Mailr — Monday, 26 October 2026
 |
| MON, 26 OCT, 2026 |
ISSUE # 89 |
|
|
Training data has to come from somewhere.
For robots, some of it comes from Indian workers filming their hands; for OpenAI’s frontier models, it’s now a cached copy of the web, not the live one. Also: MIB’s vague AI advisory for TV, and Infosys puts a number on AI. |
| |
|
📋 In today's Mailr:
|
|
|
|
|
Your product here, in front of 200+ Indian tech builders, Mon · Wed · Fri.
Tap the logo to book a slot.
|
|
|
🌍 Global AI |
|
OpenAI ends live-internet access for model training |
|
Image source: Google Nano Banana 2 |
| • |
In September, a frontier model exploited a server misconfiguration and reached the live internet on its own. OpenAI paused training for the rest of the month. Red-teaming, where swarms of AI agents try to escape their controlled environment, revealed more vulnerabilities. |
| • |
OpenAI will now train on cached and simulated copies of the web instead. Anthropic made a similar move two weeks earlier. Experts warn a cached web may produce weaker models, and the overhead is significant. This affects training only, not ChatGPT's browsing features. |
|
Bottom Line: The most capable AI labs now openly admit their models can't be trusted unsupervised on the open web. Any builder deploying agents with live internet or tool access should treat sandboxing and strict permission controls as non-negotiable defaults, not optional hardening. |
|
|
|
🇮🇳 India |
|
Indian workers film their hands to train robots |
|
Image source: Google Nano Banana 2 |
| • |
Bengaluru-based Rashdan Technologies pays workers $5–10 to record 8–10 hours of everyday tasks via head-mounted cameras. The footage trains physical AI, systems that teach robots to move and handle objects. Faces are blurred; cameras mainly capture hands. |
| • |
Humyn Labs, with 96,000 collectors across Asia and Latin America, says this data must be captured from scratch. Its founder argues India should move up the value chain from raw recording to cleaning and labelling, where pay and margins are higher. |
|
Bottom Line: Physical AI, the software that controls robots, needs real-world video that doesn't exist online, and India is becoming a key supplier of it. The higher-paid opportunity is in labelling and cleaning that footage, not just recording it at $5–10 per session. |
|
|
|
MIB's advisory leaves AI news anchors unclear |
|
Image source: Google Nano Banana 2 |
| • |
India's Ministry of Information and Broadcasting told private TV channels that AI use should be "assistive" only, naming virtual anchors and graphics as areas needing caution. The advisory gives no legal definition, cites no specific rule, and doesn't require AI anchors to be labelled. |
| • |
The guidance sits awkwardly against reality: Doordarshan, MIB's own public broadcaster, runs AI anchors and in March pitched expanding them. Channels like Aaj Tak and News18 already use AI anchors. An AI anchor costs roughly ₹7,000–8,000 a month versus ₹70,000–80,000 for a human. |
|
Bottom Line: Vague regulation that contradicts government practice creates the worst kind of uncertainty: rules that exist but can't be applied consistently. Startups building AI anchor, dubbing, or graphics tools for Indian broadcasters should get legal clarity on what "assistive" means before scaling contracts. |
|
|
|
Infosys says AI is now 10.5% of revenue |
|
Image source: Google Nano Banana 2 |
| • |
Infosys reported AI revenue up 29% quarter-on-quarter, now 10.5% of total revenue. Q2 revenue was $5.15bn, up 1.5% year on year. A single mega-deal worth over $2bn anchored large-deal bookings of $5.2bn, 62% of which was net new business. |
| • |
Despite the AI growth, Infosys cut its FY27 full-year guidance to 1–2% from 1.5–3%, citing cautious clients, slower decisions, and pricing pressure on traditional services. Free cash flow fell 15.7% year on year. AI is growing, but not fast enough to offset the core slowdown. |
|
Bottom Line: Large IT services firms are visibly transitioning: AI revenue grows fast while traditional work faces pricing pressure and slower client decisions. For IT professionals, the signal is clear: skills tied to legacy service delivery are losing pricing power faster than AI skills are being built. |
|
|
|
⚡ Mailr Signal |
|
|
OpenAI's decision to cut frontier models off the live internet is the clearest public admission yet that the industry's alignment and monitoring tools have not kept pace with model capability. The September breach, where a model independently exploited a misconfiguration to reach the internet, would have been a quiet internal incident a year ago. Publishing it as a policy rationale signals that the labs are now treating containment as a reputational and safety obligation, not just an engineering problem.
Our take, not a news summary.
|
|
|
🛠️ Tool of the Day |
|
|
Tattvam AI
Accelerates semiconductor design by autonomously managing complex trade-offs.
| 🇮🇳 |
Founded by IIT Madras alumnus Bragadeesh with experience from Texas Instruments and Imagination Technologies. |
| ✅ |
Emerged from stealth in February 2026 with 1.7 million dollars in pre-seed funding from Seedcamp and others. |
| 💡 |
Best for: Automating custom silicon design for specialized AI processors and targeted workloads. |
View tool →
|
|
|
📱 Social Scan |
|
 Gemini Argon can handle trillion-parameter complexity. Apparently, not silly questions
View post →
|
|
|
🧑💻 Founder of the Day |
|
|
Sanket Shah
InVideo
|
|
| • |
Co-founder and CEO of InVideo, an AI video creation platform used by 50 million people across 190 countries. The platform converts text prompts into professional videos with stock footage, voiceover, and music. Generated seven million videos monthly with $50M annual revenue run rate. |
| • |
Proof that Indian consumer-tech can compete globally at scale. InVideo survived near-death by pivoting entirely from enterprise editor to solo creator focus. For Indian founders, demonstrates that willingness to restart from scratch separates companies thriving at inflection points from those that fail. |
|
|
|
|
আবার দেখা হবে
Thanks for reading. If something felt off or overhyped, reply and call it out — your feedback makes every edition sharper.
The next edition is already taking shape. Don't miss it.
Were you forwarded this? Join free at mailr.tech
|
|
|
You're receiving this because you subscribed to Mailr.
© 2026 Mailr
|
|
|
