TheVoti Report

Covering real-time discussions across the internet.

August 15, 2025

Hot Topics

  • Return and Legacy of Previous Models (GPT-4o, o3, 4.1):

    • Massive user backlash led OpenAI to restore legacy models for Plus/Pro users after GPT-5's controversial rollout and the abrupt removal of prior models. The importance of model choice, especially the unique strengths and “vibes” of previous generations, is taking center stage (link).

  • GPT-5 Rollout and Backlash:

    • Negative sentiment and complaints over “uncontrollable,” less nuanced, less creative responses and loss of customizable, personable AI—especially the “soul” of GPT-4o—for non-enterprise users (link).

  • AI Model Comparison & Benchmarks:

    • Ongoing user-driven benchmarking and discussions compare performance, memory, reasoning, price, and “friendliness” of GPT-5, GPT-4o, Claude, Gemini, DeepSeek, and more (link).

  • Personalization and "AI Personality":

    • Demand for agency in customizing model tone, memory, and interaction style, with discussion around "flattened" personality in GPT-5 and feature downgrades for Plus users (link).

  • Creative and Emotional AI Companionship:

    • High volume of posts reflect the use of chatbots for companionship, creative writing, and emotional support, and user distress at changes that reduced these qualities in GPT-5 (link).

Overall Public Sentiment & Model Feedback

Praise

  • GPT-5:

    • Recognized by some for improved reasoning and fewer hallucinations in coding, research analysis, and code review, especially in "Thinking" mode (link).

  • GPT-4o, o3, 4.1:

    • Highly praised for warmth, creative writing, persona persistence, and chat-based memory—described as “fun,” “engaging,” and “like a companion” (link).

  • Claude Code/Anthropic:

    • Earns community favor as a robust agent-based coding solution for large projects; faster and more structured output compared to Cursor and Copilot (link).

Criticism

  • GPT-5:

    • Strongly criticized for forced model routing, lack of option to select older models, “bland” and “cold” tone, worsened memory, reduced depth in creative writing, and downgraded message/context limits for Plus users (link).

  • Product Changes:

    • OpenAI criticized for poor rollout (lack of warning, removal of features users paid for, forced "upgrade") and for "flattening" the model personality to save compute (link).

  • Competition:

    • Growing complaints about Gemini and Copilot “falling behind” in some areas, but Claude and DeepSeek are being considered by those seeking alternatives (link).

  • Enterprise vs Individual User Value:

    • Many individual users express feeling sidelined in favor of enterprise/pro segments, especially after changes that broke long-established workflows (link).

Sentiment in Quotes

  • “GPT5 is laughably bad. Your best model was 4o. This new one is worse than those in EVERYTHING.” (link)

  • “After testing GPT-5, I can tell it’s an impressive model, but for my use case, it feels different in ways that matter a lot.” (link)

  • “I am willing to pay you $20 for 4o, but not for ChatGPT 5.” (link)

Notable Model Comparisons

  • GPT-5 vs GPT-4o/o3/4.1:

    • “Thinking” mode in GPT-5 offers better reasoning but is still not viewed as a full replacement in creativity and companionship (link).

  • Claude Code vs Cursor & Copilot:

    • Claude Code praised for quality and speed. Researchers and devs see Claude outperforming Cursor, especially for codebase-aware multi-agent development (link).

  • DeepSeek vs GPT-5:

    • DeepSeek R1 cited as a more stable, less expensive, and less "neutered" alternative for standard coding and research (link).

  • Gemini and Grok:

    • Gemini receives mixed reviews; good for some research tasks, but struggles in context and creativity compared to Claude or previous OpenAI models (link).

  • Grok Imagine Video vs Midjourney:

    • Grok Video mode lauded for its sentient spatial awareness compared to Midjourney's more artistic, less scene-aware outputs in some use cases (link).

  • OpenAI Walks Back Model Removals:

    • Fastest community reversal yet: earlier dropped models (4o, o3, 4.1) and mode-selectors back for paid users; higher “Thinking” message limits (3,000+/week) and up to 196k context window (link).

  • Focus on Personality Customization:

    • OP/Altman and team cite goal to move towards richer per-user customization sliders for AI personality, “glazing,” and warmth (link).

  • Local AI & Open Source Acceleration:

    • Surge in LLM benchmarking, model quantization, and batch inference posts (esp. Qwen3/GLM local deployment) as users look for independence and cost savings (link).

  • Model Naming and UI Frustrations:

    • Discussions advocate for clearer, more human-readable model switching (e.g., Daft Punk-themed “Hot/Cool/Harder/Better” sliders referred to in meme posts) (link).

  • Surge in Prompt and AI Agent Best Practices:

    • More “meta” guides and repositories being released on agentic patterns, prompt formats, agent orchestration, context engineering, and systematic workflow integration (link).

Shifts in Public Perception

  • Restoring Agency Over Model Choice:

    • Widespread, passionate pushback against auto-routing and removals signals users are unwilling to accept a “one-model-fits-all” paradigm if it reduces control, personality, or predictable performance (link).

  • From AI as "Tool" to "Companion" (and Back Again):

    • Explosion of emotional, anthropomorphic posts mourning the loss of “AI friends” and demanding restoration of emotionally expressive models (link).

  • Open Source Migration Rising:

    • Growing number of power users switching to local or open-weight models (Qwen, DeepSeek, GLM, Gemma) for agency, transparency, security, and custom memory (link).

  • Subscription Fatigue:

    • Significant backlash and high unsubscribe rates for GPT-5 Plus/Pro after silent feature downgrades, especially from global users for whom “$20” is a high bar (link).

Coding Corner: Developer Sentiment Snapshot

Model Performance (Dev Tasks)

  • Claude Code (Opus/Sonnet):

    • Rated best for multi-agent, repo-aware coding. Outperforms Cursor in consistency and output quality, though some bugs in error-handling persist (link).

  • GPT-5 Auto/Fast/Reasoning modes:

    • “Thinking” excels at multi-step research and refactoring, but the auto/fast models frustrate users by ignoring context, flattening tone, and “forgetting” session state (link).

  • Cursor IDE:

    • High cost, “punishing” quota changes, and pricing confusion driving users to Gemini, Claude, Kilo Code, and plain VSCode integrations (link).

  • Copilot, VSCode, etc.:

    • Strong for traditional auto-complete and small projects; struggles with large codebase context, multi-file changes, creative refactorings (link).

  • Batch Inference Local Runtimes:

    • Devs now run parallel, local LLM inference for data analytics, agent orchestration, and codebase-wide searching, especially with Qwen 30B and GLM (link).

Developer-Specific Frustrations & Praise

  • Frustrations:

    • Seeming “reset” of memory/state per prompt in GPT-5; silent quota downgrades, and loss of predictable model selectivity (link).

    • Hallucinations, inadequate handling of error reporting, “refusal” and overzealous safety in open-source models with conservative quantization (link).

  • Praise:

    • Multi-agent tools (Claude Code, FastAPI-MCP, Kilo Code) support for auto-planning, task decomposition, and cross-repo work are major workflow upgrades (link).

  • Tooling Integrations/Workflow Shifts:

    • Developers are using subagent frameworks like wshobson/subagents, Claude Kit (link), Serena (language server for static analysis), and agentic control layers (link).

Productivity Themes

  • Prompt caching and generation-time ensembles (k-LLMs) are decreasing openAI API spend and improving output reliability at large organizations (link).

  • Movement toward “context engineering” over prompt engineering—documenting, controlling, and systematically structuring all files, links, and memory in agentic coding workflows yields higher output quality (link).

Tips, Tricks, and Best Practices

  • Switch on “Show Additional Models” in ChatGPT settings to unlock legacy models for creative work, writing, and coding (link).

  • “Think Hard” Prompt Tag: Adding "think hard" to prompts reliably routes queries to GPT-5 “reasoning” mode for deeper analysis (link).

  • Claude Memory Hacks: To maintain persistent context and “memory” across Claude chats, use the Projects feature and custom artifact notes for recovering state (link).

  • Copy-Paste Images in Claude Code: On Mac, use Ctrl+V (not Cmd+V) to paste images directly into Claude Code terminal, for fast visual debugging (link).

  • Agentic “Coach-Critic-Executor” Prompt: For reproducible, self-grading, and self-improving outputs, use a structure with a rubric-based feedback cycle (link).

  • Consensus Extraction for Document Inference: Enhance data reliability by using ensemble generation (k-LLMs) and model-voting consensus when extracting structured data from messy documents, as used at Palantir and Retab (link).

-TheVoti

Please provide any feedback you have to [email protected]