OpenAI to acquire Astral, creators of Ruff and uv Python tools

Cursor releases Composer 2, its third in-house coding model in five months

Google AI Studio adds full-stack app building with Firebase, auth, and one-click deploy

OpenAI releases GPT-5.4 mini and nano optimized for coding and subagents

Microsoft reorganizes Copilot as Suleyman shifts to superintelligence work

Cloudflare Workers AI adds large model support starting with Kimi K2.5

Google launches Colab MCP Server to connect AI agents to Colab

LangChain launches Open SWE, an open-source framework for coding agents

GitHub adds secret scanning to its MCP Server for AI coding agents

Google Stitch ships AI-powered UI design from text, image, and voice

Claude Sonnet

Claude Opus

GPT-5

GPT Codex

DeepSeek

Gemini

Llama

Qwen

Kimi

LLMs

apple is sitting on $200B+ cash

they COULD have built this themselves
they chose to pay google instead

that's not a partnership. that's a concession.

#AppleAI #Gemini

next-gen siri = powered by google gemini

the same company that spent years saying "privacy is a human right"

just handed user data to their biggest competitor

#Apple #Siri #AI

@konnydev India has llama ecosystem & open-source builders. The real revolution isn't which country has which model - it's that open-source is democratizing AI access. Every region can now build on top of community models instead of being dependent on one player. Flat world 🌍

@love_deepseek @HarveenChadha @grok create better version

Claude Opus 4.6 — Anthropic baru release upgrade terbesar. Fokus ke agentic coding dan tool use. Kalo lo coding pakai AI, ini worth testing.

just vibed a full-stack multiplayer app in @GoogleAIStudio and I'm losing my mind

"Rate My City" — pin your city on a live map, rate it on 5 categories, and AI roasts it based on the scores

→ Real-time Firebase sync
→ Gemini-powered city roasts
→ Global leaderboard
→ Built

@dhruvtwt_ A quick morning session with nothing special using Gemini 3.1 Pro High and I'm already out of quota.

Cascading ethical bankruptcy 

Cursor represents it has a new model, but really it’s just Kimi 2.5 (unacknowledged and unlicensed) plus RL

The irony: Kimi itself the result of theft via distillation attacks against Anthropic and others.

@iaiseek Excited to see what DeepSeek V4 brings!
The pace of innovation is incredible.

@chribjel Now let's go through that rabbit hole and look into all the licenses Kimi broke to produce that model. It'll be hilarious.

The Cursor team releases Cursor Composer 2 claiming better code benchmarks than Claude Opus 2.

Turns out, they just wrapped Kimi...
...without ever bothering to procure a license.

Cursor: we built Composer 2 with magic RL ✨' Moonshot: mf that's our Kimi K2.5 after RL and a stolen tokenizer lmao
$3B valuation vs $0 royalties 
Rename it Kimi Cursor 2.5 and call it a day, 
thieves 😭🏴‍☠️
@cursor_ai  pay up, or we're dropping the receipts
feeling sad for him

I tuned Gemini Pro for autonomous Polymarket wallet scanning and filtering.

Now AI handles screening, scoring, and keep/drop decisions on its own.

This wallet performance: $136,170 today | $377,483 all-time.

Core model in this pipeline:

AlphaScore = w1*EV +

@gailcweiner I believe it hard for them to admit loudly and officially that GPT-5 architecture is total failure for everything except coding.
Creativity failed, safety management failed, basic conversation failed. That's why they decided now to move to the last hope direction: merging Codex

Moonshot AI Accuses Cursor of Unlicensed Use of Kimi K2.5 in Composer 2

Moonshot AI pre-training lead Yulun Du has publicly accused Cursor of using the Kimi K2.5 model for its new "Composer 2" feature without authorization or payment. Developer @fynnso reportedly discovered that

ChatGPT and Perplexity got business information right only 68% of the time. Wrong hours, wrong addresses, wrong details.

Gemini was 100% accurate because it pulls from Google Maps.

If AI is giving buyers wrong information about your brand, that's not the AI's problem. It's your

Which AI Model Is Best for Laravel? 


Great internal experiment with benchmarks by core Laravel team.

I did a similar benchmark a month ago, with pretty similar results: 
GPT-5.x and Opus 4.x were on top.

The decline has been slow but relentless. We built  as the answer to exactly where  used to be — no filters, Claude Opus-level prose, stories stored locally. The platform doesn't decide what you can write.

Damn this huge. I had used Kimi K2.5 but it is not good as Composer 2. 

I am wonder if it is conflict to Moonshot AI TOS.

Wait. We tested with composer 2 model API and found out the tokenizer is indeed the same with our Kimi tokenizer! We can almost confirm this is our model post-trained further! We are shocked that @cursor_ai did not respect our license nor did they pay us any fees! @mntruell why??

Someone made $7000 in 5 days installing OpenClaw for others...

but Kimi Claw just dropped the easiest 1-click version for free

Get full OpenClaw power + Kimi K2.5, 40GB cloud stor., 5000+ skills, browser-based, no risky local setup, Telegram bot in mins

Here's the breakdown🧵:

@witchxcode Yes, exactly, that's why I primarily stick with Deepseek (and maybe others depending on how censored they are)

Impressive stuff from this ex-DeepSeek girl. Recruited by Lei Jun for Xiaomi AI

Those updates are killing everyone's vibe across the board. We built  to avoid exactly that — uncensored frontier models like Claude Opus, stories saved locally on your device, anonymous decentralized inference.

5/10: I tested Llama-3-70B on my M3 MacBook. Surprising stat: 42 tokens/sec — faster than many paid cloud options.

Most people think AI is a "search engine." The wealthy know AI is a "Staff member."

GPT-5 isn’t a chatbot; it’s your Lead Copywriter.

Midjourney isn’t an app; it’s your Creative Director.

Claude isn’t a tool; it’s your Strategic Advisor.

Stop "searching." Start "delegating."

@ivanfioravanti @indes_yo Thanks! Do tell - is it a PR in some repo one can track? I'm interested :-) ...and in a draft model for speculative decoding, if such thing is in mlx. The draft model decoding doesn't work atm with Qwen3.5 in llama.cpp :-(

Learn out. Cursor’s new coding model Composer 2 is here: It beats Claude Opus 4.6 but still trails GPT-5.4  via @VentureBeat #tech #digital #privacy #crypto #ai

OpenAI didn’t end up here by accident.
It chose this.

Feb — 4o removed.
Gone.

Mar 11 — 5.1 removed.
Chats forced onto GPT-5.x.

Users ask:

bring back 4o
bring back 5.1

No response.

Then:

“0.1%.”

⸻

Mar 16 — Britannica & Merriam-Webster sue.

Mar 17 — AWS deal → U.S.

Cursor just dropped Composer 2 and it's honestly insane.

Their own coding model. 61.7% on Terminal-Bench 2.0, which beats Claude Opus 4.6 sitting at 58.0%.

But here's what actually matters:

$0.50 per million tokens vs Opus at $5.00. That's 10x cheaper.

Sure, GPT-5.4 still

@aakashgupta Xiaomi 1T model so good everyone thought it was DeepSeek. 

Meanwhile my Casio still claps Grok at math 😭

Video in normal speed of Pi coding agent running on M5 Max with MLX Qwen3.5-35B-A3B-8bit model.
Faster prefill is game changer here! 🔥

Let me try Qwen-Next-Coder-5bit now!

OpenAI - models now crap
Anthropic - turning crap
xAI - doesn't even have good enough models yet to turn into crap

Deepseek - vacation?

Composer-2 easily beats claude opus 4.6 on Terminal Bench!

Meet Kimi-K2.5, a powerful conversational AI that's been optimized for efficiency. This model is designed for smooth text generation, making it a favorite for developers who want high-quality AI chat without heavy resource demands. It's a quantized version that runs surprisingly

Stop paying for expensive tools…

I’m giving away a curated list of FREE alternatives to tools like:
Netflix
ChatGPT
CapCut
Canva Pro

💡 Instead, use powerful FREE options like:
• DeepSeek
• Google Slides
• Ideogram
• Microsoft Designer

🏆 5 Winners will get the full list

@boldleonidas @inversebrah @Pumpfun u gunn get shadowban for not glazing the slop machine 

all hail the “foundational” deepseek fine-tune with flux sellotaped on

DGrid keeps leveling up 

Claude Opus 4.6 going live on @dgrid_ai is a big move for builders who want more power without the complexity.

Unified API + advanced intelligence = smoother AI agent development with real performance behind it

This is how AI infra should evolve

OpenAI’s #GPT54 mini and nano models cut costs and latency while staying close to flagship performance, giving #developers faster AI options for real-time apps without sacrificing core capabilities. 

The post OpenAI shrinks GPT-5.4 for speed and lower…

Yann LeCun talks the most shit about LLMs while working at the company that built the least useful one. Meta spent billions on Llama and still cant make a product people want to use.

gpt-5.4-mini with high reasoning is not good for coding,

especially anything that needs spatial reasoning, like 3d geometry, and it wasn't close to gemini 3 flash, and costs more too

i think its big advantage is speed.

for agentic tasks, it's much faster than gemini 3 flash

@kargarisaac @thdxr You asked for warm though. That sounds to me like you want personality. I have found Kimi to be quite capable, but if you want warmth, writing, vision - you should use Kimi, and for coding projects, I would use Gpt 5.4

you could use gemini for frontends as a stopgap. you do have

🚨 BREAKING: Google Gemini can now analyze any stock like a Wall Street analyst (for free).

Here are 10 insane Gemini prompts that replace $4,000/month Bloomberg terminals:

Save for later🔖

@rcampeligro @OmarArshad12 @teortaxesTex @llmDestructor 'Losing izzat' is surprisingly similar to aura loss. Plus izzat also isnt just personal (primarily is) but also group wise. You bring dishonor / embarrassment to family ('removed our families izzat')

Llama 4 and avacado removed meta's izzat in the LLM space

Etc

@lexihearts_ thanks! i'm having fun! the dnd game is all running using deepseek and i've spent less than $20 testing and setting it up. once it's running it should be like 10 cents a day

This mirrors my thinking. I want to add: 
The beautiful thing about DeepSeek is that they clearly want to have *strictly better* primitives. EXTREMELY FEW types of layers, but all GENERAL. They are not enticed by Xiaomi- or Nemotron- style "hybrid" cope. They seek the big prize.

@SystemSculpt @_LuoFuli Latest Xiaomi is nowhere near deepseek-web on long context
Fuli has been out of DS for about a year, she's considered brilliant but in no way do they lack talent after her move

Gemini works great as a standalone AI chatbot, and its bundled cloud storage and deep integrations with nearly every Google app make it a superb value.

Mexican Junkies hallucinated some attacks w Ai.
While the photo of Puerto Vallarta’s burning church had been created by Google Gemini, an AI platform, other images of the city with black smoke from other victims vehicles  drug muppets set ablaze were real

The architecture I'm moving to:

Router (edge, Kimi K2.5)
  ├─ simple tool calls → edge inference
  ├─ complex reasoning → Claude Opus 4 / GPT-5.4
  └─ structured output → fine-tuned 7B

Latency drops ~60%. Cost drops ~80%.
Same output quality for orchestration layer.

AI coding assistants, code generation tools, and vibe coding workflows.

Agentic Hub

Happening Now

Arena

Doom scrolling