2025: The Year in LLMs

Perceived progress in 2025 LLMs

• Some see 2025 as a major step: coding agents and reasoning modes turned LLMs from “cute demos” into tools that can meaningfully assist experts.
• Others describe the year as stagnant compared to earlier ML breakthroughs (RBMs, RNNs, early deep learning), arguing that most 2025 changes were tooling and distribution, not fundamental model advances.
• Several note that people’s baseline differs: for many, LLMs are their first exposure to 20 years of ML progress, which amplifies the sense of revolution.

Creativity, “reproducing the past,” and thinking

• One camp argues LLMs and diffusion models fundamentally sample from past data distributions, so they remix rather than create truly novel concepts; this is seen as a hard limit on scientific breakthroughs.
• Others counter that humans also mostly recombine prior knowledge, that stochastic generation can still yield meaningful novelty, and that insisting on some “magic” non-derivative creativity standard is unrealistic.
• There is ongoing disagreement about whether LLMs “think” or have any notion of truth, versus only modeling linguistic patterns.

Coding agents and developer workflows

• Many developers report large productivity gains: agents that run code, observe failures, and iterate are said to handle a majority of minor code changes and refactors in some workflows.
• Critics say generated code is brittle, architecture is poor, subtle bugs are common, and everything still requires expert review; claimed speedups are often vague or overstated.
• Reliability is framed as “good enough to be a useful assistant, nowhere near replacing a competent engineer.”

Agents, MCP, Bash, and tools

• Strong interest in architectures: MCP as a standardized tool interface vs “bash-as-universal-tool” in code execution environments.
• Some foresee MCP fading as cheap, sandboxed shells become ubiquitous; others argue MCP’s auditability, security, and interoperability make it more like REST APIs—long-lived infrastructure.
• Skills, CLIs, and custom MCP servers are all being used to connect LLMs to CRMs, JIRA, and other systems.

Economics, labor, and productivity

• Fears center on junior developer hiring drying up and potential broader knowledge-work automation; some predict manual labor will outlast white-collar work, others dispute this based on verification difficulty outside software.
• Several note that macro unemployment has barely moved, and that efficiency gains may translate into lower prices and new demand rather than mass job loss.
• Debate continues about whether measured productivity reflects any “exponential” capability gains.

Environment, data centers, and hardware

• Commenters worry about energy, water use, subsidies, and e‑waste from massive data center buildouts and GPU churn, especially in rural areas.
• Some highlight that AI demand is heavily distorting DRAM/NAND markets and fear future bailouts or “enshittification” as a few hyperscalers dominate.
• Others, especially hardware-focused participants, emphasize that AI capex is accelerating progress in semiconductors, memory, packaging, and interconnects, similar to the smartphone era.

Safety, “YOLO” practices, and harms

• Concerns about “normalization of deviance”: running coding agents with broad system access, accidental destructive actions (like deleting home directories), and the lack of mature safety culture among web-style developers.
• Various sandboxing strategies are discussed: Firejail, separate users, VMs, Docker-in-Docker, dedicated VPSs.
• There is unease about LLM-linked self-harm and “AI psychosis” cases; some see genuine risk and note labs’ mitigation efforts, others think this is moral panic compared to underlying economic stressors.

UX, slop, and user backlash

• Strong resentment toward intrusive AI chatbots on websites and in apps, which are seen as worsening UX to satisfy “we added AI” mandates and usage metrics.
• “Slop” (low-value AI-generated media) is already saturating search, music, images, and video; some predict AI labels and filtering will be needed, others doubt platforms will resist content that drives engagement and ad revenue.

Polarization, hype, and community dynamics

• The discussion reflects a wide spectrum: from “bigger than the internet” optimism to “marginally useful autocomplete” skepticism.
• Many distinguish between real, narrow utility (coding help, search assistants, document analysis) and overblown AGI narratives and corporate hype.
• Meta-discussion touches on distrust of corporate motives, previous tech bubbles (crypto, Web3, metaverse), and frustration with both LLM evangelism and total dismissal.