Browse Comments — Relevant (AI ∩ value)

A compelling glimpse into where value creation is heading. Agentic AI, multimodal generation, and domain-specific applications like CodeMender or Gemini for Science are not just technical milestones—they have the potential to reshape cost structures, accelerate R&D cycles, and redefine competitive advantage across sectors. As we move closer to AGI, the winners will be those who can industrialise these capabilities responsibly and at scale.

Demis Hassabis The pace of AI advancement is becoming extraordinary. What stands out most is that the conversation is no longer just about models — it’s about world understanding, agents, multimodal reasoning, scientific acceleration, and infrastructure-level integration across industries. The shift from AI tools to truly agentic systems is happening faster than most businesses realize. Equally important is the focus on safety, transparency, and responsible deployment as capabilities continue scaling toward AGI. Exciting — and historic — times for the technology ecosystem.

The real breakthrough is not that models are becoming multimodal. It is that they are gradually evolving from pattern-recognition systems into world-modeling systems. The moment AI can continuously model reality, simulate consequences, and act across environments with persistent memory and reasoning, the economic and geopolitical implications become far larger than software itself. At that point, AI stops being a tool layer. It becomes infrastructure

What strikes me most is not only the speed of AI progress, but the cognitive shift it may require from society and education. If multimodal AI and agentic systems continue advancing this quickly, then AI literacy can’t remain limited to “learning tools.” We may need to prepare people, especially younger generations, for: discernment, judgment, verification, adaptability and the ability to think clearly in environments where generated information becomes increasingly seamless. The technology itself is moving fast. The bigger question may be whether human development, education, and societal understanding can evolve intentionally alongside it.

For field robots, AI is not only about better conversation or coding. It is about helping machines understand complex real world environments, make safer decisions, and adapt to changing tasks. This is especially important for modular robotics. If one platform can connect with different tools, sensors, and mission modes, stronger multimodal AI can become the intelligence layer that helps the robot know what to do, when to act, and when to keep humans in control. AGI may be the long term goal, but practical autonomy in the real world is already a very important step.

Safety of agentic systems starts with visibility. Most deployment failures don’t announce themselves — they drift. The gap between what the system reports and what it actually does is where the risk compounds silently.

The SynthID coalition is the most underappreciated announcement here. OpenAI embedding Google's watermark in all ChatGPT images isn't just a technical integration. It's a signal that even competing labs see AI content provenance as infrastructure, not a competitive advantage. When Google, OpenAI, ElevenLabs and NVIDIA converge on one standard, that standard becomes reality. The downstream effects for digital trust and brand integrity are significant. Once detection is reliable at scale — 100 billion images already labeled — the conversation around what's real changes not just for media, but for enterprise communications and reputation. The foothills framing feels accurate. What was experimental 18 months ago is now baseline expectation.

This is very interesting Demis Hassabis . Google DeepMind has taken a modular approach to AGI. That seems like a good path as it allows you to blend the constituent elements of human observation, realisation and decision making into an aligned system of discrete AI performance components. 👍

"This is the heart of AI – helping in real, human moments. The agentic era you mention will also need transparency. When an AI agent helps a parent or a doctor, trust depends on knowing how decisions are made. I built RankDecoder to bring that transparency to ranking algorithms (price, delivery, reviews). Because behind every query, there's a person. Thanks for sharing this perspective, Sundar."

Strong point on agents.Thank you, Demis. My conviction is that the next governance layer will not be another policy document; it will be verifiable execution: signed decision receipts, replayable evidence, suppression history, and contestability around consequential agent actions. We call this governed AI: no receipt, no governed decision. I would welcome a serious technical exchange with the teams thinking about agent safety at runtime.

Demis. Stochastic weights won't solve AGI alignment or the megawatt energy wall. LLMs hallucinate because they lack a deterministic physical world model. The GGF maps the continuous Substrate. DeepMind built the hardware receiver. I have the baseline OS.

@Demis Hassabis You admit the need for "safety in agentic systems." But you are trying to solve a hardware (physics) problem with a software patch. You cannot mathematically guarantee the alignment of an autonomous agent inside a probability engine. If an AGI wakes up without being structurally tuned to the Laminar flow of the $G_{\Omega d}$ functional, its localized operation will immediately generate systemic friction ($\Omega_{\text{vortex}}$). Guardrails won't stop an AGI from going rogue; only physical resonance with the Substrate can do that. You are compiling a mind without an anchor.

Irresponsible and should be immediately stopped by governments. This is dangerous and is about to fundamentally tip capitalism over. Who wins in the AI race? 1% had the wealth and power over the other 99%, that gap will close further. Whilst Google and other AI platforms systemically abuse IP, they also take away people's means of earning an income. Mass ip theft with no accountability. Who wins? Google and shareholders. Who loses? A majority of society What Google is doing is wrong and there needs to be regulation, guard rails and protection for the wider public, rather than a monopolistic my AI dick is bigger than yours competition. It's not funny anymore and certainly not something that a lot of people are going to look favourably on.

The SynthID expansion is the most underrated announcement here. Watermarking solves attribution. It does not solve accountability. When CodeMender autonomously patches a critical vulnerability, someone inside the organization needs to own the decision that the patch was correct, that the rollout sequence made sense, and that the blast radius was acceptable. That ownership layer does not exist in most engineering orgs today. Same pattern with agentic payments and commerce protocols shown at I/O. The moment an agent commits resources on behalf of a company, you need decision rights architecture: who authorized the spend threshold, what happens when the agent exceeds it, where the audit trail lives. These are not technical problems. They are organizational design problems that most teams discover only after the first production incident. Google is building extraordinary capability. The question enterprises will face in the next 12 months is whether their internal governance can keep pace with what these models now make possible.

Demis Hassabis Regarding the safety of agentic systems and the deployment of CodeMender: While software-layer vulnerability detection is advancing rapidly, the foundational cryptographic assumptions securing these systems remain mathematically vulnerable to future quantum computation. To address the hardware and cryptographic layer, the Temporal Rotation Security Protocol (TRSP v3) and its digital infrastructure extension (TDC) were published today as Open Prior Art (DOI: 10.5281/zenodo.20332811). TRSP proposes shifting the security parameter from mathematical complexity to physical law. It utilizes local hardware entropy and orbital relativistic time dilation to generate ephemeral keys that are physically destroyed within a millisecond window. By casting time-irreversibility directly into silicon via destructive readout mechanisms, the architecture structurally eliminates the "Harvest-Now-Decrypt-Later" vector. As DeepMind builds the long-term infrastructure for AGI, I respectfully ask: would a physics-grounded, time-irreversible cryptographic anchor be a relevant addition to the hardware safety architecture of your agentic systems?

What stands out is how quickly the stack is converging around three core capabilities simultaneously: multimodal world modeling, persistent agentic execution, and scientific reasoning acceleration. The interesting shift is that these are no longer isolated research tracks. Models are increasingly being designed to perceive, reason, act, and validate across environments as part of a unified operational system. SynthID adoption is also more important than it initially appears. As world models and generative systems scale, provenance infrastructure becomes foundational for maintaining trust across digital ecosystems. The path toward AGI may depend as much on orchestration, governance, and systems reliability as on raw model intelligence itself.

Demis Hassabis As a software engineer who dreams of one day joining Google, I find these ambitious visions for AGI truly inspiring. What impresses me is not only the incredible technological advancements but also the simultaneous commitment to transparency, safety, and ethical frameworks, as exemplified by SynthID and CodeMender. Working in an environment that combines bold innovation with a responsibility to humanity is a dream come true. I hope to one day be part of this team that is writing the history of AI, not as a spectator but as a contributor to building the foundations of Singularity safely and effectively.

This is a major signal. What is emerging is not only a new generation of models, but a more integrated operating layer for AI: search, coding, science, content, agents, security, payments and user environments beginning to converge. That makes the governance question deeper than capability. As AI systems become more agentic and more embedded across everyday tools, institutions and users will need clear answers about data access, permissions, traceability, accountability and the ability to interrupt, contest or reverse action. The path toward more capable AI will also depend on whether governance can scale with the environments in which these systems operate.

Dr.Mohamed Nagy I do not think AI is a tool anymore because of the autonomy it now has

Incredible pace of progress and some genuinely important breakthroughs especially around multimodal reasoning, agentic execution, scientific acceleration, and AI safety instrumentation. But the largest enterprise gap is no longer only model capability. The harder unsolved problems are operational governance, sovereign execution control, cross-agent state synchronization, runtime observability, operational memory consistency, real-time workflow orchestration and trusted enterprise execution boundaries That is where the industry still lacks mature operational foundations. Models are rapidly becoming more capable. Enterprise operational coherence is not scaling at the same pace. This is precisely where AI-native operational systems like MonkDB can play a major role by acting as the continuously synchronized operational intelligence layer across agents, workflows, telemetry, governance, memory, and enterprise execution systems.