edot

Yeah, but the bloat is “correct” per the manufacturer. I view it like car parts or other things where the OEM (original equipment manufacturer) recommends certain things. Besides, they have the most training data and incentive to get their harness working as well as possible with their LLM.

Honestly? It’s the shape that makes it clearly AI. That’s the quiet admission at the heart of the problem.

I’m starting to come to the opposite approach: don’t try to customize anything, just use it vanilla, and use the best model you can afford. No AGENTS.md, no special subagents or roles, nothing but a few convenience skills which are really just textexpander. Use the harness that the LLM provider makes, and that’s it.

Making a huge custom setup is so 2025.

Not true. You can turn off training in settings on an individual plan.

No, you can’t see ships with this satellite. Too small. Besides, Russia and China have way better satellites that they’ll happily share intel from with Cuba.

Agreed. Deburring tools also frequently “catch” and dig in if you’re not careful. Especially on soft aluminum. Use a file if you’re going to do this.

Right, but she is not pathologically like the other billionaires. She married a man who became one. Same with Melinda Gates. So I guess our hope is that more billionaires will lust after young Russians or hot newscasters and get divorced, so their normal wives can do what normal people do with more money than they’d even need and help others.

What are they / you using it for in such quantities?

And OpenAI didn't try to silently degrade performance of their top model if its (extremely sensitive) safety sensors went off ...

Interesting take. I assume that, because corporations are a subset of “groups of people” that this applies to any group? The theory here being that, in the same way that LLMs produce the average of everything they’ve trained on, a group of people produces the average of every individual’s ideas and efforts (subject to things like some people having more power than others, etc.)?

I don’t think this is always true, however. It’s provably false, actually, because I can name at least one group of people that has done something unique and not sloppy. For example, the C8 Corvette is pretty sweet. So is Costco’s supply chain and employment practices. And I also think Pink Floyd’s music is pretty non-average.

Because papers are often referred to by the first author’s name, and often the first author is the primary researcher and therefore deserves the extra credit. When two or more primary authors are equally involved, they’ll often do a random ordering but annotate this so that no one thinks one did more than the others.

This is true in any power structure ever. Kings, mafia, pick-your-dictatorship, many democracies. Hurt or steal from the poor, not other rich or elite, and always make sure to kick some up to the big guys.

Yeah, and arbitrage. Arbitrage is exploiting the difference in prices of the same asset between two markets. Arbitrage is also risk-free or darn close to it. Entrepreneurship is anything but. Arbitrage is not "gee this product doesn't exist, I'll start a company and invent it and manufacture it and sell it" ...

This article feels like LLM but I can't put my finger on it. It reads like LLM where a human asked an LLM to cover up the LLM-isms. It's odd because the byline of "Drawing on Tyler Cowen & Alex Tabarrok’s Modern Principles of Economics and a century of empirical research." doesn't say who the author is, but the URL root is Alex Tabarrok's profile page.

I feel as though someone (Alex Tabarrok?) said "here's reference material - turn it into an essay".

The footer also feels how an LLM sounds when you tell it to follow specific instructions - they often seem to put too many details about how they did what they did (or how you should feel about it) into their output, kind of how a slimy salesman would explain some fancy shoes and how the cows lived before they were turned into leather: "Set in Newsreader. Layout and graphics after Edward Tufte — data-ink, direct labels, sidenotes. Colors after Sanzo Wada, A Dictionary of Color Combinations (1933). All charts are hand-drawn SVG; all numbers are from the sources above."

Some other clues: "Rent control doesn’t repeal scarcity. It just deletes the signal that manages it." "A rent is not merely a bill. It is a message flashed between strangers." "The gap between those two curves is the shortage." (I find that LLMs often use this format with slightly-misused, yet fancy-sounding words). The X between the Y is the Z. And the best part is, the graph below (which is allegedly "hand-drawn SVG" - hand-drawn by who, I ask) is straight lines, not curves ... that same graph also has the m in market overlapping one of the lines, which is a sloppy mistake an LLM does because it can't see, but a human can and should have seen because it's "hand drawn". How do you hand-draw an SVG that updates with inputs, anyways? "“Chance and favoritism” was not a rhetorical flourish. It was a prediction about who gets housing when prices can’t decide" "every fix works by weakening the control, which is the quiet admission at the heart of the policy" - humans don't write like this. I've never heard of a quiet admission at the heart of anything. Woah, interesting, if you search for "quiet admission at the heart of" on Google, you'll get a bunch of other AI articles. Neat, I found another indicator!

It's really quite an interesting topic and there probably is some real research behind it, but dude ... just put a little header at the top saying you used AI to help you write or do graphs and I'll trust you. Hiding it makes me question everything.

There are four distinct firings. Look for the little yellow poofs. The pistons move in pairs for balance (vibrations). Each piston is doing one of the four strokes (suck, squeeze, bang, blow is the crude way to remember) at a given time. None of them are doing the same thing at a given time.

You can tell these guys know nothing about LLMs or how they’re provided. I love how they show OpenRouter’s graph of token usage as if it speaks for usage across the board. DeepSeek looks like the king because people who use Anthropic and OpenAI use them on either a direct basis or AWS Bedrock …

And the bar chart for token costs, really? As if that’s information? Their sources are the API docs ffs. If they had at least modeled something to estimate token costs that would be interesting, but showing the public prices and calling it research is dumb.

There is no technological fix for the issue you raise. The issue being parents not wanting to bother raising their kids, and thus giving the state and corporations control over what they can and can’t do. That’s a cultural issue. No idea how to solve it.

You are missing half of the story. This is not “caring legislators punishing big bad tech”. This IS big bad tech. Meta has spent $2B lobbying for this. More than wanting to get kids addicted, Big Tech and the intelligence community wants perfect observability into online activities.

This is a win/win for big tech. If they don’t get age verification, they can keep getting kids addicted to propaganda and consumerism. If they do get age verification, they get to see what everyone in the world thinks and is interested in, all linked to government ID.

Edit: the one outcome big tech does not want is anonymous age verification. This is technologically extremely possible, but that would be a lose/lose for big tech because they would lose kid (aka future consuming adult) addiction AND lose perfect tracking linked to government IDs.

If everyone else jumps off a bridge, do you want to be forced to jump off a bridge?

George E.P. Box for anyone wondering. So true. “All models are wrong, but some are useful”. True of any abstraction unless you’re doing everything at the quantum level, but even then … are you sure you’ve captured all the details? But then … does it matter? As long as your model is sufficient to do the thing you need to do, it’s useful enough. Don’t make it any more complicated.

This triggered Opus 4.8 the other day for me. Said “nuke that folder” and it said I was violating TOS.

Yep, I manually listened to the meeting recordings (easy to find the spots based on the transcript timestamps) for any quotes. There are also meeting minutes and agendas with supporting docs to corroborate against (e.g. for dollar amounts). They really don’t make stuff up all the time if you root them in data.

I rolled my own. I hadn’t heard of this one, but I looked into stuff like OpenStates (now privately for-profit owned, ugh). My city just uses a Wordpress site so it’s structured enough. I’m looking at building something to ingest cities with Granicus and one other big local government meeting recorder via API whose name I forget. That should get decent coverage. There’s no way to catch the long tail of every local government’s recording process. Some cities people will just have to do manually. But it’s easy enough with LLM help.

Citizens monitoring their government is literally THE foundation of democracy (ok, maybe voting comes before it, but then you have to monitor who you voted for to see if they’re doing what you voted for).

Nope, I used a minute fraction of the technology they have, along with open records as is my right in this country, to stand up for my Fourth Amendment right to travel without creeps stalking my every move. I need to make my specific framework a bit more generic and then I'll put it here on HN. Or just offer a platform where people can bring an OR key and it can run on their city.

For some things, yes. But I'm half-assing some really cool stuff right now. Made a scraper to pull my city's meeting minutes, agendas, recordings, made transcripts. Regex for "Flock", found every mention, passed those files into a cheap model (DeepSeek V4), had an understanding of who in my city is down with building the surveillance state and who isn't. I've got research on everyone, and had emails drafted for each one based on what they said. Quotes and figures and all. I lightly polished each email and fired 'em off. Already got some replies back. Plenty more in the quiver too (pulled and analyzed CSVs of FOIA'd datasets).

If they're gonna spy on me with AI cameras, I can oppose them with AI research. :)

Agreed. I’ve cancelled all plans except OpenCode Go. OpenRouter for API spend. Feels so nice to a) not feel like I have to code when I don’t want to just because I need to make my subscription worth the cost and b) know that this level of performance won’t be yanked away. Super pleased with DeepSeek V4 Pro.

I don't quite follow his point. Is it: a) that we need a new foundational algorithm that integrates a goal (one with "taste") directly into the training step, or b) that we need to point trained models towards goals as they iterate?

If it's a), he doesn't propose such an algorithm, and I don't know how you'd do it at such a low level because how do you quantify abstract goals? Did he suggest such an algorithm and I misread? If it's b), that already exists, see AlphaEvolve or any number of things he said. Or, to be a bit of a smart-ass, just type /goal and let it rip ...

I also think he's just categorically wrong that LLMs cannot do good and novel things. And if it can, then you could just say "well that's not novel, that's derivative". A simple example, if I make up a programming language with an LLM and it works well for my purposes, then is that not novel and good? I mean, is any language other than FORTRAN not novel?

Everything is derivative and you can put an LLM in a loop to evaluate LLMs trying things. I must be misunderstanding because he's too smart to be this wrong.

Same here when I said to “nuke” a process.

Wow, this is horrible. Local LLMs are the future. Thanks, China! Seriously crazy that I’m saying that, but the American companies are being so anti-freedom they’re making the CCP look libertarian.

Also, Fable’s sensing is hypersensitive. Feels like they just have regex for phrases. No nuance. If I say I’m working on something using “GPUs to train” xyz then, will that trigger this sneaky silent screw-my-stuff-up mode?