In February 2024, I wrote about Newstake for the first time. I wanted to get out of the personalized news feed and see more of what was going on in the world. The idea was to group similar articles, remove repetition and use AI to help readers understand the context.
The developments around AI are what made me pick Newstake up again. I want to follow how capable these systems are becoming, whether we are heading towards an AI take-off, and what is happening in AI safety. By take-off I mean the possibility that AI progress accelerates sharply as the systems themselves help with research and development. I want a place where I can track this over time and compare new claims with what came before.
AI companies already publish detailed reports about their models, including evaluations, limitations and safety testing. When that material is available, it makes little sense to only read someone else's summary of it. I want to read the companies' reporting directly and keep track of what they say they have tested, what has changed and what remains unresolved.
I have since changed the approach. Newstake now collects direct sources: what people and organizations publish themselves.
Why read the same statement five times?
Imagine a company announces a new product. You open your news app and find several articles about it. They quote the same announcement, use the same pictures and tell you roughly the same thing. You could have read the announcement yourself.
The old Newstake approached this as a clustering problem. Find the articles about the same event and group them together. That helps with repetition, but it still starts with somebody else's selection of what was worth reporting.
I now want to start with the original. A company's announcement, a politician's statement or a researcher's own work. Each publication links back to its source, so you can read what was actually said.
This also changes what belongs in the feed. Original research can belong there, including research by journalists. An outlet that repeats someone else's statement does not.
Of course, a direct source can be wrong. A company wants to sell its product. A politician wants you to agree with them. Reading their own words lets you see their position more clearly, but you still have to question it. I want to make it easier to compare those positions and follow the evidence.
What is this helping us understand?
I wrote a new manifest to make the purpose of Newstake explicit.
There are big developments that I want to understand. The world order is changing. Government debt and inflation affect how we live. Aging populations and migration change our societies. AI is already changing how we work, and I expect it to reach much further.
These things are connected. A company moving production to another country can be a business decision, a response to trade policy and part of a longer change in the balance of power. Reading the announcement is a start. Understanding why it happened takes more work.
That is where I want Newstake to help. You should be able to follow an event back through earlier decisions and see how it fits into a longer history. Someone who has followed a topic for years needs different context from someone reading about it for the first time.
My objection to personalization was always that a feed quietly decides what you should see based on what keeps you clicking. Choosing a topic and asking for an explanation that fits your background is useful. I want that choice to be yours.
The manifest also puts some limits on the approach. Claims should be traceable to their sources. Common beliefs should be open to questioning. There has to be room for uncertainty and competing interests. These are things I want the product to do consistently; writing them down does not mean the work is finished.
All in on AI
I am also changing how I build and run Newstake. I am going all in on AI, with agents and automations doing the work.
In the first version, AI was mainly part of the product: clustering articles and summarizing them. Now I use it to work on the product itself. Coding agents can take a task, inspect the code, make changes and run checks. Automations can come back to recurring work without waiting for me to start each task by hand.
There are already four automations doing recurring work on Newstake:
- One runs twice a day to find and fix faulty publications.
- One checks the database load every day. If it finds a spike, it investigates and makes a fix where needed.
- One adds new entities, such as companies, institutes, products and events, to expand what Newstake covers.
- One reads submissions from the feedback page and implements the requested changes.
The feedback automation is an interesting part of this. You can report a problem or suggest a change on the website, and an agent picks it up. I do not have to copy each submission into a coding task myself.
Collecting publications is ongoing work. Sources change, feeds break and new organizations become relevant. These automations let me give agents recurring responsibilities instead of starting every task by hand.
The aim is to have agents handle the work from development through the recurring tasks needed to keep Newstake useful. My role is to set the direction, explain what I want and judge the results. I still have to decide whether the system is doing the right thing.
For research, the possibilities are especially interesting. Following a reference into an old document, finding an earlier statement and comparing it with what someone says today can take a lot of time. An agent can do that work and bring back the sources. I want to use that to answer questions that a short news article would leave open.
There is also an obvious risk: generating more text is easy. I do not want Newstake to add another layer of repetitive articles to the internet. The useful result is an explanation that helps you understand something, with sources you can inspect. Sometimes that only needs a link to the original.
Bring your own agent
I also want to build Newstake for agents. The world does not need another chat interface. If you already use ChatGPT, Claude or your own agent, you should be able to use Newstake there.
I plan to add plugins for ChatGPT and Claude, MCP servers and APIs. These will let your agent search Newstake and retrieve the original sources as part of whatever you are working on. You could ask it to follow a company or compare what different organizations have said about an event, without copying links between apps.
You bring the agent and decide what to ask. Newstake provides the sources for it to work with.
Newstake is still a work in progress. The collection is growing, and some pages have little content or sources that are temporarily unavailable. The manifest describes where I want to take it.
You can try the new version at newstake.ai. If something is missing or you find a mistake, let me know.
Header photo by Taylor Vick on Unsplash.