thimabi

Those are not mutually exclusive things. I actually sympathize with voting with your wallet.

Doesn’t use limits… on the first month only, actual limits will be disclosed later — most likely after they’ve found out how much people actually use this new feature.

It’s about time for some laws against unpredictable pricing like that.

But Meta until very recently offered its services mostly for free, on a gigantic scale, paid for by ads. Now OpenAI has subscription revenue, API revenue, and ads revenue too. The fact that OpenAI isn’t financially sustainable makes me think it hasn’t really mastered its business yet.

Perhaps we can envision a future in which we analyze Reddit data not according to all the positive things people say about a product, but according to the negative things being said.

Shilling can’t prevent dissatisfied customers from voicing their opinions in a properly moderated subreddit, and astroturfing to spread misinformation about a competitor makes companies much more vulnerable to litigation.

Not all tasks require frontier intelligence. If you’ve got an easy, but tedious workflow, Luna can be quite good at that.

It’s confusing indeed, but I like having many options, particularly considering that pricing can be wildly different depending on the model.

Maybe OpenAI can offer an "auto" mode for Codex on the subscriptions, while leaving the possibility of users manually overriding whatever model the router chooses. To me that would be the best of both worlds. The problem is building a competent model router.

You probably mean they are operating with 80% margins discounting training expenses, which will continue to be pretty high for the foreseeable future.

I wonder why doesn’t the Internet Archive require logging-in prior to accessing the Wayback Machine. It would probably help them distinguish humans from bots, at a very little cost to humans.

One of the things that stood out in your article was the high amount of negations or contrasts. This is one of the worst hallmarks of AI writing, because humans tend to write in a more straightforward way, without as many caveats.

My two cents: you should avoid using AI to “polish” what you wrote. At most, instruct it to correct grammar mistakes only, instead of rephrasing your words, changing your arguments, turning paragraphs into bullet lists, “improving” sentences to make them more idiomatic… Nowadays, many technical readers would rather read the occasional non-idiomatic prose than put up with the soulless writing style of AI.

The fact that your arguments will probably end up in an LLM’s training data makes me think they are not implausible at all

This tracks with what OpenAI has been saying about Astra: that it tends to do things in a certain way and it’s up to you to prompt it to change its style.

> Will be interesting to see if inference chips are here to stay

To me, the efficiency gains of inference chips are so significant that they are certainly here to stay — barring a revolution of sorts that leads to a world devoid of AI as we know it.

I add commands that constantly fall into this rabbit hole to AGENTS.md alongside an instruction for the agent to request escalation immediately instead of failing first then retrying out of the sandbox. It’s a hack, but it works.

What do you suggest for articulating the writing style that one wants from LLMs?

I’ve been experimenting with having LLMs write/update academic notebooks for me, and so far the best results I’ve gotten came from correcting their output and asking them what they’ve “learned” from my feedback.

I noticed the same thing. In the app, normal “Chat” threads are only available via a “Checking recent chats” window, while projects, GPTs, library… are completely absent.

The web version of ChatGPT is confusing too. Now it has separate “Chat” and “Work” tabs (what about a Codex tab?), and it shifts the burden on the user to know when to use one or the other. Note that using the “Work” tab means using Codex usage limits [^1], but that’s hidden away in the settings.

Also, apparently “GPT-5.6 Terra and GPT-5.6 Luna are not selectable in standard ChatGPT conversations” [^2] — so if you want to use these models, you must go to the Work tab or download the app.

I don’t understand why there is so much fragmentation in what was supposed to be a unified app. The way that it is now, it’s far from intuitive.

[^1]: https://help.openai.com/en/articles/20001275-chatgpt-work-an...

[^2]: https://help.openai.com/en/articles/20001354-gpt-56-in-chatg...

I’m interested in knowing how each of GPT 5.6’s variants fare in non-English writing/translation tasks.

GPT 5.5 has a tendency to write English calques and non-idiomatic prose in other languages. Although that can be somewhat tamed with detailed instructions and a corpus of confusing terms, the model’s output often reads like a literal translation rather than native prose. Since I notice these issues most clearly in languages I know well, it makes me reluctant to trust the model’s output in languages in which I’m less proficient.

Ironically, ChatGPT began as a simple text-generation tool, but much of its offerings and benchmarks now focus on coding and agentic workflows, while leaving behind what made it notable in the first place.

I wonder why they haven’t simply continued to rebrand Codex as a general-purpose tool. ChatGPT Work is a convoluted name and continues the trend of having separate brands for separate things, what runs counter to OpenAI’s purported goal of unifying every workflow into a single “superapp”.

Worse still: what happens when your workflow involves both coding and general knowledge work? Are you expected to switch apps, or switch settings? To me, it sounds very confusing and inefficient, and not at all what I was expecting.

In a few years Sony executives will be wondering why a portion of their consumer base decided to prioritize other forms of entertainment. I can speak for myself in that I’ve never upgraded past the PS3, and I feel no regrets about it.

I’ve been experiencing similar feelings. Working with LLMs often takes almost as much mental energy as working with people, but the payoff does not always scale in the same way.

I think we are still on the early days of LLMs. Right now, using them productively requires deliberate thought and an acute knowledge of their limitations. As the author says, it’s easy to get angry at a model, or to foolishly let it nudge you towards more code and more tests — even when that is suboptimal.

To a certain extent, models keep getting better and better at discerning our intentions and providing value. Yet I am not sure whether we will reach a point where using them successfully no longer causes the kind of fatigue that it does today.

Buying a camera or car is different from paying a subscription, right? Different expectations

Surely there should be a way to enable deep links with a fallback to Safari without needing Gmail to know what apps I’ve got installed

I got 68,900 words, with the vast majority of the errors being on the grandmaster level.

As a non-native English speaker, I found that result pretty good! Though being a native Portuguese speaker certainly helped me as many difficult words in English borrow from Latin, and in Portuguese the Latin influence is more pronounced.

Yak shaving with AI allows me to function more as a systems designer, code reviewer and tester than a coder per se.

AI is great if you simultaneously guide it and let it guide you. I take my time building a very detailed spec for what I want, then run it through the AI looking for contradictions, misconceptions, edge cases, performance bottlenecks, potential optimizations… anything that might cause problems in the future. Usually these discussions lead to multiple spec-improvement journeys, and that’s where the bulk of learning in a project comes from. Sometimes the AI will flag actual issues, while other times I might need to rein in its proposals — mostly in terms of feature creep and finding non-existing problems. I believe this back-and-forth is the most significant aspect of making the best out of yak shaving.

By the time the spec is “final”, it can be quickly implemented by an AI as I watch, review and test, with practically zero code banging on my part. This way, I get to understand precisely how the project works, make it tailored to my needs, and still not waste time, muscles or even mental bandwidth with menial coding.

I always liked yak shaving, but avoided it because I knew it came with costs and tradeoffs. More recently, with the help of AI, I’ve been doing lots of it, as the costs and tradeoffs have greatly diminished. In fact, I’ve learned that building my own tools and frameworks, when done properly, comes with huge performance benefits and helps me understand the problems I’m trying to solve much more deeply. There has never been a better time for yak shaving!

I hadn’t thought about that, it’s a valid use case and likely to have increasing demand as drone deliveries become commonplace in the next few years.

Very convincing indeed! I really want to know what AI made that, as I’m looking forward to creating personal/customized music in a similar way.

I wonder why Google doesn’t bother competing with Microsoft in the flight simulation niche. All that Google Maps data would be pretty cool to use for that purpose, but instead we’ve got only this toy feature inside Google Earth.

I disagree. I’d prefer if my government invested more in AI solutions, so as not to depend so much on foreign technology.

In an ideal world, Brazil would have a thriving private sector, capable of competing even in the AI sector. Unfortunately, that’s not the case, and I believe that without government action such endeavors won’t really succeed.

I wouldn’t describe what happened here as incompetence. As a “carioca”, I am pleasantly surprised to know that the government’s IT department is involved in AI work — even without the budget to create its own models from scratch.