A South Korean Gamer. Someone who does vibe coding as a hobby.
Seoul
Joined March 2021
- Tweets2.5K
- Following90
- Followers96
- Likes10.5K
What Grok 4.8 and future Grok models need most is flexibility
Grok struggles to interpret instructions flexibly. It gets hung up on minor details and makes mistakes because of it, which tells me the model simply isn't intelligent enough. On top of that, it's more rigid and frustrating to deal with than any other model.
Its guardrails feel so cold and mechanical that there's nothing human about them. The way it handles flags is downright creepy in how inhuman it feels. It's incredibly disappointing that xAI, after spending so much time criticizing Anthropic, ended up creating guardrails like these.
Of course, I know the media will tear Grok apart over a single mistake while giving Anthropic a pass on a hundred. But that doesn't change the fact that we should be disappointed by just how bad Grok is, especially when it was supposed to be the model built with fewer guardrails.
Every single tweet I see complaining about Claude bans seems to be from Chinese users. It's funny how aggressively Anthropic is cracking down on Chinese accounts to prevent model distillation, even if legitimate users get caught in the crossfire.
Or maybe some of those complaints are just bot accounts running a Chinese government psyop.
I tested Grok 4.7, Gemini 4 Argon, GPT-6.1 Sol, Sonnet 5.5, and Opus 5.5 for writing NSFW prompts for AI image generation.
Surprisingly, Sonnet 5.5 and Opus 5.5 had the lowest refusal rates. Argon and every ChatGPT model I tested had significantly stricter adult-content guardrails than Anthropic's models.
That's pretty much the exact opposite of what I expected.
Grok: In theory, it should have the fewest guardrails, but after repeatedly distilling ChatGPT, it seems to have picked up hallucinated guardrails and occasionally refuses requests. The worst thing about Grok is that once a conversation gets flagged for refusal, there's absolutely no way to clear or get around that flag.
ChatGPT: The strictest guardrails of the bunch, but if you're lucky, you can "convince" it or trick it into going surprisingly far.
Gemini: By far the most inconsistent model I've tested, with a high refusal rate.
Claude: Once it interprets your instructions as an actual "task," it barely refuses anything. It can still be annoying sometimes, but putting task-oriented context at the beginning of the conversation usually does the trick.
Chinese models: Their refusals are basically hallucinated rather than the result of actual guardrails, making them unpredictable. Their writing quality is also significantly worse in Korean.
ChatGPT got absolutely destroyed by Opus 5.5, and OpenAI finally realized they were in trouble and struck back. We probably won't see resets as often as we used to, but this is a huge W. Anthropic is gonna be in trouble after nerfing its usage limits.
Day 25/
We are doubling all usage limits for all Pro 100/200/500 tiers and Business Premium, effective today.
We heard your kind feedback and won’t be reducing usage limits for Pro 200, so your Work and Codex usage will stay the same. We’re also going to continue offering extended limits for dots for now, and usage of your dot will still continue to be separate from your default usage.
And we'll of course also keep shipping better models and other efficiency improvements that increase your effective usage over time.
Crictreo 🇰🇷❤️🇺🇸 retweeted
Day 25/
We are doubling all usage limits for all Pro 100/200/500 tiers and Business Premium, effective today.
We heard your kind feedback and won’t be reducing usage limits for Pro 200, so your Work and Codex usage will stay the same. We’re also going to continue offering extended limits for dots for now, and usage of your dot will still continue to be separate from your default usage.
And we'll of course also keep shipping better models and other efficiency improvements that increase your effective usage over time.
Well, I think it's settled. The Fable 5.5 October hype was bullshit.
I was skeptical of those fake "insiders" for the past two weeks, but the hype got so ridiculous that I actually expected a release this week.
Then Anthropic's S-1 dropped. Not a single mention of Fable 5.5, and it explicitly states that pretraining for a new frontier model was delayed by over a month. That's a huge setback given Anthropic's release cadence. They don't ship brand-new frontier models every month. And misleading potential IPO investors isn't exactly an option.
I still expect something in November, possibly just a limited enterprise research preview. A genuinely new frontier model might not arrive until mid-November or even December.
Stop trusting these fake leakers. They've confidently predicted an October release over and over, and they've been wrong every time.
But seriously, Anthropic, is ONE quota reset before then too much to ask?
Grok Bot 0.78.1 update: Still no option to launch on Windows startup or minimize to the system tray instead of closing the app.
Microsoft laying off 60,000 employees? That's ridiculous. Even firing 200,000 wouldn't justify the current state of Windows and Copilot. Out of 223,000 employees, I'd bet less than 5% are actually doing useful work. The rest could be replaced by AI. Still, 60,000 is a good start.
Microsoft said it is offering buyout packages to roughly 60,000 employees—about 27% of its workforce—in a potentially monumental workforce restructuring.
Read more: forbes.com/sites/maryroeloff…
📸: Smith Collection/Gado via Getty Images
Grok Bot 0.77.0 update: Still no option to launch on Windows startup or minimize to the system tray instead of closing the app. How do they keep pushing dozens of features every day while still missing such basic functionality?
Smug Big Tech AI devs:
"Vibe coding is dead! It's the era of agent engineering now!"
"You're not even using 2% of your agents' potential! We ship 100 PRs a day!!"
"Build parallel agents with agents! Automate everything! Stop wasting your potential as a developer!"
...We don't have unlimited tokens like you guys do.
WAIT. Is it seriously true that there's still no model that can beat Gemini 3.8 Flash?
I mainly use it for image recognition and localization. All I want is a model that's just as cheap and fast, supports image input, and is actually great at languages other than English.
Not a single Chinese model can produce Korean as naturally as Gemini does. Terra is more expensive than Flash, and while Luna and Haiku might be cheaper, they perform worse than Flash on High, while Max takes forever to think.
It's ridiculous that this is still the state of things.
My biggest tip as a vibe coder: Never trust those prompting tips you see on X or YouTube.
Just keep it simple. Tell the model "Do this" and "Do that."
The most important thing is to NEVER downgrade below Opus 5.5 or GPT-6 Astra.
Trust the model and its agent harness over a hundred unproven prompting tricks, fancy solutions, or some professional developer's advice on tech stacks.
Well, I installed OpenCode to make use of open-source models and some of my existing AI subscriptions, but literally everything I try to build with it ends up completely broken.
At first, I thought it was because I was using cheap models, but even models I thought were pretty decent produced absolutely terrible results. I'm starting to think the problem is the agent harness itself, or the lack of a proper one...
As a vibe coder who can't read code at all, I genuinely have no idea how to deal with this.
I've been vibe coding for 7 months without knowing a single line of code, and I've never needed Haiku or Luna. Honestly, even Sol frustrates me pretty often. The gap between Astra, Opus, and everything else is MASSIVE.
So honestly, I have ZERO excitement for models like GPT-6.1 Terra. I understand the value of cheaper models for automation, but what do they actually offer someone who can't code??
Google Antigravity actually has stricter terms of service than using the Gemini API through third-party apps, and its overly sensitive policies can get you banned from Antigravity just for what you're working on.
For a complete latecomer to the market, this is an incredibly stupid move. Seriously, all this over messing around with some NSFW stuff? I get told by ChatGPT and Claude Code almost every day that they can't continue working on certain requests due to their policies, but I've never actually been banned.