AI & Machine Learning Pain Points

Data Overview

1. 1

   Platform instability and unfair banning

   Platform Reliability×7

   “Banned by OpenAI after reporting a live credential hijack. They admitted in writing my account was broken. Here are 7 months of forensic receipts and 20+ cases.”

2. 2

   Unsustainable API infrastructure costs

   Pricing & Cost×7

   “infra bills are the #1 silent killer of AI startups right now... Sam Altman makes 'mic drop' offer to every Y Combinator startup”

3. 3

   New model performance regression

   Model Reliability×5

   “On most of my tasks, Gemini 3.5 Flash underperformed older Gemini variants. In the screenshot below, this is a vision emotion-detection eval with 5 runs per model: In, this eval it ended way down at 13th place”

4. 4

   Aggressive model pricing hikes

   Pricing & Cost×4

   “Google just dropped Gemini 3.5 Flash and the price hike is pretty insane... It's basically 5.5 times more expensive to run than the older 3.0 Flash model.”

5. 5

   Context preservation in long threads

   Model Reliability×4

   “after very long threads, responses start getting weirdly unfocused, repetitive, or attached to things discussed 20 - 30 messages ago. What surprised me is that restarting a fresh chat with a compressed summary often gives much better result…”

6. 6

   Loss of specific model capabilities due to sunsetting

   Model Reliability×4

   “Google's Imagen 4 and Imagen 4 Ultra are being sunset on June 30 but are essentially the only models out there that can reliably output a convincing 1990s "Disney renaissance" look”

7. 7

   Restrictive usage windows for high-tier models

   Pricing & Cost×3

   “the 5hr usage window of gpt-5.5 effectively makes it useless. This was not always the case. In fact in general I've used codex less because there's been noticeably less usage.”

8. 8

   Lack of diversity in video model aesthetics

   Model Variety×3

   “Most current video models are completely focused on realism. The few that try to handle anime usually end up producing results that look like a weird mix of 3D and realism instead of something that actually feels 2D.”

9. 9

   Agentic 'loop' and credit burn risk

   Reliability Issues×3

   “Woke up to a $500 API bill last week because a "thinking" agent got stuck in a loop overnight. Health checks were green the whole time... re-read the same docs page about 2,000 times and called it progress.”

10. 10

    Local coding model performance gaps

    Performance Gaps×3

    “as much as I've tried, I have not found any good open source free tools that can do programming tasks like Claude does (as well as claude does)”

11. 11

    Inefficient token consumption in SLMs

    Pricing & Cost×3

    “I tested llama-70b vs llama-8b for an AI agent — the "cheaper" model used 7.4x more tokens. The 8b model wasn't confident enough, so it searched knowledge base twice”

12. 12

    Hardware training compatibility issues

    Hardware / Infrastructure×3

    “getting LoRA training working on an AMD RX 9060 XT (Navi 44, RDNA4) on native Kubuntu 24.04.4. It covers everything tried, what failed and why”

13. 13

    Generalist model inaccuracy vs specialists

    Performance Gaps×3

    “accuracy of the model in some areas. I feel like it's a bit too generalist? I feel like illustrious could do a lot more in terms of following the prompt in some way.”

14. 14

    Persistent state and memory loss across sessions

    Productivity Issues×2

    “The AI forgets. Every new chat session, I have to re-explain the project architecture, our specific coding conventions, why we chose library X over Y, and the bugs we've already fixed.”

15. 15

    Hiring marketplace saturation for AI Engineers

    Career / Market×2

    “staggering to get shortlisted despite 60+ applications over the past month I’ve been building projects independently to fill skill gaps but I don’t have anyone to give me an honest perspective”

16. 16

    Product launch exhaustion and lack of utility

    Market Sentiment×2

    “Scroll through any AI/startup/building forum for five minutes and it’s everywhere: “I built this...” But the question I keep having is: Have you thought about how you’re going to get users?”

17. 17

    Occlusion failure in face-swapping

    UX / Design×2

    “problem most face swap tools quietly fail at: occlusion and preserving the natural jewelry or hair around the face”

18. 18

    Messy table data extraction in non-vision models

    Data Processing×2

    “OCR for files that contain tables, and I want to extract the actual table data... every file has a different table layout/order, so the output gets messy”

19. 19

    Security vulnerabilities in prompt guardrails

    Integrations & Safety×2

    “threat of indirect prompt injection is starting to give me gray hairs... looking at it practically, it feels a bit like a game of whack-a-mole.”

20. 20

    Safety layers as censorship/social engineering

    Ethics & UX×1

    “AI safety layers treat us all like "Apples"—and it’s damaging... any behavior that deviates from the "Apple standard" is a sin, a problem, or a psychosis that needs to be smoothed over.”