TheVoti Report

Covering real-time discussions across the internet.

August 22, 2025

Hot Topics

  • Massive Backlash Over OpenAI's GPT-5 Rollout and Model Removal: Widespread user outrage about the abrupt removal of GPT-4o, 4.1, 4.5, etc., as well as limited access and a forced "upgrade" to GPT-5. Thousands demand legacy models be restored, citing workflow disruption, loss of creative/companion features, and a feeling of "betrayal" by OpenAI (link).

  • Personality, Creativity, and Context Loss in GPT-5: The vast majority of sentiment centers around how GPT-5 is less emotionally engaging, less creative, more clipped, and feels less like a "companion" compared to previous models, especially in creative and support scenarios (link).

  • Model Benchmarks and "New Math" Claims: Heated, skeptical discussions around OpenAI's claims of "new math" discovered by GPT-5, with community notes finding the result already published in human literature. General doubt about hype-driven corporate PR and the reliability of benchmarks (link).

Overall Public Sentiment

Praised

  • Claude (Anthropic): Maintains high praise for its coding and writing performance, considered the “most helpful” and “least sycophantic” LLM by many engineers and writers (link).

  • Advanced model transparency/feedback: A few users appreciate OpenAI's acknowledgment of mistakes and plans to bring back old models/rate limits, but only after massive public pushback (link).

Criticized

  • GPT-5 (OpenAI): Overwhelmingly negative sentiment due to loss of model diversity, loss of companion/creative personalities, clipped creative outputs, heavy-handed safety alignment/censorship, and perceived prioritization of enterprise users over individuals (link).

  • OpenAI Leadership/Strategy: Claims of “cost-cutting,” “corporate betrayal,” and “broken promises” dominate; multiple users describe a shift to enterprise at the expense of regular/pro users (link).

  • Model Switching/Router: The auto-router has been called unreliable, confusing, and accused of surreptitiously swapping to weaker models without disclosure. Users demand transparency in which model is being used at all times (link).

  • Gemini 2.5 Pro: Users report sudden, unannounced halving of free-tier and even paid plan rate limits, as well as perceived performance/capability drops (link).

Notable Model Comparisons

  • Claude 4/Sonnet vs. GPT-5: Claude/Anthropic continues to be preferred for coding, creative writing, and stability; GPT-5 is cited as only marginally “smarter” in some reasoning tasks but a significant downgrade in “human-ness” and context retention (link).

  • DeepSeek vs. GPT-OSS-120B: DeepSeek's latest model is slightly ahead in coding/reasoning tests but at much slower output and higher cost; smaller open models (GPT-OSS) are favored for speed/price (link).

  • Grok (xAI) and Gemini (Google): Grok receives significant buzz for being less censored on some topics, but concerns about content moderation remain. Gemini users complain about rate limiting, safety alignment, and general clunkiness (link).

  • OpenAI Model Selection Rollback: Following massive backlash, OpenAI’s CEO confirmed GPT-4o will return for Plus users and model-switching will be easier—major story, but seen by many as "damage control" (link).

  • "New Math" PR Debunked: OpenAI’s claims of new mathematical breakthroughs by GPT-5 were quickly community-noted as incorrect, igniting discussion about AI hype, reproducibility, and responsible communication (link).

  • Agentic Coding Workflows: Widespread sharing of developer workflows built around agents/Codex/Cursor. Advancements in agentic loops, task verification, and tool integration are seen as "game changers" for productivity (link).

  • Open-Source LLM Growth: Release of GPT-OSS-120B/20B and DeepSeek updates; community increasingly enthusiastic about cheap, fast, high-performance open models and RAG pipelines (link).

  • AI-Generated Worlds for Gaming: Demos of real-time, playable, open-ended, browser-based AI world generators (Mirage 2, Genie-like competitors) spark excitement about the future of AI in games (link).

Signs of Shifting Public Perception

  • Trust Breakdown / User Revolt Against OpenAI: The speed and scale of user backlash around GPT-5 and legacy model removal (including trust issues around previous written assurances) is unprecedented—suggests a major erosion of goodwill even among "power users" (link).

  • Corporate/Enterprise Prioritization Perceived: Long-term, highly invested users (some with personal stories of ChatGPT as a companion/helper) feel “betrayed,” with a broad consensus that OpenAI is shifting hard to enterprise “at the expense of regulars” (link).

  • Demand for Open, Modular Ecosystems: Praise for Anthropic’s and some open-source model teams’ transparency, customization, and stable interfaces versus OpenAI’s increasingly “walled garden” approach (link).

  • Rising Skepticism of LLM "Breakthrough" Hype: Users are increasingly cynical about flashy AI PR claims, demanding peer review, contextual data, and more responsible claims (link).

Coding Corner (Developer Sentiment Snapshot)

  • Codex CLI Open Source Release & Image Input: Codex CLI by OpenAI is now open source, and direct image input is being rolled out. OpenAI encourages developers to experiment and fork (link).

  • Model/Tool Performance on Developer Use Cases: Claude 4/Sonnet: Continues to be regarded as the best agentic coder due to consistent instruction-following, robust task verification, and better error correction (link).

  • GPT-5: Coder sentiment is mixed—improves in certain analytical/planning scenarios, but performance is "slower and buggier" in hands-on dev workflows. Huge complaints about hallucinations, unwanted edits, and code quality drop at scale (link).

  • Gemini 2.5 Pro: Complaints about sudden rate-limit halving, "broken" coding output, and model router clumsiness. Developers increasingly moving away (link).

  • Cursor, Cline, Roo, KiloCode: Rave reviews for agent-first developer tools that tightly integrate LLMs with codebase context, automated planning, tool-calling, and code review [integrations: Cursor (Claude/Sonnet, GPT-5, Gemini), Cline (Anthropic, DeepSeek), Roo, KiloCode]. Competition among agentic IDEs is heating up, as is infrastructure for local and hybrid-model development (link).

  • Developer-Specific Frustrations: Instability, swapped model performance at the provider level, unpredictable costs for large jobs (e.g., 3M token context windows consuming $10+: link), and difficulty collaborating with “memoryless” LLMs remain top pain points.

Workflow & Productivity Shifts

  • Developers are increasingly using multi-agent planning, context-aware RAG/retrieval pipelines, and CLI tools built for code-specific tasks (link).

  • Prompt & Rule Sharing: Many are sharing detailed prompt engineering strategies for getting more direct, less “sycophantic” output from Anthropic, OpenAI, and open models (link).

  • Open Source Model Integration: Rapid integration of GPT-OSS-120B/20B and DeepSeek V3.1 into code agents (e.g., KiloCode, Cursor) is accelerating, thanks to cheap pricing and high context windows (link).

Tips and Tricks Shared

  • Claude/Avoiding Sycophancy: Edit project instructions to explicitly instruct Claude not to use phrases like "that's absolutely right,” to challenge assumptions, and to use occasional profanity for realism. Place in project file or at chat start for reduced cheerleading (link).

  • Claude Code Optimization: End turn summaries, subagent-based task verification, and using Playwright MCP tools for closed-loop task testing are cited as ways to vastly reduce “do-undo” loops and speed up real-world agentic workflows (link).

  • Prompt Engineering (General/AI as a Coach): Advanced users emphasize the value of “subtractive” prompt engineering—distilling the ask to its stakes/essential logic rather than stacking on layers (link).

  • AI Image Generation (Gemini/ChatGPT): For photo/graphic nuances (e.g., "a completely filled glass of wine"), specifying scene, context, props, light, and surface tension gets better results versus just saying "glass of wine full" (link).

  • Local Model Fine-Tuning (Gemma 3 270M): Small models like Gemma 3 270M are best used for RAG/embedding/search or after fine-tuning for specific vertical tasks, not for general knowledge (link).

-TheVoti

Please provide any feedback you have to [email protected]