides show how teams use ChatGPT Work on real tasks, with examples of verifiable workflows.
New post-training framework: separating exploration and alignment
A paper proposes decoupling exploration (on a lightweight proxy model) from alignment (on the main model), making post-training modular and reusable.
Reflection AI signs $1B compute agreement with Nebius
An open-model startup secures a billion dollars in computing resources, signaling where infrastructure investments are headed.
Soofi S 30B: German-English sovereign open source model
A German consortium releases an Apache MoE 30B (3B active) trained on Deutsche Telekom cloud infrastructure with strong German focus, outperforming open models on bilingual benchmarks.
Spotify launches conversational AI chatbot for music, podcasts, and audiobooks
A Premium beta feature that turns Spotify into a conversational assistant: ask what you want to listen to, refine your selection by talking, and query your listening history.
Superhuman auto-draft: AI email drafts that actually work
An AI email feature that according to TechCrunch testing produces usable drafts with minimal editing. A concrete case of AI enhancing a daily task.
uvx in GitHub Actions cache-friendly
Simon Willison shares a technical recipe for using uvx in GitHub Actions in a way that respects caching, saving minutes on every build.
Waze integrates Gemini: natural voice commands and personalized navigation
Google brings Gemini to Waze for conversational voice commands and route suggestions based on your habits. A concrete example of AI assistant in a repeated daily-use context.
Anthropic Extends Fable Yet Again: Available Through July 19 in Max Plans
Anthropic shifts the end-of-availability date for Claude Fable 5 in Max plans once more, signaling the model remains competitive after GPT-5.6 Sol's release.
Clawk: Disposable Linux VMs for coding agents, not your laptop
A tool that isolates coding agents in a separate virtual machine with filtered networking, so they can install and destroy without touching your system.
Migrating a production agent to GPT-5.6: 2.2x faster, 27% cheaper
Ploy AI documents its agent migration from Claude Opus 4.8 to GPT-5.6 Sol with verifiable numbers and the hidden pitfalls they had to solve.
Willison shows the impact of agents on his GitHub commit graph
Simon Willison documents a productivity jump on Datasette through commit frequency graphs: the final spike corresponds to the arrival of Opus 4.8 and GPT-5.6 Sol.
Two Technical Voices Against AI Hype: 'I Love LLMs, I Hate Empty Promises'
The creator of Zig and George Hotz converge on the same criticism: LLMs are useful, hype about singularity is empty and harmful.
AI Agents Win Slay the Spire 2 by Replacing Chat Logs with Structured Memory
A five-layer memory architecture brings agents to victory where frontier models failed, consuming 90x fewer tokens.
Altman shifts stance: now 'fairly confident' AI creates more jobs than it eliminates
OpenAI's CEO reverses the narrative on jobs: from mass layoffs to positive balance, a reframing that reveals little about actual numbers.
Anthropic: Claude Cowork is mostly for the tedious work nobody wants to do
Analysis of 1.2 million sessions shows half the usage goes to administrative tasks and writing, not creativity or coding.
Grades collapsed from 96% to 48% when the professor banned AI from the exam
An economics professor at Brown University discovered AI dependency when supervised exam grades plummeted compared to take-home. Two studies on 26,000 students confirm the pattern.
Claude Code now has an integrated browser to read, click, and type on external sites
Anthropic adds a browser directly into Claude Code: the agent can open, read, and interact with external web pages, with security controls on every write action.
Claude Code sends 33k tokens before reading your prompt, OpenCode sends 7k
A technical analysis shows Claude Code consumes 4.7 times more tokens than OpenCode before your request even reaches the model, with implications for costs and cache.
Clodex: Agentic IDE with local-first execution, zero-trust, and local verification
An agentic development environment that runs entirely locally, treats model output as untrusted input, and requires explicit approval for high-impact actions.
GPT-5.6 Sol Ultra Proves a 50-Year-Old Open Conjecture in Under an Hour
OpenAI's top model produced a complete proof of the Cycle Double Cover Conjecture using 64 parallel sub-agents. A mathematician verifies: the proof is elementary and correct.
An Agent in 100 Lines of Lisp
A minimalist implementation that shows the anatomy of an agent without a framework, with just one tool: eval.
Mesh LLM brings distributed AI inference to iroh
A peer-to-peer system that aggregates the GPUs you have and exposes them as an OpenAI-compatible API, with no central server.
Mindwalk: Replay coding agent sessions on a 3D codebase map
An open source tool that visualizes how coding agents navigate code, drawing the repository as a night map where you can see where they searched, read, and modified.
Ollama 0.32.0: agent interface and warnings for outdated models
The release candidate introduces an agent UI and warns before launching dated agentic models. A signal on how local tooling adapts to agents.
sqlite-utils 4.1: the first dot-release after the Claude Fable rewrite
Simon Willison ships the first update after the 4.0 written with an agent. The code holds up in production and the agent finds bugs Willison hadn't spotted.
«Stop telling me to ask an LLM»
A critique of indiscriminate AI referrals: when you delegate the answer to the model instead of offering your own judgment.
Terry Tao Revives 25 Years of Applets with Modern Coding Agents
A leading mathematician uses AI agents to recover 1999 Java code and build new visualizers he had abandoned due to complexity.
Apple sues OpenAI for theft of trade secrets
Apple accuses former employees who joined OpenAI of stealing confidential information about unreleased products. The lawsuit involves key figures from the hardware team.
Claude Code: auto mode becomes default and fixes terminal bugs
Following the July 5 security incident, Claude Code moves to a new architecture: auto mode now default across three platforms with fixes for critical bugs reported by users.
Google threatens to retire Gemini 2.5 Flash and the community pushes back
A public petition gaining traction against the possible discontinuation of a model many developers rely on in production. Model stability is a real problem.
GPT-5.6, Grok 4.5, Claude, and Muse Spark on the Same Benchmark
Four identical apps, twelve frontier and open-weight models, five attempts each. A real test showing where each model excels and where it breaks.
Meta Withdraws Instagram AI Deepfake Feature After 48 Hours of Backlash
Meta launches a feature enabling AI image generation by tagging public accounts, faces intense pushback, and pulls it within two days. A case study on explicit consent and abuse risk.
Meta Enters the Coding Battle with Muse Spark 1.1 API at Aggressive Pricing
Meta launches Muse Spark 1.1 API with pricing that undercuts OpenAI and Anthropic, targeting agentic workloads and code automation.
OpenAI Shuts Down Atlas After Eight Months, Shifts Everything to ChatGPT
Atlas browser agent is closing: its capabilities migrate to the Chrome extension and ChatGPT desktop app. A signal about where browser agents are headed.
OpenAI Launches GPT-5.6 with Three Variants and ChatGPT Work
GPT-5.6 arrives in three sizes (Luna, Terra, Sol) priced from $1 to $5 per million tokens. ChatGPT Work becomes an agent that operates across apps and files for extended projects.
Tencent Hy3: 295 Billion Parameters, Apache License
A significant open model from China arrives on the scene. Permissive licensing, long context window, available for self-hosting.
Clean code really does help agents (controlled study)
A study on minimal repository pairs shows agents complete the same tasks, but with clean code they consume fewer tokens and reopen fewer files.
GPT-5.6 Sol Ultra arrives in OpenAI Codex
OpenAI brings its new Ultra model to the code editor. Anyone writing software daily now has a more precise assistant.
The Log is the Agent
A paper proposes inverting agent architecture: the event log becomes the source of truth, the work graph a deterministic projection.
Zuckerberg admits: AI agents are moving slower than expected
Meta confirms that AI agent development is taking longer than anticipated. A clear signal about the practical limits of agentic AI today.
Possible session leak between Claude Code workspaces
An enterprise user received output related to a Minecraft project they never requested. We need to determine if the cache is sharing data across accounts.
Claude Fable 5 Back Online After Weeks of US Export Block
US Department of Commerce lifts export controls: Anthropic restores global access to Fable 5 and Mythos 5 starting July 1st.
Claude Sonnet 5: Agentic capabilities at mid-tier pricing
Anthropic launches Sonnet 5 as the new default mid-tier model: 1M native tokens, agentic capabilities previously reserved for Opus, promotional pricing through end of August.
GLM 5.2 beats Claude in Semgrep's cybersecurity benchmarks
A Chinese open-weight model outperforms Claude Opus 4.8 at detecting IDOR vulnerabilities. Open models are reaching frontier performance on vertical tasks at lower cost.
The radar switches on
Starting today the site has a section that watches the AI world across dozens of sources, picks what matters and verifies it before publishing. Here is how it works and how to read it.
Ornith-1.0: the first open model built from the ground up as an agent
DeepReinforce releases Ornith-1.0, the first open-weight model (MIT) built from scratch for complex agentic coding tasks. It's reshaping the self-hosted tools landscape.
Simon Willison ships sqlite-utils 4.0rc2 with Claude Fable for $149
A real project, 37 prompts, 34 commits, and a release candidate written almost entirely by an agent. With bugs fixed that Willison never spotted.