AI Weekly: BSNL × Starlink, PewDiePie's Odysseus & 27 Industry Updates
A curated breakdown of the biggest AI and tech moves this week — from India's satellite internet push to Microsoft's 7 new models, Anthropic's $65B raise, and free local AI tools.
By AI Workflows
Source video: The Indian Government Just Teamed Up With Elon Musk To Bring Internet From Space (+24 AI Updates) by Vaibhav Sisinty — summarized and expanded for developers and agencies building with AI.
Headline Story: BSNL × Starlink
India's state-run telecom operator BSNL is in early talks with Starlink to explore satellite backhaul and enterprise connectivity. No deal is signed yet — Starlink still needs full security clearance — but the strategic fit is clear:
- Rural connectivity: ~30,000 remote villages lack proper 4G/5G. Fiber to towers in the Himalayas, Northeast, and islands can take years; satellite backhaul could connect towers in weeks.
- Enterprise services: BSNL's managed networks, data centers, and voice services could be strengthened by Starlink's low-orbit satellite network.
- Distribution: Starlink lacks a large retail/enterprise footprint in India; BSNL's nationwide dealer network could accelerate go-to-market once approvals land.
What it means for builders: Reliable rural internet unlocks WhatsApp automation, cloud POS, and AI agents for businesses that were previously offline — exactly the verticals we automate at AI Workflows.
27 Updates That Matter
Local & Private AI
| Update | What happened | Why it matters |
|---|---|---|
| PewDiePie — Odysseus | Free open-source AI workspace; 50K+ GitHub stars in days | Scans your hardware, runs models locally, deep research agent, zero cloud data leakage |
| Google Gemma 4 12B | Runs on 16 GB RAM laptops, multimodal (image/audio/video) | Free on Hugging Face — no internet required for inference |
| Perplexity Hybrid Inference | Splits work between local + cloud models | Sensitive files stay on device; heavy reasoning goes to frontier models |
| NVIDIA RTX Spark | Superchip for thin Windows laptops; 128 GB shared memory | Microsoft Surface Laptop Ultra built around it — Windows competing with MacBook for AI dev |
| Hermes Agent — Tool Search | Loads only the tool needed (41% → 3% context usage) | Long-running agents without context bloat |
| Hermes Desktop App | Dedicated desktop UI for agent chats, skills, artifacts | Cleaner than keeping a browser tab open 24/7 |
Big Tech & Models
| Update | What happened | Why it matters |
|---|---|---|
| Anthropic $65B raise | Valuation ~$965B; 5 years since founding | Fastest path to $1T if it continues — Meta took 9 years |
| Microsoft 7 MAI models | MAI Thinking, Code, Image 2.5, Transcribe, Voice + Maya 200 chip | Fully licensed training data; less OpenAI dependency |
| NVIDIA Cosmos 3 | Open omni model for physical AI (robots, simulations) | Super + Nano variants on Hugging Face |
| Google Magenta RealTime 2 | Free, open-source music generation on laptop | Type beats/instruments live — democratizes music production |
| OpenAI Codex Sites & Plugins | Build microsites, dashboards, portals from prompts; 6 role plugins | 20% of Codex users are non-developers — no-code app building accelerates |
| ChatGPT Memory Upgrade | Review/steer memory summaries; auto-updates stale context ("dreaming") | Better long-term personalization without manual cleanup |
| ChatGPT Send Email | Draft and send via linked Gmail without leaving chat | Workflow compression for everyday tasks |
| Claude Cowork 2× limits | Doubled 5-hour usage until July 5 | Bigger delegated tasks in one session — limited time |
Creative & Media AI
| Update | What happened | Why it matters |
|---|---|---|
| Grok Imagine 1.5 | Cinematic video from start frame + motion prompt; up to 720p | Stage, animate, link shots with consistent look |
| ElevenLabs Dubbing v2 | Preserves tone, pace, emotion across 90+ languages | Democratizes dubbing that once cost hundreds per minute |
| ElevenLabs Music v2 | Music generation across 11 Music, API, and Creative | Games, ads, branded content from one model family |
| HeyGen frame.md | Open-source video layout spec from design.md | Brand motion rules for AI-directed video (16:9, timing, silence) |
Agents & Automation
| Update | What happened | Why it matters |
|---|---|---|
| Microsoft Scout Autopilot | Monitors inbox, Teams, calendar; pre-reads before meetings | First "autopilot" agent category inside M365 |
| OpenClaw + Microsoft MXC | Viral open-source agent sandboxed for enterprise | Companies can deploy autonomous agents with bounded access |
| Perplexity Search as Code | Agent writes a program to run all searches at once | Beats latency, dedupes results, halves cost vs sequential search |
| Ollie AI Family Assistant | WhatsApp/email/calendar agents for household tasks | Reminders, fridge recipes, flight rescheduling — family ops automation |
| Gopuff × Grok "Go" | Voice shopping assistant in delivery app | Learns preferences, auto-reorders, handles complex cart requests |
| Cursor Auto-review | Allowlisted calls run instantly; risky ones sandboxed | Agents work longer with fewer approval prompts |
| Microsoft Project Solara | Agent-first devices (badge, alarm-clock form factors) | Multimodal glance + voice — reimagining the computer beyond screens |
Product & Platform
| Update | What happened | Why it matters |
|---|---|---|
| Microsoft 365 Copilot redesign | Clean homepage, visualize tag, 50%+ faster load | Usage up 27–43% across Word/Excel/PPT/Outlook at launch |
| Odysseus tutorial | Install via Google Antigravity with zero code | Free local models via Cookbook hardware scan |
Deep Dive: Odysseus (PewDiePie's Free AI Workspace)
The standout tool this week is Odysseus — not a new foundation model, but a local-first wrapper that:
- Scans your machine and recommends which open-source models your GPU/RAM can run
- Runs entirely on-device — prompts never leave your computer
- Deep Research loops through dozens of web sources and outputs a visual HTML report
- Cookbook picks optimal free models (Qwen, Gemma) so you pay $0/month vs ChatGPT/Claude subscriptions
For agencies, this is a signal: clients increasingly want private, self-hosted AI — the same direction behind our n8n + Supabase + local model stacks.
Deep Dive: Perplexity's Two Big Moves
Hybrid Agentic Inference keeps acquisition models, whiteboard photos, and call transcripts on your Windows laptop. A local agent redacts PII before anything hits the cloud. You get frontier-model reasoning without uploading raw deal files.
Search as Code replaces the slow "search → wait → read → search again" loop. The model writes a mini program that runs parallel searches, dedupes, filters, and ranks — then writes the answer. Perplexity claims near-half cost vs competitors on research benchmarks.
What We're Watching at AI Workflows
These updates reinforce three bets we're already building on:
- Messaging-first automation — Ollie and Gopuff Go prove WhatsApp/chat-native agents are going mainstream. Our WhatsApp Restaurant Automation is ahead of this curve.
- Local + hybrid AI — Gemma 4, Odysseus, and Perplexity hybrid inference validate our self-hosted and privacy-first client work.
- Agent platforms, not just chatbots — Microsoft Scout, OpenClaw MXC, and Cursor auto-review show enterprises need sandboxed, long-running agents — exactly what we engineer with n8n and custom pipelines.
Timestamps (from source video)
- 01:19 — PewDiePie Releases Odysseus AI
- 02:22 — ChatGPT Memory Upgrades
- 03:11 — Anthropic $65 Billion Funding
- 03:34 — BSNL + Starlink Partnership
- 05:15 — Grok Imagine 1.5 Preview
- 05:56 — Claude Cowork 2× Limits
- 06:13 — NVIDIA Cosmos 3 Physical AI
- 07:29 — Seven New Microsoft MAI Models
- 08:55 — Send Emails from ChatGPT
- 09:14 — NVIDIA RTX Spark Superchip
- 10:15 — ElevenLabs Dubbing v2 Alpha
- 11:24 — Ollie AI Family Assistant
- 11:59 — Gopuff × Grok "Go"
- 13:30 — Google Magenta RealTime 2
- 14:18 — Microsoft Scout Autopilot Agent
- 14:58 — OpenClaw + Microsoft MXC
- 15:47 — OpenAI Codex Sites & Plugins
- 17:14 — HeyGen frame.md Video Layouts
- 17:59 — ElevenLabs Music v2 Out Now
- 18:36 — Perplexity Hybrid Inference
- 20:06 — Perplexity Search as Code
- 21:33 — Google Gemma 4 12B Model
- 22:12 — Microsoft 365 Copilot Redesign
- 23:42 — Cursor Auto-review Run Mode
- 24:36 — Hermes Agent Tool Search Update
- 25:33 — Hermes Desktop App Launch
- 26:16 — Microsoft Project Solara Preview
- 27:14 — Odysseus Installation Tutorial
Curated for AI Workflows readers. Original breakdown by Vaibhav Sisinty. Explore our AI Tool Hub or contact us for automation projects.