A $13 billion bet that open-source AI already won
Baseten's new valuation prices a simple thesis: the next era of AI runs on open weights, and the money is following the inference, not the lab.
Issue No. 1 · First Issue
The signal in the noise. Curated AI news, agents, and the occasional good joke.
AI News
AI found 18 diagnoses doctors had missed for years
An NEJM AI study ran patient histories back through a model and surfaced conditions that had quietly gone untreated. The result is hopeful and a little unnerving.
The lab whose model got banned just asked to govern AI
Anthropic is convening a G7 coalition on AI safety - days after one of its own releases hit a wall. Watchdog, or fox volunteering for the henhouse?
Noam Shazeer leaves Google for OpenAI
The co-author of the Transformer paper changes jerseys again. Wherever attention goes, the talent follows.
Vibe coding & agents
Pixi: drop intelligent 3D characters into your DMs
A little agent with a face, a body, and opinions - living inside the chat you already use. Strange, charming, and clearly the start of something.
The typical AI agent stack, explained
Model, memory, tools, and a loop that decides what to do next. Strip away the hype and most agents are the same four boxes wired in the same order.
Clever tech anecdotes
Pulling Fable 5 while asking to govern AI
Anthropic quietly yanked Fable 5 the same week it pitched itself as the world's AI referee. Sometimes the best safety demo is the model you decide not to ship.
The Useful Agent Corner
The Useful Agent Corner
How to actually use Claude Code & MCP
MCP - the Model Context Provider - is how you hand an agent the right context at the right moment instead of pasting walls of text. Here's the short version.
Install Claude Code
Run it in your project root. It reads your repo, so point it at real code, not a scratch folder.
Connect an MCP server
Each MCP server exposes tools and context - your database, your docs, your tickets. The agent calls them like functions.
Ask for the outcome, not the steps
Describe what 'done' looks like and let the agent plan. Review its diff, not its keystrokes.
Free opportunities
Subscribe to AI Times - free
One issue. No spam, no fluff, no 'in today's fast-paced world.' Just the signal.
AI capability development
Slowtech, by way of Tony Fadell.
The fastest tech in the world may win by helping people use less tech
Tech jokes
Because even the models should laugh at themselves.
Why did the AI go to therapy?
It had too many complex dependencies.
How many prompt engineers does it take to change a lightbulb?
Just one - but they have to ask it 'in the style of a master electrician' and promise it a tip.
Why don't neural networks ever get invited to parties?
They keep overfitting to the conversation.
What's an LLM's favorite excuse?
'I was confidently wrong, but I was confident.'
Why did the agent loop forever?
Nobody told it the task was already done.
How does a chatbot apologize?
'You're absolutely right' - then it does the exact same thing.