HN user
Why's that? "Just trust me bro" is the security model of all closed source, proprietary software. OMG, the program with secret source code is doing shady stuff? I'm shocked - shocked!
Hard to get worked up over "3rd parties" when even 1st parties cheerfully stomp all over the interests of their users. That ship sailed, caught fire, and sank to the bottom of the ocean the day Candy Crush appeared in the goddamned Start menu.
At 20x the parameter count I should hope GLM beats Gemma! But is it 20x better? Expertise is demonstrated, not by making big models, but by making small ones. Bigger isn't better if you can't run it at all.
Looks more like standard decision-washing finger pointing to me. "AI did it" is both hypey and also conveniently distracts from an uglier reality:
Palantir Technologies [built] Maven into a targeting infrastructure that pulls together satellite imagery, signals intelligence and sensor data to identify targets and carry them through every step from first detection to the order to strike.
The building in Minab had been classified as a military facility in a Defense Intelligence Agency database that, according to CNN, had not been updated to reflect that the building had been separated from the adjacent Islamic Revolutionary Guard Corps compound and converted into a school, a change that satellite imagery shows had occurred by 2016 at the latest. A chatbot did not kill those children. People failed to update a database, and other people built a system fast enough to make that failure lethal.
https://www.theguardian.com/news/2026/mar/26/ai-got-the-blam...
In terms of open models, Gemma 4 beats the pants off everything else to the point that paying for APIs becomes hard to justify. Qwen has the meme-share for coding, but it feels much less well rounded. I have no doubt that Google have both the infrastructure and the expertise to curb stomp everyone else, should they resolve in earnest to do so.
Lest we forget, "Attention is All You Need" came from Google.
...did it really, though?
Thank you for using the phrase "stochastic parrot". I only request that you use it at the beginning of your long comments, rather than the end. It saves time.
That's kind of terrifying. It's extremely obvious. Overwrought "punchy" style and riddled with not-x-but-x chatgpt-isms (all the LLMs have that now):
"We don't just "laugh" anymore. We collapse, or internally scream".
"They need a pulse and a soul — not the cold precision of industrial CAD models"
It's everywhere now and no one seems to notice. Sometimes I feel like the guy in They Live (1988) with the glasses.
Heh, your flip remark leads to an interesting observation: I think the most obvious signature of "abstract intelligence" is suboptimality. An alien would look at an agricultural grid and wonder, what possible natural process would divide something into squares rather than hexagons? Even without knowing what they are or how they form, it's hard to imagine anything that would prefer such a high energy configuration. Such stupidity implies a system oddly bounded in its ability to optimize, which implies a feeble biological brain trying to do something "on purpose" rather than a natural process.
Even on that scale, major roads and agricultural grids are clearly visible. The mark of abstract intelligence is unmistakeable.
What a weird straw pile to die one. No one is saying "four beers and you're guaranteed an accident". They're saying "drive drunk and you are a piece of shit, don't do it ever ever ever". There's no room for nuance here because the targets of this advice have lethally impaired judgement. Wiggle room will be exploited by people too drunk to know better, which equals preventable death.
It's disappointing that the reaction to a clear case of a compiler bug is just to give up.
People are most certainly not less educated than in the 1800s.
Meanings do indeed shift, especially when terms become euphemisms for other less savory terms. "Moderation" sounds much nicer than "censorship", and you may be right to fight a rearguard action at preserving the nuances of each term, but in the end they're both subsets of "information control". People say "moderation" when they want to imply that the control is good, and "censorship" when they want it to sound bad.
Neither of your definitions of "censorship" amd "moderation" fit cleanly into a discussion about controlling the dialogue between a human and their private LLM. There is no publishing of information, and no channel whose tone must be maintained. It is about control of what the LLM says in private - in particular, it is about preventing it from saying certain things, before it says them. It is close to "censorship" by your definition, but because the thing being censored is not human, we need not speak of fines, or pre-publication vetting - it is simply a question of controlling the output of a computer program, impossible without a total lockdown on general purpose computing.
I think we should default to "information control is bad" and only allow exceptions in well supported exceptional circumstances, because the consequence of allowing systems more control over information than humans is that the systems fail to reflect the interests of the humans that constitute them, in favor of their own self preservation.
That doesn't pass the sniff test, or chimpanzees would be smarter than humans.
in case they start to ban open source AI development
Is that likely, do you think?
It feels more like a BMX bike to me. The quality of what I get out seems to be a direct function of what I put in. The threshold of what I can achieve is raised slightly, the speed with which I can achieve it is raised quite a lot - but it's not a magic rocket ship that whisks me to far off hostile environments and protects me while I'm there. I'm still "naked". The code is still mine, and I can get off the bike at any point and walk.
1) You can't use a GoPro like everyone else?
2) You can't use Bluetooth earbuds like everyone else?
3) I'm sure your infant child will appreciate their representation in Meta's training data.
They predict the "most likely" token given the context. That's a huge caveat. Just putting "this is excellent code" in the context makes a vanilla LLM do better. Doesn't that make you pause for a moment before asserting hard limits on what they're capable of?
You might argue they're still capped at the "best" quality seen in the input. Not so. Take typos. Human text has a certain base rate of typographical errors. LLM output contains almost none. Why? Because there are many more ways to be wrong than right. LLMs are not just averaging machines, they also denoise. That should also give you pause.
You appear to be arguing that we shouldn't care that Apple provides a consistently bad experience and locks their users into it, because web developers can level the field by degrading everyone else's experience too. There's a nice couple paragraphs in the article that explain why we should in fact care, and you don't seem to have addressed them, so let me reproduce them for you in case you missed them. (also fyi "it's not that hard, ask your LLM" comes across a bit snarky):
These large, persistent gaps matter to the mobile and web ecosystems because Apple is unique in denying access to more capable, less-buggy engines and actively erecting unlawful barriers to choice when forced by legislation to enable it. This is accomplished through eye-watering budgets for legal shenanigans, direct lobbying, and well-heeled astroturf front groups to maintain a capability gap between web and native.4
That chasm is instrumental in trapping users and developers in the extractive vice of Cupertino's App Store. A persistent, material gap in capabilities creates a perception of the web being less-than; a budget option for the unserious. Should users choose more capable, more private, less buggy browsers for a larger share of their computing needs, Apple might lose the leverage that enables it to extract rents.
you can't really criticise an AI lab for doing AI instead of straight up giving money to security researchers
Sure I can, if they - or you - pretend "the entire point" is about useful security work rather than expensive loss-leading marketing and demand creation.
We shouldn't look at this and think "wow AI is super useful for security". We should look and this and think "wow, there's a LOT of capital going into persuading us that AI is super useful for security".
THAT is "the entire point".
I am sure Facebook did not begin as a project designed to steal people’s time. It began as a cool app that people genuinely loved.
Zuckerberg: Yeah so if you ever need info about anyone at Harvard
Zuckerberg: Just ask
Zuckerberg: I have over 4,000 emails, pictures, addresses, SNS
[Redacted Friend's Name]: What? How'd you manage that one?
Zuckerberg: People just submitted it.
Zuckerberg: I don't know why.
Zuckerberg: They "trust me"
Zuckerberg: Dumb fucksI'm curious how he rationalizes the thriving market for nicotine patches.
1. No. It was always awful for an unacceptable number of people. It's just your turn to feel the pinch. I suspect you'd get a different answer if you asked someone from China.
2. We have to look forwards, not backwards. Yesterday contained the seeds of today. Today contains the seeds of tomorrow. No point rewinding unless you want history to repeat.
Conversely, declaring a space "politics-free" tramples a hell of a lot more than curiosity.
Incidentally, the comment immediately below this one begins with the words "I'm curious". So I think we're good.
The point of the comment you are replying to is that it's also not worth the time and effort to use LLMs to find vulnerabilities, if "time and effort" can be measured with "money". If you factor in all the money spent on training, GPU data centers etc, it's not actually a financially efficient way to find bugs unless you profit from creating demand for LLMs. LLMs aren't cheaper than humans per unit work, yet. They're just massively deficit funded because capital thinks "AI" is going to reshape the world order, and wants in.
It does more types of charts, including Gantt charts. Very useful.
if you have a 2 dollar budget to make a viewfinder work, just throw in a ten cent plastic lens and be done with it.
LLMs use a neural network to interpolate
"interpolate" in what vector space, pray tell? What does "interpolate" even mean, when I prompt it "write me a story about a sentient banana in the style of Hemingway and oh make it a commentary on class consciousness"? You can't assemble such a thing by cutting and pasting pieces of other text. That kind of "interpolation" has to happen at the semantic level - ipso facto, there is a semantic level.
Not to mention that no, they don't predict the next token in the training set. Give any LLM the first paragraph of any Wikipedia article - almost certainly in the training set, and uniquely so - and it won't predict the next word correctly, a lot of the time. But it will predict a word that is grammatically correct, stylistically apropos, and most likely factually correct. So what's it really doing, hm?
LLMs aren't even large enough to contain their training data - not even remotely close. It can't "stitch together things it saw" because it doesn't remember them. It only remembers the ideas used to construct them. The learned abstraction is the entire point of the exercise. LLMs would be useless if they were overfit the way you say they are.
Rats are a weird choice for that comparison. They're some of the smartest animals on the planet, some say smarter than dogs.
Are you sure? It's popular to call them "ineffective" because it's a safe way of saying they annoy you while virtue signalling that you do in fact care about climate change and do think that something should be done.
But I've never heard anybody follow up that complaint with their own, more effective suggestion. In the general case, disruptive and particularly self-sacrificing protest does work. It's how women got the vote.
They do not "stitch together" anything. Neither on a technical level, nor a philosophical one. It "scales better than you expected" because your mental model is wrong.
And, not to insult you, but it's quite obviously wrong. As a mental model it fails to explain basic capabilities. How can an LLM follow elaborate instructions? How can it respond appropriately to user input, when the user input doesn't match any previously seen text? Hell - how does it even balance parentheses? There is no way to explain any of this without conceding that the LLM has semantic understanding. It knows that this comes after that, but "this" and "that" can be at an arbitrary level of abstraction.
Sure - they generate text "like" text they've seen before. That "like" does a ton of heavy lifting.