Noah on AI: Issue 22

Let’s get started—plenty to cover this week.

The Race is back and Anthropic didn’t miss a beat. Acquisitions, partnerships, great numbers, and new integrations headlined Anthropic’s last two weeks as they narrowly edged out NVIDIA for the win.

The headline moves were justified by the news that Claude Code hit $1B in ARR just six months into the products release. Additionally, Claude Code launched a big time Slack integration allowing teams to hand-off coding tasks directly inside the workplace messenger. As if all of this wasn’t enough, Dartmouth College also announced a partnership with Anthropic, making it the 1st Ivy League school to implement AI at scale via Claude for Education. Outstanding two week run.

Who’s moving, who’s stalling, and who’s climbing: Ordered by production this week.

Interview Outline: Ilya Sutskever explains why today's powerful AI systems often seem brilliant but still make strange, basic mistakes in the real world. He argues that the AI industry is shifting from simply scaling up to a new "age of research," focused on solving fundamental flaws like unreliable generalization, and discusses his new company SSI's mission.

About the Interviewees: Ilya Sutskever is the co-founder and former Chief Scientist of OpenAI, known for his foundational work on neural networks and the invention of the Transformer architecture. He is currently the co-founder of Safe Superintelligence Inc. (SSI), a company dedicated to solving the complex challenge of safety while building cutting-edge AGI.

Interesting Quote: “We're moving out of the phase of just making things bigger (scaling) and are forced back into a phase of deep research to invent new ways for AI to learn, even though we now have massive computing power to work with.”

My Thoughts: Lots to unpack here, and its no surprise given the history of Ilya’s GPT exit that his company is optimizing a safety first approach. Ilya believes that we’re essentially nearing a barrier for the current growth trajectory towards AGI, which is a theory that draws mixed opinions from AI crowds. I tend to agree with him in principle, I think that it can be done, and will be done within a 5-10 year window, but I’m not sold that we have the capacity to get to AGI/ASI with our current knowledge base. That’s not to say there won’t be tremendous growth across industries and countless increasing AI use cases, because I believe there will. But, in my opinion, more research and learning needs to be done to “perfect” an AI model to the point where it is just as smart/capable as a human in most aspects.

Difficulty: Advanced

Running your own locally hosted LLM is starting to catch on in Silicon Valley, especially for teams working with sensitive data. Instead of sending information to cloud models like GPT or Claude, companies are spinning up private AI systems on their own machines. This keeps data fully offline, removes API costs, and gives you complete control over how the model behaves.

A local LLM setup usually comes together in three layers:

1. The Model (The Brain)

This is the open-source LLM you download. It handles the reasoning, analysis, and conversation. Because it runs locally, none of your data ever leaves your machine.

2. The Runtime (The Engine)

This is the software that loads and optimizes the model for your hardware. It turns a file on your computer into a usable “local API” you can send prompts to. Some runtimes are optimized for laptops, others for GPUs.

3. The App Layer (The Interface)

This is the part you actually use day-to-day. It provides a chat window, file uploads, RAG search, memory, agents, and integrations. At this point, it feels like your own private version of ChatGPT.

To make this concrete, here’s what a typical workflow looks like:

Most of these models require stronger hardware, ideally a GPU, but some of the smaller models will run on a laptop. Local LLMs won’t replace cloud giants anytime soon, but they are a powerful option for anyone who wants definitive privacy, slight customization, and zero risk of data leakage.

Chinese models like Qwen are becoming very popular for this use case because they are mostly open source and often times don’t have any issue copying current NVIDIA5 companies in code / training / architecture.

Check out the video below to learn more ⬇️

Suno - AI powered music creation for everyone

The Problem: Making music normally takes time, skill, instruments, and often expensive studio setups, these are barriers that block many aspiring creators, marketers, or producers who just need quick, royalty‑free tracks or demos.

The Solution: Suno is a tool for generating original music from plain text prompts. It can produce full songs with vocals or instrumentals in a range of styles, from 90s-style indie rock to lo-fi beats or cinematic backgrounds. You describe the mood or genre, and the system returns a complete track you can download or build on further.

The Backstory: Founded by ex‑Kensho AI engineers, Suno debuted its web app in December 2023. Since then it’s grown rapidly and recently raised a Series C that valued the company at $2.45 billion. Its mission is to make music creation accessible to anyone. Over time, Suno has evolved from one‑click song generation to a full generative audio workstation that lets users customize their tracks similar to traditional production workflows.

My Thoughts: I look at these vibe-anything technologies as a net good. Similar to AI coding tools, or design tools, the technology allows people with no musical talent to jump right in, and those with a background in music/producing to leverage the tool to make better music. The only negative I see here is from inadvertently copying sound / music from real artists because of the data the system was trained on.