In partnership with

☀️ TRENDING AI NEWS

🤖 Qwen4-Max: Alibaba released a 2.4T-parameter MoE with 71B active parameters and a 1M-token context window.

🛠️ Handoff²: Hark's computer-use agent runs 2x faster than v1 and completes 82% of 478 internal web tasks.

🏢 Apple Health AI: Apple launched a study with OpenEvidence on how people use conversational AI for health questions.

🚨 Chip Exports: Congress is pushing to tighten AI chip export rules after the Nvidia spy scandal.

Seventy-one billion active parameters sitting inside a 2.4 trillion parameter model. That one ratio explains a lot about where the economics of frontier AI are heading, and it is the best place to start today.

Also on the menu: Apple wants to know whether AI health answers are actually any good, an agent that learned to stop wasting clicks, and Washington turning scandal into legislation.

🛠️ Tool of the day
 

Cursor - an AI coding editor where agents turn plain-language ideas into working code. It helps developers and non-coders build and edit software faster, and it has a free plan.

⚡ Quick hits
 
  • Indian chip startup Bharat Semi raised a $24M seed round to build AI-enabled programmable image sensors, and it targets initial commercial production in 2027. Read more →

🤓 AI Trivia
 

The answer is hiding near the bottom of today's newsletter... keep scrolling. 👇

🤖 Alibaba's Qwen4-Max Goes Big and Stays Cheap
 
A large neural network sphere with a small lit cluster of active nodes.

A 2.4T model that runs like a 71B

Alibaba released Qwen4-Max, a proprietary mixture-of-experts model with 2.4 trillion total parameters but only 71B active per token. It also ships with native vision and a 1M-token context window.

Pricing lands at $1.5 per million input tokens. Alibaba says it cuts cost per task by 62% on its Work360 workload. If you want to see what that does to your own bill, try our token calculator.

The bottom line

Sparse activation is how labs keep scaling total parameters without scaling inference cost at the same rate. The catch: this one is proprietary, so you are renting it, not running it.

The voice-to-text app that built its own speech model

Most dictation apps are a wrapper around someone else's speech model. Wispr Flow trained its own.

It's called Canto, and it was built for the way people actually talk to AI: thinking out loud, changing direction mid-sentence, half-whispering in an open office.

Wispr tested it on 10 hours of real dictations from more than 2,300 speakers. Canto had the lowest word error rate of every model tested, including models from Google, OpenAI, AssemblyAI and Deepgram. That means your prompts to ChatGPT, Claude and Cursor come through the way you said them.

Talk naturally in any app and Wispr Flow types clean, formatted text, 4x faster than typing.

🛠️ Hark's Handoff² Learns to Batch Its Clicks
 
A fast cursor batching several clicks together beside a stopwatch.

Latency-penalized RL doubles agent speed

Hark released Handoff², the computer-use agent inside Hark Pro. It averages 2x faster than v1 and completes 82% of 478 internal web tasks.

The trick is training: reinforcement learning with a latency penalty taught the agent to batch multiple actions together instead of taking one step at a time. That is a big deal for AI agents, where speed is often the thing that makes them feel unusable.

The bottom line

Note the 82% figure is on Hark's own internal tasks, so treat it as a directional signal. Still, rewarding speed in training is a pattern worth watching across every agent product.

🏢 Apple Wants to Know If AI Health Advice Actually Works
 
A smartphone with a heartbeat line and chat bubble beside a stethoscope.

A study built around real questions

Apple launched the Apple Health AI Usage Study, run with OpenEvidence, a clinical AI platform used by healthcare professionals that answers from sources like major medical journals. The new app will field everyday medical questions from participants.

The goal is to understand how people interact with conversational healthcare AI tools and how useful the responses are. It is a notable move for Apple, which has been quieter than rivals on consumer AI.

The bottom line

Millions already ask chatbots about symptoms. Rigorous data on whether those answers help or harm is thin, and a study at Apple's scale could finally change that.

🚨 Congress Targets AI Chip Exports After the Nvidia Scandal
 
An AI chip on a scale in front of a government building with a closed gate.

Both parties want tighter rules

Following the scandal involving former Nvidia CEO Jensen Huang and a suspected Chinese spy, lawmakers from both parties are backing measures to further restrict AI chip exports. They have also demanded that Huang and other Nvidia leaders testify before Congress.

Meanwhile, Elon Musk and a long list of tech bosses have signed an open letter urging Nvidia to reinstate Huang, with one report putting signatures at more than 165,000. The company is being pulled in two directions at once.

The bottom line

Export policy is now being shaped by a personnel scandal rather than a strategy document. Expect tighter, faster rules than anyone in the chip supply chain was planning for.

⚠️ AI-Related Child Abuse Reports Jump 8,400%
 
A protective shield in front of devices with a rising chart and a magnifying glass.

Deepfakes and 'nudify' apps drive the surge

Reports of AI being used to sexually exploit children have risen by more than 8,400% in two years, according to a report from the Edinburgh-based Childlight Global Child Safety Institute. It points to AI-generated abuse material, including deepfakes made with 'nudify' apps.

Researchers are calling on governments to treat the sexual exploitation of children as a preventable public health emergency. It is a stark reminder of why AI safety and deepfake policy debates are not abstract.

The bottom line

Of everything in today's issue, this is the number that should push regulators hardest. Capability gains are outpacing the guardrails, and the harm is already measurable.

Quick plug: if you need a site up fast, our sponsor 60sec.site is an AI website builder that gets you live in about a minute. And for more like this every day, visit dailyinference.com.

🌎 Trivia Reveal
 

The answer is parameters used for each token during inference! That is why Qwen4-Max can hold 2.4T parameters in total while only activating 71B for any given token, keeping compute costs far below what the headline number suggests.

💬 Quick Question
 

Would you trust an AI chatbot for a health question before calling a doctor? Hit reply and tell me where you draw the line - I read every response!

That's it for today - see you tomorrow!

🎁 Share Daily Inference
 

Know someone who would like this? Share your link. One referral unlocks the AI Tools Starter Kit, and five unlock the AI Insider Briefing.