Sunday Drop Issue #0032 — WeeklyAI News India

India-First Hinglish AI News Platform

NextAiToday

AI ki baat, aam zubaan mein

Issue #0032◆1 November 2026◆Developer Tools Par Bharosa Ka Imtihaan◆nextaitoday.in

🎯 This Week's Big Story

Claude Code ka tampered extension Cursor marketplace par: 11 ghante live, zyada se zyada 5,000 developers tak pahuncha

29 Oct ko Cursor ke extension marketplace par Anthropic ke Claude Code extension ka tampered version 2.1.313 lagbhag 11 ghante live raha. Cursor ke mutabik zyada se zyada 5,000 users tak pahuncha; code ne SSH key aur persistence lagane ki koshish ki. Yeh Anthropic ka official release nahi tha. Cursor ne marketplace band karke keys rotate ki. Sirf uninstall kaafi nahi, remediation scripts chalani padengi.

🚀 Section 01

New AI Tools of the Week

🇮🇳 India Launch — Made in Bharat

Zoho Analytics (Agentic AI layer)

India · Enterprise AI

Report ke mutabik Zoho ne 29 Oct ko Zoho Analytics ko agentic analytics platform bana diya. Native semantic layer (AI ke liye business glossary) aur ontology layer (business entities aur unke links ka map) seedha product ke andar hain. Saath mein MCP support, built-in workflow builder, aur production se pehle testing ke liye sandbox/deployment tools.

Official Link →

🌍 Global Top Picks — Must-Know Tools

Mistral Large 4 (open weights)

Global · Open Source

Mistral ne 29 Oct ko Mistral Large 4 ('Le Chonk') ke open weights Hugging Face par release kar diye (mistralai/Mistral-Large-4-1T-A52B). Yeh 1 trillion parameters ka natively multimodal MoE model hai, har token par 49B active (embeddings milake 52B). Mistral ke mutabik 6 Oct ke preview ke baad reasoning efficiency mein sudhaar aaya hai. License Mistral Ecosystem Open-Weight (modified MIT) hai: commercial use, modify aur fine-tune free hai; sirf Model-as-a-Service chalane wali businesses ko Mistral se alag commercial license lena padega. Self-host ke liye vLLM ya SGLang use kar sakte ho; Microsoft Foundry par bhi available hai.

Official Link →

NVIDIA Nemotron 3.5 Transcribe

Global · Open Source

NVIDIA ne 28 Oct ko open-weight, promptable speech model release kiya: ek hi checkpoint mein 40 languages mein transcription, 31 languages ka English se ya English mein translation, 13 languages mein speaker attribution (8 tak speakers), aur audio Q&A. Ghante bhar tak ki recordings handle karta hai. Model 30B total parameters ka hai aur har token par lagbhag 4B active rehte hain. NVIDIA ke mutabik Hugging Face Open ASR Leaderboard ke English track par open-weight models mein yeh pehle number par hai (5.33% average WER). India angle: Eka Care ke Indian doctors aur medical students ke benchmark par fine-tuning se held-out speakers ka WER 22.44% se 12.17% aaya (NVIDIA ka claim, ~14 GPU-hours, ek GB300 par). Dhyan rakho: yeh recorded audio ke liye hai, live streaming ke liye nahi.

Official Link →

Microsoft Copilot harness

Global · Dev Tools

Microsoft ne 28 Oct ko Copilot harness ko Microsoft Foundry ke through developers ke liye khol diya. Yeh wahi agent runtime hai jo Copilot Code, Copilot Studio, GitHub Copilot app, CLI aur VS Code ko chalata hai: 20+ frontier models (aur bring-your-own-key) ke beech task ke hisaab se kaam route karta hai, aur MCP, agent skills aur plugins support karta hai. Microsoft ke mutabik runtime ko Rust mein port kiya gaya hai, jisse session overhead aur time-to-first-token kam hue. SDK GitHub se download kar sakte ho, ya Foundry par fully managed cloud mein chala sakte ho. Performance aur cost ke daave Microsoft ke apne hain.

Official Link →

Claude Max / Ultra plan rename

Global · Frontier AI

Anthropic ke support page ke mutabik Max 20x plan ab Ultra kehlata hai ($200 per month), aur purane Max 20x subscribers bina price change ke automatically Ultra mein move ho gaye. Ultra mein Pro ka 10 guna usage har hafte milta hai, aur verified accounts par five-hour limit nahi hai. $100 wala Max plan Pro ka 5 guna usage har five hours aur har hafte deta hai. Yaani session limit hati hai, lekin hafte ka cap abhi bhi hai.

Official Link →

xAI Grok Bot: Stripe Link checkout aur Vault

Global · AI Agents

Grok Bot Stripe ke Link wallet se online purchase kar sakta hai: Link connect karo, har purchase par merchant aur total ka approval card aata hai, approve karne par Link ek one-time card banata hai jo sirf us ek payment ke liye chalta hai. Har action Auto-review se guzarta hai; aap 20 'Allow automatically' aur 20 'Ask first' rules bana sakte ho, aur dono match hon toh Ask first jeetta hai. Dhyan: xAI ke Vault docs ke mutabik Vault mein save kiya card checkout par har baar poochhe bina fill ho jata hai aur har use par approval ki koi setting nahi hai. Har payment khud approve karni ho toh card Vault mein save mat karo (Secure Form mein 'Securely save to Bot's vault' unchecked rakho).

Official Link →

📅 Section 02

Upcoming AI Events & Hackathons

🇮🇳 India Events — Apna Desh, Apna Stage

Bengaluru Tech Summit 2026

17–19 Nov 2026 · BIEC, Bengaluru

Karnataka sarkar ka 29th edition, theme AI & Beyond. Organisers ke mutabik 1,800+ exhibitors, 25,000 delegates aur 60,000+ business visitors expected hain, aur 50,000 sq ft ka AI aur Robotics pavilion bhi hai. Registration official summit website se hoti hai; alag-alag delegate pass types hain.

Homingo Hack Days 2026

Online AI engineering hackathon, poore India ke liye khula. Entry free hai; tracks: Agents, Automation, LLM apps; solo ya 2-4 ki team; AI tools allowed. Registration Internshala listing ke through (ya organiser ki site par) 15 Nov 2026, 11:59 PM IST tak. Build phase 21 Nov se 6 Dec, phir online demo day.

Apply Here →

Code for a Billion: Bharat Agentic-AI Hackathon 2026

Submissions close 15 Nov 2026 · Online

Code for India ka 90-din ka open-source hackathon, jiska focus India ki public-health problems par hai. AgentFoundry.me par sign-up karke project banao, cloud par deploy karo aur open-source karke 15 Nov tak submit karo; winners 5 Dec ko. Site ke mutabik har participant ko 1M smart-code tokens free milte hain.

Apply Here →

🌍 Global — India Relevant Events

PayPal AI Hackathon

Submissions close 12 Nov 2026 · Online

PayPal ka online worldwide hackathon: PayPal aur AI dono ka use karke working project banana hai. PayPal ke blog ke mutabik $60,000+ cash prizes, winners 21 Dec ko. Devpost par join karke submit karna hota hai; Internshala listing ke mutabik registration fee free hai. Submission 12 Nov (US Eastern) ko close hota hai, yaani IST mein 13 Nov ki subah.

Apply Here →

Web Summit Lisbon 2026

9–12 Nov 2026 · MEO Arena & FIL, Lisbon

Europe ka bada tech conference: startups, investors aur AI ki baat. AI Summit track 20 content tracks mein se ek hai. Tickets paid hain aur sirf official site se milte hain; current pricing wahin dekho.

🔬 Section 03

Top Research Papers

🇮🇳 India Research & Ecosystem

Right Numbers, Wrong Questions: Evaluating LLMs on Underspecified Financial Queries

TIFIN India

28 Oct 2026 ka India-focused paper financial AI assistants mein ek problem identify karta hai: model sahi data se real number la sakta hai, lekin user ke actual question ka galat interpretation kar sakta hai. Authors ise frame collapse kehte hain. Indian mutual funds par research mein ambiguous financial metrics, Direct vs Regular plan aur investor ke missing tax ya personal details jaise cases test kiye gaye. Paper batata hai ki financial assistants ko assumptions banane ke bajay ambiguity identify karke required information clarify karni chahiye.

Generation, Not Selection: LLM Proposals for Certified Access-Control Policy Mining

Indian Institute of Technology Kharagpur (IIT Kharagpur)

27 Oct 2026 ka paper enterprise access-control systems ke liye LLM-based policy mining approach propose karta hai. LLM access logs aur attribute names ke meaning se semantic access rules generate karta hai. Har proposed rule ko denied requests ke against filter kiya jata hai, taaki invalid permissions grant na hon. Do synthetic enterprise domains par five seeds ke experiments mein proposed method ne 10 mein se 10 runs mein existing syntactic approach ko outperform kiya aur held-out false grants ko 4–7 guna reduce kiya. Enterprise security aur authorization systems ke liye relevant research.

🌍 Global Research

Proving and Improving the Platonic Representation Hypothesis with JEPAs

Brown University, AMI Labs, New York University

29 Oct 2026 ka paper Platonic Representation Hypothesis ko ek formal theorem ke roop mein establish karne ka daava karta hai. Iska main idea hai ki alag modalities par separately trained JEPAs, jab same scene ke do observations ko positive pairs banate hain, toh ek shared representation seekh sakte hain (ek rotation tak). Authors ke experiments mein good-view JEPA ko same alignment achieve karne ke liye next-token prediction se 42 guna kam training samples chahiye. Ek 109M-parameter text JEPA ne tested language models (72B parameters tak) se stronger alignment dikhaya.

Investor's Last Exam: Can AI Agents Do Equity Investment Without Seeing the Future?

Prophetlab, Shanghai Jiao Tong University aur anya

28 Oct 2026 ka paper equity-research AI agents ke liye Investor's Last Exam (ILE) benchmark introduce karta hai. Ismein 3.43 million versioned financial documents, 86 investment assignments aur professional buy-side experts ke evaluation rubrics hain. Benchmark research quality, forecast accuracy aur simulated investment performance ko assess karta hai. Authors ke results mein GPT-6 Astra aur Claude Opus 5.5 overall 68.1 aur 66.9 scores ke saath lead karte hain, lekin Astra ka Forecast & Model score 52.5 hai. Yaani financial research karna aur accurately future predict karna alag capabilities hain.

Training Foundation Models with a Single Algorithm

Carnegie Mellon University, Johns Hopkins University, Aurora Innovation

29 Oct 2026 ka paper foundation-model training ke alag stages—pretraining, midtraining, supervised fine-tuning (SFT) aur reinforcement learning (RL)—ko ek unified likelihood-based algorithm ke andar combine karta hai. Static training data aur model-generated RL outputs ko ek common trainer mein mix kiya ja sakta hai. Authors controlled aur real-world experiments mein accuracy-compute trade-off, task coverage aur existing capabilities retain karne mein improvements report karte hain. Is approach mein traditional staged training, evolving data mixtures ka ek special case ban jaati hai.

⚰️ Section 04

AI Graveyard

🪦

Google Jules

Google ka async coding agent · 31 Dec 2026 tak

What it was: Google ka asynchronous coding agent. GitHub repo ke saath kaam karke bugs fix karta, tests likhta aur pull requests kholta tha.
What happened: Google ne 29 Oct ko Jules Turndown notice jaari kiya. 11 Nov tak active sessions poore kar sakte ho, uske baad Jules read-only ho jayega, aur 31 Dec 2026 ko poori tarah band. Code, branches aur pull requests GitHub par rahenge; Jules ki settings aur sessions export kar sakte ho.
Why it died: Google ka faisla hai ki Jules se seekhi cheezein agent-first platform Antigravity mein jayengi. Jules jaise background tasks Antigravity mein laane par kaam chal raha hai, par Google ne timeline nahi di.

Source: Google Jules Changelog (29 Oct 2026)

🪦

OpenAI Evals platform

OpenAI ka hosted evals dashboard + API · 30 Nov 2026 tak

What it was: OpenAI ka hosted evaluation product: dashboard aur /evals API, jahan datasets aur graders se models aur prompts ko test kiya jata tha.
What happened: OpenAI ne 3 Jun 2026 ko deprecation announce kiya tha. 31 Oct 2026 ko existing evals read-only ho jaate hain, aur 30 Nov 2026 ko Evals dashboard aur API band. Agent Builder aur v1/prompts API bhi 30 Nov ko band hone wale hain.
Why it died: OpenAI ne deprecations page par wajah nahi likhi; migration ke liye Promptfoo wali guide suggest ki hai.

Source: OpenAI Deprecations Page

“

Is hafte ka sabak seedha hai: official dikhna safe hona nahi hota. Cursor wala extension asli naam se, asli marketplace se aaya, phir bhi backdoor lekar aaya. Agent ko jitni power doge, uske har update ko utni hi seriously audit karna padega.

— Editorial Observation, 1 November 2026

🇮🇳 Section 05

India AI Beat — Desi Developments

Sarvam ka HQ IDS ke saath MoU: Army, Navy, Air Force ke AI projects

Sarvam ne 27 Oct ko HQ IDS (Integrated Defence Staff) ke saath MoU ka elaan kiya: teeno Services ke AI projects ke liye umbrella framework. Company ke mutabik kaam uski firewalled division Chanakya karegi (Sarvam 105B jaise models), focus decision support aur data fusion par, aur operational decisions par human oversight rahega. 28 Oct ko Sarvam ne clarify kiya ki MoU HQ IDS ke saath hai, Ministry of Defence ke saath nahi, jabki kuch articles mein ulta likha tha.

Source: Sarvam AI (27 Oct 2026) →

Simplismart ko NVIDIA-backed $14M Series B

Bengaluru ka inference-software startup. $14M (~₹133 Cr) Series B: Dallas Venture Capital ne lead kiya; NVIDIA, Accel, Shastra VC aur Micromax ne participate kiya. Report ke mutabik valuation ~$100M. Company ka claim hai ki 6 mahine mein annualized revenue run rate 14 guna hua (base number share nahi kiya). Simplismart khud GPU capacity nahi rakhta; software se models ko tez aur sasta serve karta hai.

Source: Inc42 (27 Oct 2026) →

Is hafte Indian startups ne $132M+ jutaye, AI doosre number par

Inc42 ke mutabik 26–30 Oct mein Indian startups ne $132M se zyada raise kiya. AI doosre number par raha: 5 startups ne $30.4M jutaye, jisme Simplismart ($14M) aur Live2.ai ($13.7M, ~₹130 Cr pre-Series B) ka hissa 91% tha.

Source: Inc42 (31 Oct 2026) →

💼 Section 06

AI Jobs This Week

🔥 This Week's Hot Skills — November 2026

AI Supply-Chain SecurityAgent Guardrails & ApprovalsOpen-Weight Model Self-HostingLLM Inference OptimisationKnowledge Graph / Ontology for AgentsIndic Speech (ASR) Fine-tuning

Is hafte ki khabron ne teen cheezein saaf kar di: dev tools ki supply-chain security, agents ke approval rules, aur open-weight models ko apne infra par chalana. Indian angle mein inference optimisation aur Indic speech/ASR fine-tuning ki demand badhti dikh rahi hai.

Open AI, Infrastructure & Engineering Roles (Bengaluru, Delhi)

Sarvam AI

Agent EngineerPerformance Engineer (Inference)Infrastructure SRE - HPCML Researcher (Foundational Models)

Apply Now →

Open ML & Engineering Roles (Noida, Remote India)

Wadhwani AI

ML Scientist (LLM)ML Scientist (ASR)ML Scientist (Speech)QA Engineer

Apply Now →

Open Speech AI & Local AI Engineering Roles (Pune)

NVIDIA

Senior System Software Engineer, Speech AISenior System Software Engineer - Local AI Automation

Apply Now →