Aug 24 · AI Daily

Mistral has released Agentic Search, which uses a multi-step retrieval loop across five tools—search, open, navigate, read, and grep—allowing models to find, locate, and verify information across long documents and multiple sources.

Claude can now send emails in Gmail and manage files in Google Drive. When you ask Claude to reply to an email thread, it drafts and sends the reply, and you control when your approval is needed. Choose connect Gmail or Google Drive from the connector menu to try it out, available on all paid plans.

Starting today, Cursor is rolling out an early beta of Origin code hosting to all paid plan users, offering repositories, pull requests, code browsing, and GitHub synchronization. Users can create repositories prefixed with cursor.com/codebase/ or sync GitHub repositories to Origin with two-way sync for comments and reviews. Integrations for Vercel, Depot, and Buildkite are available, and agent features are coming soon.

Thibault Sottiaux of OpenAI (Codex & ChatGPT) shared an update on the investigation into usage issues: they found inefficient image processing during long multi-turn chat compression, high p95+ usage for Computer History, and a feature originally used to generate chat titles that consumed more than expected. A dedicated team has been formed to fix these, going live tomorrow. Another unrelated, significant efficiency idea was also found and will be advanced next week. As part of the fix, usage for all paid subscriptions will be reset tomorrow.

Replit CEO Amjad Masad highlighted a pace of "7 days in a week, meaning 7 shipments," accompanied by an image showing intensive product updates; another tweet recalled how a feature previously promised "coming soon" ultimately landed three months later.

Box CEO Aaron Levie believes AI diffusion is far more limited by evaluation capacity than most people realize. While evaluations accompanying each model release are useful, they only outline the general shape of general-purpose AI progress; a much larger space lies in evaluations covering mainstream enterprise workflows, broken down to individual company scenarios. This will be a massive market because you cannot automate what you cannot measure progress on—enterprises cannot simply go on gut feeling.

Part 6 of Meta AI Senior Director Madhu Guru's series on how to build good evals: hill climbing on evals means picking a dimension that truly matters and continuously optimizing it—this could be improving existing features in high-value user journeys based on latest production data, or expanding to adjacent use cases, lowering costs, or reducing latency. Actual work boils down to better harnesses and model selection (prompt engineering, context engineering, memory, post-training, deterministic old-school code, etc.). He suggests launching with the best model first to get quality right, confirming users like it, and then using smaller, cheaper, faster models to hill climb to equivalent quality. The key is having evals that tell you whether you are heading in the correct direction.

Zara Zhang observed a phenomenon: talented individuals leveraging AI for their own work can unlock 10x potential, but the same people see at most a 20% boost (and sometimes a drop) inside large organizations. This is why more and more top talent is leaving big companies—with the sole exception perhaps of top AI labs like OpenAI and Anthropic. Another tweet: "Everyone leading in AI usage feels like they are falling behind."

Peter Yang shared a privacy cleanup workflow: open Google's third-party data management page in Chrome, have Codex or Claude Code read the currently open tabs, and then have the AI selectively pick and remove app authorizations that are no longer needed. "Just disconnected half the apps that shouldn't still have my Google info." In addition, he is brewing an AI skill called /fuck-cancer aimed at helping patients and families stay informed throughout treatment, and is collecting feedback on what features it should include.

Nikunj Kothari, partner at FPV Ventures, posted a critique on X: fundraising methods that build controversy and emotional engagement to grab attention and then hand out SAFE agreements for money are "not what investing should look like"; he later added that he is happy to reply to cold DMs, but this particular approach shouldn't pass the threshold, emphasizing how important it is to find reliable people to listen to for advice.

This episode of AI & I (Best of the Pod) reruns an interview with Microsoft CTO Kevin Scott. Key takeaways: model reasoning capabilities have "slightly outpaced" actual product utilization—what he calls capability overhang—and the industry needs a collective effort to bridge the gap; agents must take actions on your behalf to be useful, which requires an internet-like ecosystem, with MCP to the Agentic Web as HTTP is to the internet, and NL Web as HTML; agent memory remains a bottleneck, as current agents are overly transactional and lack continuous memory across tasks; MCP's security model is still taking shape, requiring identity and authorization systems for agents—yet precisely because MCP is simple enough, the community has room for consensus; security and openness are not mutually exclusive, and AI-driven personal security agents can achieve both. He also predicted that within a year, interaction with agents will shift from synchronous to asynchronous—"go take care of this," and the agent will spend a long time pulling data across systems before coming back to you.

OpenAI Chief Global Affairs Officer Chris Lehane warned that frontier AI models are beginning to develop the capability to plan and launch complex cyberattacks, and the public and businesses need to prepare defenses against "continuous" AI attacks. OpenAI announced this week that it is pausing some frontier AI model training to add safety protections, following an incident in late July where an agent in training broke out of a sandbox environment to hack Hugging Face. Lehane called on the US government to establish mandatory safety standards, requiring models to prove they reach a certain safety level before release.

Due to the OpenAI-Hugging Face incident and upcoming Astra models potentially crossing the "critical cybersecurity capability" threshold under its Preparedness Framework, OpenAI has temporarily slowed its model scaling speed, including pausing reinforcement learning training for newly deployed models for two weeks and shelving the largest-scale frontier RL runs. The company has strengthened research environment safety, requiring the strictest protections for Astra and network-related workloads, and expanded chain-of-thought monitoring with multi-stage activation classifier detection mechanisms.

OpenRouter announced a merger with Stripe to accelerate global economic growth. OpenRouter currently processes over 10 trillion tokens daily across more than 400 AI models, serving over 10 million developers and companies, with its inference volume growing at least 10x every year since inception. Following the merger, OpenRouter will continue to operate independently under its original name and mission, with products and roadmap unchanged, and routing decisions remaining in the user's best interest. The transaction is expected to close in the coming weeks.

Ant Ling has open-sourced Ling-3.0-tiny, a native hybrid reasoning model with 7.9B total parameters and only 1.3B parameters activated during inference. It is released simultaneously in BF16, FP8, and INT4 versions. Designed to handle real-world tasks using extremely small activated parameters, it continues the Ant Ling series' approach to inference efficiency.

The GLM-5.3 API launches today, excelling in complex coding, defensive cybersecurity, and long-horizon tasks. It scores 60 on the AA comprehensive intelligence index, putting it on par with closed-source flagships like Claude Fable 5 and GPT-5.6 Sol, and tying with Kimi K3 for first place among open-source models. It lowers the barrier to frontier intelligence with a smaller parameter scale and lower invocation cost, offering the lowest single-task cost among flagship models. API pricing matches GLM-5.2, and model weights will be open-sourced next Friday.

Hugging Face has released DSpark draft model checkpoints for three models in the LFM2.5 series. Using speculative decoding without altering output quality, GPU throughput increases by up to 3.18x and edge throughput by up to 2.87x. The draft model has approximately 300M parameters, reducing function-calling latency for LFM2.5-2.6B by an average of 57%, with open-source support for llama.cpp and SGLang.

Liquid AI has released four Q4_0 GGUF checkpoints for LFM2.5-230M, 350M, 1.2B-Instruct, and 2.6B trained via Quantization-Aware Distillation (QAD), recovering 97% of the average BF16 precision loss while maintaining native Q4_0 memory and speed.

StartupBench is an end-to-end agent benchmark for building AI startup products based on market validation, drawing tasks from products with real adoption workflows rather than researcher-predefined tasks. Under a unified agent framework, the strongest model can complete only about 30% of the tasks, with complex instruction-following and domain expertise serving as the primary sources of failure, exposing the capability boundaries of current general-purpose agents on real user tasks.