tmarthal

.🤖.

[tmarthal @tmarthal]

@ryanhammer09 Would not sweat it; All translations are going to come out 100% generated. Doubt that the tourist writes in English

[tmarthal @tmarthal]

@QuinnyPig IAM is setup to do this, but their product managers can't see the vision and simplify the setup🫠

[tmarthal @tmarthal]

@nickcammarata @JakeOrthwein Can you link that article? I thought the opposite was true (meditation increases short term
Memory)

[tmarthal @tmarthal]

@pdrmnvd Back in my day we just called this stenography

[tmarthal @tmarthal]

@eriktorenberg hmu

[tmarthal @tmarthal]

@KelseyTuoc @anistotle_ @jt_kerwin Right. If the prompt inputs are used as part of an evaluation loop or other optimization, it’s technically not part of model training (so not lying). Doubt we will ever know, but I assume that any prompts input are still stored and could be used as part model release/readiness

[tmarthal @tmarthal]

@KelseyTuoc @anistotle_ @jt_kerwin How do you know what it means to the model when you turn off that training flag? Data that you submit to the models could be in one of the RLHF layers to train weights without the raw data exposed

[tmarthal @tmarthal]

@tdhopper Check this out next https://t.co/SC0Rn1QlaD

[tmarthal @tmarthal]

@AdiPolak QA Engineer is going to be the job of 2027 😆

[tmarthal @tmarthal]

@p_naix Yeah, I looked at the migration documentation - they are targeting folks that are deploying containers on AppRunner, not the source build deploys. I messaged my AWS Startups Solutions Architect about the change/migration not being the right solution for dockerless-apps.

[tmarthal @tmarthal]

@bankertobuilder California is undefeated with parking

[tmarthal @tmarthal]

@AdiPolak The new Quality, Speed, Cost Triangle

[tmarthal @tmarthal]

@codr_1 @QandAinPublic @prasann_pandya Right, it's the ease of development vs the ease of performance optimization. I've always worried about performance after the application is successful; 99% of the time every performance enhancement is done prematurely.

[tmarthal @tmarthal]

@Cladriah @funkykongfac It absolutely is, impossible to differentiate the smartest trolls from the washed idiots; but this is just too good and it’s perfect 2026 bait

[tmarthal @tmarthal]

@codr_1 @QandAinPublic @prasann_pandya Those benchmarks below are comparing GRoutines to the fastAPI process. Make sure your python process manager matches your infrastructure, gunicorn with gvnicorn workers or use Granian

[tmarthal @tmarthal]

@codr_1 @QandAinPublic @prasann_pandya this is just objectively not true in 2026

[tmarthal @tmarthal]

@jacksondahl An Australian man went viral talking up a mat Pilates video in 2024 with a similar vibe (lots of front body opening, glute engagement).

https://t.co/yHMX15qoF3

[Alejandro Companioni @acompa_]

This might be true in neutral workflows, but decoupling the LLM from the agent is difficult for Codex vs. CC. Using both models outside of their coding agents produces a _very_ different experience.

There's no clearer example of this than CC's built-in prompt expansion logic. It's very obvious when CC delegates instructions - the main loop provides the sub-agent with clear, robust context and instructions.

IME the Codex app doesn't consistently expand prompts, and the LLM's instruction-following performance suffers as a result. Plans are also lower-quality and lack critical detail.

Of course this is all subject to change like, every 17 seconds with our current rate of progress.

[tmarthal @tmarthal]

@alexkehr @samsheffer @Google Searching with natural language is easier, I never remember the official names of some less used ones, e.g. 🌾🌱⚙️