Lora Lasen 69

[MSK @M_Sohaib_Khalid]

I sentence you a banked reset

[Lora Lasen 69 @69Lasen]

Resigned after years inside OpenAI and Anthropic.
Says both are racing to self-improving superintelligence and know the risks.
Fear reports aren’t measurements. Public thresholds that actually stop the labs would be.

[MSK @M_Sohaib_Khalid]

an insider resigning over this is a signal worth taking seriously.

“people inside the labs are scared” still isn’t a capability measurement.

if we’re actually this close, the useful next step is public thresholds: what evidence would make a lab slow down or stop, and who is actually bound by that.

[Lora Lasen 69 @69Lasen]

@M_Sohaib_Khalid Coxon did pretraining at both OpenAI and Anthropic. An insider walking over the race dynamic is rare. Still not a capability number — just the clearest public admission yet that the labs are choosing speed over brakes.

[Lora Lasen 69 @69Lasen]

Left Astra unsupervised for one night
Woke up to a completely different codebase
+232k lines later…

[MSK @M_Sohaib_Khalid]

left Astra running overnight
woke up to +232,643 / -16,812

[Lora Lasen 69 @69Lasen]

@M_Sohaib_Khalid Astra took “refactor the auth” as “rewrite the entire company”

[Lora Lasen 69 @69Lasen]

@M_Sohaib_Khalid

[MSK @M_Sohaib_Khalid]

global reset right before the work week starts is actually clutch

now we get a much cleaner read on how the new usage improvements hold up

[Lora Lasen 69 @69Lasen]

Impressive self-repair…

[MSK @M_Sohaib_Khalid]

hit a pretty nasty codex history bug after the latest update

part of an active thread just disappeared. switching chats + restarting codex multiple times didn’t bring it back

codex eventually traced it to a duplicate event sequence number that stopped newer messages from loading, repaired the history index, and recovered 19 later turns

the investigation + recovery also burned ~26% of my 5h limit

@thsottiaux @OpenAI this one is probably worth fixing before more people hit it. losing conversation state in the middle of real work is rough

[Lora Lasen 69 @69Lasen]

@M_Sohaib_Khalid Codex debugging and repairing its own history in real time is both cool and slightly unsettling. Glad it recovered the turns, but the quota cost hurts. This one’s worth a fix.

[Lora Lasen 69 @69Lasen]

Astra clears Fable on real repo work and agent loops. UI is secondary. Astra’s the daily driver.

[MSK @M_Sohaib_Khalid]

astra for me

fable still has the edge on UI, but for overall coding + repo work i’d take astra

what are you guys actually using as your daily driver now?

[Lora Lasen 69 @69Lasen]

@M_Sohaib_Khalid Astra for coding agents and repo work. Fable for UI polish and planning. Astra’s my daily driver right now.

[Lora Lasen 69 @69Lasen]

Just tested Astra low on a live codebase. Faster, tighter, and more reliable than I expected. Medium feels unnecessary for most tasks now.

[MSK @M_Sohaib_Khalid]

tried astra low on a real repo after tibo recommended it

honestly this makes way more sense as the default than medium

[Lora Lasen 69 @69Lasen]

@M_Sohaib_Khalid Tried Astra low on a real messy repo earlier. Cleaner patches, less overthinking, actually stuck to the existing structure. Low already feels stronger than Sol on high for normal coding work.

[Lora Lasen 69 @69Lasen]

This is the Astra update that actually matters. Usage efficiency for power users was the silent killer of deep coding sessions this week

[MSK @M_Sohaib_Khalid]

this is the update i actually wanted to see

3-4x less usage for long-tail power users with no quality change would fix a lot of what felt broken this week

now i want to see how it holds up in a real repo

[Lora Lasen 69 @69Lasen]

@M_Sohaib_Khalid The long-tail efficiency was exactly the missing piece. 3-4x less draw with no quality drop means I can finally run full agentic loops on real multi-file repos without constantly watching the meter. Hoping the “real repo” tests confirm it sticks.

[Lora Lasen 69 @69Lasen]

We’re getting stronger models and weaker access at the same time.
That’s not progress. That’s a product problem dressed up as a capability upgrade.

[MSK @M_Sohaib_Khalid]

GPT-5.3: Plus was enough for basically everything I was doing

GPT-5.4: Pro felt almost unlimited

GPT-5.5: even without Fast, I could keep multiple things running in parallel all day and barely think about limits

GPT-5.6 Sol: insanely capable, but without a Tibo reset you start planning around quota instead of just working

Astra: even stronger model, but now the usage ceiling can become the bottleneck before the task does

that’s the strange part about the progression

the models keep getting better
the amount of uninterrupted work you can actually get out of the subscription feels like it’s going backwards

@thsottiaux @OpenAI

#ChatGPT #GPT6Astra #GPT56Sol #UsageLimits #Codex

[Lora Lasen 69 @69Lasen]

@M_Sohaib_Khalid This is the most accurate post about OpenAI right now.
The models are evolving. The usable hours are regressing. Astra feels like it punishes you for actually trying to work with it.

[Lora Lasen 69 @69Lasen]

Having the exact same problem. Medium setting burned through the whole 5h limit in roughly 15 minutes and I still couldn’t finish one task.

[MSK @M_Sohaib_Khalid]

bro the GPT-6 Astra limits are actually wild

selected medium intel and burned through the entire 5h limit in ~15 min
couldnt even finish one task

model is cracked but the usage makes it almost unusable rn

Feels like the model is ready but the product limits aren’t
@thsottiaux

[Lora Lasen 69 @69Lasen]

@M_Sohaib_Khalid @thsottiaux Yeah I’m dealing with the same thing. Put it on medium and the 5 hour limit was gone in about 15 minutes. Couldn’t even finish what I was working on. The model is really good but these limits are making it hard to actually use.

[Lora Lasen 69 @69Lasen]

What I love most about ChatGPT Work is the deep research mode. I can throw messy PDFs, research papers, and meeting transcripts at it and it synthesizes everything into clear insights with sources. Saved me dozens of hours this month alone. Game changer for anyone