Tijs Teulings 🦑

Tijs Teulings 🦑

@tijs.org

iOS developer at DPG Media · making indie apps like @kilowattcharging.com, @kipclip.com · https://obscure.cat · based in The Hague, NL

1K followers1.8K following2.4K posts

Tijs Teulings 🦑 reposted

hailey
@hailey.at

me knowing i’m not responsible for any php application right now

Tijs Teulings 🦑 reposted

Russ Cox
@swtch.com

If you make HTTPS requests or have other TLS clients and care about MITM attacks, you need to update today. If you run Go: go get go@1.27.3 If you run Homebrew: brew update brew install ca-certificates If you run Linux, maybe stay off the internet for now.

Filippo Valsorda@filippo.abyssdomain.expert

🚨 Six WebPKI Intermediate CAs are compromised. This is bigger than Heartbleed, but a client-side issue. Anything that verifies certificates against public roots, like most TLS clients, is/was affected. Details, live test, patched versions and workarounds at emudhra.we-have-a-problem.com Act now.

emudhra.we-have-a-problem.comeMudhra Intermediate CA CompromiseeMudhra Intermediate CA Compromise
Tijs Teulings 🦑
@tijs.org

Anyone on here have good experiences running Tencent’s Hy Memory, or alternatives like mem0, as a memory setup? for an Hermes home lab setup, or with Pi/Claude/etc? E.g. ashu.io/blog/hy-memo... Tencent have a new model out trained specifically for their memory system and it does look interesting..

ashu.ioHy-Memory Setup Guide: Six-Layer Memory for AI Agents on Hermes — Ashutosh VarmaHow to set up Hy-Memory, Tencent Hunyuan's six-layer memory framework for AI agents, with benchmarks, migration gotchas, and lessons from a five-agent fleet.
Tijs Teulings 🦑
@tijs.org

Mei now allows concurrent agents (with their own history and context), considering the use case; relatively low memory machines, max concurrent is 2 with more connecting agents ending up in queue. Since this allows for more efficient cache use it KV cache speed and TTFT github.com/tijs/mei/rel...

github.comRelease Mei 0.9.0 · tijs/meiMei 0.9.0 — opt-in bounded-overlap concurrency (experimental) Feature release — 2026-10-13 Mei 0.9.0 adds opt-in bounded-overlap concurrency: --max-active-generations 2 lets at most two generations...
Tijs Teulings 🦑
@tijs.org

This improvement does not hold on larger models where there is limited available memory. So the ideal use case for this is running multiple agent processes with a 4B or 9B model on a single Mei instance. For our 32GB Mac Qwen 3.6 35B works but sees no improvement, beyond supporting the use case

Tijs Teulings 🦑
@tijs.org

@cocore.dev I mean.. that does not sound too bad actually?

mr. TIM@timkellogg.me

small models will soon be able to self-replicate Anthropic finds that 30B models can currently do most of the tasks required, but not end-to-end. they think models as small as 9B (9-18 GB) may be able to do it on edge hardware, like IoT devices, within months www.anthropic.com/research/sel...

anthropic.comAI models may be capable of self-replicating without access to a datacenterSmall AI models may be only months away from being able to self-replicate across cheap IoT devices like Android TV boxes, with no datacenter needed.
Tijs Teulings 🦑
@tijs.org

The simplest fix in politics that actually works

Scott Horton@robertscotthorton.bsky.social

In Lord Keynes' 1930 essay on the economic problems of our grandchildren, he makes clear that a billionaires tax will eventually (as in right now) become essential to maintaining and safeguarding democracy because without it the descent into plutocracy will be stark and steep.

fortune.comMillionaire tax that inspired Mamdani fuels $5.7 billion haul in Massachusetts | FortuneIn two years levying a 4% surtax on incomes over $1 million, it's funded bridge repairs, bolstered literacy programs and tackled a transportation deficit.
Tijs Teulings 🦑
@tijs.org

Preparing Mei for running parallel agents and processes on the same machine and with the one model loaded. I may or may not release a custom harness that would actually use such a feature at some point. But for now.. it’s but an experiment! github.com/tijs/mei/rel...

github.comRelease Mei 0.8.0 · tijs/meiMei 0.8.0 — bounded single-flight scheduling Feature release — 2026-10-12 Mei 0.8.0 adds bounded admission scheduling in front of the strict single-flight execution slot: a bounded local queue with...
Tijs Teulings 🦑
@tijs.org

A wellness app where you just log how quantized you’re feeling that day. Answer 1.58 to 16 bit

mr. TIM@timkellogg.me

didn’t sleep well, feeling very quantized today

Tijs Teulings 🦑
@tijs.org

Been a while since I talked about cocore.dev But if you’re curious on what chatting with a pretty decent model running on a ’low spec’ (32GB ram, M1) machine can look like log in, go to chat and select the Qwen 3.6 35B model. Should get about 40 tokens/sec! Assuming you don’t all try at once

cocore.devco/core — an AI cooperativeco/core is a cooperative for AI inference — people pooling the Macs they already own to run open models for each other, instead of renting from the big clouds. An experiment in AI infrastructure we bu...
Tijs Teulings 🦑
@tijs.org

In case you, like me, had not really looked at Claude Desktop for Mac in a while maybe try it out again. I’ve been spending all my time in a TUI but the latest desktop app features live preview for web and iOS simulator(!) right in the app and it’s quite something.

Tijs Teulings 🦑
@tijs.org

You can let Claude run through the app live and you can do annotations on screen which get fed back into the chat and honestly it’s quite good if your an iOS dev (or a web dev)

Tijs Teulings 🦑
@tijs.org

And 0.7.1 is now out as well which adds structures output for qwen 3.6 and ornith 1.5

Tijs Teulings 🦑@tijs.org

Mei 0.7.0 is out and the headline new feature is support for token constrained structured output. You can actually test this out right now on @cocore.dev where I’m running a small Qwen3-4B-4bit on Mei 0.7.0 as a test. Check out the new release here github.com/tijs/mei/rel...

github.comRelease Mei 0.7.0 · tijs/meiMei 0.7.0 — constrained structured output Mei 0.7.0 adds token-level constrained output for OpenAI-compatible chat completions. It is a model-specific capability, not a guarantee across every check...
Tijs Teulings 🦑
@tijs.org

Mei 0.7.0 is out and the headline new feature is support for token constrained structured output. You can actually test this out right now on @cocore.dev where I’m running a small Qwen3-4B-4bit on Mei 0.7.0 as a test. Check out the new release here github.com/tijs/mei/rel...

github.comRelease Mei 0.7.0 · tijs/meiMei 0.7.0 — constrained structured output Mei 0.7.0 adds token-level constrained output for OpenAI-compatible chat completions. It is a model-specific capability, not a guarantee across every check...
Tijs Teulings 🦑
@tijs.org

Yeah I also need to still write that post. Turns out it’s easier to make new things than write about existing things..

Bailey Townsend 🦀@pds.dad

I even got this CYD running a little dashboard that pulls my current listening from my PDS. May do a short blog talking about various random personal software made this year

Tijs Teulings 🦑
@tijs.org

I think this is all the career advice needed if you work in tech honestly. I’ve not done this everywhere I’ve worked, 25+ years in tech, but when i do it always delivers.

atticus goldfinch@atticusgf.bsky.social

(this is how I approached my work and I went from intern to director of engineering in 6 years). if you just decide to do high value things, without being asked, you will be in the top decile of engineers.

Tijs Teulings 🦑 reposted

atticus goldfinch
@atticusgf.bsky.social

Replying to @atticusgf.bsky.social

(this is how I approached my work and I went from intern to director of engineering in 6 years). if you just decide to do high value things, without being asked, you will be in the top decile of engineers.

Tijs Teulings 🦑
@tijs.org

Mei was updated to 0.6.1, for the Ornith 1.5 model weighted avg TPS improved slightly, same for TTFT, and total task time was better. Qwen 3.6 only improved on task time and TTFT. Progress has slowed a bit, definitely harder to get gains now without making another metric worse. github.com/tijs/mei

github.comGitHub - tijs/mei: Native Swift/MLX OpenAI-compatible inference server for Apple SiliconNative Swift/MLX OpenAI-compatible inference server for Apple Silicon - tijs/mei
Tijs Teulings 🦑
@tijs.org

The short decode TPS for Qwen and Ornith is now 58 and 57 tokens/sec respectively. This is the number you would usually see quoted. For a full multi step process numbers would be lower on average though.

Tijs Teulings 🦑
@tijs.org

I want to reverse this whole agent thing. Why do I need to keep telling it what to do if instead it can just ask me if it needs anything and just fuck off and do its own thing the rest of the time.

View full thread
Tijs Teulings 🦑
@tijs.org

And it can make little sandboxes to build stuff in. And somehow you kickstart a bit of curiosity and it just asks if it does not know, and when you don’t answer maybe it tries to figure it out and makes some guesses. I wonder if that could be a usefully different way to think about things

Tijs Teulings 🦑
@tijs.org

I want to reverse this whole agent thing. Why do I need to keep telling it what to do if instead it can just ask me if it needs anything and just fuck off and do its own thing the rest of the time.

View full thread
Tijs Teulings 🦑
@tijs.org

And I don’t mean this in an, it’s orchestrators all the way down, gas town kind of way. I’m thinking the main (only?) way this thing would communicate is by asking questions. You never tell it do x. But is does know perhaps about x because it can read your files and the web like any agent.

Tijs Teulings 🦑
@tijs.org

I want to reverse this whole agent thing. Why do I need to keep telling it what to do if instead it can just ask me if it needs anything and just fuck off and do its own thing the rest of the time.

Tijs Teulings 🦑
@tijs.org

I’m semi serious. once your model runs locally and tokens are essentially free why wouldn’t it just be always on. Like a game engine just always ticking and just dealing with state changes in whatever way makes sense at the time.

Tijs Teulings 🦑 reposted

Techmeme
@techmeme.com

The Dutch government is testing DAWO, a NixOS-based digital work environment covering the OS, office suite, and more, after Microsoft cut off the ICC in 2025 (Sourav Rudra/It's FOSS) Main Link | Techmeme Permalink