Despite high temperatures and the time-off season, the AI world never stops. This month has brought both exciting news and inspiring advancements.
Google is mapping the world, ChatGPT is getting agent mode to control users’ computers, and China turns off all AI models in the whole country – but only for a while. And for good reason.
2 July 2025 Google’s carbon emissions rise due to AI spike
An advocacy report says Google’s carbon emissions rose 65% from 2019 to 2024, and not the 51% Google reported. Using location-based data, Kairos Fellowship also finds a 1,515% jump since 2010 and a 26% rise in the past year. Google disputes the method, noting its market-based accounting follows Greenhouse Gas Protocol rules.
More can be found in The Guardian.
2 July 2025 Overusing ChatGPT accumulates cognitive debt
MIT researchers tracked EEG activity while people wrote essays with no aid, with a search engine, and with ChatGPT-like AI. Neural connectivity dropped as outside help rose; LLM users showed the weakest brain networks. AI-assisted essays looked “soulless” to teachers and their authors recalled the content poorly just minutes later. And switching from an LLM to brain-only writing left participants with reduced neural engagement, signaling “cognitive debt.” Search-engine users scored between the other groups on both brain activity and recall.
More can be found on Big Data Wire.
3 July 2025 Google adds Gemini to Docs, Slides, Sheets, Drive, and Gmail
Google added Gemini chatbots to the side panels in Docs, Slides, Sheets, Drive, and Gmail. Users can open their own Gems or ones from Google without leaving the app. Gems are built in the Gemini site, where a user describes a role and uploads files; they then run inside the workspace. Google lists uses such as brand copywriting, sales support, role summaries, and persona checks. Rollout began on 2 July, 2025 for all Workspace accounts that have Gemini panels.
More can be found in TheVerge.
8 July 2025 Chinese researchers unveil MemOS to better handle memory operations
Researchers say LLMs lack memory management. This gap creates a “memory silo” that resets models each session, preventing models from keeping long-term relationships with users and leveraging the accumulated knowledge and memories gained in the process. AI assistants then forget earlier details, so they cannot keep user preferences across conversations. To mitigate this risk, Chinese researchers introduced an operating system that considers memory an asset comparable to computing power.
More can be found in TheVerge.
9 July 2025 HuggingFace launches $299 robot to democratize AI robotics
Hugging Face bought Pollen Robotics and introduced Reachy Mini, a $299 open-source robot. The device links to the Hugging Face Hub so developers can run and share AI apps via Spaces. Reachy Mini is a DIY kit with a six-axis head, camera, microphones, speaker, optional battery, and Python control. All hardware and software designs are public; the firm will sell ready units while encouraging community builds. The plan is to spread physical AI, gather user feedback, and avoid a future dominated by black-box household robots. Scaling production, safeguarding data, and meeting demand are the next hurdles.
More can be found in VentureBeat.
10 July 2025 Gemini AI can now turn images into videos
Google uses the Veo 3 model to turn a reference image into an eight-second MP4 clip with generated sound, speech, and effects. Users can upload a picture, give motion and audio prompts, and receive a 720p 16:9 file watermarked as AI-made. The feature rolls out on web now and on mobile this week in select regions. Google’s Flow filmmaking app is expanding to 75 more countries on the same schedule.
More can be found in TheVerge.
14 July 2025 Anthropic’s Claude can now edit images in Canva
Claude now connects to Canva via the Model Context Protocol. Users can create, resize, and search through designs or templates with text prompts, then summarize them inside the chat. The link needs both paid Canva and paid Claude plans. Claude is the first assistant to run Canva tasks through MCP; similar links exist for Figma and others. A new integrations directory shows all third-party tools available to Claude users.
More can be found in TheVerge.
14 July 2025 Google releases curated notebooks collection
Each notebook comes pre-filled with source text, AI summaries, notes, and an audio overview. The launch set includes eight topics, such as longevity, parenting advice, and Shakespeare, plus material from The Atlantic and The Economist. Users can question the AI or explore the podcasts to learn more without building a notebook themselves. Google says the public sharing of notebooks is rising and plans more expert-curated sets.
More can be found in TheVerge.
16 July 2025 Meta aims to build Manhattan-sized data center
Mark Zuckerberg said Meta will spend hundreds of billions on AI development. To power this, Meta will build multi-gigawatt data centers, including Prometheus (online 2026) and Hyperion, each scaling up to about the size of Manhattan’s power footprint. Meta’s ad revenue funds the plan and capital spending for 2025 is set at $64-$72 billion. A new Superintelligence Labs division will push products like Meta AI, ad tools, and smart glasses.
More can be found in The Guardian.
15 July 2025 US Military AI contracts for Anthropic, OpenAI, Google, and xAI
The Pentagon awarded AI contracts worth up to $800 million, splitting it to $200 million each among Google, OpenAI, Anthropic, and xAI. The deals let agencies test competing models for military, intelligence, and business tasks. xAI introduced “Grok for Government,” while OpenAI and Anthropic offer similar suites. Officials say multiple vendors reduce risk but require careful integration. Observers warn that past chatbot errors show the need for strict controls.
More can be found in ArtificialIntelligence-news.
15 July 2025 Google study shows LLMs can’t withstand pressure
A DeepMind-UCL study tracked how LLMs handle confidence and feedback. Models start off sure of their initial answers yet often drop them when shown even low-quality opposing advice. If their first answer is visible, they stick to it more; if hidden, they change it more. Unlike humans, they give extra weight to disagreement rather than confirmation. Designers of multi-turn chat systems may need context-reset tools to curb these swings.
More can be found in VentureBeat.
16 July 2025 – Google AI may now make a call and ask for prices
Google Search now lets US users have AI call local businesses for price or schedule information. The tool uses Duplex with Gemini, announces itself as AI, and returns results by text or email. It supports selected services such as pet grooming, dry cleaning, and auto repair; owners may opt out. AI Pro and Ultra subscribers get higher usage limits and can test Gemini 2.5 Pro in AI Mode. Google is also trialing Deep Search within AI Mode to create multi-step, detailed reports.
More can be found in The Verge.
17 July 2025 Le Chat now offers a Voxtral voice mode for spoken, low-latency exchanges and transcription
Deep Research mode breaks down queries, gathers sources, and returns structured reports with citations. Think mode routes prompts to the Magistral model for clear, multilingual reasoning. Projects let users keep related chats, files, and settings in one folder. An image-editing link with Black Forest Labs supports text commands like removing objects or changing scenes. All the new tools are available today on Le Chat for web and mobile.
More can be found in ArtificialIntelligence-news.
17 July 2025 OpenAI launches computer-controlling agent
OpenAI launched ChatGPT Agent, a tool that runs a virtual computer to finish multi-step tasks for users. A new model, built from Operator and Deep Research, drives actions like reading calendars, booking services, shopping, and making slide decks. The agent pauses for user approval before any irreversible step and blocks financial transactions for now. Watch Mode limits work on sensitive sites and bio-chemical safeguards are active. Rollout begins for Pro, Plus, and Team plans, with enterprise access coming later.
More can be found in The Verge.
18 July 2025 Netflix admits to using AI in one of its shows for the first time
Netflix used generative AI in its new Argentinian series, El Eternauta, applying the tech to a building-collapse scene which was completed about ten times faster than in standard VFX work. Co-CEO, Ted Sarandos, said AI lets creators raise quality and cut costs, funding work that would otherwise be unaffordable. He framed the tools as aids to human artists, while unions remain wary of job impacts. The move follows strikes that secured limits on AI in Hollywood. Netflix reported $11 billion in quarterly revenue, up 16 percent, and expects its ad business to roughly double this year.
More can be found in The Guardian.
21 July 2025 Meta rejects EU’s voluntary guidelines
Meta refused to sign the EU’s voluntary code of practice for general-purpose AI. Joel Kaplan said the code creates legal uncertainty and exceeds the AI Act. The EU says signers receive reduced administrative burden.
The AI Act rules go into action on 2 August, requiring transparency on training and security risks and respect for copyright, with fines of up to 7% of annual sales. Meta argues the regime will slow frontier model deployment in Europe, echoing a letter from 45+ firms seeking a two-year delay, noting much looser U.S. policy.
More can be found in The Verge.
22 July 2025 – ChatGPT users send over 2.5 billion prompts every day
The company reached the level of 912.5 billion requests every year. The growth of the platform remains rapid, as weekly users of ChatGPT spiked from 300 million on December 2024 to over 500 million in March 2025. The data was obtained by Axios and later confirmed by OpenAI spokesperson.
More can be found in TheVerge.
22 July 2025 Gemini with Deep Think achieves gold-medal standard at the International Mathematical Olympiad
An advanced Gemini Deep Think model achieved gold-medal performance at IMO 2025, solving 5 of 6 problems for 35/42 points, with solutions graded by IMO coordinators. It improves on 2024’s AlphaProof/AlphaGeometry 2 result (4/6, 28 points) and worked end-to-end in natural language within the 4.5-hour limit. Deep Think uses parallel reasoning, reinforcement learning, and a curated math corpus with IMO hints. DeepMind will release a version to trusted testers before rolling it out to Google AI Ultra subscribers, and continues work on formal systems. The IMO confirmed solutions alone, not the system.
More can be found on the DeepMind blog.
23 July 2025 Anthropic – longer thinking makes models dumber
Anthropic reports “inverse scaling”: extending test-time reasoning can lower accuracy. Tests covered counting with distractors, regression with misleading features, complex deduction, and safety scenarios. Claude models drift toward irrelevant details; OpenAI’s o-series resists distractors but overfits framings; all models degrade on complex deduction.
More can be found in Venture Beat.
24 July 2025 Google tests new search experience
Google launched Web Guide, an opt-in Labs test in the Search “Web” tab. It uses a Gemini model and query fanout to group links, surface the two top links, generate a short summary, and sort results into AI-made categories with related questions. Users can ask more detailed questions; sources may include educational sites, YouTube, Reddit, and Quora.
More can be found in TheVerge.
25 July 2025 Google shows DeepFake detector
UC Riverside and Google developed UNITE, a detector that flags tampered with or fully synthetic videos by analyzing entire frames and motion, not just faces. It uses a transformer model built on SigLIP and “attention-diversity loss” to monitor multiple regions per frame. The team trained it on varied synthetic content and presented results at CVPR 2025.
More can be found in ScienceDaily.
27 July 2025 Indian filmmaking company releases a movie with AI-generated alternate ending
Eros Media Group will rerelease the 2013 film Raanjhanaa as Ambikapathy on August 1 with an AI-generated alternate “happy” ending, without the director’s involvement. CEO Pradeep Dwivedi says the company is evaluating its 3,000+ title library for similar opportunities for AI edits and is presenting this cut as optional, with posters noting an AI-powered ending. Director Aanand L Rai opposes the move and is exploring legal options; his studio says its partnership with Eros ended and calls for clear AI protocols.
More can be found in The Guardian.
28 July 2025 Microsoft tests AI browser in Edge
Microsoft is testing an experimental Copilot Mode in Edge. It lets Copilot search across open tabs, compare options, summarize, and make bookings, and shows up on the new tab page and in the address bar. With permission, it can use your browsing history and saved credentials to book reservations and supports voice navigation. You can turn the mode on or off; Microsoft says it will organize browsing into topic-based journeys. It’s free for a limited time with usage limits and may later require a subscription.
More can be found in TheVerge.
29 July 2025 Yelp is rolling out AI generated videos about restaurants
Yelp is rolling out AI-stitched restaurant videos nationwide on iOS, built from user photos, videos, and reviews in a TikTok-like feed. OpenAI models write scripts and topics, ElevenLabs provides voice-over, and Amazon Transcribe adds captions. Businesses can’t preview or be notified, and users can’t opt out; inaccurate or offensive clips can be reported and Yelp runs periodic audits. Videos are created only when a listing has enough content; one active video per business is shown, and display is based on personalized signals though videos aren’t personalized.
More can be found in TheVerge.
30 July 2025 DeepMind is building the most accurate earth map to exist
AlphaEarth Foundations is an AI model that builds digital maps of global land and coastal waters from multiple Earth-observation sources. It compresses data into 10-meter “embedding fields,” cutting storage ~16× and lowering errors ~23.9% versus prior methods while tracking change over time. A Space-Time Precision architecture with continuous time and SigLIP features fuses imagery, radar, lidar, and climate data to overcome clouds and gaps. Pilots include MapBiomas (Amazon monitoring) and the Global Ecosystems Atlas.
More can be found in Venture Beat.
