iiinova — where your digital self keeps evolving

Jakob S @jakobsachs.blog · 14/10/2026

On Bender herself: Fair enough i haven't seen that but i will take your characterisation. I mostly know her from the famous Stochastic Parrot Paper, and the following discourse. And www.washington.edu/news/people/... does list her as an expert on this, but i would guess then this is dated .

https://www.washington.edu/news/people/emily-bender/

200

Jakob S @jakobsachs.blog · 13/10/2026

Its truly just our blessed Americans researches vs. barbarous Chinese IP thieves

020

Jakob S @jakobsachs.blog · 13/10/2026

MFW when this, but US Department of War is fine and would never act badly :))

130

Jakob S @jakobsachs.blog · 09/10/2026

Huh so thats what happens when i let my Agent run all-week to rewrite itself in Rust

030

Jakob S @jakobsachs.blog · 09/10/2026

HUH so i guess GPT6 DOES love to neuralese itself with its subagents

110

Jakob S @jakobsachs.blog · 09/10/2026

its giving sparkle/hyper-realism AI imagegen slop

070

Jakob S @jakobsachs.blog · 06/10/2026

If i ever win the lottery, i wont tell nobody, but there will be signs ..

050

Jakob S @jakobsachs.blog · 06/10/2026

Huh.. so that's why my agent asked me if he can cleanup his .venv cache (btrfs really saving my butt here)

Terminal screenshot showing the command du -sch --apparent-size dream-rust* | tail -n 1 run in ~/.cache, which took 5 seconds. The output reads "1.3T total".

010

Jakob S @jakobsachs.blog · 06/10/2026

Love the Le Chonk Memes, but its (list most mistral models) still horrible value... The french are getting mogged, both on intelligence and cost, by Deepseek v4.1 Flash

2400

Jakob S @jakobsachs.blog · 06/10/2026

Opus 5.5 not be lazy challenge (impossible)

Agent chat showing Claude saying: "The image is 1.27 GB. I did not compare this with the old image, because I did not pull it."

020

Jakob S @jakobsachs.blog · 03/10/2026

Cat is really quite upset that I would dare move his headrest

020

Jakob S @jakobsachs.blog · 30/09/2026

help peer ? (sorry for the slop, i just found out OpenRouter has videos)

020

Jakob S @jakobsachs.blog · 28/09/2026

Most boring ass middle of the road result imaginable honestly (TW: Sama jumpscare) 12axes.vercel.app/en

110

Jakob S @jakobsachs.blog · 25/09/2026

After testing Opus 5.5 for making these cutsy explainer videos, i thought i would try the same with Astra. Told it to "make a playful explainer video of the Jev Harness we are building to play Pokemon, visually it should be GBA-esk and use approximately the Pokemon Emerald style and music"

000

Jakob S @jakobsachs.blog · 25/09/2026

Fresh book Friday :) (Yes I know its free online but physical >> PDFs for textbooks)

020

Jakob S @jakobsachs.blog · 25/09/2026

not to be that guy, but isnt this considered friendly fire ?

110

Jakob S @jakobsachs.blog · 25/09/2026

Good morning, cat pics on main

010

Jakob S @jakobsachs.blog · 24/09/2026

Absolutely insane that i can genuinely push 40k LoC/month at work. Working solo is such an advantage, and mind you this is just the BE, i think FE is roughly the same amount again. (for only one repo/service, though my main one)

Two side-by-side area charts show code growth from Aug 24 to Sep 24, 2026: src rises to 41,140 LoC and tests to 61,611 LoC, with dashed vertical markers indicating milestones/events.

001

Jakob S @jakobsachs.blog · 22/09/2026

Say about Jev what you want, over-hyped, limited-context, etc. But the fact after a week of testing it with all i could imagine, i only now have hit that 5$ initial credit..

030

Jakob S @jakobsachs.blog · 20/09/2026

after alot of trying and swearing at Sol, heres Jev playing the first minute of Pokemon Emerald. more to come tomorrow hopefullt

011

Jakob S @jakobsachs.blog · 19/09/2026

Goal for today: make Jev play my Pokemon Emerald

A screenshot of a computer desktop showing two overlapping windows: on the left, a Game Boy Advance emulator playing Pokémon Emerald displaying the intro screen for "Littleroot Town"; on the right, a terminal window running an automated code review and software testing script.

120

Jakob S @jakobsachs.blog · 17/09/2026

Ran a benchmark, having Jev play a bench of LLMs. Turns out Jev is alot smarter then all of them** and cheaper on the same input! (** i disabled reasoning for LLMs, to make time/effort more comparable between models)

Square dumbbell chart comparing Jev’s win rate against seven single-shot LLMs at Go on 9×9 and 13×13 boards. Lower means a stronger opponent. Jev’s 9×9 versus 13×13 win rates are: Claude Sonnet 5, 40% versus 33%; Mistral Large, 60% versus 83%; Kimi K3, 70% versus 83%; GPT-5.6 Luna, 80% versus 83%; GLM 5.2, 80% versus 100%; DeepSeek V4 Flash, 81% versus 50%; Gemini 3 Flash Preview, 90% versus 67%. Amber represents 9×9 and blue represents 13×13.

3121

Jakob S @jakobsachs.blog · 17/09/2026

Anecdotally, its not much worse or better against me. Just doesn't know when to quit fighting over a group

110

Jakob S @jakobsachs.blog · 17/09/2026

First iteration, Jev seems.. Not that great at Go. But considering the response time is 50ms.. Ds4.1Flash is worse

1121

Jakob S @jakobsachs.blog · 15/09/2026

Bach BWV772 visualized, the "floor line" is the lower Cantus and the notes above/on the side are the upper voice. The colors range from cool (consonant) to warm (more dissonant). Height on the wall is not absolute pitch but the interval to the base Cantus. OSU! but for barock Counterpoint

1151

Jakob S @jakobsachs.blog · 15/09/2026

Playing around with Astra and visualizing the Upper and Cantus interval. Each "rib" in the pipe is a upper note, the height inside the pipe is the interval it has between, colors are based on interval kind (3rd/6th etc.) If i get the animation smoother i'll post a recording

000

Jakob S @jakobsachs.blog · 15/09/2026

Just curious, which or these are on the level of wrong as "flat earth"? You could debate #1 and #5 for sure, but #2-4 seem fairly easy to measure/demonstrate (like the shape of the earth I suppose).

0220

Jakob S @jakobsachs.blog · 11/09/2026

Uhm im not an expert, i think its just joining primitvis likely ? I lost the original .blend but heres what astra could extract from the compiled game files

110

Jakob S @jakobsachs.blog · 11/09/2026

I wanted to stress-test Astra on its ASCII art. So i had it generate me a Möbius loop, colorized by the normal of the surface, and its also resizable/responsive! Quite hard to do this in 2D and still keep the arrow head correct, but after a few bounces between the model and me...

062

Jakob S @jakobsachs.blog · 10/09/2026

reading the sparks-of-agi paper and god damn, imagine you went back to 2023 and told the Microsoft guys what the curent SOTA is lol...

000

Jakob S @jakobsachs.blog · 05/09/2026

WTF astra is cracked in blender, it just one-shot this ? for like 0.20$ of usage ?

2100

Jakob S @jakobsachs.blog · 04/09/2026

Based Sol ? Also.. yeah what if i want to learn unsafe memory handling ??? what then ??

000

Jakob S @jakobsachs.blog · 03/09/2026

Ahh seems to be a general fuckery Also maybe Geminis status page is slopped a bit..

000

Jakob S @jakobsachs.blog · 03/09/2026

Claude, ChatGPT, Grok all down But Gemini..? Up somehow

000

Jakob S @jakobsachs.blog · 03/09/2026

How the fuck can these three all be down at the exact same time??

100

Jakob S @jakobsachs.blog · 26/08/2026

Where have i gone wrong in my life, that for research at work i literally have to type these words into my private notes...

100

Jakob S @jakobsachs.blog · 25/08/2026

Btw the doc is insanely sturdy. Base is just a CNCd block of Aluminum. I can literally hold this all by the cable or card

010

Jakob S @jakobsachs.blog · 25/08/2026

I just realized I might need a bigger PSU for my MI50, or I'll have to use one of my older Gaming Cards

110

Jakob S @jakobsachs.blog · 25/08/2026

FINALLY got my tiny chestnut, and it looks soooooo cleaaannn

110

Jakob S @jakobsachs.blog · 20/08/2026

damn claude, those are flirting words, we're like that ?

010

Jakob S @jakobsachs.blog · 19/08/2026

Not speed, actualy pico-Joule per Byte read (and maybe written) newest paper i could find arxiv.org/html/2604.16...

100

Jakob S @jakobsachs.blog · 14/08/2026

Okay i had to do some stuff with context, to make sure it finishes. Q3 got something done , but its not great. i'll retry tomorrow

000

Jakob S @jakobsachs.blog · 14/08/2026

okay i have to restart. my default vLLM setup cut of the reply at 8k reasoning (also downgrading to q3 so i can fit it on my tiny gpu)

100