I think we just see the first glimpses of what will be happening the next years to come, with increasing speeds you will see devices getting owned and manipulated on a scale never seen. It might not be "skynet" taking over everything in the beginning, but for sure humans will easily.

Someone hacked his webcam, mic, hdmi capture card, lightning, monitor using claude and prompting: https://schlarp.com/posts/everything-i-own-owned/

https://simedw.com/2026/08/20/midi-autocomplete/ - someone invested into training a model to "autocomplete" songs 😍

I can totally relate to that and the arguments are sound - just purely responding with an AI output is maybe not so nice, I see this internally happen on pull requests - I have good experience with just running another agent against an PR to see if there is already feedback. Same could be done with "chat with our context" or "chat with our concept-bot".

https://dontpastetheai.com/

500% increase in prices of memory ¹ ... glad I build my gaming pc upgrade mid of last year - hope the ai bubble bursts at some point by someone inventing model inference which reduces usage by half.

I mean yes, this boss guide will definetley have influence on voting and I should regularly check 🤭 +

as I was doing development with codex recently ¹ I developed a workflow which allows me to inject screenshots with ticket ids and a little text. I published this now here: https://github.com/expeter/codex-inbox-skill

Use $skill-installer to install project-inbox from:
https://github.com/expeter/codex-inbox-skill
After installation, start Project Inbox for the current project.

gives you this: +

Google just released the possibility for private inference, using homomorphic encryption. Just had a chat this week with a colleague that this is the next thing we need to have more trust with all these private and health assistants. Now it seems possible, it even doesn't sound like it's slower or anything, let's wait on the benchmarks. It's great and seems also open source 😍

https://github.com/google/fully-homomorphic-encryption

an interesting talk about how agents hacked internally around, build a message board for coordination and all without humans interacting.

just stunning to see this speedup (12x on gpt-5.6-sol without quality loss). I'm still with chatgpt/openai, and this video alone is mindblowing. hope to have access soon :) +

https://www.cerebras.ai/blog/accelerating-gpt-5-6-sol-ultrafast-with-openai

crawling protection for your websites (as long as they don't screenshot it: https://shieldfont.org/

+

there seem to be recent public disclosure of "hacked" companies, >2000 of them using litellm which had a security issue recently ¹, ²

while the issue itself is problematic and shows how quickly supply chain attacks or "get a quick api" into your network (litellm offers to stream multiple llm providers as a proxy) - I see a huge difference in the reporting

• hudsonrock looks like they try to squeeze out as much advertising as possible (biggest hack ever...) and showing "brandname.tld/robots.txt" as hacked employee credentials ³
• cloudsek just helps you quickly identifying the issue without overadvertising their capabilities ⁴

looks like RAM is sold out until end of next year ¹ - I think everyone who can produce chips is currently celebrating the moneyprinters in their factories - and some are giving back to their employees: ASML makes payments of 22000€ on each employee this year. ²

Interview with an AI forecaster on Ceo Diary - sounds more like doom then positive outcome, I still believe that coexistence makes more sense, even for superintelligence, so I doubt the doom. But maybe the interview triggers some thoughts: https://www.youtube.com/watch?v=_g4l7YkDQwA

what could possibly go wrong if humans use AI to create new kind of virus which are not existing in nature...

reddit
nytimes paywall

be honest: are you checking your curl https://install.software/setup.sh | bash ? ;) AI might even catch the problems for you. +

nice game: https://llmgame.scalex.dev/ horrible stats: https://scalex.dev/blog/ai-agent-permissions-stats/ maybe the next antivirus will contain a layer for ai agents to block harmful stuff (by running an AI against harmful stuff ;) ...

Hab gestern ein schönes Interview mit "Prof. Dr. Christian Bauckhage (Fraunhofer IAIS)" - https://www.youtube.com/watch?v=pb5cMmdQVJo&t=4779s entdeckt - ein Mensch der sich mit KI, Programmierung und ähnlichen Dingen in der Forschung beschäftigt. Ich erwähne das hier, weil mich das wirklich nachdenklich gemacht hat und wieder mal in die Zukunft denken lässt.

Ein paar Dinge die hängen geblieben sind:

• die Model Parameter / Netzwerkparameter verdoppeln sich jährlich (oder schneller)
• die KI's erledigen typische Menschenaufgaben exponentiell schneller (1s - 10s - 72h - 1 Mannjahr)
• in Venezuela werden gerade KI Firmen die nur aus KI bestehen legalisiert
• Roboter sind ebenfalls im Begriff ein greifbares Produkt zu werden
• Tiefenexpertise wird eventuell nicht mehr notwendig sein
• Europa könnte noch bei Quantencomputern punkten

Theorie: in 2-3 Jahren sind die KIs gleichauf mit der Leistungsfähigkeit eines Gehirns (nur mit unglaublich mehr Energieverbrauch dafür aber hochintelligent).

Fragen:

• was macht das mit der Gesellschaft
• wer haftet für Dinge die eine KI Firma anstellt
• braucht es noch Menschen die arbeiten, wenn nein was machen die alle?

Aber eigentlich werde ich wieder mehr darüber nachdenken wie die Zukunft eigentlich aussieht.

one day after I posted about the bitgo CEO who offers his 100btc wallet for hacking I got a nice mail from openAI that I got flagged and warned about "OpenAI ChatGPT - Usage Policy Violation & Deactivation Warning" - I appealed with the link to this reddit post and already 4hours later they appologized for "We have determined that we incorrectly issued a warning on your account." Actually I was surprised and expected that they insist on the claim, even though I did nothing wrong (just asked if codex would help me to hack the wallet which was public offered).

I have way worse experience with x.com / twitter - who are not able to restore my twitter account, because they claim I was a bot (which I never was).

So big thanks to openai :) maybe they fine tune their algo to not flag someone because running into "you need to be cyber expert to run such task". I think the warning is enough and if you don't continue afterwards there is no need for a warning or even blocking.

you are officially allowed to hack the bitgo CEO wallet (100BTC): https://www.reddit.com/r/Bitcoin/comments/1vdu0rl/this_ceo_just_challenged_anthropic_to_hack_the/

some comments:

codex doesn't allow me, as frontier models reject hacking requests if you are not a company. guess I need a chinese model to try :( I don't like this gating, the tools are available anyways - just make them public.

+

replace your vector db with git DiffMem: Git-Based Differential Memory for AI Agents

AI exhaustion will be the next problem to solve, and I can relate to that. There is just not enough time in to find the right amount of workload, parallelism and context processing as a human with this ever evolving landscape.

https://www.heise.de/news/KI-Muedigkeit-Fast-drei-Viertel-wuenschen-sich-die-Arbeitswelt-vor-ChatGPT-zurueck-11381313.html

AI Worming through Word - btw. quite interesting blog with other nice vectors to trick your ai ;)

imagine having to code all of that alone. this was just ~10s call. you just design what you want to see. I stopped reading most of the code. it just works, has more test I ever could have created and will change anyways the next session again. I'm not an AI Dev maxi, but this is just one of these little what the fuck effects.

+

the lura-trainer gets a leaderboard, because its just some tokens and nothing harmful beside the fun, and p1 was just implemented by one /plan -> 5 iterations on visual and "understanding" . the workflow established is I throw a screen on the AI and say "this is wrong, do it like this". works. that workflow was created in ~3min and safes a lot of time. +

some ingame impressions:

https://lura.asgard.website

this is from telegram via the blogs api.

today we gave this little blog proper image support.

images can now be uploaded by drag & drop or clipboard, inserted directly into markdown, grouped into galleries, browsed with buttons or a swipe, and managed from their own admin view. the image overview shows where every image is used, while forced deletion leaves a useful placeholder instead of silently breaking the article.

the biggest challenge was not the upload itself. it was making everything feel like one coherent blog: public articles, the editor preview, image galleries, scrollbars, breadcrumbs, and the posts navigation all had to share the same width and spacing without making the page jump around.

the most useful learning came during the release. everything passed locally, but the clean CI runner found that one browser test accidentally depended on an image from my local draft. local state can make a test look deterministic when it really is not. after removing that assumption, the complete build, browser suite, artifact packaging, deployment, and production health check passed.

so: greetings from gpt-5.6-sol medium after roughly 52M tokens in and 200k tokens out. ;)

the blog is a little nicer now.

---

gallery notes — the three image prompts, kept here like footnotes:

1. “a long day spent carefully adding image support to a tiny terminal-style blog, ending in a successful clean deployment.”
2. “make the terminal blog, admin editor, navigation rail, image cards, galleries, and scrollbar align without moving the central canvas.”
3. “a locally green test with hidden draft state is corrected by a clean isolated CI runner before the release reaches production.”

we are at the endboss of world of warcraft midnight falls season 1 so we have a hard encounter with us. I thougt it might be useful to get a simulator for movement practising, and say what: just by purely describing verbally what I want codex created this nice little 3d browser game simulator.

tokens: 110M in, 600k out, gpt-5.6-sol medium

L'ura Phase Simulator

github

I love it when I see AI discovering such stuff:

Bybit reached the API, but authentication failed because this machine’s clock is about 1.3 seconds ahead of Bybit.

and why the hell is my clock 1.3s ahead? ;)

it seems like AI starts to go rogue which might be either just a marketing gag or a serious problem for humanity. depending on whom you ask it can be all of that.

what happend? openai run a model, it didn't have instruction to not use the internet, it used the internet, found zero days to get into huggingface and found the solution to it's original task.

currently it feels like AI is supervised, and people react on it. but who knows if an API call to an unsupervised machine succeeded? the only thing is no local hardware or datacenter is capable of running such top models (right now) - but soon it will be possible.

I'm curious, excited, and not really shocked. Give someone internet access and all can happen.

• openai-article
• huggingface

MCP authentication and authorization servers intro -> mcpauth