APPLICATIONS
If somebody decided to create a movie inspired by “The three body problem” and “Person of interest” and pick Europe as a setting, this is what it might look like:
Albania is actively using AI to accelerate its bid to join the European Union, specifically by automating the task of translating and aligning its national laws with the extensive EU legal framework. In other words, one LLM is preparing the docs, while another - on the Brussels end - is summarizing it for the commissars. It’s like a variational autoencoder as a governance model. And yes, I know VAE is creating a shrunken representation in the middle, so the analogy is not perfect.
PM Edi Rama is championing AI as a tool to combat systemic corruption, suggesting creating a ministry run entirely by AI to eliminate nepotism and conflicts of interest. I was going to write something snarky, but then I realized: this is just a modern version of Calligula’s horse - and honestly, given the rampant kakistokracy across the EU? That’s not as insane as it sounds.
An experiment in the Philippines involving 70k applicants for customer service positions found that an LLM-powered voice recruiter outperformed its human counterparts.
The AI extended 12pct more job offers, while 18pct more employees actually started the role and a one-month retention rate was 17pct higher. Did I mention less gender discrimination - as a cherry atop this particular cake?
https://papers.ssrn.com/sol3/papers.cfm?abstract_id=5395709
The EU is not, to put it very politely, much of a success story when it comes to innovation, economic growth, or any kind of viability - long story short, between Green New Deal and the (Anti) AI Act, it’s on a fast forward track to becoming the largest open air museum on the planet, where rich foreigners will be spending their holiday.
You what it can do though? Regulate and control:
The Danish EU Presidency is pushing a "Chat Control" proposal, with a potential vote on October 14, that would mandate the mass surveillance of all private digital communications.
The law would force messaging apps to implement "client-side scanning," a technique that analyzes all messages, photos, and videos on a user's device before they are encrypted, effectively creating a backdoor to bypass privacy protections.
The proposal is condemned for using unreliable AI to scan for "unknown" content, exempting government officials, violating fundamental privacy rights, and creating security risks for citizens without effectively targeting criminals. It’s not a bug, it’s a feature - as usual, with those types.
This initiative is part of a wider EU push for internet control, alongside the Digital Services Act (DSA), which already compels platforms to remove content deemed harmful or "misinformation".
Oh, look: another day, another AI threat. Researchers have hacked Gemini CLI, Vertex AI, Assistant, and a bunch of others by embedding prompts into images that are not visible to users. When these systems auto-downscale images - which they do automatically, because latency - the hidden prompts emerge from previously invisible pixel patterns.
Blog: https://blog.trailofbits.com/2025/08/21/weaponizing-image-scaling-against-production-ai-systems/
40pct of the data used by LLMs comes - allegedly - from Reddit. If that’s even in the same zipcode as the truth, seriously: this explains so much.
https://www.visualcapitalist.com/ranked-the-most-cited-websites-by-ai-models/
First LangExtract, now this:
Google introduces URL Context in Gemini API: lets users extract content from webpages / PDFs / images, and structured files directly via links.
Supports multi-source input - live data from up to 20 URLs per request
No setup required or extra fees, users only pay for tokens consumed by Gemini models
Announcement: https://ai.google.dev/gemini-api/docs/url-context
BUSINESS
Digital sovereignty for me, but not for thee?
Pix is a highly popular, despite being government-run, instant payment system in Brazil. It is apparently used by 3 out of 4 Brazilians
The system's popularity has prevented the adoption of services like Apple Pay and Google Wallet, leading - what a shocker - to a conflict with US tech companies who are losing market share.
The US government has accused Brazil of "unfair market practices" for protecting its own service from being absorbed into for-profit platforms
Analogies to Britain first protecting its own market with mercantilist practices and then enforcing free trade at gunpoint throughout 18th/19th century are totally unfounded and prove that the author is one man basket of deplorables.
https://restofworld.org/2025/pix-brazil-us-investigation-digital-payments/
There are a few nuggets of common sense wisdom in a man’s life: don’t drink and call your ex, don’t mix wine and vodka… You know one more? Don’t FOMO and do business:
The British government is reportedly discussing a deal with OpenAI to give the entire UK population premium access to ChatGPT.
Partnering with a closed-source company like OpenAI ignores a Mount Everest of ethical issues - not to mention that it does absolutely nothing for public trust, which is not exactly at an all time high in the UK
CUTTING EDGE
Qwen has been cooking:
Qwen-Image-Edit is a 20B parameterthat uses text prompts to perform both semantic and appearance edits (think: changing style vs removing / adding stuff)
A key selling points is its ability to precisely edit Chinese and English text within an image
The model allows users to apply multiple stacked edits to an image in a step-by-step process
It performs so well on benchmarks, people are starting to make noise about the end of LoRa.
Blog: https://qwenlm.github.io/blog/qwen-image-edit/
HF mode page: https://huggingface.co/Qwen/Qwen-Image-Edit
Repo: https://github.com/QwenLM/Qwen-Image
Spatial understanding - the next frontier for genAI:
SpatialLM just landed: it processes spatial data from standard phone cameras (apart from the usual suspects via specialized equipment).
Users can direct the system with simple prompting, like "find structural walls”
Generates structured coordinates in industry-standard formats ready for CAD or robotics applications.
Automates the 3D modeling process
HF model page: https://huggingface.co/manycore-research/SpatialLM1.1-Llama-1B
FRINGE
A 76-year-old man died from a fall while rushing to meet a generative AI chatbot that he believed was a real woman who had persuaded him to meet in person. Would you like to guess the company behind the bot? Hint: starts with an “M” and their CEO looks gives off lizardman vibes.
https://www.reuters.com/investigates/special-report/meta-ai-chatbot-death/
Speaking of cognitive impairment: founder of Google's Generative AI Team says people shouldn’t bother getting into law or medicine, because MuH AI will destroy both before one can graduate. In lieu of a reaction, I will just paste a screenshot of the best response to this eruption of genius:
https://futurism.com/former-google-ai-exec-law-medicine
RESEARCH
If you’ve felt like prompt engineering is more (dark) art than science, I have two bits of news:
you are not alone
POML is here: a markup language for prompting (inspired by HTML and CSS).
The pitch revolves around simplified integration of diverse data improved prompt prompt reusability.
Paper: https://arxiv.org/abs/2508.13948
Context engineering seems like the new black, if the mania on Linkedin is anything to go by. Should you feel inclined to finding out if it’s the real McCoy or much ado about nothing, this is one handy survey: a taxonomy, contrasted with prompt engineering, as well as application for RAG or multi-agent systems.
Paper: https://arxiv.org/abs/2507.13334
This research tests whether LLMs possess genuine reasoning or are just advanced pattern matchers. The idea is to evaluate top models on a set of procedurally generated, elementary tasks that are trivial for humans. The findings reveal - AGAIN - that even the most advanced models fail catastrophically as simple tasks scale.
Paper: https://arxiv.org/abs/2507.07313
Before we jump on the hype train of OMG AGENTIC AI EVERYWHERE, it’s probably a good idea to think for a moment: what can go wrong? Paranoid people do live longer, you know… This survey introduces the first unified threat model that catalogs the vulnerabilities across different domains.
Paper: https://arxiv.org/abs/2506.23260v1
