Hey bcherny, I'm confused as to what's happening here. The linked issue was closed, with you seeming to imply there's no actual problem, people are just misunderstanding the hidden reasoning summaries and the change to the default effort level.
But here you seem to be saying there is a bug, with adaptive reasoning under-allocating. Is this a separate issue from the linked one? If not, wouldn't it help to respond to the linked issue acknowledging a model issue and telling people to disable adaptive reasoning for now? Not everyone is going to be reading comments on HN.
It's better PR to close issues and tell users they're holding it wrong, and meanwhile quietly fix the issue in the background. Also possibly safer for legal reasons.
Isn’t that what they just did here? Close Stella’s Issue, cross post to hn, then completely sidestep an observation users are making, and attack the analyst of transcripts with a straw man attack blaming… thinking summaries….
There's a 5 hour difference between the replies, and new data that came in, so the posts aren't really in conflict.
Also it doesn't sound like they know "there's a model issue", so opening it now would be premature. Maybe they just read it wrong, do better to let a few others verify first, then reopen.
The strategic playbook of the web era said: Get a huge userbase of normies, then figure out how to monetise them (usually via advertising). OpenAI stumbled into the userbase via ChatGPT, but it's unclear if the strategy or the economics apply to AI. Anthropic tried to compete in the consumer market, but couldn't, so focussed on coding and enterprise, and it looks like that's actually turning into a smart choice, at least right now, because it turns out people will pay subscription costs for agents that do their job for them.
There are three possible paths that sort of substantiate current valuations:
1) Business: LLMs become essential to every company, and you become rich by selling the best enterprise tools to everyone.
2) Consumer: LLMs cannibalize search and a good chunk of the internet, so people end up interacting with your AI assistant instead of opening any websites. You start serving ads and take Google's lunch.
3) Superhuman AGI: you beat everyone else to the punch to build a life form superior to humans, this doesn't end up in a disaster, and you then steal underpants, ???, profit.
Anthropic is clearly betting on #1. Google decided to beat everyone else to #2, and they can probably do it better and more cheaply than others because of their existing infra and the way they're plugged into people's digital lives. And OpenAI... I guess banked on #3 and this is perhaps looking less certain now?
If you include a word like innovating in quotes it typically implies that you're quoting it from the link. It can also signify irony, but in a context like HN where we're discussing a published article, it's often ambiguous.
As for Java, I'd agree that its pace of advance was pretty glacial during the Sun era, but from what I've seen has picked up considerably since the Oracle acquisition and Brian Goetz became architect.
And however bad Java is, it's nothing compared to JavaScript. It takes a decade just to add new a library function, and every new syntax proposal is DOA.
Cute idea, but you're never gonna blow your token budget on output. Input tokens are the bottleneck, because the agent's ingesting swathes of skills, directory trees, code files, tool outputs, etc. The output is generally a few hundred lines of code and a bit of natural language explanation.
Pretty neat site you've got there. You should submit it to Show HN. I had fun clicking around - it's like TVTropes, except the examples make me angry, lol.
It would be pretty fun to train an LLM on this site and then have it flag my comments before I get downvoted, haha.
Thanks! I want to do something similar to your LLM suggestion, the endgame is tooling for forums and individuals to improve the quality of discourse. More broadly, I think LLMs and recent advancements now make it possible to assist with self improvement (e.g., see former startup Humu’s nudges but for everyone instead of just B2B)
Good point and it's actually worse than that : the thinking tokens aren't affected by this at all (the model still reasons normally internally). Only the visible output that gets compressed into caveman... and maybe the model actually need more thinking tokens to figure out how to rephrase its answer into caveman style
Grug says you can tune how much each model thinks. Is not caveman but similar. also thinking is trained with RL so tends to be efficient, less fluffy. Also model (as seen locally) always drafts answer inside thinking then output repeats, change to caveman is not really extra effort.
Did Anthropic run Claude in a loop and tell it "work on our pricing and usage strategy", or something?
As far as I can tell, they now have Free, Pro, Max x5, Max x20, Team, and Enterprise tiers. Plus session usage, weekly usage, extra usage (up to a spend limit (set by you) and/or a monthly cap (set by Anthropic)), and now usage bundles, which are extra usage but with a lower price and a fixed amount.
I don't know, but it really feels like there's an LLM churning away somewhere, endlessly tinkering on a PRICING.md file that is gradually accreting strategic slop.
> I don't know how long this pace will last. I suspect that bugs are reported faster than they are written, so we could in fact be purging a long backlog
Hopefully these same tools will also help catch security bugs at the point they're written. Maybe one day we'll reach a point where the discovery of new, live vulnerabilities is extremely rare?
Around 70% of security vulnerabilities are about memory safety and only exist because software is written in C and C++. Because most vulnerabilities are in newly written code, Google has found that simply starting writing new code in Rust (rather than trying to rewrite existing codebases) quickly brings the number of found vulnerabilities down drastically.
You can't just write Rust in a part of the codebase that's all C/C++. Tools for checking the newly written C/C++ code for issues will still be valuable for a very long time.
You actually can? A Rust-written function that exports a C ABI and calls C ABI functions interops just fine with C. Of course that's all unsafe (unless you're doing pure value-based programming and not calling any foreign code), so you don't get much of a safety gain at the single-function level.
If you're going to swap out one function in a chain of functions for a Rust version, you're destroying your codebase. You simply can't replace one tiny piece of code in a large codebase with a version in a different language. Doing so would be insane.
C ABI is not C++ ABI. People often write C/C++ but they're completely different languages. C++ is much higher level and modern. C++ is closer to Rust than it is to C.
It comes from all his reporters being teenagers in developing countries with older models, and people using SOTA models who know how to qualify a potential vulnerability having much bigger fish to fry than curl. curl is a meaningful target, but it's in nobody's top tier.
No, this is false. For Rust codebases that aren't doing high-peformance data structures, C interop, or bare-metal stuff, it's typical to write no unsafe code at all. I'm not sure who told you otherwise, but they have no idea what they're talking about.
It's the classic "misunderstanding" that UB or buggy unsafe code could in theory corrupt any part of your running application (which is technically true), and interpreting this to mean that any codebase with at least one instance of UB / buggy unsafe code (which is ~100% of codebases) is safety-wise equivalent to a codebase with zero safety check - as all the safety checks are obviously complete lies and therefore pointless time-wasters.
Which obviously isn't how it works in practice, just like how C doesn't delete all the files on your computer when your program contains any form of signed integer overflow, even though it technically could as that is totally allowed according to the language spec.
If you're talking about Rust codebases, I'm pretty sure that writing sound unsafe code is at least feasible. It's not easy, and it should be avoided if at all possible, but saying that 100% of those codebases are unsound is pessimistic.
One feasible approach is to use "storytelling" as described here: https://www.ralfj.de/blog/2026/03/13/inline-asm.html That's talking about inline assembly, but in principle any other unsafe feature could be similarly modeled.
It's not impossible, it is just highly unlikely that you'll never write a single safety-related bug - especially in nontrivial applications and in mixed C-plus-Rust codebases. For every single bug-free codebase there will be thousands containing undiscovered subtle-but-usually-harmless bugs.
After all, if humans were able to routinely write bug-free code, why even worry about unsoundness and UB in C? Surely having developers write safe C code would be easier than trying to get a massive ecosystem to adopt a completely new and not exactly trivial programming language?
Rust is not really "completely new" for a good C/C++ coder, it just cleans up the syntax a bit (for easier machine-parsing) and focuses on enforcing the guidelines you need to write safe code. This actually explains much of its success. The fact that this also makes it a nice enough high-level language for the Python/Ruby/JavaScript etc. crowd is a bit of a happy accident, not something that's inherent to it.
Good developers only write unsafe rust when there is good reason to. There are a lot of bad developers that add unsafe anytime they don't understand a Rust error, and then don't take it out when that doesn't fix the problem (hopefully just a minority, but I've seen it).
Now that Dye is gone, I still hold out hope that Apple will change direction and start fixing their UI. But that fact that it got this bad in the first place implies things are seriously broken at a senior leadership level.
Or just consider the “asshole dinner guest” trope that appears in so many TV shows and movies. They will always be talking too loudly and gesticulating/pointing with their cutlery.
A new long-form article, published every day, like clockwork, since the 1st of June 2025. All about AI, all attributed to the same author.
Congrats to him on finally getting this slop to the top of HN, I guess. Shame on everybody here for upvoting it.
reply