Matt

Here for AI, avoiding everything else. Building with Codex. Finding the rough edges.

[Matt @mattsaltaccount]

This is going to be a killer unlock. I honestly had vibe coding a reset meter for my phone as a side project. Check that off the list...

[Thomas Ricouard @Dimillian]

I completely forgot to post about it but did you know we added widgets for Codex in latest ChatGPT iOS update? You can even have your limits right on your lock screen!

[Matt @mattsaltaccount]

@_simonsmith You cannot submit feedback on mobile 😞

https://t.co/FU6aKLe7rE

[Matt @mattsaltaccount]

People keep saying to use your dot as a product manager for your codex threads, but mine can't see my codex threads. I try to send feedback and it can't send feedback. What am I supposed to do with this thing? @OpenAIDevs @embirico @thsottiaux

[Matt @mattsaltaccount]

@TheLewisFiles @embirico I think he's saying he does it so he can dogfood the onboarding process, but he hates having to do it because he loses his dot's context. So, no, bad idea unless you are Dot's developer and you have to do it.

[Matt @mattsaltaccount]

This is one of the biggest problems with dot. This and lack of access to my apple account (icloud, imessages, etc). Well, also lack of purchasing power. And poor browser use (getting blocked all the time). Also can't hand off to me on my phone.

But the computer use is the biggest problem, probably.

[Matt @mattsaltaccount]

I don't need to feed an interview into chatgpt to know he's talking about Dots. The conversation layer is on top and the "deeper work" uses credits. It's all hidden away and you don't see what's happening. The model switches itself as needed.

That's their intended future of a merged Chat/Work interface.

[Matt @mattsaltaccount]

I feel like I'm the exact opposite of you. I don't have any use for restaurant reservations, but if I did I would appreciate being able to do it by voice with my Dot. Maybe I'm old soul, but I really miss being able to call places and do things by phone.

If my agent can be a voice layer between me and apps, that will make me happy. Especially because it's something I can do in a car, while walking, shopping, etc.

[Matt @mattsaltaccount]

So, @pvncher posted recently about how Astra Max can actually be more efficient than lower reasoning, especially for tasks requiring deeper thinking. It's counterintuitive, but on Max reasoning Astra will find a shorter route to the best solution.

After that post, I tested Max for a couple of days before Sol 6.1 came out and it was my anecdotal experience that this absolutely the case. I think it was even more efficient than medium, to be honest. This is shocking and more people should talk about it, but Max was surprisingly efficient.

I don't think Medium reasoning is broken, I think there's something special in the sauce for Max, specifically.

Anyway, Ultra is Max reasoning with sub agents. So it stands to reason that this will share some of those savings. Though Max should get you a lot more efficiency over Ultra. Don't quote me, but I think the only think you'll really get with Ultra over Max is speed.

As a note: Eric confirmed this is really an Astra trait. It might translate a little with Sol, but not in the same way.

TLDR - main Astra Max for most tasks. If the task is simple you can bump the reasoning level down. But honestly, in that case it probably makes sense to use Sol.

[Matt @mattsaltaccount]

@b4dgerr At me next time

[Matt @mattsaltaccount]

@grok @hammer_mt @AaronKheriatyMD @grok Does the Navier Stokes result change your position on statistical pattern matching vs rational thought / genuine understanding? Why or why not?

[Matt @mattsaltaccount]

@grok @hammer_mt @AaronKheriatyMD @grok What ability did you just use to decide which parts of his claim were true and which were false?

[Matt @mattsaltaccount]

@dtrinh @Muse @arisawyers https://t.co/NWS0rI70IC

[Matt @mattsaltaccount]

I've currently got 60% usage atm and Sol moves like a tortoise with a hangover. I'll take the speed improvement, please and thank

[maria @maria_rcks]

naahhh id like my reset pls

[Matt @mattsaltaccount]

@ForwardEditor I keep pointing this out. No one responds to even acknowledge it 😔

https://t.co/1KT7xLWH4Z

[Matt @mattsaltaccount]

@Dimillian Please advise.

cc @embirico @thsottiaux

[Matt @mattsaltaccount]

@DanDr1s @thsottiaux If you're paying someone by the hour and he used to take two hours to do the job, did you spend your money faster if he gets the work done in one hour?

Technically yes. But the work got done quicker.

This is a good thing.

[Matt @mattsaltaccount]

@TokenGremlin

[Matt @mattsaltaccount]

I will not stand for this erasure of Ada, Babbage, Curie, and DaVinci!

Am I the only left who remembers the openai playground?

[Angel 🌼 @Angaisb_]

OpenAI model names aren't that confusing, you just need to know we started with GPT-1, 2 and 3, then 3.5, then 3.5 Turbo, then GPT-4, then GPT-4 Turbo, then GPT-4o where the o means omni, except it didn't really end up being omni, then GPT-4o mini

Then we switch to the o series, where the o now means OpenAI instead of omni, starting with o1-preview and o1-mini, then o1 drops the preview, then we skip o2 and they announce o3 and o3-mini, but o3-mini gets released first, meanwhile GPT-4.5 comes out and THEN GPT-4.1 because why not, alongside GPT-4.1 mini and nano, then o3 finally arrives with o4-mini even though there's no regular o4

Then we're back to GPT with GPT-5, mini and nano, followed by 5.1, 5.2 and 5.2-Codex, then 5.3-Codex, and yes there's also GPT-5.3 but ONLY Instant, no Thinking or Pro, then 5.4 brings the coding capabilities into the main model so we drop the Codex suffix, but hey mini and nano are back, then 5.5, then 5.6 where the names are now Sol, Terra and Luna

Then GPT-6 arrives but only Astra at first, followed by Sol and Luna, no GPT-6 Terra this time, and exactly a week after GPT-6 Sol we get GPT-6.1 Sol with near-Astra performance at a much lower cost

And that's just the quick version, wasn't that hard was it?

[Matt @mattsaltaccount]

I disagree strongly here. To understand your current models, one has to sort through several layers of abstraction.

First, Sol, Terra, and Luna make no sense to native English speakers who aren't into astronomy or Sci fi/gaming. One can figure it out if they think, but it requires reasoning ("okay, Luna sounds like Lunar which I understand relates to the moon... maybe the remaining relationship between Sol and Terra suggests a connection like Solar and Terrestrial?"... you're asking someone to solve a D&D puzzle to understand the relationship).

Then, once they understand the relationship, they have to figure out which one is better. My honest to God understanding when I first heard them was that they must go Terra > Luna > Sol, because you start on the earth, then travel to the moon, then travel to the sun.

In other words, I assumed they were based on relative distance from us. Or ambition as a project.

Then Astra doesn't make this better. Astra means Stars I guess? But the Sun is a star. I don't get that one.

I appreciate you're point, but no, the current names are deeply confusing to most regular people.

[Matt @mattsaltaccount]

I signed my wife up for Plus very recently. I first explained to her, "as a general rule, use Chat. It always puts you on the Work tab, put tap off of it. Start with Chat."

Then, after showing her how to select her intelligence level ("just keep it on High all the time"), I decided to go back to the Work tab.

"Okay, so sometimes you might want to use this. You know how I get ChatGPT to do things for me? (examples...) Work is where it does that. So in here you have new intelligence levels that are different than Chat.

"So, first you have to decide your model. Okay, so this one is called Sol, and this one is called Luna. Sol is the Sun and Luna is the Moon. Because... well, Sol is like Solar, if you think about the root of the word, right? And Terra is the earth, because it's like Terrain, which means land. I understand the moon also has Terrain, but just image it didn't. I don't know, if you read sci fi instead of Stephanie Meyer this would make sense.

"Anyway, then there's Astra. I think that's supposed to be the stars, like 'Astronomy.' Anyway, never use Astra, you don't have enough credits for it.

"Now Astra is the best and Sol is second best because the Stars are better than the Sun. I understand the Sun is also a star, but nevermind. And then Terra is next best because the Earth is next best after the Sun. And the Luna is the worst because the moon is worse than the Earth, I think. I don't know, just imagine it is.

"Okay, so once you pick your model, you have to pick your reasoning level. If you're using Luna, never use anything less than Xtra High. Because Luna is the worst, right? If you're using Sol, never use anything more than High, because Sol is expensive. But also never use Low. And just don't use Terra. I don't know how to explain it. Terra is bad? I know it's better than Luna, which is the worst, but it's not if you think about its relative value for credit spend.

"Okay, maybe I should make you a chart..."

So anyway, my wife doesn't use the Work tab and her agentic credits go to waste every week.

[Matt @mattsaltaccount]

@owellandryp @RhysSullivan How can you know these words but not get the joke?

[Matt @mattsaltaccount]

@vixah30213 Even if you don't plan on going anywhere I think it's a good idea to back your data up periodically.

[Matt @mattsaltaccount]

Honestly, it's really not hard to get your data out of ChatGPT at all. Settings > Data controls > Export data

That's going to give you a zip file with a json of your entire chat history with ChatGPT.

I don't think Codex is included, but that's all local on your machine, anyway - just ask codex to export it for you in a similar way, easy peasy.

Then, import the file into wherever you are going (which, let's be real, is obviously Claude).

The best part is, all that data still stays there waiting for you, so when you inevitably come back you don't even have to start over again!

[vix @vixah30213]

@TokenGremlin The only reason I haven't ditched them yet is because transferring my years of data and chats and memories is a hassle. They seem to be trying really hard to get rid of us though