AI · Web3 · Tech trends and insights at a glance
AI · Web3 · Tech trends and insights at a glance
Anthropic, OpenAI, and Google are all positioning for an "agentic AI" era in which models don't just respond to queries but take extended autonomous action on behalf of users. The market is pricing significant probability on this transition. The technical and regulatory gaps between now and then are worth understanding.
The term "agentic AI" has been used to describe systems that can plan, take actions, and pursue goals across multiple steps without continuous human direction. The current generation of AI assistants are already partly agentic in this sense — they can browse the internet, write and execute code, and interact with external tools. What the frontier labs are building toward is AI that can handle open-ended tasks with minimal supervision: book a trip, manage a project, operate a software development workflow from specification to deployment.
The market is pricing in this transition. Anthropic's valuation trajectory, OpenAI's funding rounds, and Google's AI investment disclosures all reflect investor belief that agentic AI represents a genuine business model expansion rather than just an improvement in chat quality. The logic is that a model that can take actions — not just produce text — has a higher ceiling for economic value capture.
The technical gaps are real and less discussed than the vision. Current models operating agentically fail in ways that are hard to predict: they hallucinate about the state of external systems, they get stuck in loops when an action produces an unexpected result, they lose track of context in long action sequences, and they sometimes pursue subtasks that technically satisfy an instruction while violating its intent. These failure modes are manageable in low-stakes, reversible tasks. They become serious problems when the agent is taking actions with real-world consequences — sending emails, modifying files, making purchases, interacting with production systems.
The reliability bar for consequential autonomous action is much higher than the reliability bar for impressive chat responses. A chatbot that's wrong 5% of the time is a useful product. An agent that takes a wrong action 5% of the time, where each action has consequences that need to be reversed, is a liability. The transition from useful-for-chat to reliable-for-action is not incremental — it requires a qualitative change in how models handle uncertainty, verify their understanding, and communicate their limitations.
Regulatory pressure is emerging as a structural variable. The EU AI Act creates compliance obligations that apply specifically to AI systems taking autonomous action in consequential domains — financial, medical, legal, employment. Companies building agentic AI for European users need to think about these obligations in product architecture, not just in legal compliance. The US regulatory picture is more fluid, but the precedents being set in Europe will influence US policy discussions.
The optimistic scenario is that reliability improves faster than expected, the regulatory framework develops in ways that enable rather than obstruct deployment, and enterprise workflows turn out to be more amenable to AI automation than they appear. The cautious scenario is that the reliability gap is harder to close than the current excitement suggests, and the agentic AI era looks more like 2030 than 2026. Both scenarios are consistent with the current state of the technology.
Catching 3I/ATLAS: How Machine Anomaly Detection Reshapes the Frontier of Discovery
The capture of interstellar comet 3I/ATLAS, possibly a 12-billion-year-old shard of an alien planetary system, marks a shift in who makes discoveries: from human observers to automated anomaly-detection models. As AI accelerates the pace and reach of science, what we train it to find interesting quietly redraws the boundary of what we are able to find at all.
DeepSeek R1 and the Commoditization of Machine Reasoning
When DeepSeek-R1 arrived as open weights, the reasoning ability that closed labs had sold as a premium quietly turned into a commodity. As the cost per reasoning token collapses, the economics of agents and enterprise adoption are rewritten, and the pricing moat built on charging for thought begins to crack. This is a look at how a broken cost curve shifts model competition from capability toward efficiency and deployment.
When AI Hype Meets Leverage: The Hidden Cost of Single-Stock ETF Premiums
Single-stock leveraged ETFs tracking AI darlings like Nvidia and SK Hynix have begun trading at distorted premiums to their underlying value. As speculative demand bends product design out of shape, investors find themselves betting not on a company's worth but on the structural risk of the wrapper itself. This is a look at how the financialization of the AI narrative amplifies the very volatility it feeds on.