solid_fuel

I’m certain the next big frontier model will also be able to improve the weights of comparatively tiny models. That does not improve the large frontier model.

> This is the "recursive self-improvement" they've been warning about.

Well, except for the self part since it improved the performance of a smaller, dumber network.

And the recursive part, since it couldn't improve the performance of a larger network.

Shipping more feature doesn't make you more secure. Taking the time to evaluate and test what you are shipping does. The obvious conclusion is that if you want to go all-in on AI for some reason, you do so by using LLMs to do more testing in the same amount of time you would have used before. You supplement human expertise with LLMS.

You don't just yell at people to ship features faster, with less testing.

We used to understand that code was harder to read than it is to write. It follows, directly, that better engineering discipline means shipping LESS code. Less to support, less to understand, means a product that is easier to maintain and extend.

Now, token calculators spew out code at a pace no human can match, which completely removes the barrier of effort - which was one of the largest natural counters to shipping massive amounts of slop.

I'll be impressed by a coding model when it writes less code than a human, and spends more time refactoring than it does just adding new layers. Until then, this vibecoding is just madness.

There's a lot of ground between

"can technically read your emails but is motivated to protect your data"

and

"is an advertising company that mines your emails to sell ads targeting you, specifically".

> How do you compete if your competition is willing to move fast and break things?

You focus on providing something that isn't broken. It's really that simple.

If the docs are vibecoded too that would explain why there's a link to the compromised "not official" cursor version of the extension here https://code.claude.com/docs/en/vs-code which they supposedly aren't responsible for.

It sure seems like they've been vibecoding with everyone asleep at the wheel for a while.

If you look at job listings right now, there's a ton of "we apply an LLM to your business use case" type startups that really don't have good products or fundamentals, but that are getting a lot of funding because "AI".

They use LLMs, often have flashy demos, but can't keep things aligned and struggle to keep the models on task. E.g. they have burned a ton of capital and don't have a lot to show for it.

The ones that went all-in on vibe coding are even worse off.

If those startups die, the big inference providers lose a lot of customers.

Abolishing a criminal organization and jailing the unqualified thugs who have spent the last 2 years terrorizing and murdering citizens in the street does not mean "quit enforcing immigration".

ICE was created in 2002 as part of the notorious Homeland Security Act. Do you think immigration law was just... not enforced before that?

Hint: Immigration law was previously handled by an agency called the Immigration and Naturalization Service, with customs enforcement being handled by US Customs. ICE has shown complete disregard for the Constitution, a common issue with the agencies created by the Homeland Security Act.

Dismantling ICE does not mean stopping border enforcement. It means decentralizing that power and devolving immigration enforcement to agencies better equipped to handle it while protecting the rights of American Citizens.

If you go back and read the comment you are responding to, you will see this text:

> I assure you, it is perfectly possible to enforce border and immigration policies without engaging in wanton cruelty or violations of civil rights.

The false equivalence, then, is right here:

> If there’s no immigration enforcement, that’s functionally equivalent to having open borders.

Where you again just skip over the enumerated criticisms and pretend that all border enforcement must be done in such an aggressive and violent manner without regard for constitutionally guaranteed rights.

If you don't want to acknowledge the difference between "enforce immigration law in a civil and reasonable manner" and "terrorize neighborhoods and execute citizens to make a point", then I can't help you.

> The Minnesota 15 are not accused of ordinary protest. The article gave no evidence that these people hadn't committed any crimes it just went on and on about investigative capabilities of HSI.

Actually, the way the criminal justice system works here in America is that you are presumed innocent until proven guilty. It's not the job of the newspaper to exonerate the victims of government overreach.

Do you think the only way to enforce civil infractions, like immigration violations, is by raiding an apartment building by helicopter at 1 AM?

Do you think the appropriate way to handle people SUSPECTED of having an expired visa is by throwing them in an overseas prison without trial?

Do you think it's RIGHT to throw pregnant women into a prison camp and deny them medical care, just because something was filled out wrong on their immigration paperwork?

I assure you, it is perfectly possible to enforce border and immigration policies without engaging in wanton cruelty or violations of civil rights.

Here's some advice for future debates: drop the petty whataboutisms, quit making strawmen to attack, and address the actual criticisms.

Literally none of them, but it won't matter to the "both sides" types or the right wingers because it's about their feelings not about the facts of what actually happened during COVID or how many people the TEMPORARY restrictions saved.

>> The DoJ is literally run by trump’s personal lawyer. <<

> As to the political makeup at DOJ, information about political donations at federal agencies is public.

I'm sure many of the rank-and-file at the DOJ DO donate to democrats these days, after all the republican party has demonstrated rank criminality across the last ~ 15 years, which is only getting worse. Anyone at the DOJ who still cares about criminal justice and the rule of law is going to likely support the only remaining party that isn't bending the knee to criminals and con men.

It still doesn't change my original criticism - that the DOJ is run by Trump's personal lawyer - and it doesn't make your whataboutism suddenly relevant.

> And it's not because nobody before Trump ever did anything rising to the level of paying off a porn star and calling it "legal expenses." It's because we have had a taboo against prosecuting politicians.

Are you somehow under the impression that is the actual complaint about illegal behavior and weaponization of the government? You won't be able to contribute meaningfully to this discussion if you insist on pretending that the hush money payment is the limit of the criminality being discussed here.

Please go do some research into the recent actions of the DoJ before trying to defend them like this, irrelevant whataboutisms will not move this conversation forward.

As expected, you ignore the actual complaint and just draw specious comparisons, making unsubstantiated claims that "most" of the people at the DoJ are totally democrats, trust me dude.

I'll just leave this here, since this was the actual complaint: https://en.wikipedia.org/wiki/Todd_Blanche

After the crap they pulled in the latest Brazilian election, there is no reason to give Facebook the benefit of the doubt. They are obviously and intentionally trying to use their extreme influence over online communities to influence elections.

Watch for more fuckery during the current US election season.

The DoJ is literally run by trump’s personal lawyer. You can play this game where you pretend it’s all the same but the rest of us have eyes too.

Wow, maybe the republicans should have considered the consequences before letting the Trump administration weaponize the DOJ.

They didn’t, and they own the consequences.

Weird to get all bent out of shape about actual, research backed science in a Scientific American article just because your gut feeling says otherwise.

I'm the opposite - usually good at skimming, close reading takes more effort. I find LLM generated text is basically unintelligible for both.

It you don’t think national secrets should be sold on ebay to personally enrich whoever happens to hold office, good news: we are in the middle of election season and now is a great time to do something about it.

Let this serve as a reminder to vote! Removing this dementia addled criminal and his supporters from power is the patriotic thing to do if you care about this country and its future.

> But the way it is going, if AI is writing all the code and no human needs to read it anymore, does it really need to be maintainable?

It's more important than ever, because code is just a way of expressing intent. If the code becomes overly complex, it de-coheres, and intent is lost across the system you are building.

This is why vibe coded projects turn into slop and eventually stop iterating: at some point that actual intent is too vague and contradictory because of the many layers of code that have been added.

> For me, I can't read books at all, because I forget what was said two paragraphs before. I bet people that are good at skimming text have less problems with AI generated text.

Actually LLM generated text is almost impossible to skim. It's too flowery and the actual meaning is spread too thinly across the text, which means that you can't just read the "important" sentences of a paragraph.

For those who can't view the link due to the continuing degradation of the twitter UX, this is just the world's richest man making vague threats to our elected representatives about "coming after the ones who pull the strings" over a recent report that democrats in congress intend to hold him accountable for his crimes at Doge and elsewhere.

All I have to say is: If you can't do the time, don't do the crime.

He should have worried about consequences when he was busy illegally dismantling USAID and terrorizing our civil servants.

Only in the context of something like RAG - there, you can attach provenance information to data included in the prompt.

With training data, however, no such annotation is possible.

I would rather keep things on postgres but everything wanting its own postgres container is really annoying. Postgres already has support for multiple databases, users, and access controls. I would personally prefer to just set up one postgres instance with proper backups and have everything run against that, unfortunately some self-hosting projects have fairly poor support for that.

Well yes, but that's the point entirely - you can easily put a binary into a docker container, or a jail, or on a VM. But when something is shipped as 5 different docker containers it is MUCH harder to extract it into something you can run without docker or in a different configuration (like sharing a postgres instance between multiple applications).

One distribution method is flexible and can be deployed however you want, the other prescribes solutions and makes any customization difficult.

You might not like hearing it, but it is true.

Code has never been the bottleneck for a skilled engineer, and if you value high quality then you need to spend a lot of time reading over the slop LLMs produce to turn it into something hardened and reliable. So, for a skilled engineer, LLMs have a very marginal benefit in terms of delivery speed.

This is backed up by LLM productivity studies.

> Here the data model is a direct copy, down to the terminology.

That’s not surprising to me - they designed and built it using their LLMs, of course it’s duplicating prior work. That’s essentially the core of what LLMs do, after all.