KevinSimback

Coding is the undeniable PMF use case for AI, but most knowledge work still has a ways to go
Coding works because all the context lives in a git repo that is versioned, structured, in one place, and usually with a test suite that tells you what is correct
Knowledge work on the other hand is based on information that resides in a bunch of different places - Slack, email, various systems, and often in people's heads
So if you want knowledge work to be automated like code, you need a "context repo" or as often called a "company brain"
But a company brain that just takes files, puts them somewhere

Putting an LLM as judge is the simplest and easiest way to create a loop
1. Do the thing you already do by prompting AI
2. Use another model to evaluate the output and give it feedback
3. Iterate until both are satisfied
Only see the final output once the loop is complete
Since you're using an LLM as judge, this is generally best for non-deterministic tasks like research, design, building specs, etc.
Using Looper to do this will make it super easy by refining your prompt for a better constructed loop, adding the judge, and setting the exit criteria so it doesn't run unnecessary iterations
Now

PTSD induced, iykyk

Just departing Italy amidst this heat wave, I really don’t get why AC is even a debate
Literally everyone - locals included - are complaining
Sweltering indoors just makes people irritable, install the AC already

If you’re not US or China, I don’t know how you sit back and become dependents on them for the intelligence that will define the future
You’d think it would be a national priority to have a strategy here
Of course, easier said than done - you need big $ and talent to compete with the frontier labs
The easiest first step is to build sovereign data centers and host the open weights models or post-trained versions - at least you own some form of the intelligence
The UAE and Saudi Arabia seems to be going this path and making moves towards training
The most impactful next step would be focusing on

If you’re doing diligence on a new startup, you now need to assess their “AI nativeness” alongside team, product and market
Why? If they’re not at the forefront of using AI, it puts them at risk of not executing fast enough
It’s not a hard pass filter, but it is signal - here’s a few early tells:
1. Did they give you a .md file, alongside a deck, that you can feed to your LLM?
2. Did they build an interactive HTML deck or just send a PDF/docsend?
3. Do they talk about proprietary skills, data, evals or methods they’re using to build their product?
4. Do they ship significant product updates be

If you’re diligence a new startup, you now need to assess their “AI nativeness” alongside team, product and market
Why? If they’re not at the forefront of using AI, it puts them at risk of not executing fast enough
It’s not a hard pass filter, but it is signal - here’s a few early tells:
1. Did they give you a .md file, alongside a deck, that you can feed to your LLM?
2. Did they build an interactive HTML deck or just send a PDF/docsend?
3. Do they talk about proprietary skills, data, evals or methods they’re using to build their product?
4. Do they ship significant product updates between mee

Riddle me this:
What happens when GLM-6 comes out and has Fable/Mythos capabilities?
It’s not a matter of if, it’s when
And understanding the scenarios and implications are incredibly important

A big problem with crypto right now is you have to assume every new token project is going to rug, either intentionally or not
But for decentralized AI to work, you need some coordination mechanisms and tokens make most sense
How do we reconcile this?

Meta is starting to look like the biggest own goal of the AI era

Fable came and went - was it good?
GLM 5.2 just dropped - is it good?
Most will answer anecdotally, but we should all be creating our own personal evals
Pick a few structured workflows that you commonly do, build an eval set, then run them with any new model as a first task

The consumer inference conundrum:
I see a lot of consumer-oriented inference projects working to address the high cost of frontier APIs
Methods vary, but the premise is to provide cheaper inference vs aggregators like OpenRouter
The challenge here is one of timing
Today, consumers get highly subsidized frontier AI
Plenty of options for consumers to get SOTA inference plans starting at $10 and the $200/mo plans offer an insane amount of usage
There just isn't nearly as much consumer demand for API pay-as-you-go inference given the existence of these sub plans
*enterprises are a different story