In Canada, any person or corporation in the country can make a request for general records to any agency of the federal government through the Access to Information Act (the equivalent in the United States is the Freedom of Information Act). The government provides a searchable database of completed requests, but includes only a summary of the request and the number of pages of responsive material. The actual documents turned over are not included. However, completed request packages may be informally re-requested, and should you do so, someone from the relevant agency will (usually) send them to you eventually.

This re-request process has its limits. It can takes weeks or months for the documents to be sent, and the database itself only goes back to January 2020 (they used to delete records older than two years, but stopped doing this some time after 2020). Occasionally, they will never send the documents at all, and all you can do is either re-request them again or open a formal access to information request (which will cost you $5).

Making it easier to access completed access to information requests is why the Investigative Journalism Foundation built Open By Default, “the biggest database of internal government documents never before made publicly accessible”. It includes documents from completed access to information requests obtained using both automated (presumably the re-request form) and manual processes (donations from trusted partners, particularly of documents from before the online re-request form was available). The files are cleaned and OCRed into one beautiful, searchable database.

Time zones are hard. As a well-known Computerphile video so eloquently puts it:

What you learn after dealing with time zones, is that what you do is you put away your code, you don’t try and write anything to deal with this. You look at the people who have been there before you. You look at the first people, the people who have dealt with this before, the people who have built the spaghetti code, and you thank them very much for making it open source, and you give them credit, and you take what they have made and you put it in your program, and you never ever look at it again. Because that way lies madness.

The Canadian province of British Columbia recently decided to switch to permanent daylight time. I wanted to see if this update made it to the IANA Time Zone Database yet. Luckily, we can now view updates to this database as commits on GitHub. And there it was in the news file!

I’ve perused the tz repository before, and I always learn something interesting. For example, during WWII Britain adopted double summer time, adding two hours to the clock in the summer and one hour in the winter. The bulk of the comments in the database are dedicated to documenting this extensive history of time zone changes across the world.

Today, I wanted to share this piece from economist Scott Cunningham (Baylor University), who wrote about how AI is widening the gap between research and publishing. Or, in economics terms (emphasis mine):

But what happens when the same productivity shock hits a system where the bottleneck was never really production in the first place, but rather was a hierarchical journal structure that depended immensely on editor time, skill, discretion and voluntary workers with the same talents called referees for screening quality deemed sufficient for publication?

The post mentions the Autonomous Policy Evaluation project—the end-to-end AI paper pipeline I wrote about a few weeks ago—and discusses the likely consequences of this flood of AI-generated papers. Assuming the number of publication slots in reputable journals is relatively fixed, AI-generated papers should add a very large amount of mass to the left side of the paper quality distribution. Acceptance rates will plummet and journals may rely on other signals of quality (name recognition, pedigree, institution) to thin the herd before actually reviewing content. As always, the rich get richer!

But this is imperfect, not to mention unfair, and so desk rejection gets noisier: good papers get killed by tired editors and marginally lower quality papers slip through to referees. It’s a cascading failure: volume breaks editors, broken editing wastes referees, wasted referees slow science.

…

As mentioned last week, I’ve been meaning to test out a personal agent from the Claw-like ecosystem. I settled on testing out Zeroclaw, a popular and lightweight OpenClaw alternative that should run well on my Raspberry Pi 4 4GB.

I wanted to harden my setup as much as possible and opted to running everything in Docker. I started with the official Docker compose file and added my OpenRouter key. I brought up the pre-built container image and tried sending the basic “Hello” message to the agent using the CLI. However, I got error because the automatically generated config file defaulted to a version of Claude Sonnet 4 that wasn’t available on OpenRouter. I switched to claude-sonnet-4.6 and then gpt-oss-20b (for much cheaper testing).

The Zeroclaw web gateway was a bit of a mess. Of the features I tried, only memory management and the basic status dashboard worked. Trying to talk to the agent through the web interface would give me a black screen (here’s someone complaining about the same error). I’m still being charged for the tokens, though! The cost tracker always displayed zero, even as I sent CLI and Telegram messages (more on that soon). The configuration editor gave me an error and so did the diagnostics tool.

The project docs/wiki were helpful for figuring things out, but development is running so far ahead of releases that a bunch of the features referred to aren’t available in the current stable version (v0.1.7, from last week). This includes getting and setting specific config options from the CLI and resetting the gateway pairing token. To use these features, you have to compile yourself.

It’s been one month since my first post on Big Muddy. There were a few factors driving my decision to start this project:

• I wanted to get into the habit of writing every day.
• My “random interesting links” folder was overflowing, but I wasn’t doing anything with these links.
• My admiration for Simon Willison’s work and his suggestion for everyone to start a blog to share what they learn.

As the saying goes, writing is thinking. Instead of allowing interesting articles, tools, and bits of knowledge to languish in a “temporary” bookmarks folder, I could actually engage with and learn from the material by writing something about each item and make it easier to re-find later. This also forces me to curate the links and ideas that are actually worth saving, since writing a post, even a short one, takes a lot more effort than just throwing a link into a folder.

I figured I might as well share the results with the world, since someone else might find this information useful. And I’m helping to write my ideas and preferences into the next generations of LLMs, I guess.

I’ve made exactly one post per day since starting this blog, which was my goal when I set out. A handful of these posts are pre-written the day before (if I know I won’t have the opportunity to write a post the next day), but most are written the day of. Most are short (Bash tells me just over 190 words on average, though this is slightly inflated by Markdown formatting). Some are very perfunctory, just a link with a few words, when I really needed to get a post out for the day. At the start of this project, I cut my “temporary” bookmarks folder to zero. It has now been replaced with a backlog of links I want to write about on this blog.

Twitter’s “Today’s News” section is a mix of real news, very minor stories (usually discussion of a random AI-related post), nonsense trends, and barely disguised marketing.

The algorithm behind it seems pretty easy to manipulate.

This trending topic revolves around an explosive DM warning of imminent 25% layoffs at a FAANG company:

Here is the original post, which comes from an account called Tech Layoff Tracker (@TechLayoffLover):

There is no reason to believe this post is real. The account, created this month (February 2026), made its first post 7 hours ago. The post in question was made 5 hours ago, or 2 hours after the account’s very first post. Of course, the account carries an utterly meaningless blue “verified” checkmark.

But despite all this, the news summary puts “Tech Layoff Tracker” right in the headline, as if it’s a known reliable source and not an account (most likely) created the same day as the summary itself!

I recently came across the following policy on the submission page of an academic journal:

Use of Artificial Intelligence (AI) tools: One of the goals of Spectrum is to stimulate critical thinking and skill development among authors and reviewers alike. Spectrum discourages the submission of content generated by artificial intelligence (AI)-assisted technologies (such as chatGPT and similar tools). This includes tools that generate text, data, images, figures, or other materials, as well as tools that are used to summarize and synthesize sources. Authors should be aware that such tools are vulnerable to factual inaccuracies, biases, and logical fallacies, and may pose risks to privacy, confidentiality, and copyright.

If authors choose to submit work created with the assistance of AI tools, such use must be disclosed and described in the submission. The disclosure must include: 1) what system was used, 2) who used it, 3) the time/date of the use, 4) the prompt(s) used to generate the content, and 5) the content in the submission that resulted from use of AI tools. The output from the AI system should also be submitted as supplementary material. Authors must accept full responsibility for the accuracy and integrity of the submission. AI systems do not meet the criteria for authorship, and should not be listed as a co-author.

…

Chrys Bader has created ClawCharts to track the popularity and growth of OpenClaw and its growing number of competitors.

I have an unused Raspberry Pi 4 4GB that I’ve been meaning to test one of these Claw-like personal agents on (locked down to prevent the security nightmare scenarios we’ve seen play out since OpenClaw took off).

OpenClaw is a bit of a resource hog (which is why so many people are running out to buy Mac Minis), so I’ve been looking at the list of lightweight competitors. There is no obvious reason to prefer one over the other, so I’ll probably go with the fast-growing ZeroClaw.

ZeroClaw offers OAuth connectors for OpenAI and Anthropic subscription plans, but presently neither company is clear on whether this usage is permissible or not. Anthropic recently blew up the OpenClaw community by updating their docs to specifically ban using OAuth outside of Claude Code. An Anthropic employee partially walked this back on Twitter, but there is still no clear statement whether this use case is permitted. Regarding the use of OAuth from OpenAI for OpenClaw (specifically, GPT Codex), Peter Steinberger, creator of OpenClaw, stated on Twitter: “that already works, OAI publicly said that”. No one can seem to find this public statement, but it’s worth noting that Steinberger himself is now an OpenAI employee. So, will you get banned for using your ChatGPT Plus/Pro or Claude Pro/Max subscriptions with OpenClaw? Nobody knows.

To follow on yesterday’s post on AI-produced research, here is a reflection on “vibe researching” from Prof. Joshua Gans of the University of Toronto’s Rotman School of Management. Since the release of the first “reasoning” models in late 2024, he has gone all in on experimenting with AI-first research.

One of the key takeaways is that he found himself pursuing low quality ideas to completion more often, precisely because the cost of choosing to continue to pursue a questionable idea has been lowered. Sycophancy is a problem, too. With an AI cheerleader, it is easy to convince yourself you have a result when you do not.

Those ideas were all fine but not high quality, and what is worse, I didn’t realise that they weren’t that significant until external referees said so. I didn’t realise it because they were reasonably hard to do, and I was happy to have solved them.

I will note that (human) peer reviewers cannot be the levee that stops the flood of middling AI research: the system of uncompensated labour that undergirds all of academic publishing is already strained to bursting, as every editor desperate to find referees for a paper will tell you.

Prof. Gans concludes his year-long experiment in “vibe researching” was a failure, despite publishing many working papers and publishing a handful of them:

Prof. David Yanagizawa-Drott from the Social Catalyst Lab at the University of Zurich has launched Project APE (Autonomous Policy Evaluation), an end-to-end AI pipeline to generate policy evaluation papers. The vast majority of policies around the world are never rigorously evaluated, so it would certainly be useful if we were able to do so in an automated fashion.

Claude Code is the heart of the project, but other models are used to review the outputs and provide journal-style referee reports. All the coding is done in R (though Python is called in some scripts). Currently, judging is done by Gemini 3 Flash to compare against published research in top economics journals:

Blind comparison: An LLM judge compares two papers without knowing which is AI-generated Position swapping: Each pair is judged twice with paper order swapped to control for bias TrueSkill ratings: Papers accumulate skill ratings that update after each match

The project’s home page lists the AI’s current “win rate” at 3.5% in head-to-head matchups against human-written papers.

Prof. Yanagizawa-Drott says “Currently it requires at a minimum some initial human input for each paper,” although he does not specify exactly what. If we look at initialization.json that can be found in each paper’s directory, we see the following questions with user-provided inputs:

1. Policy domain: What policy area interests you?
2. Method: Which identification method?
3. Data era: Modern or historical data?
4. API keys: Did you configure data API keys?
5. External review: Include external model reviews?
6. Risk appetite: Exploration vs exploitation?
7. Other preferences: Any other preferences or constraints?

The code, reviews, manuscript, and even the results of the initial idea generation process are all available on GitHub. Their immediate goal is to generate a sample of 1,000 papers and run human evaluations on them (at time of posting, there are 264 papers in the GitHub repository).

It’s no secret that Canadian journalism (like journalism everywhere) is in trouble. Newsrooms face a steady stream of layoffs despite a couple hundred million Canadian dollars of direct and indirect government subsidies every year. The vast majority of outlets eligible for these subsidies take advantage of them, and combined they can subsidize half of a journalist’s salary. News organizations are desperate to diversify their revenue streams.

The Hub is a right-leaning publication launched in 2021 with a focus on policy and politics. Notably, the outlet declines or donates their subsidies, citing a valid concern that the scale of such subsidies threaten the perceived trustworthiness and independence of the media.

In late January 2026, The Hub launched NewsBox, an AI-powered publishing tool. NewsBox aims to make it easier for creators to transform their content (written, audio, or video) into other formats, such as speeches, essays, or talking points, while maintaining the author’s distinct voice. You can see examples of the tool’s output on new articles in The Hub, each of which is accompanied by an AI-generated summary and list of quotes at the top of the page. There is also a “Hub AI” chatbot in the sidebar of every article.

The app very much uses The Hub’s branding, prominently featuring the outlet’s co-creators, who also created NewsBox. While their pitch talks about preserving creators’ voices to avoid the “soulless prose” and “slop” outputted by ChatGPT and similar tools, I have to wonder if tighter integration of AI into the news and opinion side of the operation will raise its own issues with trust. The Hub has always been fairly tech-friendly, including a longstanding sponsorship by Meta.