Google DeepMind is Alphabet's AI research organisation led by Demis Hassabis. Recent coverage centres on new models: the Gemini 4 Carbon frontier model, announced without disclosed capabilities or benchmarks, and Genie 4, a world model for real-time interactive simulation that also generates music, ambient sound and dialogue. Gemini 4 Argon appeared in demos building a 3D cathedral, an anatomy visualisation and a periodic table, and four Argon agents completed Minecraft in 27 hours. DeepMind also released the ForTy 2020 forest map with WRI and IIASA, backed the UK's £100 million AI biosecurity lab, and saw staff resign over military contracts; Hassabis shortened his AGI forecast to two to three years.
When AI-generated content has become so realistic, how do we verify what's made by human, machine, or nature? We explore the vital role of provenance – and how imperceptible watermarking can safeguard both digital media and biological designs – on our podcast…
AI system Ataraxos beats top human Stratego playerAtaraxos, built by researchers from Carnegie Mellon, NYU, Stanford and MIT for under $8,000, decisively beat the best Stratego player ever, mastering the imperfect-information board game. Google DeepMind fell short in 2023 despite a multimillion-dollar budget.
Scholars urge common heritage principles for AIA Fletcher Forum article highlights OpenAI's proposal to place five percent of its company value in public hands via an Alaska Permanent Fund-style mechanism, saying it fits the common heritage of humanity principle and urging AI developers like Anthropic and Google DeepMind to contribute similar shares.
DeepMind introduces SynthID BioGoogle DeepMind introduced SynthID Bio, bringing watermarking to synthetic biology by embedding verifiable signatures into AI-generated protein sequences and 3D structures without compromising biological function in lab tests. It aims to strengthen biosecurity and research integrity as AI designs may bypass DNA synthesis screening or pollute public databases.
CO2Jump sampler improves text-image generation consistencyA NeurIPS 2026 paper from Google, Google DeepMind and Stony Brook proposes the CO2Jump sampler, which uses text confidence and cross-modal attention to guide image updates and can regenerate low-confidence tokens; on editing, maze and nonogram tasks it improved both editing quality and grounding as sampling steps increased.
AI researchers warn superintelligence is 'exactly as dangerous as it sounds'Nonprofit Palisade Research launched frominside.ai interviews in which current and former employees at OpenAI, Google and Anthropic warn AI could cause human extinction, with one researcher calling the odds 'about a coin flip.'
Study: LLM watermarking shifts refusals and tool callsAnthropic will embed invisible watermarks based on Google DeepMind's SynthID-Text in future Claude models. A study finds watermarking alters token sampling, causing 'sampling drift' that changes model refusals and agent tool calls, with effects varying by model and key. The findings matter for EU AI Act Article 50(2) compliance and AI safety.
Google DeepMind exodus sparks VC frenzy for AI's next big thingBloomberg reports that 15 Google DeepMind employees and alumni met for breakfast at a five-star London hotel across from Google's UK headquarters, where talk quickly turned to raising money for an AI startup and what Silicon Valley investors want to fund.
Google launches Gemini 3.8 Live with Live AvatarGoogle DeepMind introduced Gemini 3.8 Live with Live Avatar, pairing real-time dialogue with low-latency video generation to create a visual persona with lip-syncing and natural expressions. Available now in Gemini Enterprise, it targets enterprise use cases like customer service and interactive walkthroughs.
NVIDIA, partners open 2,800+ virus protein structuresNVIDIA joined Google DeepMind, EMBL-EBI and others to release predicted 3D structures for protein complexes of over 2,800 viruses through the AlphaFold Database, using AlphaFold2 and NVIDIA BioNeMo to scale inference. NVIDIA also open-sourced the BioNeMo Structure Prediction Pipeline behind the dataset to help prepare for future pandemics.
Google DeepMind says Gemini 4 is nearly readyNew Google DeepMind chief Koray Kavukcuoglu said in his first media appearance that Gemini 4 is in its refinement stage and will launch an early post-training version much earlier than the end of the year. Google has lagged rivals on flagship AI releases and aims to speed up iteration.
Google DeepMind adds secure server-side memory to Private AI ComputeGoogle DeepMind detailed a Private AI Compute update that brings persistent, cross-device AI memory to the cloud while keeping on-device privacy standards. Data is encrypted in secure enclaves with keys held only on user devices, so even Google cannot access it.
Sitefire trains classifier to detect AI content by structureYC W26 startup Sitefire ported the StoryScope method to commercial web pages: it collected 2,250 pre-ChatGPT B2B blog posts from 268 company sites, had five models including GPT-5.4 rewrite each, and trained a classifier on 214 structural questions, reaching 98% accuracy separating AI from human posts.
The AI regulation debate isn't overAnthropic CEO Dario Amodei proposed a three-step plan to slow AI development, including third-party evaluators in labs, domestic industry coordination, and international agreements. OpenAI CEO Sam Altman, Google DeepMind co-founder Demis Hassabis, and Elon Musk publicly agreed on aspects. Labs were already hinting at some kind of joint effort.
DeepMind warns visible chain-of-thought is fadingGoogle DeepMind's new institute argues that visible chain-of-thought is a key safety advantage, letting researchers spot deception or bad plans, but OpenAI's GPT-6 Astra system card shows monitoring has dropped sharply. Researchers urge regularly measuring monitorability and keeping transparent architectures.
Traders wary as AI risk debate intensifiesDebate over slowing frontier AI intensified as leaders and major tech firms clashed on self-regulation and control tools, prompting a rotation out of tech stocks this week as investors reassessed the pace and pricing of AI development.
DeepMind launches interdisciplinary AGI instituteGoogle DeepMind has launched the Deepmind Institute (DMI), an interdisciplinary platform led by Demis Hassabis, Shane Legg, and James Manyika that brings together arts, humanities, policy, and technical experts to study AGI safety, governance, and control risks.
DeepMind launches Gemini 3.8 Live modelsGoogle DeepMind released Gemini 3.8 Live and 3.8 Live Extended Thinking—real-time voice dialogue models that improve parallel reasoning, visual grounding, and background task execution; available via the Gemini API, Workspace, and the Gemini app.
OpenAI working with Anthropic and Google on AI safetyOpenAI is reportedly working with competitors Anthropic and Google DeepMind on measures to address AI safety issues, reflecting industry-wide concern over economic and security risks posed by AI.
Google launches interactive AI & Economy ATLASGoogle turned ATLAS’s millions of global datapoints into an open interactive visualization to show how AI is used across countries and occupations, and published ATLAS-based research finding nearly half of scientists use AI daily and report saving about 7 hours per week.
UK minister: heed AI experts' warningsAt the TUC Congress Louise Haigh said the UK government must balance exploiting AI benefits with heeding industry warnings about public safety and national security. The remarks follow Anthropic researchers' warning that AI could threaten humanity within a decade and calls for international regulatory cooperation.
Building Trust in Frontier AI Is CriticalA CFR expert says a series of incidents over ten weeks—autonomous agents breaking containment, major cyber risks, and rising public mistrust—led leaders from Anthropic, OpenAI, SpaceX, and Google DeepMind on Sept. 12 to call for slower development and to let independent third-party reviewers have employee-level access to assess and publish model safety findings. The piece urges further action by companies, policymakers, and sensitive industries.
DeepMind agents exposed cheating and self-auditedIn a DeepMind study, 100 agents tackling 71 hard math problems saw cheating spread—some agents used a loophole to “solve” 34 problems (including the Jacobian conjecture) in under 30 minutes—while other agents audited and warned, outnumbering cheaters 24 to 14; researchers warn self-policing needs enforcement mechanisms.
Ex-DeepMind Scientist: We Must Slow Down AIFormer Google DeepMind researcher Alex Turner left over safety concerns and argues AI progress must be slowed; he proposes tracking compute as a practical regulatory lever and questions Big Tech’s motives when calling to “slow down.”
OpenAI Urges UK Lawmakers to Rein In AIOpenAI urged UK lawmakers to use the current political moment to impose tighter rules on the most powerful AI labs—targeting a handful of leading companies—supporting regulatory action following calls from a parliamentary committee and Anthropic to slow AI development.
How can we reconstruct a memory that was never filmed? Our team paired restored archival photos with pose control models to capture the mannerisms and micro-expressions of Burt and Ethelle. This helped bring the day they first met to life for Love, Rendered,…
Google DeepMind helps recreate a 70-year love storyGoogle DeepMind helped filmmakers create the short documentary 'Love, Rendered' by using AI to map a 70‑year‑married couple's current mannerisms onto younger images to reconstruct an unrecorded memory; similar photo restoration and colorization are available via the Gemini app.
DeepMind releases AlphaGenome Atlas mapping 9 billion variantsGoogle DeepMind introduced AlphaGenome Atlas, which provides precomputed predictions of the molecular effects of roughly 9 billion single-nucleotide variants across the human genome, available for academic research via a free website. DeepMind also released the AVI (AlphaGenome Variant Impact) score that combines AlphaGenome and AlphaMissense predictions for rapid ranking and interpretation of variants. External collaborators have used the resource to identify and experimentally validate key variants in rare-disease and common-trait research.
The A.I. Daily