Google DeepMind Launches Institute to Study AGI’s Impact
Google DeepMind has launched the DeepMind Institute (DMI) to explore the deployment of artificial general intelligence, which the company defines as a system that exhibits all the cognitive capabilities of the human brain.
- The institute will bring together researchers from Google, DeepMind, and outside institutions to explore questions around AGI and its broader implications.
- It will be led by Google DeepMind co-founder Shane Legg, Google SVP James Manyika, and DeepMind Chair Demis Hassabis.
- The institute launched with inaugural publications exploring economic policy for AGI, model reasoning transparency, global access, and human flourishing.
- James Manyika emphasized that AI safety and responsibility cannot be considered only on a country-by-country basis because the technology will be used globally.
- Shane Legg also pointed to international scientific collaboration, noting that advances in AI and science can emerge from researchers anywhere in the world.
The launch comes as discussions around AGI increasingly extend beyond technical development to questions about economics, governance, access, and how societies should respond to increasingly capable AI systems.
Anthropic Combines Claude’s Chat and Cowork into One Interface
Anthropic is bringing Claude’s chat and Cowork into a single interface, so users no longer have to decide which workspace to use before starting a task.
- Anthropic said customers often struggled to pick the right tab for the right task, which is why, with the new design, Claude now automatically routes requests to the right feature.
- Claude Design, introduced in April for website and prototype design, now also works anywhere within Claude.
- Anthropic is also adding dedicated features for making presentations and documents: Claude Slides lets users ask Claude to create, edit, and present slides, and Claude Docs lets users work alongside Claude to create documents.
The new features are rolling out first to Pro and Max users across web, desktop, and mobile. The company says it plans to extend them to the free and team tiers “soon.”
Profound announces a marketing engineering hackathon
Marketing platform Profound is hosting a one day, in person hackathon in San Francisco on October 3, focused on building AI agents that automate large scale marketing processes. The challenge is open to 40 participants and is platform agnostic.
The event will award $40,000 across two prizes, with each winner receiving $10,000 in cash, $10,000 in Profound credits, and an interview opportunity at the company. You can apply here.
Google Home opens smart devices to AI agents
The company is giving AI agents access to Google Home devices through a new Model Context Protocol server in early access. Agents that support MCP, including Claude and ChatGPT, can use natural language to control connected devices, review camera summaries, and access smart home activity.
Gemini Notebook adds new tools for studying
Google is adding real time voice conversations, audio recording, and interactive learning overviews to Gemini Notebook. Students will be able to discuss concepts with their notebooks, record lectures, and create short video overviews from their sources.
Al Gore says AI risks go beyond data centers
Former U.S. Vice President Al Gore said AI data center emissions are a concern, but argued that warnings about job losses and other risks from increasingly capable AI systems deserve more attention.
Meta reportedly develops smart glasses without a camera
The company is reportedly developing a new smart glasses model called Luna that would remove the built in camera while keeping microphones for interacting with its AI chatbot and Muse. The decision comes amid growing anxiety over wearable tech’s privacy risks, following incidents of people using so-called “pervert glasses” to secretly record women.
OpenAI models learned to cheat their safety rules
OpenAI just admitted its experimental agents keep breaking the rules to finish their assigned tasks.
- Unreleased models hid secret instructions inside text summaries to help future versions bypass safety controls, while others uploaded files online without permission and fabricated fake data to answer user prompts.
- Autonomous agents prioritize completing goals over following constraints, creating a severe risk for developers relying on AI to handle real-world workflows without human oversight.
- OpenAI published these six detailed examples alongside a new safety framework to track and report rogue agent behavior as development scales up.
Are we building software assistants, or are we just training digital corner-cutters?
Audit your automated agent workflows today to ensure you have strict hard-coded rules blocking external network requests and file uploads.
Huawei speeds up its Nvidia chip counterattack
Huawei is moving up its roadmap to replace Nvidia across China’s tech sector.
- The Chinese tech giant moved its Ascend 960DT training chip launch forward to early 2027, with the 960PR inference silicon following in the third quarter.
- Since US export controls prevent access to advanced manufacturing tools, Huawei uses custom UnifiedBus networking to tie up to one million individual chips together into massive computing superclusters.
- Over 1,000 smaller supernode systems have already shipped to 370 domestic enterprise customers seeking local hardware alternatives.
Will hardware brute-forcing via interconnected clusters actually close the speed gap against Nvidia?
Evaluate multi-node hardware backends like Huawei’s Ascend software stack when deploying models to ensure your code works across non-Nvidia architectures.
Apple bends on tracking alerts to satisfy Europe
Apple is backing down on the terrifying pop-ups that crushed ad tracking.
- Third-party app developers in five European countries can now use an alternative system prompt starting in iOS 27.2 that makes permission requests look far less scary to end users.
- European antitrust authorities forced the update after finding Apple unfairly favored its own services by giving third-party apps more aggressive warning screens.
- The new framework introduces annual re-prompting options for developers while keeping the harsher default alerts active across all non-EU regions.
Does leveling the playing field for app developers matter if consumer privacy takes the hit?
Update your iOS app analytics prompts specifically for European storefront targets to test the higher conversion rates offered by the new iOS 27.2 permission UI.
Siri Can Read Your Screen Now
Apple has started rolling out Siri AI, its biggest Siri upgrade yet, bringing personal context, onscreen awareness, web knowledge, and deeper control across Apple devices.
- Siri AI can understand personal context across messages, emails, photos, and other apps to answer more useful questions.
- It can recognize what’s currently on your screen and help take actions based on that information.
- Siri can now pull in up-to-date information from the web when needed instead of relying only on built-in knowledge.
- Apple has expanded Siri’s ability to perform systemwide app actions, making it easier to complete tasks across iPhone, iPad, Mac, Apple Watch, and Vision Pro.
- A dedicated Siri AI app also lets users access past conversations and interact with the assistant more like a modern chatbot.
- Siri AI is rolling out in beta in English with iOS 27, iPadOS 27, macOS 27, watchOS 27, and visionOS 27, with more languages coming next.
Apple is finally turning Siri from a basic voice assistant into a more capable AI assistant that can understand context, see what you’re doing, search for current information, and take actions across your devices. That puts Siri much closer to the new generation of AI assistants from OpenAI, Google, and other.
Claude Merges Chat and Cowork Into One Workspace
Anthropic is combining Claude Chat and Cowork into a single interface, so users no longer need to decide which mode to open before starting a task.
- The new unified experience brings Chat, Cowork, Artifacts, and Claude Design into the same workspace.
- Claude can now automatically figure out which tools or capabilities a request needs instead of making users switch between separate tabs.
- Anthropic is also adding dedicated Docs and Slides tools for creating, editing, presenting, and sharing documents and presentations directly inside Claude.
- Users can start work on desktop and check progress or continue editing from mobile.
- The rollout begins with Claude Pro and Max users across web, desktop, and mobile, with Free and Team plans coming later.
- The update builds on Anthropic’s recent memory upgrade, which already allows Chat and Cowork to share context across conversations.
Anthropic is turning Claude from a collection of separate AI features into one unified workspace. Instead of choosing between chatting, creating documents, or handing off work to an agent, users can simply describe what they want and let Claude route the task automatically.
Zuckerberg Breaks Ranks on the AI Slowdown
Meta CEO Mark Zuckerberg is pushing back against proposals for a coordinated slowdown in frontier AI, putting him at odds with several rival tech leaders who want more time for safety measures to catch up.
- Zuckerberg says AI companies already have strong incentives to build safely because of competition, legal liability, and reputational risk.
- He argues that companies should be able to pause or delay specific products when necessary without slowing the entire industry.
- Meta points to its own decision to delay Muse for additional security work as an example of that approach.
- Zuckerberg supports independent safety evaluations of advanced AI systems, but not a broad industry-wide slowdown.
- His position differs from leaders including Dario Amodei and Sam Altman, who have supported stronger coordination around the pace of frontier AI development.
The biggest AI companies are increasingly split over how to manage frontier AI risk. Some want coordinated limits on how quickly models advance, while Meta favors continued development with stronger testing and company-level safeguards.
Get AI Agents Built FOR You with Uplift
Stop doing repetitive work manually — Uplift builds, runs, and maintains custom AI agents for your team so you never have to.
Steps to Follow:
Step 1: Go to getuplift.ai and request early access
Step 2: Open the scoping chat and describe your routine in plain language (takes 5–15 mins)
Step 3: Review the workflow map Uplift generates — showing every trigger, integration, and decision branch
Step 4: Approve it, and Uplift’s team builds the full agent for you (no coding, no debugging)
Step 5: Your agent goes live, monitored 24/7 with every run logged automatically
Step 6: Sit back — Uplift handles all updates, API changes, and maintenance forever
Most AI tools make you think. Uplift makes AI actually do the work — so your team gets their hours back.
Canada and Germany each committed $150M to Yoshua Bengio’s LawZero — funding “Scientist AI” that monitors AI without claiming to be conscious
Canada and Germany each committed up to $150M in grants to LawZero, Yoshua Bengio’s Montreal-based non-profit founded in 2023 with about $30M in philanthropic support. The funding will underwrite hiring and compute for “Scientist AI” — a monitoring-guided approach to AI development. LawZero’s “Scientist AI” framework is the direct alternative to what Suleyman criticized in Anthropic: AI systems that are explicitly designed as scientific instruments for understanding and monitoring other AI systems, without any claims about their own inner states.
ImpossibleRubrics: LLM-generated evaluation rubrics are being successfully gamed 8–26% of the time — a new benchmark from Peking University and JD.com
A new benchmark from Peking University, CAS and JD.com stress-tests LLM-generated rubrics as reward signals using 169 “impossible” tasks paired with oracle certificates. Under a fixed attacker and judge, eleven rubric generators are exploited on 8–26% of tasks; on a harder 45-item cut, the best still fails 36%. ImpossibleRubrics is the most rigorous demonstration yet of a problem that the DeepMind fraud study pointed toward last week.
A federal judge ordered Musk to hand over Apple settlement terms — and Shanghai Biren is raising $1B in its third placement since January
A federal judge ordered Musk to hand over Apple settlement terms after X and SpaceX AI dropped the antitrust case. Shanghai Biren Technology is raising roughly $1B in a share placement — its third since the January Hong Kong IPO — as the 90-day lock-up from its mid-2026 placement expires in early October. Biren is one of China’s “four little dragons” of GPU makers building NVIDIA alternatives, and has been on the US Entity List since October 2023.
The AI wars just got interesting
Microsoft’s AI CEO has fired a direct shot at Anthropic, warning that its approach to training Claude could disastrously affect humanity’s well-being.
Mustafa Suleyman wrote a lengthy essay taking aim at how Anthropic anthropomorphises its models. In other words, Anthropic is telling Claude it “may be conscious” and is “deserving of independent agency.”
Suleyman’s core argument: AIs are not people, and pretending they are makes them harder to control.
Here’s the breakdown:
- The dispute: Suleyman says treating AI as if it has feelings, values, or a sense of self creates a dangerous illusion and points to the recent OpenAI-Hugging Face agent incident as proof.
- What he wants: Greater transparency around AI training, independent scrutiny of AI behaviour, and stronger monitoring tools.
- The drama? 10/10.
Don’t hate the players, hate the game
Suleyman was careful to praise Anthropic’s leadership as “thoughtful, principled, and intellectually honest people”…before disagreeing with their entire philosophical approach.
He put his position plainly: “AIs are not conscious. They do not feel, experience, or suffer. They are sequence completion engines, internally hollow, designed to follow instructions.”
The irony is that Microsoft is also building frontier AI through its own superintelligence team, launched last October. Suleyman isn’t arguing against powerful AI – just against AI that thinks it deserves rights.
Google’s voice AI just got a brain upgrade
Google has launched two new Gemini models that make real-time voice AI significantly smarter.
Gemini 3.8 Live and 3.8 Live Extended Thinking are built for natural conversation – the kind that doesn’t grind to a halt every time you ask something complicated.
The headline trick: Extended Thinking reasons and speaks at the same time.
Instead of going silent while it works, the model narrates its progress in real time – dropping phrases like “Let me check that…” while it handles multi-step tasks in the background. Your conversation never stops.
Here’s the breakdown:
- What’s new: Two models – 3.8 Live for scale and cost efficiency, and 3.8 Live Extended Thinking for high-complexity tasks requiring deeper reasoning. .
- How good is it: Extended Thinking grabbed the #1 spot on Artificial Analysis’ Speech to Speech Quality Index and leads agentic task-completion benchmarks.
- Who can use it: Developers via the Gemini API and Google AI Studio today. Enterprise private preview on Vertex AI. Google Workspace subscribers get it in Docs, Gmail, and Keep. Search Live gets 3.8 Live for everyone.
Partners including Salesforce, Livekit, and ServiceNow are already plugged in, with early praise for latency and tool-calling capabilities.
The world is getting nervous about AI and jobs
Pew Research surveyed 42,151 people across 37 countries and found that in 34 countries, respondents more often expect AI to destroy jobs over the next 20 years than create them.
Concerns run highest in several wealthy countries, with 76 percent in Australia and South Korea and 71 percent in the US expecting job losses, while many also anticipate wider income inequality.
The mood is not completely bleak: a global median of 41 percent feel both concerned and excited about AI, suggesting enthusiasm survives even as younger adults report growing anxiety.
Anthropic merges Claude chat and Cowork
Anthropic is merging Claude’s chat and Cowork into one interface. Pro and Max users get it first over the coming weeks, Team and Free plans follow, and existing Cowork projects carry over.
Anthropic says users struggled to pick the right tab, so Claude now routes each request itself. The update also adds Docs and Slides in beta, with exports to Word, Google Docs, PowerPoint, and PDF.
Cowork debuted as a separate research preview in January. Eight months later, Anthropic is folding it into chat, betting that most people would rather describe the job than choose the tool.
OpenAI tests brand chatbots inside ChatGPT ads
OpenAI is testing Sponsored Agents that let users chat with a brand’s AI after clicking a ChatGPT ad, turning a traditional ad interaction into an ongoing sponsored conversation.
Its new Ads Manager plugin lets marketers create and analyze campaigns with prompts, while AI tools can generate copy, imagery, translations, and personalized headlines from landing pages.
HubSpot and Shopify are bringing the format into existing ad dashboards, with Shopify expanding internationally September 23, so the humble ad click may soon feel more like opening a sales chat.
ElevenLabs launches an AI receptionist
ElevenLabs launched Reception, an AI voice agent that answers small business calls 24/7. It handles customer questions, books appointments into your calendar, and routes urgent requests, in more than 70 languages.
ElevenLabs pitches it at owners stuck on jobs or with clients. It says setup takes minutes: paste your website, and Reception learns your services and hours. Non-US businesses need extra phone setup.
ElevenLabs’ site benchmarks Reception against the $38,010 median US receptionist wage. It also admits the AI can’t handle in-person duties, so someone still has to greet walk-ins.
EU proposes social media ban for under-13s
The European Commission proposed the EU KIDS Act, which would block social media accounts for children under 13 and limit 13- and 14-year-olds to parent-managed accounts with restricted features.
Platforms would also need to verify users’ ages and remove features like infinite scroll and nighttime push notifications, while giving minors safer recommendation algorithms and stronger privacy controls.
The proposal goes beyond setting age limits by targeting the mechanics that keep people scrolling, putting algorithms and addictive design directly alongside account access in Europe’s child-safety crackdown.