When AI is smart enough
is that a world, or a matrix?
The newsletter for the technically curious. Updates, tool reviews, and lay of the land from an exited founder turned investor and forever tinkerer.
Hey folks,
GPT-5.1 is out—only in ChatGPT for now, API access is coming soon. We get two new models: GPT-5.1-Thinking and GPT-5.1-Instant (which can also think if needed, but just a little). The focus is on two things: vibes and safety, not smartness. OpenAI claims it has better vibes than 4o when chatting, and it’s safer than GPT-5. ChatGPT is also getting new default personalities that you can pick from. Here’s how they all respond to the same question.
This year, we have seen a bunch of research previews that generate navigational environments from images/prompts. Marble by World Labs is another one on the list. It’s not interactive like some others, but it’s available to use as a product, and I’m really impressed by the quality.
Realistic video was one of the key trends in 2025 (a noticable improvement from 2024’s “still looks weird/fake” videos). We’ll see a lot of these realistic/interactive worlds next year with likely a lot more interest in VR headsets—Samsung just launched one couple weeks ago, Steam is also launching one for games next year.
- Keshav
ElevenLabs has a new speech-to-text model - Scribe v2. It’s better than Gemini 2.5 Flash in accuracy and latency. They also released an option to license Iconic Voices (like Dr Maya Angelou and Sir Michael Caine). ElevenLabs is one of the few companies in the AI x Creatives space that’s not alienating the existing talent in their business.
Anthropic is also building data centres now with a $50B investement, starting from Texas and New York. In the meantime, they are letting software engineers (who know nothing about robotics) teach robot dogs to play fetch.
Eliminate AI anxiety. Build with confidence. AI agents are the future—but centralized systems create costly risks. With Airia, prototype and orchestrate AI agents securely while creating a resilient, decentralized AI ecosystem for your organization. Get started today at airia.com*
🌐 What I’m consuming
Satya Nadella answers hard questions from Dwarkesh and Dylan Patel.
AI adopters produced 39% more code merges, with no sign of a decrease in quality.
AI glossary - Simple definitions for key AI terms with visualisations.
How to improve Claude’s frontend design outputs using Skills.
Building a Character AI clone with Google AI Studio in 50 minutes.
Multi-agent systems can handle more complex tasks, but are they worth navigating dozens of moving parts and infra costs that add up fast? Yep, if you’ve got a guide to follow. The team at Galileo has compiled just that, proven tips to simplify the process, and they’re giving it away for free.*
⚙️ Tools and demos
CData Connect AI – Connect any of your data sources to AI for real-time enterprise data connectivity with MCP to make AI work for you.*
Deep Research in Gemini now connects to your workspace account, letting you give it access to your Gmail, Drive and more. Both Google Ads & Google Analytics are getting an AI advisor. Plus, Gemini Live can now speak faster or use accents.
Scribe Optimize - See how work happens in your organisation, identify bottlenecks and get real steps to improve each workflow.
Product Intelligence by Pylon - Turn your customer feedback into a product strategy
Delphi - Create a digital version of you, just by being interviewed.
LM Arena has a new Code Arena to test how models plan, scaffold, debug, and build real web apps step-by-step. Try it here or check out the leaderboard:
🥣 Dev dish
A deep research demo using Claude Agents SDK.
pi - A radically simple and opinionated coding agent with multi-model support. (disclaimer)
captions.events - An open-source template for broadcasting real-time transcripts.
SWE-fficiency - A new eval with 498 optimisation tasks to test if AI models can speed up real GitHub repos on real workloads.
💰 Who got that bag?
Parallel AI (the web search company by ex-Twitter CEO) raised a $100M series A.
Wisdom AI has raised a $50M Series A funding for its AI data analyst.
Tavus has raised $40M Series B for making PALS.
🍦 Afters
NYT might still get their hands on 20M random ChatGPT chats.
Meta released an ASR model that covers 1600 languages and an Ads recommendation model inspired by LLM-scale techniques.
That’s it for today. Feel free to comment and share your thoughts. 👋
Read about me and Ben’s Bites
📷 thumbnail creds: @keshavatearth,
Thanks to today’s sponsors who made this newsletter possible :)
Airia, Galelio and Cdata.
Wanna partner with us? Last few slots left for the rest of the year.