Startups & Ventures

🔍 Google speeds up Gemma-4 with multi-token drafts

Google introduced Multi-Token Prediction (MTP) drafters for Gemma-4, letting it predict several tokens at once instead of one by one. This cuts down the slow step-by-step token generation typical of transformer models.

A small auxiliary model drafts multiple tokens ahead. The main model then verifies this batch in a single pass. If a token is wrong, it gets replaced and the draft restarts from that point. This saves multiple passes through the main model.

MTP boosts generation speed by about 3 times without quality loss. Google released all drafters, so this faster mode is ready for testing.

📊@tech

🪐 Milky Way Photo Contest Breaks Record

The Milky Way Photographer of the Year contest got over 6,500 entries from 15 countries, a new record.

The final selection includes 25 photos from 12 countries.

The winner collection is on Capture the Atlas’s site.

📊@tech

🤖 Anthropic agents now self-improve during idle time

Anthropic launched a new feature called dreaming in its research preview. It activates when agents are idle between sessions, letting them analyze past interactions to spot repeated errors, user preferences, and workflows.

The agent updates its memory by adding effective patterns and rules based on this analysis. Users can choose automatic updates or require manual approval.

This should improve agent performance and reduce resource use by keeping memory compact. The feature is currently limited to Claude Managed Agents.

📊@tech

🔍 Boston Dynamics shows new Atlas video

Boston Dynamics released a video of its humanoid robot Atlas demonstrating advanced mobility and manipulation. The robot may use Google’s Gemini Robotics technology.

Boston Dynamics and Google started partnering this year to develop next-gen humanoids. The video highlights progress but gives few details.

📊@tech

🍏 Apple to let users pick AI models in iOS 27

Apple plans to let users choose which AI model powers Apple Intelligence in iOS 27, iPadOS 27, and macOS 27. Third-party chatbots will connect as extensions via App Store apps. Users can select their preferred model in settings, affecting Siri, Writing Tools, Image Playground, and other system features.

Currently, ChatGPT is the only third-party model integrated. Apple is testing Google and Anthropic models too. Different Siri voices can be assigned to different AI models for responses.

This move shifts Apple from building its own AI to hosting multiple models in its ecosystem. The announcement is expected this fall.

📊@tech

🤖 Anthropic launches 10 financial AI agents

Anthropic launched 10 AI agents that turn Claude Code into a Wall Street analyst.

They analyze reports, build financial models, and conduct market analysis.

The agents manage personal finances and integrate with Excel, PowerPoint, and Word.

All agents are available now.

📊@tech

🚀 NASA tests lithium thruster reaching 120 kW for Mars missions

NASA fired a lithium-fed electric thruster at 120 kW, the highest power level for this type of system in the US.

The prototype uses electromagnetic acceleration of lithium plasma and consumes up to 90% less propellant than chemical rockets. It produces steady thrust over time, suited for deep space travel.

The next goal is scaling to 500 kW to 1 MW per thruster. A human mission to Mars would need 2 to 4 MW and over 23,000 hours of operation.

The focus is shifting to high-power electric propulsion for long-duration spaceflight.

📊@tech

😺 OpenAI rolls out GPT Instant 5.5 with shorter and more accurate responses

OpenAI has updated GPT Instant to version 5.5 and is rolling it out to all ChatGPT users.

The model produces shorter answers and reduces hallucinations while improving overall reasoning quality.

OpenAI also updated the memory interface to show which stored memories influenced each response.

The update focuses on clarity and transparency in everyday use.

📊@tech

🎵 Vinyl Player Controlled by Construction Crane

Swedish designer Love Hultén built a vinyl player where the tonearm is replaced by a construction crane. The user manually controls the crane via an aluminum panel with buttons and knobs, moving the needle left, right, up, and down.

The needle’s precise movements are powered by motors and Arduino-based sensors. There is no automatic playback, each listen requires physically guiding the stylus over the record.

This custom-made player was created for a collector and won’t be mass-produced. It works perfectly as an art piece.

📊@tech