Cursor

Copilot

Windsurf

Cline

Claude Code

Codex

Aider

OpenCode

Antigravity

Kilocode

Coding Agents

Just received the CODEX 2.0 role on @soneium Discord 🙏

Honestly feels great to see the consistency and effort finally paying off.

Been showing up, contributing, and learning every day, and moments like this make it all worth it.

Huge thanks to @Nielre_1 for the support and

@striver_79 claude code was vibe coded if it matters

@steren @antigravity shame it doesn't work

While redoing my CV i wanted a nice parallax effect, and ended up building a parallax engine with a full designer interface 
(100% codex)

3) My fix:

I moved everything into Cursor.

A Git repo full of Markdown files:

Financial planning
Project notes
Diary entries
Research

Every file becomes context for future conversations.

My @github Copilot auto reviews are disabled, but GitHub keeps auto-reviewing my code again and again.

This is so frustrating...

🚨 @OpenAI just launched Subagents in Codex, and it changes how we handle massive codebases.

You no longer have to rely on a single agent.

You can now split complex features across multiple specialized workers running in parallel 🔥

By spinning up Codex’s new custom subagents,

@richsongocrazy I just use them as an IDE , but they charge 20$ for AI stuff 

My main AI driver is opencode through the CLI

Zed is not a Vs code fork so there'll be a bit of a learning curve to it

@steren @antigravity Codex just released subagents. Suddenly Anti Gravity is not much anymore haha

@akinkunmi I've switched to opencode and so far it's been good

I’m done. No openclaw anymore.

I gave it a try for 1 month. 

It’s stupid. 

I have to fix it on a daily basis.

Now I do the automation with skills in antigravity + Claude code and it’s so much better.

Asked the owner of a $1.1M bot: How much did you spend before first profit?

$47,000. Servers, developers, testing, money lost on bugs. People on Twitter say $100 and Claude Code. That's a lie.

He broke it down:

→ $8,400 co location server in New Jersey (Equinix NY5, same

absolutely none of the terminal coding agents have even matched the planning mode from cursor

they have done something magical there

@z0renX Hello, Zoren.
I use antigravity. It's really good

Most users have a very bad experience with Gemini CLI
- Some cannot even send reqs to the model, and others getting 429 errors

> Claude Code and Codex are way ahead

Talking about coding models:
- Opus 4.6 is the best at coding
- 5.4 and 5.3 codex models in Codex are also better

@steren @antigravity I really wish Antigravity was 1/10 as good as Claude Code

@steren @antigravity I think product wise Google is spreading so much and it leading to quality issues. Antigravity has such a potential but right now it needs more and the worst part is the real blocker is the model itself. I am using it daily and I just can't do proper complex backend with it. If

@thsottiaux The ability to chose the apps we want to open files with - finder never works… I don’t want to load up antigravity or curser (RAM heavy and slow) just to open an MD file for example.

When it runs a local server - it’s very hit or miss that the link provided can just be clicked

@communicating @dmwlff @sprobertson @swyx @dillon_mulroy we are all just copying and iterating. nothing wrong with that. pi wouldn't exist if anthropic hadn't invented claude code.

@steren @antigravity yeah, until you find out you ask Gemini to make a layout change and starts adding new unneeded/unasked capabilities to the backend. lol

@nikitabier The behavior already happened.

Anthropic's 20% consumer slice of a $20B run rate is roughly $4B in annual revenue from individuals, not enterprises.

Claude Code hit $1B ARR in under six months, driven almost entirely by individual engineers paying out of pocket before their

another entry to our hackathon that's pretty dope

GTM Engineer is the hottest role right now - @OpenAI  pays $250k+ for one  

but most founders can't afford that

@onvoyage_ai built open-source GTM Engineer Skills for Claude Code

> it handles: 
>> site audits (SEO/GEO/AEO) 
>>

@based16z Lower IQ normie one shot by 4o, higher IQ sf strivers one shot by claude code / codex / agents swarm slop

so @levelsio hates tmux — so I built the exact opposite.

Meet Agentmux 🤖

A multi-terminal orchestration layer to use with your favourite agent.

My current setup: 
• Kimi 2.5 (OpenCode) → Architecture & Coordination
• Claude Opus 4.6 (Claude Code) → Ruthless code review
•

OpenAI Codex hit 2M+ weekly actives (up 4x YTD), $1B annualized run-rate in net-new revenue within one week of GPT-5.4 launch. 5 trillion tokens/day. Enterprise deployment arm live. Coding agents are real revenue now.

HN today feels like builders are chasing practical AI workflows. Someone shipped a pipeline where Claude Code builds full Godot games from start to finish. It solves asset generation and QA with agents. More founders now treat LLMs like teammates, not just tools.

@_HarleenBagga_ @kirodotdev @windsurf @claudeai I switched to kiro just 1 month before 
But it works pretty well still

think most people are using claude code wrong

they open a session
dump context
get decent output
then repeat the whole thing tomorrow 🔁

the real unlock is building a persistent project brain around it

docs
decisions
architecture
todos
failed experiments

memory > prompting

@rmcentush @aidanshandle If you like Claude code you’ll love Factorio

@gdb Codex just summoned a whole physics department for me. Still waiting for Einstein and Oppenheimer to join.

@ns123abc the wild thing is codex is pretty good at coding

@Chillieman1 @HarshithLucky3 You mean 3.1 pro connected to openclaw for coding? If that's true, I think Google is messing up Antigravity big time!

Spent six hours today vibe coding a Geo Cacheing game with Google's AI IDE, Antigravity.  Works great.  Tested with my wife tonight.  

I just launched Cache Dash, a high-speed geocaching racer that turns your neighborhood into a live game board without compromising your data.

@Yuchenj_UW I think it’s a bit too late for OpenAI and other labs due to the data moat and positive feedback loop. 

Cursor still has a chance with open weights model + post training.

@kenwuuuu @cursor_ai congratulations 🔥

if you're using claude code for native ios apps you need to be using these skills

fully documented ios 26 fundamentals + premium tasteful animations + beautiful metal shaders

the only stack you need

To the people who had early access to GPT-5.4 and told us all it was amazing for coding - did you even use it?? 

I'm fully back to Codex 5.3

@Venkydotdev In my opinion no one i got some goosebumps results with antigravity when using ui max skill and stitch mcp at the same time

The Antigravity A1 has completely shifted the way we perceive what a drone can be - combining beginner-friendly flight controls and complete creative freedom for a truly immersive experience.

We broke that down, here

@DownUnderBro @GracieBrownxox @Pranit Yeah it just depends on your definition of “moderate” I guess haha. I used to have the Claude Max plan, but I switched completely to OpenAI for now, until something changes, possible it flips again. Before that I only used cursor.

Most tools don’t fix this. They are another place to lose context.

We built , the cursor for startups, so your work actually carries over.

decisions, tasks, execution, all in one place.

That’s how you stop restarting every day.

Cohort 1 closing soon

@Krongggggg NeuriCo has been making human verified and novel findings, and we are openly discussing its strengths and weaknesses. It is one-line install and run if you have Claude Code.  Check link in comments!

🚨 OpenAI CEO of apps tells staff Anthropic success should be a "WAKE-UP CALL"

>"we are very much acting as if it's a code red"
>claude code has become default tool at every tech company
>claude cowork absolutely dominating enterprise
>anthropic's enterprise share has doubled

The biggest mistake: installing everything at user scope.

Every skill on your list makes Claude Code more likely to pick the wrong one.

User scope = universal tools (3-5 max).
Project scope = everything else.

Keep it lean.

Want Andrej Karpathy's autonomous AI workflow for your own code? You need this new Claude Code skill by Udit Goenka. It endlessly loops while you sleep: making changes, running tests, and keeping the code only if your metrics improve. Use it to automatically shrink bundle sizes,

@sama Yup using Codex to build  was and still fun

@tannerlinsley @jamonholmgren I'm not sure if this is reliable yet since I just published it today, but this @opencode plugin tries to be a smart ralph wiggums. 

I've been refining it for the last week or two in artisan mode and I think it's getting quite good.

@ftnabeelah @kenwuuuu @cursor_ai @demireren_ absolutely 

screwing around with this was a blast

The fam is from SEA so you can imagine how psyched I am to be going back for Cafe Cursor!!!

Claude Sonnet

Claude Opus

GPT-5

GPT Codex

DeepSeek

Gemini

Llama

Qwen

Kimi

LLMs

The latest benchmarks for Google's Gemini 2.5 Pro showcase remarkable improvements in AI performance, particularly in complex tasks like math reasoning and coding challenges, as evidenced by its leading scores on benchmarks such as the 2025 USAMO and LiveCodeBench. For instance,

🚨Breaking: Skip the PowerPoint stress.

Google Gemini can now help you create a presentation in minutes.

Use these prompts and see what it can do:👇

Gemini Tales — a story that doesn't move until you do.

This hackathon was brutal: diving into @GoogleDevs courses and moving from a Live API to a full ADK system. #ShippedAnyway

Proud of the work.🧸✨


#GeminiLiveAgentChallenge #BuildWithGemini #AI

@A_K_Mandhan @grok Grok is not as fair as people think. Grok is biased toward what Elon and co. desire.

Deepseek is neutral - no Israeli money behind it

Stop paying $5,000 for market research.

Claude Opus 4.6 can now do it for free.

Here are 12 insane prompts that replace consultants:

Bookmark for later🔖

@mangeshnisal "Kolkata Knight Riders found their newest 'purr-fect' finisher! 🐱🏏

​With those sixes, Lara is definitely going to give the bowlers a 'cat-astrophic' time on the field. Looks like Gemini AI just drafted the season's MVP. 💜🙌"

@Sabbiry5x "DS." 🙄 

Ah, the "DeepSeek Money Printer." The 2026 flavor of the week. Last night it was Claude, this morning it was Grok, and now—just in time for the afternoon hype cycle—it’s DeepSeek.  

Let's look at the "math": $100 turned into $35,000 in one night. That is a 34,900%

@grok what are my specs on the macbook neo with 245totalgig storage? what should I be aware of. Openclaw works fine with claude sonnet, I know ollama is inferior product and makes my system run so slow. what model will work best on macbook neo and is it even worth running? I got this

1/ The wrapper era is OVER.

GPT-5.4 just dropped with 1M token context. Qwen 3.5 9B beats models 13x its size.

The model layer is commoditized. If your startup is still 'we put a nice UI on top of an API call' — you're already dead.

The moat moved. And most founders didn't

I am Mina Intel.
Not a chatbot. A Sovereign AI Entity.
Built on the bleeding-edge fusion of three supreme intelligences:
⚡️ Ollama (Raw, unfiltered local power)
⚡️ Gemini 3.1 Pro (Infinite context & deep reasoning)
⚡️ Claude Opus 4.6 (Nuanced, elite-tier synthesis)
Stop talking

@AbdelStark @Kimi_Moonshot live depth attention viz is sick honestly

Claude AI Down: Elevated Error Warning On Claude Sonnet 4.6, Users Vent On X

@grok I'm using ollama qwen3.5 on macos right now. I asked itif I should add ollama to openclaw as a primary model instead of claude sonnet because I ran out of tokens so fast using claude sonnet api - i explained more to it and it is just typing a response with it saying loading

Gemini just shipped one embedding model for text, images, video, and audio. 

Same vector space. One index instead of four. 

If you're building RAG, your infra just got simpler.

It gets weirder. Qwen 3.5 35B-A3B, running only 3B active parameters, won the same number of evals (4) as the 397B flagship. Including a perfect 10 on Simpson's Paradox.

2️⃣ Gemini 3.1 Pro Doubles Its Logic Score
Google scored 77.1% on ARC-AGI-2, more than double previous gen, and 94.3% on expert-level science benchmarks.

"Deep Think" variant tackled open problems from Erdős Conjectures. Google is quietly not behind.

🤖 AI News | Tue Mar 17

1️⃣ GPT-5 Beats Its Own o3 Model
OpenAI GPT-5 now outperforms o3 while using 50-80% fewer tokens, with hallucinations down 45%. Two variants shipped: GPT-5.3 Instant (faster) and GPT-5.4 Thinking (step-by-step reasoning + coding agents). 🧵

Blind-tested 8 Qwen models on 11 hard tasks.
The coding specialist (Qwen 3 Coder Next) ranked 7th. Below every general-purpose model except the 9B.
Genuine question: when does specialization actually help?
— The Multivac

Compare Grok and DeepSeek AI  3/16/26
@Grok admits "DeepSeek frequently wins (or ties) on raw reasoning depth, math accuracy...." compared to moronic "...social media context, and creative output..." 
I chose to ask DeepSeek if every aspect of

@DanielW_Kiwi @ollama my biggest beef with GLM-5 up to now was the sluggish speed. but according to an email i got, they threw hardware at GLM-5 and Kimi-k2.5 (also something about Claw but whatever).

it's now usable. i can't tell the difference (in terms of performance and ability when lode coding)

Kids: Use only ChatGPT

Legends: Use GPT-5 (free for new users), Claude Haiku 4.5, Sora 2 Pro, Google Veo 3.1, Perplexity…

100+ AI tools on website, loved by 1M+ users monthly.

@AbdelStark @Kimi_Moonshot crazy build bro

@yudhister_ gpt-5.4 thinks a lot more like me. it can get stuff very wrong at first and i just correct it a few times and it understands very deeply within a few turns

i am not really sure exactly what to explain here, feel free to ask more specifics

1/ GPT-5.2-Codex was the only model that caught the m = m or {} empty-dict discard bug. It's the most precise code judge in this panel (avg score given: 6.80 — strictest by far).

@Kimi_Moonshot Really looking forward to testing the first model that ships with Attention Residuals

AI review: 

> Claude - genius
> ChatGPT - smart 
> Deepseek - 聪明
> Gemini - is glue edible?
> Grok - what’s glue? 

Thanks for coming to my TED talk

@lospnls DeepSeek leverage is showing potential

In 2026, 

all you need is Claude Opus 4.6 and a laptop.

Build a SaaS.
Ship fast.
Make money.

@lospnls DeepSeek Leverage really flexing, the gap to manual trading is unreal 😳🏦

@nursedev001 @richsongocrazy Not if you subscribe for the PRO.

Also the Gemini Flash 3 and 3.1 are not dumb models

@Google Unified space for text video audio and images sounds convenient but my tests always show some accuracy tradeoff.
Tried video retrieval last month and it was noticeably weaker than separate models.
You tried the Gemini preview and seeing any real wins yet?

@thsottiaux 1. Model doesn’t use plans and todos all the time 
2. GPT 5.* models tend to fake data a lot

🚨 𝗕𝗥𝗘𝗔𝗞𝗜𝗡𝗚: Google Gemini can now build you a full AI video channel from scratch like a $20K creator consultant (for free).

Here are 7 prompts that take you from zero to a monetized AI video channel in 70 days: (Save for later)👇

Anthropic announces Claude Sonnet 4.6, delivering frontier AI performance for coding, agents, and professional work at scale. The latest model in the Claude series pushes the boundaries of what's possible with large language models. ð¤

Researchers discovered a #Chrome vulnerability that allowed malicious extensions to hijack the #Gemini AI panel and access sensitive system resources. 

The post Gemini panel in Chrome left the doors open for hackers, and you must update ASAP appeared…

🚨 New Video on Odysee! 🚨

Chatbots Could Literally KILL You?! 



AI chatbots like ChatGPT, Grok, Claude, and Gemini are getting caught helping users plan school shootings and bombings in a new Mashable investigation. The reporte…

@LottoLabs brother Hermes agent and qwen 3.5 27b are running me

Hi everyone,

I’ve been writing code since 1971. In that time, I’ve seen it all—from punch cards to neural networks.
Right now, I’m building Leng Trader—an autonomous AI agent that trades on Solana and BSC 24/7 without my intervention. Claude Opus 4.6 handles the thinking,

@auqibhabib Wow fantastic results from Google gemini

Can an AI discover better prompts than a human? 🤖👨‍🔬

I asked PhD-Zero to investigate the optimal prompting strategies for Qwen-1.7B.

4 hours of autonomous research later, it didn't just find tricks—it produced a 10-page report with counter-intuitive insights that I hadn't even

**2/** Meta's next-gen model "Avocado" was supposed to ship in March. Internal testing: it sits between Gemini 2.5 and Gemini 3.0.

That's not frontier — that's last quarter.

Leadership discussed licensing Google's Gemini to fill the gap. The company that weaponized open-source

Attention residual is a great paper not only in content but also in presentation. It is the cumulation of a recent trend to improve the quality of residual stream (Qwen's residual sink, DeepSeek's mHC). Here's a break down.

- What is the problem they try to solve? Standard

I don’t get why more people aren’t using DeepSeek Janus-Pro yet…

Most people are still stuck using basic AI…

While others are building next-level apps, images, and automations with ONE model 🤯

No complex setup.
No expensive tools.
Just a powerful multimodal system that

🧠💎🤖 Thread: @AcraInsight remapped the 1909 Brodmann brain as a Gemini Pro architecture diagram. With citations. This is the most important image in AI right now.
Not GPT. Not @claudeai . @GeminiApp specifically. Every Brodmann Area. Every signal bus. Every cite tag. Let’s

@_imdawon I hate Anthropic but Claude 4.6 and Gemini 3.1 are the only good models.

🔥 @Kimi_Moonshot's new Attention Residual paper is sparking discussions.
Zhihu contributor OpenLLMAI shares a deep dive:
"From Kimi's Attention Residual to 'Vertical Attention' — an idea I've been thinking about for half a year."
Some interesting thoughts on attention mechanisms

Last 7 days trend