Jasper

Out with a UV light when exploring the Outback near Lake Mungo.

90 years is awesome, but still sad. RIP Mr Koenig. May you surf on an ion storm to wherever you want to go.

Looking forward to Stevie Wonder in about 20 minutes. (ZAG Arena in Hannover)

Ok. Ich bin ehrlich. War nicht so der Bringer. Ich zeige Euch gleich ein paar Modelle. 3 Händler. Teilweise Mondpreise.

There are apparently around 6k active podcasts about AI (based on sampling podcastindex.org).

startup as hypercert collector

I’m exactly like Sisyphus except instead of pushing a boulder up a mountain I have to repack the tupperware drawer in my house.

“Are you lazy?” used to be an interview question.

I’m worse than the humans in wall-e

Do not engage. Do not engage. Do not engage… .

First (Claude based) testing confirms Embedding Gemma 2 is not better or faster if one only needs text embeddings for retrieval or clustering (Google also did not claim that). For music audio, MuQ-MuLan seems to remain the way to go?! (Multimodal is ofc very useful for many other purposes.)

An era reaching its end.

A good fresh Star Trek take on those fake isekai title jokes

Just randomly remembered how amazed I was 25 years ago, visiting the United States, that my hosts never turned off their Mac. In Germany you absolutely had to shut down your PC back then.

“Slop” still getting red underline. Come on, some automated pipeline!

divepool.com went out of sync with the new mission: help individuals to discover others and themselves. I have one night to bring it up to speed. Planning to introduce audio tracks as a whole new content type, and using plays, (dis-)likes, dates and other signals globally to enhance recommendations.

Tried out the hand prototype to product line today, and it went pretty well on the first pass with Opus 5.5. Data wise everything is right, but needs some more polish to _feel_ right.

Could a PDS be text only, and not support blobs? /cc @pds.dad

„What is love of one's country; is it hate of one's uncountry? Then it's not a good thing. Is it simply self-love? That's a good thing, but one mustn't make a virtue of it, or a profession“ (Le Guin, The Left hand of darkness)

The voting window is now closed! We’ll be announcing the results soon.

I feel seen by the restroom marketing at onethousand.ai workshop. Identifying slightly more with the wrong side here😅. (Definitely tending towards isolating myself in projects.)

For a friends trip I bought and burned my ibla.sh 7+ ratings to five CDs. And people even liked some of the songs😎. (You can use my (teal) plays and (divepool) ratings, too, if you can find them in my repo. Seems at least one other person started rating, too, which is a great success in my book.)

throwing slop grenades

Primary motivator should not become maxing out your token budget. It should remain building what you want to build.

Yes.

🎶 One More Time.... ....we’re getting the room ready for tonight’s Elixir Berlin Meetup and we’ve all got Daft Punk in our heads. Because tonight, we’re hosting the #Elixir Berlin community at bitcrowd one more time, the very last time 💜 Looking forward to seeing you there!

Bro, Weyland-Yutani needs to stop naming all their ships after ominous sounding historical stuff. You're not going to have problems with xenomorphs on something called the Star Goose, I'll tell you that much.

has anyone had success with giving their LLM harness access to something else to interface with the OS instead of 'bash'? i'd prefer some stricter, typed language that signals mistakes early. but it still has to work in a REPL context and not be too verbose

No login needed. Just rate a couple of tracks and see what happens. ibla.sh

Idea of ibla.sh is to rate tracks to get better recommendations, publishing both ratings and plays back to atproto. (Ultimately, I want this data to be used by @divepool.com to make the account discovery feed even better. So it’s not only about the tracks but the people you discover through them.)

Does anybody want to join me for lunch at Maharani (Grindelhäuser, Hamburg) today?

Moved to Codium. VS Code has many improvements over Sublime, and I missed them. Basically I only wanted to get rid of the agent integration, which Codium provides. vscodium.com

Love and hate everything about this at the same time😅.

Solid first draft. Always considered “home schooling” more libertarian than conservative, though, and it needs having kids🤔.

Working on claim/“requirement without intent” extraction from legacy code, and whether the intent should be supplemented somehow or not. Boring yet interesting stuff. Embedding/Graph/Keyword layers playing supporting roles, and those in turn have always interested me for B2C recommendation purposes.

Didn’t know solo John Lennon had other great songs. “One thing you can’t hide, is when you’re crippled inside.” youtu.be/EX7roXRmkOU

to achieve parallelism, all the layer activation data goes over the interconnect for every token (fan-out). for 93 layers BF16 that's about 93*7168*2*2 ~2.5MB + some overhead per token RDMA over TB5 is 6-8GB/s which is magnitudes slower than memory (~800GB/s), but latency is negligible

Is my Bluesky follow feed dying? Yes. isblueskydying.com/skorfmann.com So whom should I follow? Mostly agentic / ai topics or somehow local (Hamburg metropolitan area)

Since my tool supports session to session messaging now, I tried some coordinator/delegation workflow, but couldn't make it work. So far, me working multiple topic based sessions in parallel (multitasking) seems to still work better. Ofc might be my setup, but switching back to multitasking for now.

Watched a short play about sentient garbage that was not garbage at all. I was constantly afraid of being eaten though: “Lačna čna (Hungry gry)”, “gry” apparently being a typo in the booklet. hinundweg.jetzt/event/lacna-...

@smartgit.bsky.social, please implement ASAP, thanks: smartgit.userecho.com/communities/... (or tell me where the setting is hiding :D).