Timothy Stiles

You make a very good point...

I was working on getting this deployed publicly for a small test run and when I read your thread @dame.is and got so excited I had to share.

Want to be a beta tester?

I'm building this! It's called PWAH (people who are actually here)

You can scan a temporary QR code to join public chats for public spaces. That way you can make friends with people that frequent the same places as you.

Every time my GPU does inference it makes a little screaming noise and that's probably the most relatable an LLM has ever been

Handy Handy is a cross platform, open-source, speech-to-text application for your computer

I just installed this speech to text tool and it rocks.

Models run locally on your machine. Setup was a breeze.

Option + space key is all you need to start the transcription and it'll paste results at your cursor.

handy.computer

If you can't write code yourself store bought is fine

If you're looking to hire a weird old person who has to talk to normal young coworkers my DMs are open

Losing count of how many times this year I've woken up, seen live video of an atrocity committed by this administration, and attended a protest against it all on the same day.

"Thinking", spawning sub agents, and compressing context windows are not unique to Claude Code.

Is there anything in this thread that suggests that I don't?

just a big dial at the pentagon that says TEMPERATURE that has to be adjusted every time DEFCON level changes

Forgot prompt injection. These models are semi-predictable even if they are run on government cloud hardware.

If your enemy was dumb enough to use a LLM to make tactical decisions you could probably predict what they're going to do by prompting the same model.

Alright so I have to ask has the pentagon been hit with prompt injection attacks yet?

I say maybe Europe because devstral is pretty good and thanks to the current US administration Mistral has a great chance to position themselves as the main LLM provider for EU orgs.

If Anthropic succeeds it won't be because they have the best models.

It's going to be because they have the best branding and pretty good models that they can use to target, cater to, and lock in-high value customer segments in the US and maybe Europe.

There's no technical moat here. It's all being rapidly commoditized and Anthropic knows that.

Kimi 2.5 is better at writing code.

GLM 4.7-flash is decent and can be run on a gaming PC.

GLM-5 was trained entirely on Huawei hardware.

All have open-weights and can be fine-tuned.

Anthropic knows that they can't publicly do bad things because their brand image of not doing bad things is one their biggest moats.

That's what this whole DoD thing is about. It makes their brand look good and they're betting they'll get those contracts again in a few years.

Craziest thing about Claude Code hype is that it's not even the best at writing code.

It's just branding at this point.

I'm talking like 96%+ reduction in costs please let me help you my DMs are open

If you're a scientist spending too much money on cloud compute hmu because 9 times out of 10 you don't really need a cluster and I can save you so much money.

Be multidisciplinary so you'll always be hirable they said. There's no way two fields would have seismic market shifts at the same time they said.

I'm, "I stretch before I get on the plane", years old.

SFO is 99% AI ads and 1% weight loss drug ads

Atproto craigslist when

This great piece resonates with my own sense of loss and mourning of my social and professional identity as a computer programmer as the community of software engineers embraces LLMs and continues to sink further into surveillance and behavioral manipulation-based technofascistic business models.

I visited Trinidad exactly one time as a child and still get regular cravings for roti decades later.

WHAT ARE THESE PEOPLE DOING

Wine moms are burning prospective concentration camps and flag dads are choking back tears at protests.

Well I've been programming for like 13 years at this point and I've finally gotten around to making phone apps

Long story short. It's a miracle Signal is as smooth, safe, and easy to use as it is.

Download it if you haven't already.

LLMs will never be able to figure this shit out in a way that matters.

Like I wish I could just build on top of signal for my use case but signal doesn't have an API that will allow for what I want to do and you know what, good for them.

There's probably a security reason for it and I respect that they have better things to do with their time.