WEARABLES

Harvard dropouts launch $249 AI glasses that never stop listening

The two Harvard students who sparked global privacy debates with facial recognition glasses are back, and this time they want to record every conversation you have. AnhPhu Nguyen and Caine Ardayfio, the duo behind the controversial I-XRAY project that could instantly dox strangers, have raised $1 million for Halo X — smart glasses that continuously listen, transcribe and analyze everything around you.

The $249 glasses feature only a display and microphone, deliberately avoiding cameras after their earlier privacy nightmare. "The AI listens to every conversation you have and uses that knowledge to tell you what to say … kinda like IRL Cluely," Ardayfio told TechCrunch. The glasses pop up information like math calculations or word definitions in real-time, powered by Google's Gemini and Perplexity.

This launch comes as the always-on AI wearable space has exploded beyond the failures since we first covered this space. Remember Friend.com? That $99 AI companion necklace launched by Avi Schiffmann pivoted from a productivity tool called Tab into pure emotional companionship. Unlike Halo's productivity focus, Friend deliberately avoids work applications — it just wants to be your digital buddy.

The competitive landscape has intensified dramatically since then. Meta has doubled down on its Ray-Ban partnership, investing $3.5 billion in EssilorLuxottica for nearly a 3% stake, with plans to grow that stake to 5%. The Ray-Ban Meta glasses have sold over 2 million units since late 2023, validating consumer appetite for smart eyewear when done right.

Privacy advocates warn that Halo normalizes covert recording. We just covered Otter.ai’s class action lawsuit, which is basically for a digital version of Halo. "I would also be very concerned about where the recorded data is being kept, how it is being stored, and who has access to it," Eva Galperin from the Electronic Frontier Foundation told TechCrunch. The glasses record everything, transcribe it, then delete audio — but twelve states require consent from all parties being recorded.

The same people who proved facial recognition glasses could instantly dox strangers now want us to trust them with always-on audio recording. Removing the camera doesn't fix the fundamental problem — it just makes the surveillance harder to detect.

Unlike Meta's Ray-Ban glasses that flash recording indicators, Halo deliberately hides its activity. This makes consent much harder to obtain since other people can't know they're being recorded unless told verbally. The founders acknowledge the legal issues surrounding two-party consent laws but shift responsibility to users.

TOGETHER WITH SONAR

From strengths to flaws: What your LLM’s code isn't telling you

What does your LLM’s coding “personality” say about your code? Based on an analysis of over 4,442 Java programming assignments completed by five leading LLMs, this report from Sonar goes beyond typical performance benchmarks to understand LLM strengths, shared flaws, and unique styles. Plus:

• Coding archetypes for each LLM analyzed

• Hidden quality & security risks of using LLMs for code generation

• How to select the best LLM model for your needs

Download the free report today and see how you can use AI responsibly in your dev workflow.

FRONTIER LABS

Hundreds of thousands of Grok conversations now searchable on Google

Over 300,000 private conversations between users and Elon Musk's Grok chatbot are now publicly accessible through Google Search, exposing personal information, passwords, and requests for illegal content.

The leak occurred through Grok's "share" feature, which creates unique URLs that users can send to others. However, these URLs are being indexed by search engines including Google, Bing and DuckDuckGo, making conversations publicly viewable to anyone who searches for them.

Forbes reported there's no warning to users about this indexing, and the exposed conversations contain concerning content including questions about hacking cryptocurrency wallets, bomb-making instructions, and personal information like passwords and medical details.

• Over 370,000 conversations have been indexed across search engines

• Users likely believed their shared conversations would remain private

• Some chats contain sensitive personal and potentially illegal content

This follows a similar incident last month when ChatGPT shared conversations became searchable on Google, which OpenAI described as a "short-lived experiment" (corporate speak that sounds suspiciously like "the dog ate my homework").

If you've shared any Grok conversations, assume they're now permanently viewable by anyone with a search engine. Deleting them from your account won't remove them from Google's index.

TOGETHER WITH EIGHT SLEEP

Does Your Sleep Suck? Here’s How To Fix It..

Struggling to get – and more importantly, stay – asleep is one of the most common problems facing people everywhere. Even worse, the trickle-down effects from bad sleep can interfere with everything from your job and focus to your physical and mental health… which is exactly why we’re so excited by the latest from Eight Sleep.

Their revolutionary new Pod 5 smart mattress cover puts the secret to better sleep in your hands. With night-changing precision temperature regulation, automatic elevation adjustment, sound control, and sleep tracking metrics, the Pod 5 has been clinically proven to improve quality sleep, giving you up to an hour of additional sleep every night. That makes a big difference on your focus, energy levels, and everything in between.

If you’re tired of subpar sleep, you can use code DEEPVIEW to get $350 off your very Pod 5 Ultra right here – complete with a 30-day trial and no-hassle returns.

SCIENCE

NASA's new AI can predict solar storms two hours before they hit Earth

NASA and IBM have released Surya, an open-source AI model that can predict dangerous solar flares up to two hours in advance — potentially doubling current warning times for space weather events that threaten satellites, astronauts and power grids.

The model was trained on over a decade of data from NASA's Solar Dynamics Observatory, creating a dataset exceeding 250 terabytes. Surya analyzes solar imagery across multiple wavelengths to detect patterns that precede solar flares and coronal mass ejections — events that can disrupt radio communications, damage satellites and endanger astronauts with radiation bursts.

"It can predict the solar flare's shape, the position in the sun, the intensity," said Juan Bernabe-Moreno, the IBM AI researcher who led the project. While scientists can easily identify when solar flares are likely, pinpointing exact timing has remained elusive.

The stakes are significant. Minor solar storms cause regional radio blackouts every few weeks, but a major solar superstorm could knock satellites out of orbit and collapse electrical grids. Some solar scientists believe Earth is overdue for such an event.

• Two hours may seem brief, but every moment counts for protecting critical infrastructure

• The model can identify flare location, intensity and shape before eruption

• IBM researchers hope to connect solar weather patterns with Earth weather phenomena like lightning

Built as a foundation model similar to ChatGPT, Surya could tackle multiple solar physics challenges beyond flare prediction. Researchers believe it may help unlock broader understanding of stellar behavior, using our sun as "a laboratory" for studying other stars across the universe.

LINKS

• OpenAI logged its first $1 billion month

• NSA’s acting director tried to save top scientist from purge

• Trump eyes stakes in other chip makers that received CHIPS Act funds

• Character.ai in talks to sell or raise money, a year after founders depart

• OpenAI-challenger “Manus” projects annual sales of $90m

• Vantage Data Centers plans $25b AI campus in Texas

• Google announces a ton of stuff at “Made by Google 2025”

• US tech stocks hit by concerns over future of AI boom

• Canva begins share sale at $42b valuation in road to IPO

• Anthropic bundles Claude Code into enterprise plans

• We must build AI for people; not to be a person

• FieldAI raises over $400m to make robot "brains"

• Sudowrite: “The non-judgemental AI writing partner you always wanted.”

• Arcads: Make top performing ads with AI

• Obsidian Bases: Turn any set of notes into a powerful database

• Warestack: Create custom release protection in plain English for your team’s code changes

• Anysphere: Account Executive

• Perplexity: AI Software Engineer - Agent Platform

• Synthesia: Accounts Receivable & Collection Specialist

• Canva: TikTok Lead, US

GAMES