Get hand-picked daily summaries of the best, most informative AI articles from around the web. Browse recent summaries, then jump to the original source when you want the full article.
Chrome 149 adds a "Select from screen" feature to Gemini, letting users send highlighted text or images directly to the AI. Google also opened Gemini 2.5 Flash's computer-use API to developers for automating multi-step browser tasks.
A $2.3B startup, General Intuition, has built a single AI model powering both a office-navigating robot (trained on just 8 minutes of real-world data) and a Fortnite-playing agent, using action data rather than video footage.
DeepSeek Flash slashes browser agent costs 100x by generating reusable executable code instead of repeatedly calling frontier models, reducing inference calls and leveraging cheaper, cached tokens.
Naveen Rao launched Unconventional AI to cut AI power consumption 1,000x using oscillator-based chips. Its first model, Un-0, is an image generator currently running on software simulation while hardware is in development.
Gemini 3.5 Flash can automate Android tasks by analyzing screenshots and sending ADB commands like taps, swipes, and text input via Python, with potential iOS support by swapping one compatibility layer.
Games disclosing AI use on Steam receive ~52.6% fewer reviews, indicating commercial consequences. This likely reflects player perceptions of low-effort development, though poor game quality may also contribute.
This archive is built from SMRTR newsletter stories. Subscribe for hand-picked stories without the extra noise.
This free Claude skill prevents memory files from bloating and causing inconsistent AI behavior. It interviews you line-by-line, flags contradictions, shows diffs, and only applies changes you approve.
HoprLabs is a Python CLI tool that simulates AI training math, estimating model size, memory, VRAM, training time, and token budgets while flagging config risks—before you commit resources to actual training.
Rippling's new Data Cloud tracks employee AI spending ROI, catching cases like a $30K/year Claude subscription with little value and flagging low-quality AI output. It also launched same-day payroll banking, rivaling Ramp.
Anthropic accused Alibaba of using ~25,000 fake accounts to conduct 28.8 million exchanges with Claude AI, stealing its capabilities without paying R&D costs. Anthropic is now urging Congress to update antitrust laws, tighten...
Avon and Somerset Police secretly deployed 20+ crime-prediction algorithms using sensitive data like mental health records, with one model achieving under 10% accuracy. The UK is now investing £75 million to expand such AI...
The Trump administration has asked OpenAI to delay GPT-5.6's public release, limiting access to select partners pending government review—mirroring Anthropic's restricted Claude Mythos rollout over fears that powerful AI could...
AI revenue is growing rapidly but barely covers massive infrastructure costs. Future growth hinges on falling prices, rising demand, and AI delivering measurable value per output.
Unconventional AI's image generator Un-0 uses thousands of coupled oscillators instead of neural networks to produce images, scoring 6.74 FID on standard benchmarks while targeting ~1,000x lower energy consumption than...
Public distrust of AI has surged due to overhyped job-loss predictions, forced workplace adoption, environmental concerns from data centers, and long-standing Big Tech skepticism—making the backlash increasingly politically...
AI agents like OpenClaw and Hermes are moving beyond advice-giving to actively handling tasks like emails, invoices, and scheduling via platforms like HubSpot or Slack — but granting such access to sensitive data carries serious...
A developer built a fully local, privacy-first search engine that uses Gemma 4 to transcribe and describe video, audio, and text, converting everything into searchable vectors via Qdrant — no cloud or API keys required.
Sakana AI launched Fugu, a model-blending "router," ten days after the US forced Anthropic to shut down its top AI models globally, positioning Japan as a serious AI competitor.
A 75-year-old built a professional website in under a week using Assistant Bot, which lets users update sites via text instructions — no coding required. This effectively eliminates the technical barrier to web accessibility.
A tiny Llama2 AI model runs inside MIT's Scratch by converting C inference code into Scratch blocks, generating short stories token-by-token in a sprite's speech bubble.
Valence is a REST API platform that detects human emotions from voice in real time or asynchronously, enabling developers to easily integrate emotion-aware features into their applications.
Mistral's OCR 4 converts documents into structured, labeled data with location, type, and confidence scores across 170 languages. At $4 per 1,000 pages, it supports on-premise deployment, appealing to European enterprises with...
OntoCortex lets an AI agent request actions in plain language while a separate deterministic layer validates and executes them, keeping the AI isolated from databases. This blocks prompt injection attacks and allows backend...
AI agent use in customer service surged from 39% to 66% in one year, with 70% of companies seeing ROI within 60 days. Salesforce's new pay-per-resolution model could accelerate adoption further.