In this edition we’ll be covering…
OpenAI and Broadcom's chip collaboration that could reshape AI infrastructure
A tutorial on how Superwhisper lets you ditch the keyboard entirely
Andrej Karpathy's nanochat that puts ChatGPT in your hands for $100
5 trending AI signals from California to Figure's humanoid robots
3 AI tools to supercharge your workflow
And much more…
The Latest in AI
OpenAI and Broadcom Team Up to Build Custom AI Chips
OpenAI just made a move that could reshape the entire AI infrastructure game. They're partnering with Broadcom to design and deploy custom AI chips, and it looks like OpenAI’s betting big on controlling their own hardware destiny…
The collaboration basically means OpenAI designs the brains, Broadcom builds them, and together they're planning to pump out 12,000 units annually from their BotQ facility in San Jose. Eventually? They're eyeing 100,000 units over four years.
The details:
OpenAI is NOT just buying off-the-shelf anymore. They're taking lessons learned from building models like GPT-4 and o3 and baking that intelligence directly into silicon.
While OpenAI handles the architecture, Broadcom deploys the chips and packs them with their suite of Ethernet, PCIe, and optical connectivity products. It's a full-stack solution.
Last month, NVIDIA dropped $100 billion into OpenAI for data centers. Earlier this month, AMD traded 160 million shares (10% of the company) to get OpenAI to commit to six gigawatts of their upcoming MI450 chips. OpenAI is basically speed-dating the entire chip industry.
So What?
OpenAI is playing 4D chess while others play checkers. By designing custom chips, they can optimize for their specific workloads rather than making compromises with general-purpose hardware.
This move signals they're super serious about vertical integration — controlling everything from the model architecture down to the silicon that runs it.
🧠 You're reading this newsletter because you know AI matters…
But reading about AI ≠ actually using AI.
Here's the gap: You consume AI content, but you're not building meaningful products or workflows with it yet.
Inside our learning platform, you’ll have access to:
100+ hands-on lessons (not just theory)
A complete roadmap to take you from 0 → 1 with AI
Prompts you'll actually use today
Step-by-step automations for your workflows
(PLUS) weekly updates as AI evolves
(PLUS) 20+ new lessons released each month
All created by industry veterans, all for the price of a coffee a month. Start with a free trial! 👇
Tool Spotlight
Never Type Again with Superwhisper
Tired of typing? Meet Superwhisper, the AI-powered speech-to-text tool that understands what you want to say and how you want to say it. Our team has been loving this lately.
Superwhisper works on both iPhone and Mac, turning your spoken words into polished emails, crisp notes, or whatever format you need. No more hunting and pecking on your keyboard.
Here's how you can start using it:
Download and set up Superwhisper on your device
Available for both Mac and iPhone
New users get 15 minutes of free Pro access to test it out
Set your preferred language and choose your AI model (cloud or local)
Edit the settings to fit your needs.
Start talking instead of typing:
Works in any app where you can type or paste text
Supports 100+ languages with on-the-fly translation to English
Industry Intel
Karpathy Drops Nanochat
Andrej Karpathy just released nanochat, and it's basically the CliffsNotes version of building your own ChatGPT. If you've ever wanted to understand how these systems actually work (or just want to flex that you trained your own LLM), this is your ticket.
The GitHub repo packs the entire pipeline — tokenizer training, transformer pretraining, fine-tuning, and inference — into about 8,000 lines of code. That's right, everything you need to go from raw data to a working chatbot, all in one place.
Here's what makes nanochat different:
You can train a working model for around $100 — Spin up an 8xH100 GPU node and in roughly 4 hours, you've got a small ChatGPT clone that can actually hold conversations. Scale it up to 12 hours and you'll surpass GPT-2's benchmark performance.
It's built for learning, not just using — Karpathy designed this as the capstone project for LLM101n, an undergraduate-level class at his company Eureka Labs. It's meant to be hackable, readable, and forkable, so you can actually understand what's happening under the hood.
Full transparency, minimal abstraction — Unlike productionized systems where everything is hidden behind APIs, nanochat shows you the whole stack. From training the tokenizer in Rust to implementing KV caching for efficient inference, it's all there to explore and modify.
So What?
One of the godfathers of AI (Karpathy) is literally democratizing AI model training for anyone with curiosity and a credit card. While big labs are racing to build the most powerful models, Karpathy is making sure people understand how they actually work…
Quick Bites
Stay updated with our favorite highlights, dive in for a full flavor of the coverage!
California becomes the first state to regulate AI companion chatbots with SB 243, requiring companies to implement safety protocols including age verification, self-harm detection, and warnings that interactions are artificially generated, effective January 1, 2026.
Thinking Machines Lab co-founder Andrew Tulloch departed Mira Murati's AI startup to join Meta, where reports suggest he was offered a compensation package that could have been worth up to $1.5 billion over at least six years.
Google's Nano Banana image editing model is expanding beyond the Gemini app to Google Search, NotebookLM, and soon Photos, bringing advanced AI editing capabilities to more users after generating over 5 billion images to date.
Google introduces Speech-to-Retrieval (S2R), a new approach that maps spoken queries directly to embeddings and retrieves information without first converting speech to text, significantly outperforming traditional ASR systems.
Figure 03 humanoid robot launches as the "Model T of robots" — designed for mass production with 12,000 units planned annually, featuring improved hands that can sense just a few grams of pressure and built-in cameras for better grasping tasks.
Trending Tools
🤖 nanochat - Andrej Karpathy's minimal ChatGPT clone that packages the entire training pipeline in 8,000 lines of code.
⚡ Microsoft Amplifier - A complete AI development environment with 20+ specialized agents (zen-architect, bug-hunter, security-guardian), parallel worktree systems, and knowledge extraction to supercharge AI-assisted coding.
🔐 RedPill - Confidential AI gateway providing TEE-encrypted access to 200+ models including GPT-5, DeepSeek V3, and Qwen2.5, ensuring your queries stay encrypted and are never stored by providers.
Until we Type Again…
Thank you for reading yet another edition of Digestibly!
Reyhan
Early AI leader, advises Fortune 500 companies on AI Development, LLMOps, AI strategy, speaks (for fun) on practical AI. Turns cutting-edge theory into workflows teams can ship today.
Kevin
Ex-AWS SageMaker/Bedrock lead. Shipped infra powering 1M+ devices and $2.5M ARR. Obsessed with high-throughput, low-latency systems— and brings that discipline to every Digestibly release.