Browse Comments

This could become a very important turning point for edge AI. As models become smaller, more efficient, and increasingly optimized for local hardware, AI may gradually move from massive centralized data centers toward personal devices and local inference systems. The long-term implications for privacy, latency, and AI accessibility could be enormous.

The shift from isolated tutorials to open-sourcing actual enterprise-grade, production-ready architectures is exactly what the AI engineering ecosystem needs right now, Abhishek Veeramalla. Building a basic wrapper is easy, but managing persistent memory, multi-agent reasoning, and scalable vector DB implementations in the real world is where the real friction lies. Oracle open-sourcing a bluepri…

This has to be managed but there is always a missing part to this dicusssion - the potential net positive. If the obligations or reven rules are right, AI can be applied to generate exponential carbon savings across industry, serices and user activities.

Everything will move this direction. Local models running for regulated industries and those seeking privacy for IP security. It is the same reason Punky Tiger Labs created the PCIe card to turn any desktop PC into an AI machine.

What feels impossible today often becomes normal surprisingly fast in technology. The pace of AI hardware advancement right now is honestly incredible.

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 …

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 i…

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,…

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 c…

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 robo…

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…

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 tea…

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 ge…

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?…

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 …

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 digita…