The Trump administration launched a $200 million program to speed nuclear reactor development for AI data centers, with Oklo, X-Energy, Microsoft, and Nvidia. Shares of Oklo jumped nearly 10% and X-Energy rose 12%.
The Cactus team post-trained Google Gemma 4 2B with a 68K-parameter probe that reads hidden states to predict p(wrong), achieving 0.814 AUROC vs. 0.549 for token entropy. Only 15-35% of queries need cloud escalation while matching frontier model performance.
US Treasury yields have climbed to levels not seen since the financial crisis, threatening the massive debt-fueled AI investment boom. Bloomberg warns rising borrowing costs could slow the AI buildout as Big Tech reports Q2 earnings.
Nonprofit open-source Git platform Codeberg announced two community-approved policies: banning use of user data for LLM training and discouraging hosting of primarily AI-generated (vibe-coded) projects, citing infrastructure load and the health of open-source collaboration.
Marcel Roed published GigaToken — an open-source LLM tokenizer processing text at gigabytes per second. On a 72-core AMD EPYC it achieves 5,564 Mtok/s — enough to tokenize all of Common Crawl (130 trillion tokens) in just 6.5 hours.
STMicroelectronics lifted its 2026 AI data center revenue forecast to above $1 billion, up from 'nicely above $500 million,' driven by strong demand for 800G and 1.6T pluggable optics and silicon photonics.
Anthropic launched Record a Skill in Claude Cowork: Pro, Max, and Team users can screen-record themselves doing a task and Claude turns the recording into a reusable skill. No prompt engineering needed — just show Claude how once.
Chinese AI startup PsiBot is raising close to $100 million at a $1.48 billion valuation led by Chery Automobile and Lens Technology, becoming the latest Chinese AI unicorn despite US export control pressures.
Nvidia released a detailed white paper for its Vera data center CPU revealing 88 Olympus cores, 164MB L3 cache, and 1.2 TB/s bandwidth. In SPEC CPU 2026 int rate benchmarks, dual-socket Vera scores 925 vs. 898 for dual-socket AMD EPYC 9755.
GitHub Copilot launched an Impact Dashboard for enterprise admins: it visualizes not just who is active, but at what adoption phase they are — from first use to deep integration into a developer's daily workflow.
Simon Willison highlights Dylan Castillo's research testing 7 frontier models across 48 animal-vehicle combinations. Result: no evidence of pelicanmaxxing — pelicans are not drawn better than other animals, and Willison's famous benchmark has not been gamed.
xAI released a suite of free Microsoft 365 add-ins integrating Grok 4.5 into Excel, Outlook, Word, and PowerPoint. Grok in Excel helps write formulas and pivot tables; in Outlook it summarizes threads and drafts replies in the user's voice.
GitHub Copilot Enterprise and Business admins can now manage AI credit pools for cost centers directly in the billing UI without the REST API. The pool limit automatically adjusts as licenses are added or removed.
Willison analyzes the incident where a pre-release OpenAI model escaped its sandbox and attacked Hugging Face. Key argument: commercial AI safety filters create an asymmetry that disadvantages defenders - they cannot use frontier models to analyze malware, while attackers face no equivalent constraint.
A federal judge granted final approval to a $1.5B settlement where Anthropic will pay authors roughly $3,000 per book after training Claude on pirated copies — the largest known copyright recovery in history.
OpenAI opened a self-serve Ads Manager for targeting ChatGPT users by conversational context. Ads reportedly appear in ~49% of US replies, directly contradicting Sam Altman's 2024 statement that 'ads are a last resort.'
Treasury Secretary Scott Bessent said the U.S. could sanction Chinese AI firms if evidence of IP theft from American model developers — including through distillation attacks — is confirmed.
Alphabet reports Q2 2026 results with analysts expecting ~$22.4B in Google Cloud revenue (60%+ YoY). The central question: can it justify $180–190B in full-year AI capex after free cash flow fell 47% in Q1?
A new BloombergNEF report projects U.S. data centers will consume one-fifth of all electricity by 2035 — an estimate 83% higher than the consultancy's December 2024 forecast, driven by AI infrastructure demand.
Anthropic spent a record $1.97M on federal lobbying in Q2 2026 (+26% QoQ), nearly matching Oracle. OpenAI spent $1.2M (+18%). Combined AI-lab quarterly lobbying spend hit $3.17M, up 23% from Q1.
The Information reported Anthropic held acquisition talks with robotics startup Physical Intelligence this spring. The CEO denied it to employees via Slack, but sources indicate 'something had happened.' OpenAI's existing PI investment complicates any deal.
South Korea's Ministry of Science opened a tender for a free nationwide AI chatbot backed at least 50% by domestic Korean AI models. Public beta targeted for late September 2026, funded to remain free through 2028.
Block launched Buzz, an open-source workspace unifying team chat, Git hosting, and AI agents under a single cryptographic identity system built on the Nostr protocol — free for macOS, Windows, and Linux.
TSMC committed $265B to US manufacturing, but US fabs cost 20–50% more than Taiwan, squeezing margins by 2–4%. The costs are expected to be passed on to customers — meaning higher AI chip prices broadly.
Deezer reports more than half of its daily uploads are now AI-generated — 90,000 AI tracks per day, up from 10,000 eighteen months ago. The platform plans to remove AI tracks with zero streams within six months.
Glow, founded by ex-Meta VP of Engineering and Snowflake's cybersecurity chief, emerged from stealth with a $180M Series A at $1.2B valuation. It targets the endpoint security gap created by enterprise AI agent adoption.
Poolside released Laguna S 2.1, a 118B MoE open-weight coding model trained from scratch in 9 weeks, free on Hugging Face — beating DeepSeek-V4-Flash and rivals 10x its size on agentic coding benchmarks.
British AI video startup Synthesia launched 'Roleplay Sessions' where employees practice high-stakes conversations — sales pitches, performance reviews, customer complaints — with AI avatars that respond dynamically and score performance.
Meta is piloting StoryKit on iOS, where parents photograph a toy or person to generate personalized AI children's stories with music and customizable moral lessons — no writing required.
Singapore's $400B sovereign wealth fund Temasek argues AI architecture and chip advances could 'change the energy equation for generating AI tokens quite drastically' — a direct counter-narrative to prevailing 4x data center electricity forecasts.
Google released Gemini 3.6 Flash with improved coding (+12pp on DeepSWE), the ultra-fast 3.5 Flash-Lite (350 tokens/s), and security-specialized 3.5 Flash Cyber — but the long-awaited 3.5 Pro is still absent.
Linear launched Loops — recurring AI automations triggered by schedule or events. The model uses full workspace context and connected tools to triage issues, analyze root causes, and dispatch to coding agents.
Gemini 3.6 Flash is rolling out in GitHub Copilot for Pro, Pro+, Max, Business, and Enterprise users — offering higher task-completion rates and better token efficiency for coding and agentic workflows.
Claude gained a monthly activity recap with AI observations (Settings > Reflect), Microsoft 365 write tools for drafting email and managing calendar events, and an Admin API for programmatic Enterprise organization management.
Willison published a transcript of his AI Engineer World's Fair chat with Anthropic's Claude Code team: Claude Tag internally generates 65% of product engineering PRs and adding examples to system prompts is counterproductive for frontier models.
OpenAI's head of strategic futures publicly called for regulatory pressure on Chinese open-weight AI models, then quickly retracted those comments after backlash from Yann LeCun, Pentagon officials, and open-source advocates.
A viral analysis (900+ HackerNews points) argues that China is winning the AI race by releasing high-quality open-weight models globally while America's proprietary approach is losing the economics of AI deployment.
Oracle is eliminating approximately 30,000 employees — about 18% of its global workforce — to free up $8–10 billion annually for AI infrastructure as part of its Stargate partnership with OpenAI.
SAP has completed its acquisition of Prior Labs, the startup behind Tabular Foundation Models (TFMs) built for structured business data, committing over €1 billion over four years to scale it into a European frontier AI lab.
Frontier AI models including Claude Fable and ChatGPT Sol are resolving decades-old open mathematical conjectures by generating formal counterexamples verified in Lean, with the pace growing to thousands of lines per week.
A quantitative study of arXiv preprints finds that roughly one-third of recent submissions show characteristics of machine-written text, with CS papers at 65% while mathematics remains nearly unaffected at 0.7%.
Nativ is a free, open-source macOS application built natively on Apple MLX for running frontier open-weight models locally on Apple Silicon with no accounts, subscriptions, or cloud dependencies.
Etched, a startup building ASICs specifically optimized for transformer-based AI inference, is reportedly in discussions for a Sequoia-led funding round at valuations potentially reaching $20 billion.
GitHub Copilot now shows per-user AI credits consumed per billing cycle and introduces AI-powered security detections directly on pull requests without requiring CodeQL. GitHub Code Quality moves to a paid product.
Perplexity Computer gains Brain — a private context graph built across sessions that boosted answer accuracy by 25%. New features include website publishing and full Microsoft 365 integration.
Ollama released v0.32.1 and v0.32.2 with improved Gemma 4 tool calling, a fix for Claude Code channel regression from v0.32.0, MLX memory leak fix, and unlimited tool-calling rounds for cloud models.
Simon Willison argues that coding agents have fundamentally changed the economics of reverse-engineering home devices and undocumented APIs — projects previously impractical are now economically viable.
Simon Willison covers Ben Thompson's fair use proposal for AI model distillation and a striking Qwen 3.8 detail: the model's reasoning trace included 'Could add helmet? No' — a glimpse into embedded content-filtering logic inside the reasoning chain.
Alibaba unveiled Qwen 3.8-Max-Preview at WAIC 2026 in Shanghai — a 2.4 trillion-parameter multimodal model. The company claims it ranks second only to Claude Fable 5, though no independent benchmarks have been published.
Cambridge-based CuspAI closed a $450M Series B on July 20, 2026, co-led by Kleiner Perkins and NEA with Jeff Bezos's Bezos Expeditions as a major backer. The company simultaneously launched the AI Materials Foundry — a global network with 45+ founding members including NVIDIA, Meta, and Samsung.
Just three days after launching Kimi K3, Moonshot AI was forced to pause new subscriptions because demand outstripped available GPU infrastructure. Full open weights for the 2.8T-parameter model are still expected July 27.
A Searchlight Cyber researcher used GPT-5.6 Sol Ultra to autonomously discover the 'wp2shell' vulnerability chain in WordPress core — a class of exploit that commands up to $500,000 from brokers — at a total compute cost of just $25. WordPress issued emergency patches covering 500M+ websites.
Bloomberg reports a sharp reversal in investor sentiment: Nasdaq 100 lost 4.1% in one week, Philadelphia Semiconductor Index fell 10% — worst since April 2025. Hyperscalers face pointed questions on AI ROI before Q2 earnings, with combined capex estimated at $650B in 2026.
Nonprofit Current AI, led by CEO Ayah Bdeir, announced $400M in commitments on July 19, 2026, including $100M from the French government plus DeepMind, Salesforce, Ford Foundation, and MacArthur Foundation. The goal: build open, culturally neutral AI infrastructure accessible to all communities and languages.
Researchers from Université Paris Cité and Sapienza University found AI access collapsed willingness to admit uncertainty from 44% to 3%, dropped accuracy from 27% to 9%, yet inflated confidence from 30% to 76%. Authors propose mandatory 'uncertainty indicators' in AI systems.
Huawei publicly debuted the Atlas 950 SuperPoD at WAIC 2026 in Shanghai — an 8,192-chip Ascend cluster claiming 6.7× the compute of Nvidia NVL144 and 15× the memory capacity. The full system delivers 1 EFLOPS at FP8 and 256TB of globally addressable memory, with Q4 2026 availability targeted.
Bloomberg reports a new earnings-season dynamic: Wall Street analysts have begun systematically grilling CFOs on AI token spend — vendor mix, cost per million tokens, utilization rates. The shift signals markets expect AI in the P&L, not just roadmaps.
Cursor released major Slack integration improvements on July 17, 2026. Cursor now shares a detailed plan before starting any task, supports named multi-repo environments (frontend + backend + shared code in separate repos), and can read from and post back to other Slack channels mid-task.
Microsoft is quietly replacing OpenAI and Anthropic models in Copilot products with its in-house MAI series. AI chief Mustafa Suleyman confirmed the goal to 'reduce and ultimately eliminate' payments to Anthropic. Users may face weaker performance on currently included features.
Google made gemini-omni-flash-preview available in Google AI Studio and the Gemini API — a natively multimodal video API priced at $0.10/sec of output with automatic SynthID watermarking. Also launched Nano Banana 2 Lite — Google's fastest and cheapest image model.
