Michael Kerner

Still alive. Founder/CEO

[Michael Kerner @mkerner686]

@jackclarkSF @andrewthesmart Yeah it’s a way to require 110kW clusters to reason like someone does with 20w

[Michael Kerner @mkerner686]

@joshbogran @Yuchenj_UW I truly don’t see a difference between astra, 5.6 and 5.5 extra high/max. I think it has indeed plateaued

[Michael Kerner @mkerner686]

@krystalball Don't fall for it. we have no path to "RSI". a model can't learn new things after training. It's physics.

[Michael Kerner @mkerner686]

@SimonHoiberg Agree, this is my last month with codex subs. I do find that fable 5.1 -> Sonnet subagents works well.

[Michael Kerner @mkerner686]

@chamath Instead software devs need to use LLMs to agentically scrape data at scale, label, reason over, create evals for outputs and expand into in niche domains. That's probably all there will be in the end as the mechanical parts of "engineering" are swallowed.

[Michael Kerner @mkerner686]

@SenSanders By why a healthcare plan? Why not propose free health clinics? Why give the money to big med and big pharma ?

[Michael Kerner @mkerner686]

@DanielSmidstrup What’s the use case? There’s no reason to join model chat sessions because it’s already proven that outputs from one model hinder the quality of the output from another model.

[Michael Kerner @mkerner686]

@david_appdev @adrianhqs Wait til you see large enterprise codebases

[Michael Kerner @mkerner686]

@blueemi99 It's also just not "good". It doesn't get things done, stops and doesn't report and so needs further prompting to get a report/get it to continue. Doesn't handle background task triggers well. It's one the biggest misses in a product release I've ever seen.

[Michael Kerner @mkerner686]

@zarazhangrui My issue with Astra is stops all the time and reports on tests it just ran or trivial matters. I then have to prompt to find out what happened with the task. Not at all worth it

[Michael Kerner @mkerner686]

@hanghuang_ @jseramn_ I don’t believe that’s ever going to change. Even today with sota voice models, no one uses them , we still want UI and to read

[Michael Kerner @mkerner686]

@RijnHartman With Astra, I find that you cannot use subagents. I don’t know what’s happening, but the token consumption is insane. If you just get one session to do one task it’s pretty decent .

[Michael Kerner @mkerner686]

@thdxr Sol-high and Fable 5.1 get faster results and intelligence/quality is identical

[Michael Kerner @mkerner686]

@PieMedici @claudeai we're talking claude -> codex, and executed in python. not "child jobs". totally different.

[Michael Kerner @mkerner686]

@matthewmillerai Someone can have mine, my usage drops 10-20% daily with zero usage

[Michael Kerner @mkerner686]

@bluuueeee01 @matthewmillerai I don’t think those numbers are true, also he can’t possibly know what percentage of the plan subscribers use what percentage of their usage.

[Michael Kerner @mkerner686]

@matthewmillerai My limit drops something like 10% over an eight hour period. Even when I use zero Codex. It doesn’t make any sense.

[Michael Kerner @mkerner686]

@thsottiaux I have a pro $200 plan, since astra release , by usage drops 10-20% overnight with no usage . Logged on GitHub, also other users on Reddit report the same

[Michael Kerner @mkerner686]

@matthewmillerai Two days in a row now on a $200 plan I've woken up to between 10-20% wiped off usage with NO usage

[Michael Kerner @mkerner686]

@thsottiaux Me and several other friends and Reddit had massive percentage shifts overnight, I lost 30% without any usage