ankit219

Because your subscription depends on the very API business.

Anthropic's cogs is rent of buying x amount of h100s. cost of a marginal query for them is almost zero until the batch fills up and they need a new cluster. So, API clusters are usually built for peak load with low utilization (filled batch) at any given time. Given AI's peak demand is extremely spiky they end up with low utilization numbers for API support.

Your subscription is supposed to use that free capacity. Hence, the token costs are not that high, hence you could buy that. But it needs careful management that you dont overload the system. There is a claude code telemetry which identifies the request as lower priority than API (and probably decide on queueing + caching too). If your harness makes 10 parallel calls everytime you query, and not manage context as well as claude code, its overwhelming the system, degrading the performance for others too. And if everyone just wants to use subscription and you have no api takers, the price of subscription is not sustainable anyway. In a way you are relying on others' generosity for the cheap usage you get.

Its reasonable for a company to unilaterally decide how they monetize their extra capacity, and its not unjustified to care. You are not purchasing the promise of X tokens with a subscription purchase for that you need api.

Thats a doublespeak. You are talking about an act with a legal basis. You follow and comply with those regulations lest you break the law.

If it was a voluntary standard or code, maybe you could conform. The thread talks about the ACT and COMPLIANCE in its heading. You follow the regulation, you comply with it, you don't conform to a mandatory set of rules anyway. If you do, that is literally orwellian (because it implies a mandatory legal obidience is same as a voluntary fitting in).

No its not. You follow the rules or you adhere to them.

You change your behavior or belief system to match a group's standard/norm (conform).

Conform and follow/adhere are quite different, and convey the thought process underneath.

It was a joke on the name. Nothing to do with the content. Conform to newspeak is what 1984 is famous for.

TIL you cant take a joke on the name, and you have to invoke a stated premise to feel better. If you really thing that is the only thing the act is doing (without realizing the carve outs it has for facial recognition etc.) then i have a tower bridge to sell to you.

Rules and regulations always assume it would be used by a rational actor, but should be written such that someone irrational would not be able to misuse it. That is teh premise of 1984. But you do you .

This was always against the terms of service. Took too long to crack down on it.

There seems to be a lot of FUD on how anthropic prices it. but i think many devs and harness builders are intentionally misrespresenting things to build pressure.

API capacity has specific SLAs and the infra is for peak capacity expectation. API demand is notoriously spiky, especially when you don't know who is going to be using it via the third party. At any given time you would have 20%-30% utilization, given the api is mostly used by tools like Cursor who have a system of routing by themselves. For subs, they can use the unused capacity (with high priority to API requests). The marginal cost of a query is not as high especially given kv cache and continuous batching (and with new parallelism techniques). With subs, the number is expected, the concurrency can be modelled, and hence the price per token is lower.

With third party APIs reverse engineering that, it breaks the whole thing in two ways - one more pressure via subs. Claude code rarely parallelizes requests while Cline does it every time (for example). It also cancels out the peak capacity estimation. In a way you could argue that same users doing it via subs woudl have been doing it via api too, but given the pricing differences, not the same. This ends up affecting people legitimately using the subscription.

Calling it "Conform" is very 1984esque. There is also a 2025 book (a dystopian romance) called the same: https://www.goodreads.com/book/show/223239535-conform

I am saying that if you force Apple to move away from integrated devices to something which has to be generic (modular like Intel) which does not know which hardware it will pair up with and hence needs baseline performance, it turns into android to a large extent. Other businesses may have legit incentives to reach those customers, but unless Apple makes drastic changes to current setup, the software would not support other manufacturers to the same extent. So they will go to EU and then Apple will write a more generic code - to ensure all manufactures are similarly supported, and it takes them away from integrated system they currently have. Competition wins, but customers dont.

Modular in the sense you have to support multiple hardwares (of different kinds) instead of just one. Eventually you arrive at a place where software is good enough, and hardware + kernels cannot do the exact heavy lifting that is happening today in conjunction. Not the intel level but directionally similar kind of tradeoffs.

modular does not mean in terms of how the library is architected, but in terms of how many vendors/customers it needs to support. Airpods' hardware is built and then kernels are written in a way to compliment each other and get the most out of the system. With another set of headphones with a different chip, there is a very good chance that code written today would not be optimal because other builders could manufacture different things based on the same spec. You cannot bring everything to software, nor can you have hardware doing everything. Tradeoffs would be needed.

The issue comes in second order effects. If third party headphones are given access and then the experience is not as good, they complain that Apple hasnt open up the spec enough, and it just results in Apple being forced to be modular in their approach.

Its mostly true when the integrating company cares for the user experience. Which apple clearly does.

The example you shared is the opposite. I am imagining a kernel today written in a manner that airpods would be able to use it to extract the max out of it. Now, it has to support 10 other third party pods, so at the minimum, kernel would be more generalized.

A company making an integrated experience would inevitably provide a better experience/performance than a company asked to build for 100s of devices with different spec. That Apple did not want to open it up is a separate discussion.

My contention is this: expecting a third party provider to be able to provide the same experience as the first party is an impractical goal. Even pushing companies towards that means a lot of second order effects where everyone ends up like Intel or Windows for that matter. We already have android on that level.

You can have a reasonable requirement where Apple should not be able to block other companies from providing similar services based on an iphone. But clearly the directive here is that Apple's competing products should not be better based on better integration, which can only go in one direction. Apple degrades its own products to comply. Yes, competition wins, but consumers lose. In this case specifically - consumers who would want to choose Apple, better experiences would not be able to simply because Apple cannot ensure the level of software/hardware alignment as it works today if the same software is written with modular hardware in mind.

Are we learning the wrong lessons? Integrated always works better than modular components. Here, Apple is being asked to enable their versions of software for third party devices, which do not have the same hardware assumptions as Apple did. (Apple will not release the exact hardware spec for airpods anyway). This means the newer version will be designed modularly, with some tradeoffs to enable the "same" kind of access to third party. Then there is a caveat that it there is even a bit of experience change from 1st party to third party access, it will be complained about and investigated. so, the way fwd is designing with third party in mind, and that almost always leads to bloat and substandard experience for end user.

Probably better would have been just simpler access, even if not the integrated experience like. But that would lead to complains from third party manufacturers.

Think its a combination of factors.

- Google cloud is setup for big organizations. Not for individuals. All cloud providers are pretty much confusing in a similar way. - India has specific rules re cybersecurity and financial regulations that Google has to comply. (mandatory id verification and kyc compliance). Others have asked for an id check too.

From what confused me, if OP wanted to use a model, the easier way would have been to pay cursor/windsurf etc. and select that model. Usually that is how people try out a new model. Trying out a specific way means going through the norms every country imposes, and bloat in case of legacy products.

AWS and Azure have come up with their own models. If their future versions hit close to sota and people want to use it, many would end up in a similar loop (and woudl be easier to just use it from the aggregators).

Maybe, not lie about what they are doing? That is not unreasonable as a base expectation.

you already sent the prompt to gemini api - and they likely recorded it. So in a way they can access it anyway. Posting here or not would not matter in that aspect.

Openai, deepmind, anthropic, xai. who is the fifth frontier lab?

because they put restrictions. now they cannot. because all the restrictions meant saying no to some legit things as well - inevitably. but then they got sued, laws got passed, to not be monopolistic, and still secure the users. this is the aspect of tech saying no when the thing being asked is impossible but people assume because they dont want to do it for whatever reason.

Google is about to find out the next step of this chain - give access to everyone, don't gatekeep / do checks, and yet take responsibility for anything that goes wrong.

"You should open up the tool, put no restrictions, and yet ensure that it is safe and secure" is an impossible task for anyone.

The thing is that if you are giving professional advice in US - legal, financial, medical - the other party can sue you for wrong or misleading advice. In that scenario, this leaves Openai exposed to a lawsuit, and this change seemingly eliminates that.

They can block the site on a mac if they dont want you to purchase via a mac. They can block any browser agent they want if they dont like the traffic. They can also close your account if they have some reasons. They dont need a reason either. You agree to a terms of service when you make an account, and that usually means giving them broad leeway.

I think its more to do with obfuscation than simple agent buying stuff. Perplexity disguises its agents as chrome browser, and that messes up things at Amazon's end. Not saying its good or bad, but think of ad relevance, then recommendations effectiveness, placements etc. because by default an agent is going to click more and more randomly than a user.

Then, when the agent buys stuff, my guess is that returns are a higher %age (because naturally agents have no preference model so end users might not like what they purchased eventually and returned). I don't know if enough people returned for it to matter, but as a user, if an agent is making a purchase and i know its super easy to return, i won't check everything agent did, just buy and then return if I dont like it.

I don't think it's always a price comparison or margin thing at this point. Given their margins and volume, the agent purchases woudl need to be an order of magnitude more to matter on that front.

I agree it's a workaround. Ideally the model should follow instructions directly, or check before running another server to see if it's starting. Though training cannot cover every usecase and different devs work differently, so i guess its acceptable as long as its on track and can do the work.

create a instruction.md file with yaml like structure on top. put all the instructions you are giving repeatedly there. (eg: "a dev server is always running, just test your thing", "use uv", "never install anything outside of a venv") When you start a session, always emphasize this file as a holy bible to follow. Improves performance, and every few messages keep reminding. that yaml summary on top (see skills.md file for reference) is what these models are RLd on, so works better.

The headline is provocative. The issue is simple: Meta has a recommended feed and another feed which is reverse chronological. The option is hidden and not default, and all the NL judge is asking is for that option to be preserved and not be reset everytime user opens the app/website. Instagram never had this option, this does not say whether they have to implement it too.

I would be curious if the order stands as for curation as well. Someone could have 1000s of friends, and you cant show posts from everyone in a reverse chronological order for a good ux.

Location: San Francisco

Remote: okay

Willing to relocate: Not outside bay area

Technologies: More of a generalist. can code. Currently working on RL posttraining. In the past worked on agentic flows, and some foundational level research.

Résumé/CV: here is linkedin: https://www.linkedin.com/in/ankitmaloo/

Email: ankit21.iitk@gmail.com

I have worked on AI projects for the past three years. Mostly on my own. I like the foundational level research, and i do write about how the tech is evolving at ankitmaloo.com. In the past, i had a startup in AI in 2018. I have worked on early stage GTM at various companies, and can help with figuring out scaling. I like working on impactful products.

I see your point. I thought the first one was already known when deepseek came out. Perplexity team showed how they removed this kind of bias via finetuning and their finetune could answer sensitive questions. I mistakenly thought you went for the second since that part is new and interesting.

Your claim and the original claim are vastly different. Refusing to assist is not the same as "writing less secure code". This is clearly a filter before the request goes to the model. In the article's case, the claim seems to be that the model knowingly generated insecure code because it was for groups china disfavors.

Hi! Thank you for the clarification. I was just saying it might be possible in the future (in a way you can determine how much compute - which model - a specific query needs today as well). And the experience has definitely improved w router so kudos on that. I don't know what the final form factor of ads would be (i imagine it turning out to be a win win win scenario than say you show ads at the expense of quality. This is a google level opportunity to invent something new) just that it seems from the outside you guys are preparing for monetization by ads given the large userbase you have and virtually no competition at chatgpt usage level.

The router introduced in gpt-5 is probably the biggest signal. A router, while determining which model to route query, can determine how much $$ a query is worth. (Query here is conversation). This helps decide the amount of compute openai should spend on it. High value queries -> more chances of affiliate links + in context ads.

Then, the way memory profile is stored is a clear way to mirror personalization. Ads work best when they are personalized as opposed to contextual or generic. (Google ads are personalized based on your profile and context). And then the change in branding from being the intelligent agent to being a companion app. (and hiring of fidji sumo). There are more things here, i just cited a very high level overview, but people have written detailed blogs on it. I personally think affiliate links they can earn from aligns the incentive for everyone. They are a kind of ads, and thats the direction they are marching towards .