rfw300

Germany’s austerity policy after 2008 may be one of the largest economic blunders in history. It would be one thing if they merely committed self-harm, but they also used their pull in the EU to drag the rest of the continent down with them.

Reminded of Matt Yglesias’s excellent headline from 2010: Angela Merkel Lucky the Bar for “Worst German Leader” is Very High

if a malicious actor can weaponize an agent to do their bidding

In my experience, human employees are much more vulnerable to this particular weakness than frontier agents (i.e. phishing attacks).

I understand that the’ve written zero lines of code for this application, but would it kill them to write a few lines of the blog post by hand?

Forcing readers to wade through an unceasing string of LLM clichés demonstrates the opposite of the point you’re trying to make—that the consumers of your work are worse off because you exercised no human judgment in creating it.

A law professor studying AI has an affiliation with the center at their university that studies applications of AI? Scandalous!

A chapeau is not "just like another title basically". It's a lead-in, a phrase which acts as the grammatical start of a sentence which the following subsections finish. For instance, the text in the first paragraph of 18 U.S.C § 3632(a) which ends in an em-dash is a chapeau. Taking pride in work you have not done and not bothered to understand is perplexing.

The author (author's operator?) does not understand the data they are working with. And in doing so, they inadvertently make the case against their own "dark factory" nonsense.

For one, nothing about this project makes "every law" a commit. It just takes the _annual_ snapshots published by the House clerk and diffs chunks of those files against each other. A project which actually traced the edits in each annual snapshot to a specific passed bill would be incredibly cool (and is probably tractable now for the first time with current AI agents). This is not that!

All this does, as far as I can tell, is parse a set of well-structured XML files into chunks and commit those chunks to Git. It's not literally nothing, but it's something that the author's own README credits multiple people doing years ago with ~100 line Python scripts.

I don't mean to be overly harsh. But this is exactly the problem with treating your software as a "factory": you release something you do not understand, in a domain you did not care to learn. And we are all the poorer for it.

What is a "truly new task"? Does there exist such a thing? What's an example of one?

Everything we do builds on top of what's already been done. When I write a new program, I'm composing a bunch of heuristics and tricks I've learned from previous programs. When a mathematician approaches an open problem, they use the tactics they've developed from their experience. When Newton derived the laws of physics, he stood on the shoulders of giants. Sure, some approaches are more or less novel, but it's a difference in degree, not kind. There's no magical firebreak to separate what AI is doing or will do, and the things the most talented humans do.

I don't understand why their "Instant Grep + roundtrip to us-east-1" is so slow. First of all, the round-trip latency should not be nearly so bad to us-east-1. But second, and much more importantly, the LLM runs in the cloud. Shouldn't you just situate the LLM, agent runtime, and regex index in the same region? Wouldn't that be faster than round-tripping to the user's local machine?

On those terms, they also wasted a lot of cash. 90% of it went to candidates who lost (or opposing candidates who won).

In fact, looking at the blog post, the agent orchestrating 16 GPUs is half as efficient as the agent using 1 GPU in GPU-time. Since it uses 16 GPUs to reach the same result as 1 GPU in 1/8 of the time.

Yeah, assuming there's no active monitoring during the training runs, you can trivially give the agent an abstraction which turns "1 GPU" into "16 GPUs" that just so happens to take 16x the wall-clock time to run.

Do you have a sense of whether these validation loss improvements are leading to generalized performance uplifts? From afar I can't tell whether these are broadly useful new ideas or just industrialized overfitting on a particular (model, dataset, hardware) tuple.

Super interesting study. One curious thing I've noticed is that coding agents tend to increase the code complexity of a project, but simultaneously massively reduce the cost of that code complexity.

If a module becomes unsustainably complex, I can ask Claude questions about it, have it write tests and scripts that empirically demonstrate the code's behavior, and worse comes to worst, rip out that code entirely and replace it with something better in a fraction of the time it used to take.

That's not to say complexity isn't bad anymore—the paper's findings on diminishing returns on velocity seem well-grounded and plausible. But while the newest (post-Nov. 2025) models often make inadvisable design decisions, they rarely do things that are outright wrong or hallucinated anymore. That makes them much more useful for cleaning up old messes.

I don’t necessarily endorse the author’s broad conclusions about “AI”, but I will say that the Spotify DJ specifically is an enragingly bad product. Nothing close to the utility of Claude Code.

I've no problem with the intuition. But I would hope for a lot more focus in the marketing materials on proving the (statistical) correctness of the implementation. 15% better inference speed is not worth it to use a completely unknown inference engine not tested across a wide range of generation scenarios.

OK... we need way more information than this to validate this claim! I can run Qwen-8B at 1 billion tokens per second if you don't check the model's output quality. No information is given about the source code, correctness, batching, benchmark results, quantization, etc. etc. etc.

More likely: this is a transitional phase where our previously hard problems become easy, and we will soon set our sights on new and much harder problems. The pinnacle of creative achievement in the universe is probably not 2010s B2B SaaS.

It is entirely possible, however, that human beings will not be the primary drivers of progress on those problems.

I did, and yet I also felt more relaxed reading it than I am reading most blog entries posted on here. I didn't feel like I had to guard against my time being wasted by vacuous LLM fiction.

Being wealthy solves virtually all problems of consumption, so the invisible hand provides new problems to serve the market need. Beautiful, really.

Why should it be? The agent session is a messy intermediate output, not an artifact that should be part of the final product. If the "why" of a code change is important, have your agent write a commit message or a documentation file that is polished and intended for consumption.

Making those tools first-class primitives is good for (human) UX: you see the diffs inline, you can add custom rules and hooks that trigger on certain files being edited, etc.

If I had to bet, there will be some kind of face-saving climbdown by the end of next week. But all I can do right now is read the words on the page.

I don't think he got it backwards, at least if Hegseth's statement is accurate. AWS, GCP, etc. all do business with DoD. If they, as DoD contractors, are no longer allowed to do business with Anthropic, then presumably they have to stop re-selling or hosting Anthropic's models to anyone.

More generally, Anthropic's reliability track record for a company which claims to have solved coding is astonishingly poor. Just look at their status page - https://status.claude.com/ - multiple severe incidents, every day. And that's to say nothing of the constant stream of bugs for simple behavior in the desktop app, Claude Code, their various IDE integrations, the tools they offer in the API, and so on.

Their models are so good that they make dealing with the rest all worth it. But if I were a non-research engineer at Anthropic, I wouldn't strut around gloating. I'd hide my head in a paper bag.

I have little doubt where things are going, but the irony of the way they communicate versus the quality of their actual product is palpable.

Claude Code (the product, not the underlying model) has been one of the buggiest, least polished products I have ever used. And it's not exactly rocket science to begin with. Maybe they should try writing slightly less than 100% of their code with AI?

It also strikes me as being in competition with, you know, a group chat.

Well if only we had such thoughtful minds in the education space. How did no one ever think of "why don't we just give up on the children"?

This strikes me as entirely logical in the short run, and an insane way of packaging software that we will certainly regret in the long run.

Relevant xkcd: https://xkcd.com/1958/

Interesting project, but the lack of any actual benchmark results on existing models/agents is disappointing.