When I went browsing through their GitHub I was surprised at how little web-specific code they have. It's basically just their React Native mobile app and a tiny go server. I understand that with a small team they've got to prioritize, I do hope at some point they implement server-side rendering for when you click on a direct link to a post.
So far each big, unpopular decision by Bezos has led to significant cancellations of paid subscriptions. Since it's happened more than once I don't think it's safe to assume the remaining subscribers are happy either.
Bezos cancels planned endorsement of Harris: 300,000 paid subscribers cancel in period up to election. (400,000 new paid subscribers over the period, but they offered significant promotional discounts)
Bezos tweets change to opinion policy: 75,000 paid subscribers cancel within four days
With in the order of 2.5 million paid subscribers before all this started that's significant losses.
Look at the article currently promoted at the top of Post opinion page: "Trump is off to a good start with an AI action plan" https://archive.is/ERCme
Regardless of what you think of the quality of that opinion, it took very little effort to make.
Compare the sources they used to the work it would take go out on the ground and do novel research:
- Their own news article about it (itself based on press releases and an off-the-record comment that obviously would have come from someone in the White House press office assigned to promote the press release)
- Their own past opinion pieces
- Reuters.com
- WhiteHouse.gov
- Online govt statistics
- CNN.com
- NeurIPS' blog
- Columbia Business School blog
- Matthew Yglesias' blog
- Greg Lukanioff's blog
I could have found those sources based on vague memories of tweets I've seen by following journalists on Bluesky and a few hours of googling. I suspect they did the same, except they used X instead.
It feels analogous to what would happen if you put me in front of a broken project without source control that I've never seen before and asked me to fix it without giving me enough time to actually understand it. It starts from errors and bugs, guesses corresponding source code, and tries to narrowly fix just that snippet. Generally it favors deleting, but not specifically deleting new code.
I would have thought it could record a log of its actions and use that log to think about undoing. I would also think it could annotate lines with git blame so it knows undoing wouldn't involve changing anything more than say a day old. Unfortunately that isn't consistent with what I've seen.
I just make a WIP git commit and run git commit -A --amend --no-edit after manually reviewing each unit of work.
Edit: I also wish Claude implemented undo at a higher level instead of relying on the model. Some combination of snapshotting the whole repo and tracking operations that have precise inverses. But I understand that would have drawbacks.
maybe a system prompt to tell it to do checkpoints. Stash is one way, another is jujutsu with git backend. If this was Claude Code then hooks would be the perfect place to put that logic.
I have a gut feeling that writing this kind of MCP server can't be the future of software development. I'd expect a two year old AI model to need this kind of handholding, but I don't understand why it's still necessary.
Couldn't any modern AI model know that Zig docs are relevant to the question, figure out how to find the docs, write some code to parse it, and guess how frequently to update it's cache?
I expect there to be plenty of problems AI can't write for the foreseeable future but they have a very different vibe from this.
Edit: I just asked Claude Sonnet 4 to pretend it has a tool that makes docs available that has an update frequency parameter. It said the zig stdlib should be updated weekly but the Java stdlib would only need quarterly. Seems reasonable to me.
That's exactly what zig-mcp does. It gives the LLM tools like search_std_lib, get_std_lib_item, and it decides when and how to use them based on context
Stop thinking of models as all-knowing oracles; they aren't and shouldn't be. They are reasoning engines. You will always need to provide relevant and up to date context for them to work with as ground level truth, and the quality of your results will be a function of the density and quality of that context, just as a human developer is far more effective with access to web search and documentation.
> Second date has capabilities of network eavesdropping, MiTM, and code injection
This is probably a dumb question but doesn't that require an SSL cert? Obviously the NSA can get someone to issue a cert for a domain they don't own but wouldn't that be visible?
Couldn't you have every user device log the SSL certs it sees to detect this attack? What about CT?
I think they meant relative to the best other approach, which is Reducto’s given that they are the creators of the benchmark:
Reducto's own model currently outperforms Gemini Flash 2.0 on this benchmark (0.90 vs 0.84). However, as we review the lower-performing examples, most discrepancies turn out to be minor structural variations that would not materially affect an LLM’s understanding of the table.
Why should we question that? The most obvious answer would be that women who have a hard time giving birth are more likely to get painkillers and less likely to want a second child.
Anyone who cares can find those satellites again using the same techniques. I guarantee the Chinese government knows their exact orbital parameters. There are no stealth satellites.
Probably. The Iranians certainly have the technical capability to detect those satellites, at least well enough to calculate the orbital parameters. So it would be safe to assume that they pass that intelligence on to the various terrorist groups that they sponsor.
