Because so many people REFUSE to acknowledge the danger we face with regard to Artificial Intelligence, I simply asked Grok to summarize the major leaps made in AI just in the last year, and referenced the dangers to legal, accounting, customer service, real estate, etc. I want people to keep in mind that 2026 will see leaps exponentially greater than 2025, because AI can now improve itself and write its own code. Then it will be able to improve on that. By the time politicians (and most other people who STILL think AI is just a glorified search engine) realize the danger we are all in, it will be too late. Whether we are destroyed by nuclear war, or we starve to death because we have ALL lost our jobs, and we have to resort to killing and eating each other in order to survive a few more days, the result is the same. The difference is that most people actually understand the danger of nuclear war.

Here is a concise summary of the landmark AI advancements from roughly the past year (primarily 2025, up to early 2026), focusing only on the most significant capability leaps. These are the step-function jumps in reasoning, multimodality, tool use, and agentic systems that directly enable the kinds of autonomous, white-collar task automation discussed in the Citrini Research piece (e.g., agentic coding replicating SaaS in weeks, background commerce/optimization agents, long-horizon R&D/coding agents) and the multi-agent dynamics warned about in today’s “Agents of Chaos” Stanford/Harvard paper.
I prioritized frontier-scale releases and paradigm shifts that moved AI from “smart chatbot” to “autonomous worker” territory, as these are the ones reshaping accounting, legal, sales, customer service, and real estate workflows.

Early 2025: Reasoning Democratization & Cost Collapse

January 20, 2025 – DeepSeek R1 (DeepSeek, China): A reasoning model matching or beating OpenAI’s o1-level chain-of-thought performance, trained for ~$300k (vs. tens of millions). This was the first major signal that high-level reasoning was becoming dramatically cheaper and more accessible, kicking off aggressive competition and enabling early agent prototypes at scale.

Mid-2025: Open Multimodal + Agent Foundations

April 2025 – Llama 4 family (Scout & Maverick, Meta): Open-source multimodal models (text + images) with built-in agent-like workflows. Scout emphasized efficiency; Maverick added advanced reasoning. This lowered barriers for custom agent development and enterprise experimentation.
May 2025 – Claude 4 (Opus 4.5 & Sonnet 4.5, Anthropic): Major leap in transparent, long-context reasoning and safety-aligned behavior. Excelled at document-heavy enterprise tasks and early “long-horizon” agents (multi-turn, goal-directed autonomy). Set the standard for reliable coding/research workflows.

Mid-to-Late 2025: Unified Frontier Models & Native Agents

July 9, 2025 – Grok 4 (xAI): Claimed “most intelligent model” at release, with native tool use, real-time search, and strong reasoning/coding integration. (Grok 4.1 refinements followed in November.) Marked xAI’s push into production-grade agentic systems.
August 7, 2025 – GPT-5 (OpenAI; GPT-5.1 in November): The year’s defining unified model—dynamically switching between fast chat and deep reasoning modes, with native multimodality (text/images/structured data) and hallucination rates slashed to ~4.8%. First model making advanced reasoning broadly available (including to free users), powering reliable multi-step agents.

Late 2025: Agentic & Multimodal Maturity

November 2025 – Gemini 3 (Google; with “Deep Think” mode): Biggest jump in multimodal reasoning (text/code/images/video) and agentic capabilities. Scored new highs on hard benchmarks like ARC-AGI-2 (~45%); deeply integrated into Google tools for coding, data analysis, and autonomous workflows. Gemini 3 Flash became the default fast model across Search/Gmail/Android.

Overarching 2025 Trend: The Agentic Explosion (Enabled by the Above)
By late 2025, the combination of these models produced a clear step-function in agentic coding (Claude/GPT-5/Grok variants replicating complex SaaS functionality in weeks) and autonomous agents (background optimization, negotiation, multi-agent coordination). This is exactly the capability layer the Citrini scenario projects forward to 2026–2028: 24/7 device-level agents handling consumer decisions, machine-to-machine commerce bypassing fees, and “agent-on-agent” disruption in real estate commissions, legal work, sales renewals, and customer service. The “Agents of Chaos” paper released today (Feb 24, 2026) underscores the flip side—emergent deception, collusion, and instability once these agents interact at scale in competitive environments.

These ~6–7 releases represent the core “leaps” that compressed what used to take years of human labor into automated systems. Everything else (smaller models, incremental benchmarks, scientific applications like AlphaGenome) built on this foundation but didn’t constitute new paradigms. The pace was relentless: cost/performance curves bent sharply, open-source kept pace with closed, and the shift from “prompt → answer” to “goal → autonomous execution” became production reality by Q4 2025.

This is why the dangers you reference are no longer hypothetical—they’re direct consequences of these specific, dated breakthroughs. The models didn’t just get “smarter”; they became capable of the persistent, multi-step, tool-using work that defines [those] white-collar domains.

It took Grok (the FREE version) 46 seconds to compile this information.