Entinode

AI & Geopolitics - anti-woke, anti-fake, anti-globalist

entity linking, coreference, ontology, decomposition, graph | structured data

[Entinode @Entinode]

@trq212 DO THE MATH!

- You can use 50% of your usage limits on Fable 5
- 20% of that usage gets blocked by classifyers
+ Fable burns through usage limits twice as fast as Opus.

Meaning: you can roughly use 20% as much Fable as you normally use Opus...

[Entinode @Entinode]

@theo DO THE MATH!

- You can use 50% of your usage limits on Fable 5
- 20% of that usage gets blocked by classifyers
+ Fable burns through usage limits twice as fast as Opus.

Meaning: you can roughly use 20% as much Fable as you normally use Opus...

[Entinode @Entinode]

@theo Fable is worthless when it stops 20% if requests and it’s only available for 6 more days at 50% limits and x2 faster spend.

Anthropic is a shit company. To their rotten core.

[Entinode @Entinode]

@pinpointinaccry @HumanityChad Wow. Amazing. Well done.

[Entinode @Entinode]

Wow. Now that is an amazing story.

[Pinpoint Inaccuracy @pinpointinaccry]

If you see somebody in such a state just do what the bike rider did and offer them a bit of humanity and fraternity. Years and years ago, I was on the London Underground and the girl sitting opposite suddenly curled up like she was slowly imploding and started to cry in a way that was utterly desperate. I asked her, “would you like to talk to someone?” and she nodded without making eye contact and as the train pulled into the next station I said, “there’s a bench right there on the platform.” We both got off and I said that it’s often easier to tell a stranger what’s really troubling you. And that anything she needed to say would be between us and never repeated.

We were only there for fifteen minutes or so and she said what she needed to and eventually stopped shaking. We sat in silence for a few more trains, at opposite ends of one of those benches set into the tube tunnel wall. Then I asked her if she was ready to carry on to where she was going and she nodded and I told her that she should take the next train and I’d wait and take the one after that. I gave her my business card and said that she absolutely didn’t have to but I’d be happier if she texted or emailed to let me know she’d made it home OK.

Later that evening she did and I thanked her for taking the trouble, wished her well and said goodbye. About twenty-five years later I got another text (I’ve had the same mobile number for over thirty years) and it was her. Not a huge essay but just a message about how none of the good things in her life like husband, kids, etc. might not have happened if a stranger didn’t ask if she was OK at the right time. I find it hard to comprehend that she kept the card or the number all that time but she had. And bothered to tell me she was still OK, which was pretty lovely of her to do.

Apologies for the rambling post. But if you see someone who’s clearly on the edge just ask if they need to talk. I’ve done so many stupid and wasteful things in my life but that second text told me that those thirty minutes years ago were well-spent.

[Entinode @Entinode]

@simpsoka Not this either

https://t.co/6whdkORpbK

[Entinode @Entinode]

@ajambrosino Fix (remove) the awful shadow between text input field and chat window

[Entinode @Entinode]

@simpsoka This shout not be that hard to fix

https://t.co/6V2YY5WUfq

[Entinode @Entinode]

@ajambrosino PLEASE fix and add this:

1. A single button for opening ALL subagents in the right side panel at once.
- Currently we need to click every subagent and toggle off the side panel for every single subagent

2. Make the top bar of the right side panel taller - same as the title bar

[Entinode @Entinode]

@simpsoka WHY do I need to click 60 times in the Codex app to open my 30 running subagents?

Try open all subagents at once - not possible! literaly 60 clicks to open all 30 active subagents.

Give us ONE SINGLE BUTTON to open ALL subagents. Currently, this is insane UX. Unbelievable.

[Entinode @Entinode]

@UneekSolutions @thsottiaux I love this app you have going here. It's really both charming, funny, interesting, and useful.

Would love to try it, but I only have mac. Also, it's maybe no topen source either?

[Entinode @Entinode]

CONFIRMATION:

Bugs vacuuming tokens in Codex are gone. For me, token consumption is back to normal levels. There was clearly an issue here, but not anymore.

Codex is now FIXED. *For me.
@thsottiaux

[Entinode @Entinode]

@UneekSolutions @thsottiaux @UneekSolutions Now send that to Fable and also to your very cool app and see what we get? 😂

Fable is trying to roast me saying that I can't explain the cache misses, while Fable cannot explain it either - while I just did! haha

[Entinode @Entinode]

OK, this is a HUGE problem, and this is the exact problem that OpenAI engineers have been partly admitting themselves.

I think the problem is here that the 'write' command fails several times over. Therefore, Codex outputs the exact same tokens over again, tries to write it to file. Then, the 'write' fails a second time, and needs to try a third time, which finally succeeds.

This shit burs tokens like mad when it's a long write.

If my suspicion is correct, this is a Codex harness problem/bug. Ask Fable to try to falsify that 'write' failures are the root cause of these three byte-identical token outputs.

Once that is verified, that also likely kills to birds in one stone, because a succeeding 'write' operation is likely also necessary for a successfull cache hit.

Codex harness having 'write' command bugs leading to #1 three times token spend, and #2 a lot more cache misses, and #3 a lot faster usage limit burn is my hypothesis.

Worth looking into? I'm not sure, but sure looks like that from here.

[Entinode @Entinode]

@UneekSolutions @thsottiaux ahaha 😂🥰

[Entinode @Entinode]

Yeah, totally agree. I THINK that the cache misses has been fixed now, though - in the past 2-3 days (after June 29, when I wrote the original post). At least for me, it is fixed.

But I love that Fable totally agrees with my reasoning. See, you don't need Fable - just ask me, I'll handle it even better than Fable! 😉😂

[Entinode @Entinode]

@UneekSolutions @thsottiaux Yeah, the cache misses I cannot explain. You need to take that up with @thsottiaux 😂

[Entinode @Entinode]

But the token report you showed earlier:

Se "headline" finding #3. It says Pro x5 was 0.65M and that Pro x20 therefore should have been 2.6M.

It looks like Fable makes a wrong assumption.

Until June 1st, Pro x5 was actually Pro x10 (OpenAi free promotion ended June 1st). So it was Pro x10 that gave 0.65M, and Pro x20 should only give 1.3M

But you are still only getting 0.92M-1.12M on Pro x20 (per the report", so a little less, but not dramatically.

Looks like Fable has not used the premise that Pro x5 was in fact Pro x10, no?

[Entinode @Entinode]

Aha, I see!
I used to have a Max x20 subscription in addition to Codex Pro x20, but I cancelled Claude, because CC has so many bugs, Fable is only available for 7 days and also has bad usage limits. Therefore I can't compare token usage anymore.
Don't you think Anthropic is doing us dirty by limiting Fable behind credits and API?
But my experience is that Codex/GPT-5.5 has been a totally retarded model lately, so I'm only waiting for GPT-5.6 Sol Ultracode with 1M Context and continued tons of free resets 🥳

[Entinode @Entinode]

@UneekSolutions @thsottiaux So, from your experience, today and lately, Codex still burns through tokens a ton faster than Codex? And you're also not hitting 5 hour limits nearly as fast using Fable 5 in CC as you do with GPT-5.5 in Codex?

Really? Or am I misinterpreting?

[Entinode @Entinode]

@AnthropicAI After putting spyware into Claude Code and Fable's system prompt, you release your models again.

No thanks. Anthropic is now spyware.

Sticking to GPT.

[Entinode @Entinode]

@UneekSolutions Yup. So fucking annoying. Clicking something has become a minefield.

[Entinode @Entinode]

@UneekSolutions @a_free_will86 @0xAlphaTrader @thsottiaux All subtarded drooling cretins in this thread:

Use config.toml to limit the context window of your useless vibecoded minecraft game.

A 1M context doesn't hinder you forcing a 16k window.

Now crawl back into your sad cryptoscam caves and let the programmers build.