shubhankar raj

Applied AI Enginner @tap |last acc 16.2k followers insta @subh.tly(disabled)

First drafts: code & tweets. Final draft: the future.

[shubhankar raj @subhtly]

No cloud. No account. No API key. Draw Things just open-sourced Local Code for Mac — an AI assistant that runs on your own machine. Catch: the good models need real RAM. 16GB minimum, 64GB+ for the flagship DeepSeek model.

[shubhankar raj @subhtly]

1 trillion params. ₹0 to download. Mistral just open-weighted Large 4 (Oct 29) — 49B active per token, 1M context.
the catch: the FP8 weights are a 1.06 TB download, built for an 8-GPU server. free to download, not free to run.

[shubhankar raj @subhtly]

40 languages. 8 speakers. ₹0 to download.
NVIDIA's Nemotron 3.5 Transcribe (Oct 28) turns recorded audio into text, translates 31 languages, and labels who said what.
catch: recorded audio only. no live streams, no real-time captions.

[shubhankar raj @subhtly]

Google's best model just launched. You can't use it.

Gemini 4 Carbon (Oct 28) is #1 on the Intelligence Index, finds + patches real vulnerabilities. But it goes to a small group of cyber defenders first — no date for everyone else. $2/$10 per 1M tokens when it lands.

[shubhankar raj @subhtly]

GitHub just stopped making you pick a coding model.
Copilot Fusion (live Oct 28) routes the task itself — Standard, High or Max.
GitHub claims Max nears Opus 5.5 at half the cost.
Catch: no separate fee, but you pay per model it uses. 30% off for paid plans till Nov 30.

[shubhankar raj @subhtly]

your AI coding agent can delete files, read secrets, and install shady packages — before you even notice.

HOL Guard checks each action first: allow, ask, or block. free, local, no account needed.

catch: it only covers the agents it supports, so check yours first.

[shubhankar raj @subhtly]

Google just announced Genie 4 — a world you can walk through, not a video you watch.
type a prompt, steer it live, characters talk back, sound included.
catch: research preview for trusted testers only. rest of us waitlist.
closest thing yet to a playable AI world.

[shubhankar raj @subhtly]

15 seconds. that's how long MorBizAI's free audit takes to check if your site shows up in ChatGPT and Google AI answers.

it also scores SEO, speed and privacy. catch: the audit is just step one — the content pipeline behind it is the real product.

[shubhankar raj @subhtly]

Anthropic just shipped a setting that lets Claude Code delete files without asking.

Even the old --dangerously-skip-permissions flag still blocked rm -rf ~. The new env var removes that last guard.

Use it only in a throwaway container. Not on your main machine.

[shubhankar raj @subhtly]

No server. No API bill. Opera's WebAI puts a free AI model inside the browser — out Oct 27 as an early preview in Opera One, running Google's Gemma locally. catch: each site needs your permission, and speed depends on your hardware.

[shubhankar raj @subhtly]

No account, no signup — free AI chat. Swiss lab Giotto launched Alto today: free anonymous chat on European cloud, not US servers. catch: self-hosting the flagship takes 8 GPUs, so free really means their cloud.

[shubhankar raj @subhtly]

OpenAI's new model is half the price. And you still can't use it for free.

GPT-6.1 Terra launched Oct 26: $1/$5 per million tokens, 50% cheaper than Sol. But it's Plus/Pro/Enterprise/Edu only in Work and Codex — free plan not included.

Cheaper for devs, not for you.

[shubhankar raj @subhtly]

33B params. 512K context. ₹0 to download.
Nanbeige4.3-33B just dropped on Hugging Face under Apache 2.0, built for coding agents.
the catch: full weights need ~66GB, and its benchmark wins are self-reported. still, free to verify yourself.

[shubhankar raj @subhtly]

Canva built its own image model, and it's free on the free plan.
Craft in Dream Lab plans the layout before it draws, so text lands where you put it. Up to 2K, 30+ ratios.
Catch: Canva won't say how many images your free allowance covers.

[shubhankar raj @subhtly]

0 words of text. that's what Block's new voice AI Dessa was trained on.
human conversation audio only, no pretrained model. it listens and talks at the same time.
catch: you can't use it yet. waitlist only, researchers first.

[shubhankar raj @subhtly]

I unfold my Duo approximately 6 to 10 times a day. Mostly just leave it unfolded at my desk.

[shubhankar raj @subhtly]

NO CLIENTS? 6 MONTHS BUILDING, STILL $0 REVENUE.

[shubhankar raj @subhtly]

stop pasting your database schema into chat.

google's Data Agent Kit is free: MCP tools for 15+ data services + 37 skills, so your agent in claude code, codex, cursor or antigravity can see your tables, queries and job logs.

catch: the kit is free, the cloud usage isn't.

[shubhankar raj @subhtly]

JUST IN: Abliterated DeepSeek-V4.1-Flash @ $0.20 IN / $0.80 OUT, plus enhanced with cyber RL for exploit development. Now 88.1% on CyberGym. Match Daybreak performance with no approval required.

[shubhankar raj @subhtly]

10,000 free Neurons a day. that's Cloudflare Workers AI on the free plan.

catch: it's one shared pool across every model, and free accounts just stop when it runs out. pick the small model first or a busy day burns it fast.

then it's $0.011 per 1,000 Neurons on Paid.