Tero Tasanen

AI | Local LLMs | Open Source

[Tero Tasanen @ttasanen]

Quick test of Ornith-1.0-35B: it crashed and got stuck in a loop multiple times, even on simple tool calls.

There’s clearly potential here, but it’s not quite ready yet.

[Tero Tasanen @ttasanen]

Spot-on: more AI agents don’t parallelize your brain - you’re the serial bottleneck. Brilliant framing and advice. Every dev should read it.

[Addy Osmani @addyosmani]

https://t.co/0DyFIXJueI

[Mario Zechner @badlogicgames]

i use AI all the time as well and I support this message.

don't leave your brain at the door. use agents as a force multiplier, not a replacement of your wet ware.

[antirez @antirez]

Anthropic did a big strategic error. Normally they compare their models with their old models. Instead today, now that everybody knows how strong GPT 5.5 is at coding, they put it in the mix, basically showing all their customers that the benchmarks can't be trusted.

[Tero Tasanen @ttasanen]

@badlogicgames Yes!

[Tero Tasanen @ttasanen]

GPT5.5 excels in understanding and actually implementing working features in large and complex code bases.

Can wait to try Opus4.7

[DHH @dhh]

For complicated agent work, it's amazing how much GPT5.5 has improved. I found 5.2 to be very far behind Opus. Now using Opus 4.7 after 5.5 feels like a big step backwards. Gotta love this level of competion! Strong comeback for OpenAI.

[Tero Tasanen @ttasanen]

@0xSero How? 😯

[Tero Tasanen @ttasanen]

@puresight I think just by starting to use it. Pi is very slim at the core so most of the learning, for me at least, was just realizing how little I actually need from the harness. Just let the LLMs do the magic.

[Tero Tasanen @ttasanen]

Pi is the best harness IMO.

Tiny core, no bloat, system prompt <1k tokens.

By default Pi gives the model just four tools: Read, Write, Edit and Bash.

[Tero Tasanen @ttasanen]

This can be really useful. Great job @antirez

[antirez @antirez]

We have the new Array type merged into Redis! https://t.co/fmw4WIH4tI
If you have a use case or want to implement agents shared skills knowledge via ARGREP, this is the right moment to start :)

[Tero Tasanen @ttasanen]

@badlogicgames @matteocollina Thank you!

[Tero Tasanen @ttasanen]

@antirez That is crazy 🤯

[Tero Tasanen @ttasanen]

@0xSero Great work 🫡

[Tero Tasanen @ttasanen]

@spendergrsec Touches user-page pinning, refcounting, completion notification and failure mode includes "leaking pinned pages". 😰

I hope it's just a bug fix pushed to net before net-next.

[Tero Tasanen @ttasanen]

@ranilmukesh @vllm_project This just got merged

[Tero Tasanen @ttasanen]

The MTP branch just got merged in llama.cpp 💪

Time to start testing this out. Should give a nice performance boost especially with Unsloth's gguf models.

https://t.co/OCnYawxUKf

[Tero Tasanen @ttasanen]

Pi is also by far the most efficient harness token wise. This makes a big difference especially with local LLMs where performance is limited or with pay-per-token API providers.

[Tero Tasanen @ttasanen]

Mistral Medium 3.5, Dense 128B model with 256k context by @MistralAI . It's open weights and released under a Modified MIT License.

A solid western alternative for Chinese models like Qwen and DeepSeek. Holds up quite well in coding but still trails on efficiency and reasoning.

https://t.co/4iyK3cIE6w

[Tero Tasanen @ttasanen]

@0xSero Congrats 🫡

[Tero Tasanen @ttasanen]

@0xSero Yes