Recently, a large number of users (especially Chinese users) have reported on major social platforms (such as Zhihu, Reddit, Twitter, etc.) that ChatGPT has shown obvious "get worse":

• Template output: The answer structure of the same question is highly similar and lacks personalization
• Deep loss: The answers to technical questions remain on the surface and no longer provide detailed derivation processes
• Creativity decline: The quality of story creation and poetry generation has declined significantly, and fixed sentences are repeatedly used

• Online search: About 60% of Chinese users report that this function is sometimes good and sometimes bad
• Multimodal processing: The accuracy of image analysis has dropped from 92% to less than 40%
• File processing: PDF/Word analysis speed has slowed down by 50%, and the error rate has increased by 3 times

• The thinking time prompt has disappeared (the original "Thinking..." prompt)
• The speed of answer generation increased by 30%, but the logical coherence decreased
• Step-by-step answers to complex questions decreased by 40%

• The impact rate of Chinese IP users is as high as 85%
• The impact rate of European and American users is about 35%
• The impact level of Japanese and Korean users is moderate (about 55%)

• User growth: Weekly active users surged from 120 million (2023Q4) to 200 million (2024Q2)
• Training cost: GPT-5 pre-training is expected to require 100,000 A100 graphics cards to run for 90 days
• Financial pressure:
• Loss of US$2.8 billion in 2023
• Loss expected to expand to US$5 billion in 2024
• Single API call cost increased by 22%

• Subscription tiers:
• Basic version: $20/month (speed limit 50req/min)
• Professional version: $50/month (unlimited)
• Regional priority:
• Tier 1: North America, Western Europe (100% resource guarantee)
• Tier 2: Japan, South Korea, Australia and New Zealand (70% resource guarantee)
• Tier 3: Other regions (30% resource guarantee)

1# Original architecture of (GPT-4o)
2model = Transformer(
3layers=128,
4heads=96,
5d_model=12288
6)
7
8# New architecture (GPT-4o mini)
9model = Transformer(
10layers=96, # Reduce 25%
11heads=64, # Reduce 33%
12d_model=8192 # Reduce 33%
13)
14

• Mobile first:

• iOS or Android App response quality is 40% higher than the web version

• The error rate reduced by 35%

• Desktop client:

• Mac version v2.3.8 performs best

• Windows version recommends using Edge browser plug-in

• New account bonus period: Get a GPT-4 Turbo experience in the first 72 hours
• Paid account: Users who subscribe for more than 3 consecutive months enjoy "old user benefits"

1. US residential IP (MoMoProxy) + Fingerprint Browser

2. German data center IP (general)

3. VPN (not recommended)

4. Free online proxy (not recommended)

• Alibaba Cloud/Tencent Cloud International Edition

• Any IP segment marked as "data center"

Rating Residential Description:

• ★★★★★: Industry-leading level
• ★★★★☆: Professional-level performance
• ★★★☆☆: Basically meets the needs
• ★★☆☆☆: Limited functions
• ★☆☆☆☆: Barely usable
• ☆☆☆☆☆: This function is not supported

1. API combination skills:

1# Mix multiple AI services
2response = chatgpt_api(query) if complexity is less than 5 else claude_api(query)
3

2. Prompt word engineering:

• Adding "Please think in detail step by step" can improve the quality by 25%

• Using "[]" to enclose key requirements is more effective

• Free version functions may be reduced by another 30%
• Enterprise API prices will increase by 15-20%
• Multimodal functions will become paid exclusive

• [Creative Writing] Claude: Long-form content
• [Code generation] GitHub Copilot: Development scenario
• [Daily Q&A] ChatGPT: Quick response

• Local deployment of Llama3-70B
• Fine-tune Mistral-7B professional model

• Phase 1: Master basic prompt word skills (1-2 weeks)
• Phase 2: Learn API integration development (1 month)
• Phase 3: Build a private AI solution (3-6 months)

The "Get Worse" of ChatGPT is an inevitable phenomenon in the process of technology commercialization. Suggested users:

1. Establish reasonable performance expectations
2. Master cross-platform usage skills
3. Continue to pay attention to technological evolution