Menno

vindictive, vicious cruelty designed to humiliate and criminalize every trans person in kansas

horrifying fascist shit reminiscent of the runup to the literal holocaust

Yes thank you! I do have atemporal decay system I'll have to figure out how to integrate recency

Mamdani’s Vanquished Foes Are Plotting a Comeback, if Not Revenge

A bunch of billionaires and corporations pooling their money to try to keep a popularly elected mayor's agenda in check are not exactly making the case that the city and state shouldn't tax the hell out them.

But the problem is that in my employee handbook it explicitly states that anything I build at work ,(or using company resources)is owned by the company.

So I can only bring my side work to actual work if I build something entirely distinct so they can't claim company solution came first.

I mean yes. But my company needs to have record breaking returns every quarter they can waste those funds on me!
/S

I don't need a powerful model for this and even if I did they are dirt cheap but I think I can use a local model.

Outside of work I'm building a memory management system for my agents because I think it will be useful (and I'm building towards my own idea of an openclaw).

It uses text embedding models to convert memories to vectors for semantic search and 💡

I am working on a tool at work that will help us streamline our reporting process. Part of this process is using an AI to address ambiguity. It's very good but at scale it will be an issue. And the issue isn't complex, just too complex for stuff like Fuzz or other "normal" models.

At work my tech stack is super limited. I also have to be mindful of costs. No, worse than you're thinking. It look 12 years of consecutive people in my position asking for it for us to get approval for a Jira pilot.

This bleeds everywhere

Working on a side project last night and I had a breakthrough for a work problem that's been bugging me for months.

This is why companies should encourage, not stifle, people having side work

Don't watch the state of the union. Don't skeet clips. Ignore that shit stain

a man is standing in front of a large monster in the fog holding a wreath . Alt: a man is standing in front of a large alien behind a screen. From Arrival

I think Narn but that's tv

Movies is Arrival

As a daily pusher of a wheelchair, trust me, this is a very big deal

Tldr:
-LLMs are a tool and can be a valuable step in process (not just only step with most public models)
-Built in checks and hard coding can help make hallucinations something that is not significant, but this requires you knowing the process

Now instead of having to review 10% of the data manually, you're down to 3% or maybe less. This can be tens of hours saved.

The people who are like "I just fed it to AI and told it to do it" are almost certainly lying, but is wildly valuable (even simple models) for ambiguity cleaning

So you process the entire database using python and get 90% of the way there. The remaining 10% you pass off to an LLM with very specific instructions and tests. It responds and gives you outcomes+ confidence. You can build python to test the (smaller) edge cases it responds with.

I'll answer my way

So python or similar code is still unmatched when it comes to data sorting, where LLMs can help us when you hit ambiguity. Instead of having to hard code edge cases or use models like Levenshtein distance as the source of truth you use LLM as backups

Favorite example of this is the number of people I see saying they're using a conversational LLM to generate long tail keywords for SEO.

Like... Y'all listening to yourself?

Yeah definately. but at least in my industry (media) the other big problem is that people are seeing everything through lens of "next Google" and thinking that it will be used just LIKE google. Despite literally every report showing this isn't the case.

What I am finding is that so many people think that AI models start and end with "What can i ask in the form of a question to get output" and ... nothing else.

SOmeone talking that Claude Code was inferior to a few "specialized" AI media tools because they didn't consider you could... build tools

Elon can be excused for not understanding how taxes work considering he doesn't pay them

once again. politicians highlighting cool and good shit they and the government does is awesome

Lee. Because my parents read something about names needing five syllables and so that is what they liked the most.

And the best.. black raspberries.

Rude

And this is before you get to the issue of latency meaning each of these satellites will be their own system and they can't reliably share tasks.

AI is an arms race over speed. So Elon is proposing either launching a ton of very expensive tech that will fail into space. OR he's recommending launching even more expensive tech into space that can't run frontier models.

You have cosmic radiation, micrometeorites, debris. The earth offers some protection but you still need shielding and ideally tech that's hardened against radiation. This tech is slower than earth bound versions

AI models need serious hardware because they need a lot of very expensive ram and/or gpus so that they can keep the entire model in memory.

These are highly advanced pieces of tech.

You know what's not good for highly advanced tech? Space.

AI compute in space via microsatellites is an idea so bad only Elon could think of it.

In this and this alone he is a singular talent.