jameshart
jameshart on Hacker News
Just because a game could graphically have been displayed on an N64 doesn’t mean the game could be built. Vampire Survivors has hundreds of units on screen at a time firing hundreds of projectiles and weapon effects; running those on a memory bandwidth-constrained 93MHz single core processor in 4MB of RAM would be, uh… a challenge.
How are the weights even being mapped into token space as input? Are they bypassing tokenization so the weights just come in as what must seem like gibberish? Are they stringified as numerals then tokenized?
Ah yes, I heard about this ‘pirate code’.
Only I got the impression the Code is more what you'd call guidelines than actual rules
To be fair the Spanish word for security is the same as the Spanish word for safety: seguridad. It could as easily be a ‘safety’ alarm and a mistranslation.
‘AI’ isn’t giving unsourced answers. Hastily cobbled together chatbots based on LLMs, inadequately prompted and without adequate research tools, are being put on the internet
by humans and giving unsourced answers.
Well prompted LLMs with appropriate tools can do exactly what OP asked for. So AI can do this. That the AI-based products people are being given don’t is not a limitation of AI, it is a failure of humans.
What if the ICMP packets are coming from
inside the datacenter?
Consider a shared or cloud hosting scenario for example.
Then maybe phrase it less dismissively?
‘It’s in the dataset’ sort of implies ‘the model is just reproducing output it was trained on’.
In this case you’re saying ‘the model is repeating tasks it was trained to do’
That’s the difference between accusing someone of just regurgitating something they read, vs praising them for having prepared well for an exam.
That rather appears to be what they have done, given that the company with lowest market cap per employee has a circle full of people.
Well then ‘it’s in the dataset’ doesn’t matter. An agent can just pull down the code as part of its research when it’s tasked with a similar rewrite. That seems far more useful.
Games work in very different ways. Some are narratives where unless you complete them it isn’t really true to say you’ve fully played them. Others are essentially skill-acquisition challenges - you don’t have to finish Sonic the Hedgehog or ‘beat’ Civilization to be able to say you’ve played them. And others are goalless sandboxes - the Sims or Minecraft, where what you get out of it depends entirely on what you put into it.
This sort of ‘it’s probably in the dataset’ thinking is a bit of a relic from the ‘stochastic parrot’ era of LLM critique.
Reinforcement learning doesn’t work by training an LLM on a series of inputs like ‘rewrite python in befunge’ paired with a prepared model implementation. It doesn’t matter if an open source implementation was in the dataset that trained the language model. All it got from that was an appreciation for syntax and language idiom; when it is generating code it’s doing so according to the approaches and patterns that it’s been reinforced into following in the context of a coding prompt - regurgitating existing codebases is undesirable so it is a behavior it will have been reinforced against.
If you manage to reverse engineer the hardware of your iPhone I don’t believe Apple can prevent you doing anything you like with it.
“I want root access to my phone” is a nonsensical statement on some level. It’s a device. The fact you have mistaken it for a Unix machine is your own delusion.
Did people used to get like this about pocket calculators or digital watches? “Why am I forced to be limited to only the manufacturer’s firmware?”
‘A regulator once told one business that it was okay to do one thing’ is very rarely any indication that that same regulator would take the same position for another business in another situation. Let alone that regulators in general will take the same position.
Yes that would never happen to the internet
It’s more that both the club and the area are independently named after the site of the stadium where the club used to play. The club would still be called Crystal Palace even if the name had died out for the area of town.
Anthropic and OpenAI models are available via a number of different channels eg AWS, GitHub copilot, Google Cloud… using those kinds of platforms behind a vendor-neutral tool gives you the ability to swap models out while maintaining the same workflow.
Crystal Palace is also a bit weird - named after a building that was moved, then burnt down, which also gave its name to an ill defined area of London that straddles three boroughs, which is not where the club plays any more…
If you want similar benefits in more urban environments, consider property next to a cemetery. The neighbors are quiet and the lot is very unlikely to be turned into a strip mall.
Codex and Claude should not be seen as ‘the standard’
They are vendor offerings but you don’t have to take them.
Whole piece has that Claudish sense of a complete absence of theory of mind. The LLM has all this background knowledge in context - it’s just read all the historical facts about what happened and who said what when - so it jumps around trying to thread a narrative among those facts but it has no sense of what the reader is likely to be familiar with or what is novel; there’s no storytelling.
It could just be bad writing by someone who knows the subject well and is done explaining all the background needed to understand their thesis. Or it could be extracted from chapter 10 of a book where the prior 9 chapters laid all the groundwork so you can start weaving together a thesis like this from the pieces you’ve already planted.
But it reads a lot like Claude.
Why? Don’t you think the steady improvement in pelican bicycling SVG art is itself a worthy field of endeavor for LLMs to be optimized for?
Yes, Simon has made pelicans a benchmark and now models have grown very good at rendering them. Is there a reason to believe that this isn’t manifesting as useful skill in general in SVG generation?
Is it me or is there maybe some carryover happening where the pelicans are portrayed as expending different degrees of effort too? Like, the high effort pelican has a determined look on his face.
Is the prompt context telling the models what their effort level is, and the ambient presence of that concept of ‘high effort’ in the context window is leaking through and affecting the portrayal?
I’m just saying we don’t yet know what works best long term for LLM managed code. Not suggesting that what LLMs produce today is optimized for that, and ‘human-readable’ is for sure a good default place to start.
Most store doors are open much of the day
If someone walks in and unplugs your refrigerator and causes you to lose all the stock in it, your insurance is going to cover you. You don’t need to show you locked the power sockets.
This is true, but we also don’t know the long term architectural tradeoffs that are right for an LLM-maintained codebase.
For human-maintained code you manage duplication and locality of changes in the codebase because you want concepts to fit inside a human’s working memory. Keeping method bodies short is so they fit in one code window on the screen.
LLMs’ working memory works in very different ways; they see codebases not as windows of code but as blips of grep results. It might be that it’s easier for an LLM to just ensure five pieces of code are all modified in sync with one another than to safely implement a DRY refactoring.
People have gotten very used to breaking down tasks into separate things they know how to do on a computer. Asking an LLM to do each of those individual tasks is easy but feels both lazy and also ‘so what?’ - I knew how to do that task.
The thing people are slow in realizing is that you don’t need to break the task down, and you don’t need to know how to do the task. You don’t even need to have the software installed on your computer to do the task, or even know that software to do the task exists - a sufficiently motivated LLM can pull down the needed open source tools and use them, or even build its own.
You are perhaps mistaken here. The FAA allows passengers to take guns on planes provided they are documented and secured and checked into the hold. What the FAA is concerned about is people being personally armed on planes and there are obvious reasons for that.
They are largely unrelated to the reasons countries might have for being nervous about merchant vessels packing anti aircraft weaponry.
You are I think getting lost in your analogy and in the process conflating two different things:
- the reasons countries might want to prevent people from entering with weapons
Vs
- the reasons countries might have for being nervous about armed merchant vessels in their navigational waters