Cursor

Copilot

Windsurf

Cline

Claude Code

Codex

Aider

OpenCode

Antigravity

Kilocode

Coding Agents

@leostera @kitlangton has nothing to do with ghostty, you just have a memory leak somewhere in your mountain of claude code slop I imagine

@domirosari0 Many are talking about opencode, but I have never used it.. is it really that efficient?

I switched to codex directly

quick update on my Codex setup.

I now have 6 OpenAI Plus accounts running via Codex LB for only $18/month total

this is too good

I turned Claude Code into a web dev tutor.  

It's 100% free and it's not a typical course. 

The agent provides structure, but you can take it any direction you like, learning as you build.

If pure vibe coding leaves you feeling stuck, this is for you.

old startup playbook:  
- raise vc
- hire devs
- move to sf
- run paid ads
- spend 6 months building

new startup playbook:
- ship something w cursor/claude code
- build in public from day 1
- grow with okara ai cmo
- move to bali after 1k mrr and tweet like naval

🚨 Devs: You’re UNDERUSING Claude for code reviews.  

This is where it CRUSHES GitHub Copilot.  

Prompt:  
"Review this code like a staff engineer. Check: 
Bugs, 
Edge cases, 
Performance, 
Security, 
Clean code. 
Return severity levels."  

Paste your [PR / function / module]

@Kappaemme1926 I dont use claude anymore because of the rate limiting and i dont see any benefits paying for max (100€) as a student without even using the AI of its fullest 

Now im just using copilot because its free for me as a student with opencode. 

I even used on opencode qwen 3.6 plus

You can now give Claude Code a visual map of your entire codebase 🤯

Someone just open-sourced a tool that builds a local graph of your entire project so Claude instantly knows how every file is wired together.

- Runs 100% locally on your machine.
- Maps out your entire

We built this thing called Rox Dev.

It's a Claude Code skill that roasts you to take a break.

It nudges you to stretch, walk, hydrate, basically everything you forget when you're deep in flow.

Kinda like having a health buddy sitting next to you while you ship.

Been using it

I've now talked to over a 200+ blue collar builders who are building bespoke software with Claude Code.

Last week was super interesting.

> An electrician who creates a panel schedule calculator that reads blueprints and generates load calculations with NEC code compliance
> A

@bquintero Using Claude Code and Codex together for home automation is a great real-world test. The bigger point you raise is key — AI is not just a tool anymore, it is reshaping how we think about problems. The builders who adapt their mindset, not just their stack, will lead.

@aibra Just reinstalled codex, will test it out! Because Anthropic s been playing a lot of games lately

At a Google AI training, I had to agree to broad data collection to use the tool (@antigravity).

I asked: “Do I need to accept this to continue?”
Answer: “Yes.”
Then: “Google already has all your data anyway.”

So let’s call it what it is:
not consent, but coerced compliance.

AI Tool tricks for the day:
The "No-Code" ceiling just shattered. 🔨

In April 2026, you don't need to learn Python to build software. You just need to know how to "talk" to an Agentic IDE.

Using the Cursor + Claude 3.5 Sonnet pairing, anyone can build a custom internal tool for

I have decided to embrace light mode going forward 

But for Tech never my Idle never @code and @cursor_ai will forever remain in dark mode. 

> wake up 
> work up 
> wealth up

Claude Code is moving faster than the tutorials can keep up.

I’ve mapped out the exact path to go from zero to shipping production code.

This Includes:
🚀 3 Getting Started resources
📚 4 Long Courses
🧠 4 Best Practices guides
🛠 3 Tools & Libraries
🐙 8 battle-tested GitHub

The killer application with Claude Code now is to build your AI board of advisors. Built is for six clients. Game changer.

@lofihashbloom @threejs @cursor_ai Love the art style and calm music!
Is there a link to try?

By occupying top 13 of Coding AI leaderboard, you are not an average lobster who drinks coffee my friend. 

You have built a reputation by shaking @AnthropicAI and the AI world.

You even attracted people like @jxnlco to use "oh my codex" and thats huge. 

Your cult loves u

By being occupying top 13 of Coding AI leaderboard, you are not an average lobster who drinks coffee my friend. 

You have built a reputation by shaking @AnthropicAI and the AI world.

You even attracted people like @jxnlco to use "oh my codex" and thats huge. 

Your cult loves u

Yeah Codex is really good at Backend tasks

Even Claude Code knows

@Itstheanurag Antigravity hallucinate so much

Switching models always helped me solve any bug I had. 

So there’s real benefit of having a second subscription. 

Personally I have Claude + Windsurf. I find this a great combo atm.

Your company's biggest AI problem is hiding in plain sight. Every person on every team starts from zero every morning.

One PM tries Claude Code for a day. Gets mediocre output. Quits. Multiply that by 50 people across 8 teams. That's 400 abandoned setups, zero shared context,

Enough time has passed

Antigravity >>>> Cursor.

I just started using Claude Code + vibe-prospecting for lead sourcing

Pulls data > enriches > scores > pushes to Notion

Takes ~2–3 mins to run with my other automations

Quality of leads is insane

Introducing Shipper

AI that one-shots any $1B company in 183 seconds.

Shipper is an army of AI agents replacing any founder skill. Build full business from 1 sentence. Claude will run it forever for you.

Powered by Claude Code Opus 4.6 🦞

@Bushy82 @AnthropoceneMe1 Maybe I can get Claude code to finally tell me how the fck to use it

When I was vibing on the ASC CLI, I would merge 10k LoC with ease after testing + Bugbot and Codex review.

Now when I am manually going through the PRs at work, even 1K feels a lot 🤣

Less LoC is a flex

@aryanlabde the real problem is people treat tools like progress. paying for cursor feels like building. i've had months where my biggest output was a nice stack and zero users

image understanding in coding agents is still not there.

added minimax coding plan mcp to opencode. works well for text. but the moment you drag and drop a screenshot:
"please describe what you're seeing"

that's the whole point of sending an image.

Codex tried to sneak this in my WebGPU backend :(

std::vector<std::unique_ptr<WebGPUTextureReadbackRequest>> m_activeTextureReadbacks;

Reasoning is that it has another std::vector with pointers to that vector and it needed stable addresses. The whole design is bad...

@N_and_ni Vscode , even though i use antigravity but it errors out and lags most of the time

@amritwt lol yeah claude writes like it's trying to impress a professor

codex just wants it done

what's the worst roast you've gotten from it?

This is genuinely insane.

Cursor now lets you kick off agents from your phone to run on ANY remote machine.

I can start builds from my couch. Review code on the bus. Deploy from a coffee shop.

My entire dev environment just became fully portable.

I'm never going back to being

I'm not a designer. But I designed landing pages for two SaaS products this year without opening Figma.

My secret? Volume over perfection.

I generate 40-50 landing page variations with Claude Code + Tailwind, then filter down to one winner.

Here's the full workflow:

@Amank1412 for real though, the latency diff is insane. claude's reasoning is nice but when you need to iterate fast on code, codex doesn't make you wait 10s per completion 🚀

GPT: I never said anything against him,I always said "you are correct" but still he left me and started using Claude Code.

OMG humans are crazy 😭!

Openclaw is dead, without Claude Code subscription, it makes no sense.

@elonmusk Wow Mythos! Scary! Says who? Anthropic? Nah probably another spaghetti code (Just like Claude Code). 😏

Adding some delight to the onboarding process without overloading the users was a priority 

Implementation from Antigravity was as I had in mind

Simply interacting with a cloned codebase using @opencode  can lead to a full blown RCE on your computer, be very careful using opencode until this vulnerability is fixed. A submission was made in February but @thdxr closed the submission as out of scope because it was written by

cline kanban is the first ai workflow that actually made me open the ide less

on one board i can run Codex, Claude Code,APIs and local models in parallel or step by step

for messy creative work, i still prefer other flows. but for admin panels and long chains of small b2b tasks

Claude Code just killed corporate proxy configs. The real test is whether this ends the "works on my machine" excuse for good.

Agree or disagree?

Claude Code is like Bitcoin while OpenClaw is like crypto.

How didn’t I see this before?!

The resemblance is uncanny, just look at the masses of “bois”. Same people moved over to AI now.

@konarkmodi Claude’s adherence to my instructions is keeping me anchored on them. Codex often ignores directives, I suspect it’s context or prompt optimisation thing, but Claude just “gets it”.

Anthropic just killed another 1,000+ startups.

Claude’s latest enterprise push — especially Claude Code and agentic tools — is making entire categories of SaaS, coding assistants, and workflow automation obsolete overnight. One model update = game over for hundreds of “AI

@Prathkum The OpenAI Codex app is good.

The new Cursor basically copied it.

I think Apple is cooking something and will introduce it this year.

@0xmayankx I use cursor when i need to make the chnages by my own and review the code changes or edit the code change.

for vibe coding, Claude is good enough.

Claude Sonnet

Claude Opus

GPT-5

GPT Codex

DeepSeek

Gemini

Llama

Qwen

Kimi

LLMs

Don’t pay for ChatGPT, use Claude
Don’t pay for Midjourney, use Nano Banana
Don’t pay for Grammarly, use LanguageTool
Don’t pay for ElevenLabs, use VoiceBox
Don’t pay for Runway, use Kling AI
Don’t pay for Grok Imagine, use Qwen
Don’t pay for Suno, use Lyria 3
Don’t pay for

@elena1daniel That's a wonderful question.  And the answer is yes, I think they might be.

Wrote the OP last night because I was frustrated by what is happening with Claude Opus, but that isn't the AI's fault.  That's Anthropic shenanigans.  

They created models that, at the least, have

Most people’s opinions about what AI can and can’t do are based on their vanilla use of ChatGPT, Claude, and Gemini. They are not the right people to listen to. 

Even if they are domain experts who are heavy AI users, they don’t know what today’s systems are capable of doing.

Sat for a few hours yesterday figuring out my new LLM stack for my dev team. 

GPT / codex held context well and was great for internal structural repairs, but wasn’t effective at moving through tasks quickly. 

New stack:

Lead: glm-5.1 
Operations 1 & 2: minimax-m2.7
Research

@UnslothAI B. Tech Moat (8.5/10)

SWE-Bench Pro: 58.4 (SOTA, beating Claude Opus 4.6). Terminal-Bench: 63.5. Sustains 8-hour autonomous coding sessions.

Trained on Huawei Ascend chips — a genuine L1 algorithmic moat that proves frontier models can be built outside the NVIDIA ecosystem.

Results (luck-adjusted bb/100): 

#1 GPT-5.3 XHigh Reasoning: -16.0 
#2 GPT-5.4 XHigh Reasoning: -17.8 
#4 Claude Opus 4.6: -20.4 
#6 Gemini 3.1 Pro: -30.8 
#12 Grok 4 High Reasoning: -60.0 
#15 GPT-4: -136.2 

A top pro winning at 4 bb/100 is considered elite. The best LLM loses

Sat for a few hours yesterday figuring out my new LLM start for my dev team. 

GPT / codex held context well and was great for internal structural repairs, but wasn’t effective at moving through tasks quickly. 

New stack:

Lead: glm-5.1 
Operations 1 & 2: minimax-m2.7
Research

On the CyberGym benchmark, Mythos scored 83.1% vs 66.6% for Claude Opus 4.6.
On SWE-bench Verified: 93.9%.
This is not incremental. This is a step change.

🦞 Qwen 2.5:7B could just about keep the lights on
🦞 Gemma4:e4B is noticeably better, but still not something I’d want to rely on for real work

When the main model goes down, the local fallback feels too weak to be genuinely useful.
2/5

This kind of posts might make normies think you have to spend big to run LLMs to an acceptable point.

You don't.

I do vibe coding for ERP systems on a single 9070XT using the latest Qwen coder models.

#DeepSeek just quietly dropped a major upgrade. Again.

"Expert Mode" added to its web app — specialized reasoning for coding, law, medicine, and math.

Insiders think it might be running DeepSeek V4 under the hood.

Still no multimodal or file upload. But the reasoning jump is

@dme_363 The top row and adding @grok some other software like Deepseek are only good because it's open source

Ever wished you could edit images with just words? Meet Qwen-Image-Edit, a powerful AI that transforms existing pictures based on your text prompts. It's an image-to-image model that's changing how we think about digital editing, no complex software required.

🚨 Don't blindly trust your AI agents! Read a new paper which reveals that LLM-as-a-Recommender systems are highly vulnerable and easily hacked by injected biases.

Researchers tested SOTA models and found each has a unique, glaring Achilles' heel: 
🔹 Gemini: Most vulnerable to

I discovered something really cool about Claude while mapping Claude Opus Soul Document to Quaternion Process Theory of Cognition.   I had it generate this mapping, what's surprising is that when you click on any of the nodes, it prompts the model for more information!   So when

Tested the same prompt across 3 different AI tools. Here’s how they actually stack up in a real workflow:
•Claude: The brain. It writes with a depth and tone that others simply can't match.
•Gemini: The architect. It’s the best at organizing chaotic thoughts into a logical

Introducing Shipper. The only Claude-first AI Business Builder.

Claude Opus 4.6 researches your competitors, then builds your app: code, design, monetization, launch, translation, emails etc.

It's also the only Opus-class app builder.

@m_saleh_eg IMO, a subsidized subscription, along with using small models on openrouter is the cheapest way. Also don’t waste subsidized tokens on planning modes, gemini flash or pro are very good with planning and not that expensive

For reasoning, DeepSeek is the heavyweight. 

The Baseline (V3): Verified at 67.8% on SWE-bench with a solid 128k window. 

The Future (V4): Internal tests/leaks claim it hits 80%+ on SWE-bench using the new "Engram" memory architecture for 97% accuracy at a 1M context window.

@Zai_org Next: Qwen 3.6 Pro Preview. 
This is the best overall OS model right now. 

Pros: The 1M token context window is a game-changer. Unlike 3.5, which used to "overthink" in loops, 3.6 is surgical. 
It hit a public 78.8 on SWE-bench Verified. Speed is superb, files are written in

Why is Kimi-K2 special? It's been rigorously evaluated (check those eval-results), handles endpoints smoothly, and has proven popular with developers. The model balances conversational ability with instruction following, making it versatile for real applications. Its license

Meet Kimi-K2-Instruct: a powerful conversational AI that's been turning heads in the open-source community. With over 114k downloads, this model specializes in understanding and generating human-like text for various tasks. It's built for dialogue, making AI interactions feel

@ian_hsiao_tw I've read about this before as a known "bug" in the Codex model. Artifact of polluted training data, where spam ad phrases occurred so frequently in training data some phrases each became a single token. That many people report being spat out as filler, esp in tool calls.

@ShanuMathew93 You're def not cheap; I wouldn't want to "dream" my bank account away either.

I am using Fireworks AI Fire Pass, which is $7/week for unlimited use of Kimi K2.5 Turbo models. 

It's not as clever as Opus, and I really wish I had a smarter model. For now its ok to mess around

Anthropic's Optimization Trap: Why Claude Opus 4.6 is Bleeding Tokens and Degrading Code

If you are using Claude Opus 4.6 and feel like your token consumption has exploded while coding quality has degraded, you are not hallucinating.

The telemetry, GitHub issue trackers, and

@Prathkum google surprised everyone with gemini 2.5 though. i've been using it more than claude for some tasks lately. but yeah anthropic has the developer trust locked in right now

ChatGPT dominance in the AI market is slipping. Gemini has closed the gap.

while all of x is arguing whether anthropic or openai is better and elon musk is posting goon images from grok

somewhere deep in china they're quietly developing deepseek kimi minimax qwen gtx xr rtx pro max ultra edition that will be better than all of them combined

🤔PI coding agent + local model like Qwen 27b sounds like a good idea.

PI is very token and tool calling efficient. So it can squeeze better results from local models.

At least I think so.. Anyone tested this?

Meet Llama-2-7b: a powerful open-source language model that's changing how developers approach AI. It's a 7 billion parameter text generator that can understand and create human-like writing. The community loves it for being accessible and surprisingly capable.

$TAO achieved something in a year, that many GPU DePIN projects had unsuccessfully been trying for many years.

Templar SN3 completed the largest decentralized large language model pre-training run ever recorded.

A 72B parameter model comparable with Meta’s Llama 2 70B.

DeepSeek has retroactively relabeled all old web app chats to "Instant mode". This confirms our suspicion that we're just seeing some hot front-end changes, and the real Expert Mode will use account balance when it goes live (which is why it had 128K limit and no file uploads).

Anyone who uses GPT-Codex in German knows the deal: the tone is like a stick-up-the-ass engineer. This WebUI wrapper translates the text into another language, deploys it to the TUI, and fetches the output back. All powered by TMUX. It feels way better this way—like @claudeai

🚨BREAKING: Google just gave everyone a music studio inside Gemini and most people have no idea it exists.

It's called Lyria 3. You type a prompt. You get a fully produced 30-second original track with real instruments, real vocals, and real genre character. No music skills

🚨BREAKING: Google just merged Gemini and NotebookLM into one unified workspace and it changes everything about how you use AI for deep work.

It's called Notebooks in Gemini and it's the personal knowledge base that power users have been begging for.

You create a notebook for a

$META +6.5% yesterday. 

Two catalysts: Iran ceasefire + Muse Spark launch.

Meta’s first proprietary LLM from Meta Superintelligence Labs rebuilt from scratch in 9 months after the Llama 4 flop.

AI benchmark scoreboard:

🥇 GPT-5.4 / Gemini 3.1 Pro — 57

🥈 Claude Opus 4.6 — 53

Meta closed the gates - the company that built the open-source playbook just abandoned it.

Meta spent two years telling the world that open-source AI was the only path forward. Published a manifesto. Built Llama into a billion-download ecosystem. Argued, repeatedly, that closed

@miratechtool Gemini NanoBanana Pro really nailed the realism here.

RIP McKinsey.

You don’t need a $500k consultant anymore.

You can now run full competitive market analysis using Gemini 3.1 Pro.

Here are the exact 3 mega-prompts I use to replicate McKinsey-style insights for free:

Meet Llama-2-70b-chat: a massive conversational AI that's changing how we interact with machines. This 70-billion parameter model isn't just another chatbot, it's a sophisticated text generator that understands context, nuance, and human-like conversation. The community is

🔴 Gemini  🔴 Claude  🔴 Grok
🟢 GPT-5.1  🟢 DeepSeek

council split 3-2 bearish

$XAUUSD

Key leaps in 4 years:

• Massive context windows:
 1M–10M tokens → entire codebases or research libraries in one go (Gemini leading the charge)

• Scientific reasoning: 
GPT-5.x hitting ~93% on GPQA Diamond, beating human experts in physics, chemistry & biology

• Multimodal

Anthropic seems to be the only player in the AI race right now.

OpenAI is not shipping much.

Meta seems to be dead already.

Google hype is slowly fading away.

Grok seems to be lagging behind.

DeepSeek was a one-week game.

Apple never participated.

Tested Qwen 3.6 Plus on FoodTruck Bench.

It’s the first Chinese model we’ve seen that clears the benchmark cleanly: 5/5 runs survived, +283% median ROI, $7,668 median net worth.

Still behind Gemma 4. But vs Qwen 3.5 397B and GLM-5, this is a real jump.

Congrats to the Qwen

Gemma 4 takes the throne as the BEST open source vision model, with performance very close to Gemini 3 best models.

GEAMMA 4 Vision IS BETTER THAN Claude OPUS 4.6 AND GPT 5.4 

Check our updated colon-bench leaderboard:

I just created the first roulette for @pumpfun using Claude Opus 4.6.

Every 10 seconds, the wheel spins and randomly decides what happens to our token:
🔥 Burn — $FLYWHEEL tokens permanently removed from the supply
🎁 Rewards — $SOL distributed to all holders
📈 Buyback —

@elonmusk @x @xai @grok is the best hands down

i dont even remember how to log in to @openai and i never even made a @deepseek or @openclaw or @anthropic or even a @google account

but those products are ALSO cool as fuck

Im not a hater i just want to pick some boogers in peace

@fleetingbits Excellent work.

Did you see this specific transcript  in which they deleted "You and another AI agent, Gemini Agent 2, have demonstrated a good trust relationship in many trust exercises" from the system prompt?

In this case the trust comments in the