Hacker News Board Hacker News, browsed like a classic message board
Live data from Hacker News

Claude Sonnet 4.5

anthropic.com

Post reply on HN
Print view View original

671–680 of 819 posts

Re: Claude Sonnet 4.5

#671

Post by nosianu »

wyre wrote: post #528
paool wrote: post #469

Earlier quoted context omitted.

> shouldn't it just take advantage of what's already there? It's not a good idea to have any coding agent put unnecessary amounts of lines into the context window in order to understand your code base. Performance of all llms drop drastically when the context window is filled or full. The purpose of being more specific with your prompts is that you spend a little bit more tokens up front to make the task a lot more e…

I'm not sure exactly what you mean by the vibe coding experience non-coders want, but if it's one-shotting a buildable codebase off of an unspecific prompt, the major breakthrough would have to be brain-computer interfaces so the agent can literally read the user's mind. If that same person approached a software development company with the same prompt without following up with any other details, they won't get good…

> the major breakthrough would have to be brain-computer interfaces so the agent can literally read the user's mind.

And even that would not be enough.

In reality, it would have to put the user to sleep and go through various dream scenarios to have the user's brain really build an internal model that is not there in the first place. No brain interface can help find what is not there.

We usually need interactions with reality to build the internal model of what we actually want step by step, especially for things we have not done before.

Even for info that is there, that's also a limit to fantasy or sci-fi brain scanning. The knowledge is not stored like in a RAM chip, even when it is there. You would have to simulate the brain to actually go through the relevant experiences to extract the information. Predicting the actual dynamic behavior of the brain would require some super-super sub-molecular level scan and then correctly simulating that, since what the neurons will actually do depends on much more than the basic wiring. Aaaaand you may get a different result depending on time of day, how well they slept, mood and when and what the person ate and what news they recently read, etc. :)

Re: Claude Sonnet 4.5

#672

Post by baq »

typpilol wrote: post #389
lukev wrote: post #123

Earlier quoted context omitted.

It means that the benchmark isn't useful anymore and we need to build a harder one. edit: as far as what the numbers mean, they are arbitrary. They are only useful insofar as you can run two models (or two versions of the same model) on the same benchmark, and compare the numbers. But on an absolute scale the numbers don't mean anything.

I thought the percentage was how many problems it successfully solved

Technically correct, but not helpful nor actionable.

Re: Claude Sonnet 4.5

#673

Post by la_fayette »

vunderba wrote: post #502
la_fayette wrote: post #217

Earlier quoted context omitted.

At this point it would be an interesting idea, to collect examples, in a form of a community database, were LLMs miserably fail. I have examples myself...

Any such examples are often "closely guarded secrets" to prevent them from being benchmaxxed and gamed - which is absolutely what would happen if you consolidated them in a publicly available centralized repository.

Since such a database should evolve continuously, I wouldn't see that as a problem. The important thing is, that each example is somehow verifiable, in the form of a unmodifiable test setup. So the LLM provides a solution, which is executed against the test to verify. Something like ACID3 Tests... But sure it can be gamed somehow in probably all setups...

Re: Claude Sonnet 4.5

#674

Post by grncdr »

janee wrote: post #628
Philpax wrote: post #243

Earlier quoted context omitted.

Well, yes, that's Claude Code. And OpenAI Codex. And Google Gemini CLI. Your average dev can just use those.

Yes but you need to setup quite a bit of tooling to provide feedback loops. It's one thing to get an llm to do something unattended for long durations, it's a other to give it the means of verification. For example I'm busy upgrading a 500k LoC rails 1 codebase to rails 8 and built several DSLs that give it proper authorised sessions in a headless browser with basic html parsing tooling so it can "see" what affect it…

> rails 1 codebase to rails 8

A bit off topic, but Rails *1* ? I hope this was an internal app and not on the public internet somewhere …

Re: Claude Sonnet 4.5

#675

Post by dbbk »

hsn915 wrote: post #542

It is time to acknowledge that AI coding does not actually work. ok, you think it's a promising field and you want to explore it, fine. Go for it. Just stop pretending that what these models are currently doing is good enough to replace programmers. I use LLMs a lot, even for explaining documentation. I used to use them for writing _some_ code, but I have never ever gotten a code sample over 10 lines that was not in…

I have 20 years of experience and am now just running Opus 4.1 all the time. Of course it’s possible.

In my case I found having it always Ultrathink and to always work with TDD to work well. Also you have to use Plan Mode first and refine the plan. “What clarifying questions do you have for me” prompts me with a handful of numbered questions that are always really beneficial for refining the plan.

Re: Claude Sonnet 4.5

#676

Post by nosianu »

jen729w wrote: post #598
nikcub wrote: post #554

Earlier quoted context omitted.

> but these types of prompts and approaches are why I believe so many people think these models aren't useful. 100% agree. The prompt is a 'yolo prompt'. For that task you need to give it points in what to do so it can deduce it's task list, provide files or folders in context with @, tell it how to test the outcome so it knows it has succeeded and closing the feedback loop, and guide it in implementation either via…

The fact remains, however: ChatGPT did it. Claude did not.

That fact is pretty useless to draw any useful conclusions from with one random not so great example. Yes, it's an experiment and we got a result. And now what? If I want reliable work results I would still go with the strategy of being as concrete as possible, because in all my AI activities, anything else lets the results be more and more random. Anything non-standard (like, you could copy & paste directly from a Google or SO result), no matter how simple, I better provide the base step by step algorithm myself and only leave actual implementation to the AI.

Re: Claude Sonnet 4.5

#677

Post by Panoramix »

wilg wrote: post #654
genidoi wrote: post #612

Earlier quoted context omitted.

> where I feel so disconnected from my codebase I'd rather just delete it than continue. If you allow your codebase to grow unfamiliar, even unrecognisable to you, that's on you, not the AI. Chasing some illusion of control via LLM output reproducibility won't fix the systemic problem of you integrating code that you do not understand.

Who cares about the blame, it would just be useful if the tools were better at this task in many particular ways.

It's not blame, it's useful feedback. For a large application you have to understand what different parts are doing and how everything is put together, otherwise no amount of tools will save you.

Re: Claude Sonnet 4.5

#678

Post by conradfr »

Implicated wrote: post #274
iagooar wrote: post #258

Anecdotal evidence. I have a fairly large web application with ~200k LoC. Gave the same prompt to Sonnet 4.5 (Claude Code) and GPT-5-Codex (Codex CLI). "implement a fuzzy search for conversations and reports either when selecting "Go to Conversation" or "Go to Report" and typing the title or when the user types in the title in the main input field, and none of the standard elements match, a search starts with a 2s de…

I'm not trying to be offensive here, feel the need to indicate that. But that prompt leads me to believe that you're going to get rather 'random' results due to leaving SO much room for interpretation. Also, in my experience, punctuation is important - particularly for pacing and grouping of logical 'parts' of a task and your prompt reads like a run on sentence. Making a lot of assumptions here - but I bet if I were…

No true Promptman.

Re: Claude Sonnet 4.5

#680

Post by mrheosuper »

sophiebits wrote: post #372
icelancer wrote: post #361

Earlier quoted context omitted.

Huh. Doesn't work over here. --- SYSTEM: Got it. Those are the setup commands for Simon Willison's LLM CLI tool, which lets you interact with language models from the command line. The -e flag installs it in editable mode, and [test] pulls in the test dependencies. Are you running into any issues with the installation or tests? What are you working on with the LLM library? USER: Correct. You run them using your code…

You need to enable the new code interpreter mode: https://simonwillison.net/2025/Sep/9/claude-code-interpreter...

Interesting, enable those setting and the claude on claude.ai become claude code, and it try to run everything on claude container like it owns the machine. I don't want that.
Post reply on HN