Discover the AI Tools

Satya Nadella just sent a clear message to Microsoft's staff: when it comes to AI safety, we should take all the time we need. In an internal memo, he pushed for third-party testing and outlined a 37-page code of conduct. But what does this mean for the race to build smarter machines?

Shanghai AI Lab just open-sourced Shusen-S2, a multimodal foundation model that gives scientific research a serious boost. It comes with a plug-and-play memory module—think of it as an external brain you can swap in for different fields—and it excels at long-term reasoning and tool use. In biology, chemistry, and materials science, it goes head-to-head with top closed-source models. This could be a game-changer for open-source AI in science.

DeepSeek's official desktop client for Harness is almost here, with macOS and Windows support. But should you run V4.1 Flash or the built-in model? And what's the ideal chain-of-thought level? We break down the details.

The Gates Foundation is putting at least $1 billion over the next two years into making AI work for everyone, not just the privileged. In its latest Goalkeepers report, the foundation warns that without deliberate effort, AI could deepen the gap between rich and poor. The money will target education, health, agriculture, and digital infrastructure—with a focus on local languages and community control. It's a big bet on a simple idea: technology should serve the many, not the few.

Bill Gates just pledged $1 billion to make AI work for the world's poorest. In a candid chat, he weighed in on the US-China model race, why open-source matters, and that humanoid robot Olympics in China that left him genuinely amazed. Here's what he had to say.

Anthropic, OpenAI, and Google have been quietly meeting since July to create an AI industry standards body, insiders say. The talks gain urgency after Anthropic CEO Dario Amodei's call to slow frontier development. But not everyone's on board—Cohere's CEO calls it 'a cartel's design,' and the White House effort stalled. As the debate heats up, one question remains: who gets to write the rules?

9月15日，微信AI助手小微因隐私问题冲上热搜。微信火速澄清：「偷窥聊天记录」纯属谣言。小微只能看到你本人能看的内容，已删朋友圈绝不回看，聊天记录也是用完即焚。当AI深入社交场景，隐私边界究竟在哪？

Major companies like NVIDIA and Booz Allen Hamilton are pulling back from using top AI models from Anthropic and OpenAI, citing data security worries. A policy change at Anthropic that retains customer data for up to 30 days has sparked a trust crisis. Now, firms are isolating sensitive work and turning to their own models, like NVIDIA's Nemotron. Microsoft sees a new market in private cloud solutions.

StepZenith just dropped five audio models that are shaking up the voice AI world. From real-time conversations to music creation, these models are already topping global rankings. Curious about what they can do? Let's break it down.

A Canadian AI researcher says we need to hit the brakes on cutting-edge AI—permanently. Slowing down isn't enough, he warns, because current safety tests are easily fooled. Meanwhile, industry leaders are calling for a pause, but critics argue their plans fall short. Is humanity running out of time?

StepZen unveiled its StepAudio 3 series on September 15, featuring five voice models that target real-time chat, speech recognition, voice generation, audio creation, and music. Several models grabbed top spots on Artificial Analysis leaderboards. But can they really change how we interact with machines? The lineup promises more natural conversations and easier audio production, all available now on StepZen's open platform.

Anthropic quietly rolled out Claude Opus 5.2 in a night-time gray release. Developers report blazing speed, cleaner output, and an end to the model's lazy habit of asking users to 'continue.' A secret probe question reveals who's in the test, while leaked plans hint at even more powerful models waiting in the wings.

Anthropic quietly rolled out Opus 5.2 in Claude Code, and developers only spotted it by sniffing network packets. Meanwhile, an internal report hints at a secret 'nuclear-level' Model 2 that's already writing most of the company's code and could replace 85% of the research team. The message? AI is eating its own dog food—fast.

Apple's new Siri might not be just Siri. Code sleuths found iOS 27 and macOS Golden Gate can plug in Claude or GPT-5.6. Two mechanisms let third-party models either assist Siri or take over completely. The EU's Digital Markets Act is likely the nudge. But for now, it's all under the hood—no user access yet.

Ant Group just open-sourced SingProbe, a lightweight safety net for large models. It taps into the model's own hidden states to flag risks in real time—with less than 0.5% extra computing cost. Think of it as a built-in alarm that catches harmful outputs before they happen. Already compatible with 29 mainstream models, it's a big step for AI safety, especially in sensitive areas like healthcare.

At the All-In Summit, Nvidia's Jensen Huang and President Trump pushed back against calls to pause AI development. Huang argued that slowing down plays into the hands of those who want to hold the industry back. Trump said the US won't be swayed by political pressure or China, and Huang couldn't agree more.

OpenAI quietly hires hundreds of contract workers to read your ChatGPT conversations, according to a new investigation. They rate responses on a 1-7 scale, aiming to curb sycophancy and other quirks. But does that mean your private thoughts are being eyeballed by strangers? Privacy filters exist, but mistakes happen.

Instagram is putting a paywall on its AI photo editing tools. Use them too much, and you'll be asked to subscribe to Meta One for $7.99 a month—or wait until tomorrow. It's a bold move that pits Meta against the likes of Photoshop and VSCO, but will users actually pay?

Ever wished Siri could do more? New code hidden in Apple's system hints that third-party AI models like Claude and GPT-5.6 might soon power Siri. Developers found two ways Apple could let these models handle tasks Siri can't—or even replace its brain entirely. The EU's Digital Markets Act might be the nudge Apple needed. Here's what we know so far.

Apple just dropped iOS 27, watchOS 27, iPadOS 27, and visionOS 27, and the star of the show is a long-awaited Siri AI makeover. The assistant now runs on all four platforms—though it's still in English beta. New features range from a Liquid Glass opacity slider to eye-controlled notifications on Vision Pro. Curious what else is new? Let's dive in.