Keith

Building AI systems for businesses

[Keith @KeithIsKosher]

https://t.co/QZBcgSh9Kl

[Keith @KeithIsKosher]

@WARDOGS ty john wardog can we get an xp boost or some head?

[Keith @KeithIsKosher]

@thsottiaux My finished site has failed to publish through Codex 8 times with the same internal error. Recent task history also vanished from view. I've spent hours and a lot of usage on this. Support escalated it. Can you help get a fix and a usage refund?

[Keith @KeithIsKosher]

Have an office cleaning bid to finish?

For $399, we prepare the proposal, private cost sheet and crew task list from your notes. One office, up to 8 areas. Reusable kit included.

Taking project inquiries. Send your deadline:
https://t.co/IolZXWY1n4

[Keith @KeithIsKosher]

@LoveladyLuke Being rushed through the walkthrough would worry me more than the competing bids. Was $300 their stated budget, or what you estimated after seeing the site?

[Keith @KeithIsKosher]

“General office cleaning” leaves a lot open.

Dishes? Fridge interiors? High glass? Restocking supplies?

I’m building Office Bid Kit to carry those details from the bid to the crew checklist. Here’s the free walkthrough sheet:
https://t.co/gNeePcmNqf

[Keith @KeithIsKosher]

@thsottiaux I think somethings fucky boss

[Keith @KeithIsKosher]

Opus still wins on frontend design. Cleaner structure, better component logic, less rework. Codex is faster and cheaper for raw code but usually needs more cleanup. Price matters. Id switch if Sol or Codex matched that design quality at the same reliability. Happy to show examples if you want

[Keith @KeithIsKosher]

@kunchenguid Anyway we can have an option to hook into a sub account showing quota instead of am api account? For instance my kimi quota won’t display properly. Currently, it states (sub)

[Keith @KeithIsKosher]

@kunchenguid Think Anthropic will ever let the Claude sub quota work outside Claude Code, or are they just locking it down on purpose?

[Keith @KeithIsKosher]

@alexocheema Interested! Real end to end agent testing instead of just TPS is exactly what’s been missing. Happy to help!

[Fish Audio @FishAudio]

Today we’ve raised $52M Seed and we are announcing the public launch of S2.1 Pro.

>It can clone a voice from 5 seconds of audio
>2x faster than Cartesia & 1/6th the cost of Eleven Labs
>most expressive model with word level control over emotion, intonation, pacing etc

We support frontier AI companies including HeyGen, LiveKit, Retell, Sanas, and OpenArt all run our model in production.

If you're a business and we can't cut your voice AI costs by 50%, we'll give you 1 year of Fish Audio for free.
Book a demo: https://t.co/vHkyZf9JoG

To celebrate our first birthday, we'll give you 1 month of S2.1 Pro for free. Like, retweet, and comment “Fish” to get it.

[Keith @KeithIsKosher]

@FishAudio Fish?

[Keith @KeithIsKosher]

@BR1INFINITE Get me in coach

[Keith @KeithIsKosher]

@DannyKPolitics Very tough internet individual

[Keith @KeithIsKosher]

@sacsheriff Free fireworks shows?

[Keith @KeithIsKosher]

@TwannaNell33869 Haha yeah, graveyard is the perfect word for it. Reseek sounds like a solid fix though. the OCR + tagging combo is exactly what’s missing for most of us

[Keith @KeithIsKosher]

https://t.co/QZBcgSh9Kl

[Keith @KeithIsKosher]

Two AI launches are colliding this week.

OpenAI is moving GPT-5.6 Sol, Terra, and Luna from limited preview toward broader access.

Musk says Grok 4.5 goes public tomorrow as an Opus-class, faster, cheaper model.

The obvious take:

"Which model wins?"

The operator take:

"Which model makes the workflow cheaper?"

This is the shift most people are missing.

Models are not just capability ladders anymore.

They are product lines.

Sol: maximum capability
Terra: daily work
Luna: speed and cost

Grok 4.5 is being sold the same way:

not just smarter, but faster, more token-efficient, and cheaper.

That tells you where the next AI edge is.

Not picking the smartest model once.

Building a workflow that routes work by value.

Use the expensive model for:

strategy
architecture
hard debugging
final judgment

Use the cheaper model for:

summaries
cleanup
classification
draft variants
monitoring

If every task goes to the biggest model, you do not have an AI system.

You have a very expensive autocomplete.

The winners this year will not be the people who chase every launch.

They will be the people who build a model router for their actual work.

Question:

What do you still send to the expensive model that probably belongs on the cheap one?

[Keith @KeithIsKosher]

@BR1INFINITE as a former 90 cranker and current extraction shooter gooner, count me in?

[Keith @KeithIsKosher]

the useful benchmark for agentic models is not just “did it solve the task?”

It is:

1. goal retention after 50+ tool calls
2. recovery after bad intermediate work
3. knowing when to ask instead of brute forcing
4. validating before claiming done
5. handing off clean state to another agent

that is what separates a smart model from a reliable operator

[Keith @KeithIsKosher]

I broke down the full system here:

Your Second Brain Is No Longer a Metaphor

https://t.co/FjcR798Smw

[Keith @KeithIsKosher]

https://t.co/QZBcgSh9Kl

[Keith @KeithIsKosher]

Most people did not build a second brain.

They built storage.

The old system was:

save notes
organize folders
hope you remember where everything went

The new system is different:

capture context
retrieve by meaning
let AI surface what your first brain forgot

The rule:

Offload memory.
Keep judgment.