I decided on my OAI subscription when comparing mainly due to reports like those. GPT-Sol has been happily assisting me in reverse engineering efforts without any issues. I very strongly dislike an AI model or company deciding what my morals are for me.
I've only started hitting the anthropic blocks in the last week or so, I get much more blocks from openai/codex. This plus the recent reductions in usage allowance on both are finally pushing me to look elsewhere. I've had the highest level x20 subs for both since they came out, but I'm getting 5x the value/price from DeepSeek at this point.
Claude is still working for me on a mashup of two very mainstream games. It has had to do (and continues to) extensive reverse engineering and quite a bit of "hmm the game assets are encrypted, let me decompile it and find the decryption key". Not sure if it just refuses new projects or I got lucky and managed to avoid all the tripwires so far. I don't plan on sticking around to have it suddenly start refusing to work for me, so I'll also be switching to OpenAI once my month ends. Hopefully it keeps working for the rest of my remaining time!
Jokes on me if Codex refuses right out of the box, so I'll start low on the usage tiers and slowly upgrade as it proves it can be trusted to continue the work.
It looks like they wanted to kill most of the reverse engineering by cutting down on low effort recomps. It had got to the point where people who knew nothing of coding were recompiling major games just by throwing tokens at it. I'm neither condemning nor supporting it, just reporting.
Someone on Reddit made a good point. Even a short fence stops most people from treading on your lawn.
In other words, people who try to divide and conquer recompilations will likely still succeed... for now.
It would be easy enough for OpenAI to block this kind of use, both in their TOS and technically. It seems like there would be plenty of other reasons not to support unlimited chats long term.
I believe his point it, that exploitative- and abusive behaviour as well as "inventive interpretations" of agreements by the few, ends up punishing the majority.
It should not be necessary to set up technical or written guards against every possible act of bad behaviour.
Instead be nice. Don't be that **ole who steals the whole candy bucket on Halloween just because you can.
I think the core issue isn't the abuse, but the very existence of a service that claims to provide infinite of something. Infinite is always BS because resources aren't infinite.
It's just like "all you can eat" places kicking out people after they've been in the restaurant for a while. We should discourage these marketing tricks because they are always a lie.
Regarding Halloween, we stopped leaving unattended buckets out because of bad actors taking everything. Now we hand out the candy ourselves. Our home no longer provides "all you can grab" treats...
How is this small neighbourhood morality applicable to a globally available Internet service? They're not giving out free candy, they're trying to make money and get more customers by offering unlimited chats. If unlimited chats stop being worth it for them, they remove them. At no point of the process I can see them at a board meeting crying about people not being nice.
OpenAI played stupid games, won stupid prizes. They have designed a subscription where the optimal behaviour is to spend down limits as soon as possible, because if you don't their random resets, which make up almost half the value of subscriptions sometimes, will be wasted.
But to do THAT you need to max out every 5 hour block. Of course people are going to start abusing unlimited chat in the mean time and start optimising the ability to burn tokens as fast as they can.
Isn't it toxic to brag about it online like he does? It wasn't a feature many were even aware of, but instead of placing it as responsible disclosure, he places it as: "I ruined it for all of you, but I had a lot of fun doing it".
Trillion dollar company repeatedly does things in their own self interest at the expense of the rest of society. Individual does something in their own self-interest at the expense of trillion dollar company.
You're not wrong, but only in the sense that OpenAI and plenty of other powerful companies are full of people doing pretty much the same thing, and they cause a lot more "nice things" to disappear than individuals like the one you're talking about could ever cause to get taken away in their lifetime.
> Since ChatGPT chats cannot spawn or control subchats/subganets so I had to be a bit creative in order to do multi-agent development and orchestration.
Multi-agent is mostly dead. One big agent loop running GPT6.1 absolutely dominates any weird ideas like this now. I still use sub agents for some tasks where I want to compress the action space and context. For example, testing a workflow that requires 100+ tool interactions for each logical test. Without a subagent it's really hard to guarantee a test makes it E2E. I have some very "persuasive" methods for long horizon tests that would break the main loop's dynamics.
We may not feel so compelled to exploit the system if we could figure out how to not be so wasteful. Multi agent schemes can bloat token usage by orders of magnitude. They force us to consider consuming less ideal tokens from contexts and vendors with higher variance.
HN loses its marbles when I suggest it's economically feasible to prepay for API tokens and use these as the exclusive means of performing AI work. I get done with a thousand tokens what my peers cannot do with a million. It doesn't matter how many hammers or workers you have if you don't have the aesthetic preferences and domain expertise to employ them correctly.
Wow that's a huge take. Seems like you're on your own with this one.. everyone I know is scaling up the number of agents they're able to keep fed with good, useful work.
Everyone I know doing things like this has actual abysmal work output. They spend all their time playing with their toys and no time actually producing useful output.
For me Opus 5.5 will frequently make several thousand LoC where 10 would suffice. One time I got fed up and told it to repro the "bugs" it was solving for with puppeteer, and lo and behold, it could not do it overnight. Hard to believe that people let this run wild on their codebase, if they do, they are making a huge untenable mess.
This matches with my experience as well, especially with pull request reviews. Stock command gives me 10 "issues" and once you start asking to fix them, it will gladly tell you this is not the issue. I use the following prompt now and it significantly reduces the amount of noise.
/code-review <PR#>
versus
Review the changes in <branch> against <base>. Verify your finding and offer the fixes if the finding are valid.
That's a big bold claim. Multi-agents are still very useful for tokenOps and context management. You don't need sol to investigate codebase or the datasets, terra or luna would do it cheaper. Clear context for certain tasks also still matters.
Probably not. It’s more likely it’s due to OpenAI having some problems raising funds which forces them to be more efficient with their capital: https://news.ycombinator.com/item?id=50312922
While I'm sure OpenAI won't miss OPs business, the "Claude here I come" is pretty indicative of the value in subs at the moment.
Claude seems to have higher limits by default instead of playing the random reset games.
In a world where subscription value isn't obfuscated by unlimited chat, the choice for a power user is obvious, especially since it took a month after Sonnet 5.5 for OpenAI to get Terra 6.1 out, using Sol for everything just ate through the limits.
37 comments
[ 0.19 ms ] story [ 18.3 ms ] threadWe had 2 major threads in the past 48 hours complaining about this:
* https://news.ycombinator.com/item?id=50319358 (AI/ML work with large models)
* https://news.ycombinator.com/item?id=50320255 (reverse engineering)
The threads are full of upset comments!
The only non-OAI options are open models and I guess Gemini 4 now.
Jokes on me if Codex refuses right out of the box, so I'll start low on the usage tiers and slowly upgrade as it proves it can be trusted to continue the work.
Someone on Reddit made a good point. Even a short fence stops most people from treading on your lawn.
In other words, people who try to divide and conquer recompilations will likely still succeed... for now.
Anyone whose buainess wont be competitivw without a cloud model is going to lose their business.
Any government letting cloud models consume its private infrastructure is yielding its authority to private intetests.
In the US, that’s considered a feature.
I've been doing that and it seems like quite a deal.
Seems like the doors are (slightly?) open right now. OpenCode just announced that users can use their Claude subscription:
- https://x.com/opencode/status/2116198602523513081
- https://github.com/anomalyco/opencode/pull/58782#issuecommen...
> I am simply making the best use of what OpenAI has made available to me
People like this are why you can't have good things.
It should not be necessary to set up technical or written guards against every possible act of bad behaviour.
Instead be nice. Don't be that **ole who steals the whole candy bucket on Halloween just because you can.
> Do not attribute to malice that which can be attributed to really weird usecases.
see also, xkcd.com/1172 (spacebar heating)
I think the core issue isn't the abuse, but the very existence of a service that claims to provide infinite of something. Infinite is always BS because resources aren't infinite.
It's just like "all you can eat" places kicking out people after they've been in the restaurant for a while. We should discourage these marketing tricks because they are always a lie.
Regarding Halloween, we stopped leaving unattended buckets out because of bad actors taking everything. Now we hand out the candy ourselves. Our home no longer provides "all you can grab" treats...
* Lets people use whatever approach they prefer (- and if someone wants to keep their existing setup they can - with reasonable limits)
* Minimizes risk of users getting banned, including due to A) activity they didn’t know was disallowed, or B) false positives from OAI
* Heads off different future loopholes
* Avoids adding more TOS complexity
* Avoids users being uncertain if borderline activities will get them banned
But to do THAT you need to max out every 5 hour block. Of course people are going to start abusing unlimited chat in the mean time and start optimising the ability to burn tokens as fast as they can.
You're not wrong, but only in the sense that OpenAI and plenty of other powerful companies are full of people doing pretty much the same thing, and they cause a lot more "nice things" to disappear than individuals like the one you're talking about could ever cause to get taken away in their lifetime.
Multi-agent is mostly dead. One big agent loop running GPT6.1 absolutely dominates any weird ideas like this now. I still use sub agents for some tasks where I want to compress the action space and context. For example, testing a workflow that requires 100+ tool interactions for each logical test. Without a subagent it's really hard to guarantee a test makes it E2E. I have some very "persuasive" methods for long horizon tests that would break the main loop's dynamics.
We may not feel so compelled to exploit the system if we could figure out how to not be so wasteful. Multi agent schemes can bloat token usage by orders of magnitude. They force us to consider consuming less ideal tokens from contexts and vendors with higher variance.
HN loses its marbles when I suggest it's economically feasible to prepay for API tokens and use these as the exclusive means of performing AI work. I get done with a thousand tokens what my peers cannot do with a million. It doesn't matter how many hammers or workers you have if you don't have the aesthetic preferences and domain expertise to employ them correctly.
Wow that's a huge take. Seems like you're on your own with this one.. everyone I know is scaling up the number of agents they're able to keep fed with good, useful work.
Only you can decide if that speed is worth the cost.
I agree. It's possible future AIs will be able to salvage those code bases though.
/code-review <PR#>
versus
Review the changes in <branch> against <base>. Verify your finding and offer the fixes if the finding are valid.
Claude seems to have higher limits by default instead of playing the random reset games.
In a world where subscription value isn't obfuscated by unlimited chat, the choice for a power user is obvious, especially since it took a month after Sonnet 5.5 for OpenAI to get Terra 6.1 out, using Sol for everything just ate through the limits.