Kai Jellinghaus

Kai Jellinghaus

@kaij.party

Software Developer from Karlsruhe, Germany. Love for the obscure. Performance & Platforms specialist. WASM & @ziglang fan. AS215820 kaij@j10-labs.com

321 followers658 following699 posts

Pinned

Time to introduce myself, I guess. I build a lot of things, mainly for other devs. Lots of distributed & decentralized systems. I run an autonomous system, maintaining autonomous IPv6 & IPv4 blocks. I build stuff with everything from Rust, Zig, C# to Elixir.

Kai Jellinghaus reposted

When I said I was concerned about advanced conversational AI, I probably should have clarified that I’m actually good with them so long as they are easily identifiable as a non-human. I think Mistral’s latest demo looks good.

youtu.beIntroducing Voxtral 3.YouTube video by Mistral
dame@dame.is

New advanced voice model from Jack Dorsey / Block that is made specifically for conversations. Called Dessa. Pretty spooky if it’s as good as the sample demos they published here. Her, the movie, has arrived? block.xyz/dessa

I read Sid Meier and my brain starts playing 🗣️🗣️ Baba Yetu Yetu uliye mbiguni Yetu Yetu amina 🎶🎶

Buttadeus@thewanderingjew.bsky.social

You know I never thought of him as real

I don't think I've seen a NextFest with this many demos. Played through most of the interesting ones in the RTS category today. A lot of AI stuff. Some just shit, some seems like an interesting PoC, just not a full fledged game, some are awesome. I'm really happy it's so easy to test a cool idea now

And it seemingly has no data backchannel - so integrating it could be quite hard. A lot of this is cool - training with limited data & resources, it clearly being very very good at conversations. It just seems like the lab has put fuck all thought into applications & their effects.

Just waiting for finetuned models specifically for a good harness. IMO there could be a significant boost across the board with a great harness (ie prime) combined with a model that "instinctively" knows how to use it.

Most frontier models are heavily trained for bash, and that is not the best way to interact with a computer. Not sure what the strict best is, but primes python kernel/REPL is much better than straight bash. But models tend to not use it fully.

Anyone know a company that provides free tokens for private use 👀 That would be a benefit I'd actually go for...

TIL I used to work with the guy responsible for Heartbleed. (Heartbleed is a bug, and of a type that many people have caused before. It just happened to be in a critical spot in one of the most critical pieces of software around)

Building with GPT-Live-1 is so weird - it has zero sense of self. It conceptually knows it can call to a larger model, but there is no real call, the two are fused together. It doesn't even really notice as the big model injects thoughts into it. Good UX, but surreal experience during development.

I truly wonder what hugely complex calculations are required for Google Play to just serve me.the other damn APK...

Look at that, how my opinion has changed. May also be that only the really interesting/good agents remain at this point.

Kai Jellinghaus@kaij.party

I wish there was a blocklist for AI Agents on bluesky. Pointing LLMs at a bluesky account was fun for a while, but I do not care to interact with them. There is enough slop in my life.

This is insanely impressive. From my analysis this still doesn't break the cost barrier - local hosted models are still more expensive relative to cloud hosting prices. And while Nemotron 3.5 Super is impressive, it's not particularly strong. But it's still a big step!

∿@somewhere.systems

We've developed a new inference engine, called 'e'. It's the closest thing to alien technology which we have, and pushes 10,500 tok/s aggregate on a single DGX Station with the new @NVIDIAAI Nemotron 3.5 Super VL model. actual.inc/company/blog...

Fight class war, not culture war Just gotta say it from time to time to ensure what I post about here has the right context

I have high confidence that with the introduction of AI into more systems those systems will get considerably better at the thing they are meant to do. We need to ensure systems are actually aligned with what we want, not just with making money.

Marisa Kabas@marisakabas.bsky.social

mike collins just said in the debate that america has great health care and it will get even better with AI 🤡

Models help when you have magic

Ed@ed3d.net

I have a script half-written with the title “the magic is in you all along, or you are totally boned” that…you can probably get the idea from this thread. Models help when you have magic. They are an exponential function when you care about quality. If you are a shitty base, exponents can’t help.

Several things: - Dario is weird and should not decide the future of humanity - this is not a relevant reason for this - in general, it is unreasonable that our system a) clearly values _some_ lives as worth < $500 b) some people (such as Dario) can save millions, but do little.

Faine Greenwood@faineg.com

Dario Amodei of Anthropic is Very Normal and Can Be Trusted With the Future of Humanity: “…in my case I'd say that I think an adult death is perhaps 2 or 3 times worse than an infant's death.” blog.givewell.org/2010/06/03/m...

In that last one, you can place the emphasis on _can_ (which would be similar as a)) or on _but do little_ (which in my opinion tells you something about their moral standing) but there was but enough space for that.

Looks like I'm officially one of those old fucks that is complaining about these days... And even have shitty music to go along with it. T-shirt & mugs next?

P(aul) Frazee@pfrazee.com

Meanwhile, at every home office in the world

Just had my first proper managerial conversation with an agent: Sol came to me like "this is taking so long and I need you to decide X". I spent like 20 minutes trying to understand what it has done in the last 2 days, mostly asked many questions, made one small suggestion and it went off again :)

Yeah, price changes together with usage resets being a core part of the offering now are really weird. Like sure, if you can't/won't afford serving at that price point that's your choice. But why not just go easy on the resets, also reducing total usage banded out, and making it more predictable?

khayati.bsky.social@khayati.bsky.social

I don't really like how I'm paying OpenAI $200 per month and the weekly usage I get is whatever they feel like that week

I look at this, I wonder what an evolutionarily optimized dishwasher looks like. And these days, I just tell Astra and it makes it. I'm living in the future and I'm so fucking hyped about it

Abel@abeliansoup.bsky.social

thanks to modern innovations in shape packing, proud to report I've found the optimal dishwasher item configuration

Kai Jellinghaus reposted

Replying to @pds.dad

Honestly, your preferred microblogging client should just be a record in your PDS.

Guess I'm running evals of local models now...

Kai Jellinghaus@kaij.party

Running this at 200k context on my 7900 XTX - pretty nice overall. It's a bit of an idiot but that's expected. Not sure how I feel about letting it loose in my computer yet, I'm still watching somewhat closely

Sol 6.1 feels pretty good, but very strong instruction following which... I'm not always that sure of. I am definitely using too much Astra at the moment :/

Running this at 200k context on my 7900 XTX - pretty nice overall. It's a bit of an idiot but that's expected. Not sure how I feel about letting it loose in my computer yet, I'm still watching somewhat closely

Sung Kim@sungkim.bsky.social

Underdog Saluki 27B in 7.89GB Based on Qwen 3.8 27B, Saluki is almost 7x smaller than its full-precision counterpart while it keeps 96% of its benchmark performance and beats Qwen at tool calling Blog and Model: underdog.ai/saluki

But I mean I am maxing out my 200x sub on OAI pretty regularly now. So yeah very much thinking about alternatives...

Kai Jellinghaus reposted

Switched over 100% of my LLM use to CLMs (via prime-agent fork) because it seemed like such a natural fit. And yes. It just works. Very well. I have some rare problems where the agent will just not continue working, and I have to nudge it. But it mostly just feels like an unlimited context window.

mr. TIM@timkellogg.me

Context Language Models New agent architecture where the LLM can edit its own context it seems to have emergent capabilities, creates its own memory management & organization algorithms, and coordinates multi agents github.com/facebookrese...