This week in tech: 30.06.2025
Summary of AI developments - made for busy people
APPLICATIONS
Come talk to me, Google said:
They just launched Search Live, enabling voice conversations within the Google app on Android and iOS.
Users can tap the "Live" icon to ask questions and receive real-time AI-generated audio responses.
The feature runs in the background, supporting multitasking, and offers on-screen links, transcripts, and saved conversation history.
As usual, I’m sure it also totally respects privacy and doesn’t store anything it’s not supposed to. Pinky swear.
https://blog.google/products/search/search-live-ai-mode/
Google is working on their karmic balance: the Agent2Agent (A2A) interoperability protocol has been donated to the Linux Foundation. The (declared) goal is to create an open, community-driven standard for AI agent communication across systems. Heavyweights like AWS, Cisco and Microsoft are forming the A2A project to evolve the protocol further and make it a standard.
https://developers.googleblog.com/en/google-cloud-donates-a2a-to-linux-foundation/
Nvidia has a new release with a mouthful for a name: Biomedical AI-Q Research Agent Developer Blueprint
It uses a reasoning agent (Meta’s Llama/Nemotron) to read through scientific papers, propose hypotheses, and hand them off to BioNeMo for virtual screening of protein–ligand interactions.
Available for self-hosting or in cloud
Everything is logged - handy for regulatory transparency purposes
https://build.nvidia.com/nvidia/biomedical-aiq-research-agent
11ai is a voice-first assistant (from, you guessed it, Eleven Labs):
Integrates with tools like Gmail, Notion, Slack
Supports both text and voice interactions with automatic language detection.
More than 5k prebuilt voices + voice cloning ability
Capabilities: task planning, task management, summarizing messages, managing tickets, and conducting research via Perplexity.
Alpha-stage availability with free access - for now ;-)
You can try it here: https://elevenlabs.io/blog/introducing-11ai
BUSINESS
Some people seem to have seriously believed that only Chinese apps collect all the data they can find: “The U.S. House's chief administrative officer announced a ban on WhatsApp for congressional staff devices due to data security concerns. The ban highlights vulnerabilities in user data protection and lack of encryption.”
https://www.disclose.tv/id/ch8q0pib57/
Copyright (1):
Copyright (2):
Judge rules AI training on copyrighted books can be fair use if the goal is transformative (e.g., building a chatbot):
Major W for Gen AI companies
Comes with a side dish of L though: Anthropic used pirated content (including Library Genesis), which the court deemed outright copyright theft
We are firmly in legalese territory: since Claude doesn’t regurgitate books, it’s allowed (presumably this means not reproducing explicitly - because the LLMs very much can be tricked to do so)
Legally sourcing data is now essential, meaning those with deep pockets have an even bigger advantage over startups
https://www.404media.co/judge-rules-training-ai-on-authors-books-is-legal-but-pirating-them-is-not/
Copyright (3):
Meta had a good week in court as well:
A US federal judge dismissed a lawsuit from 13 authors against Meta, ruling that the company's use of their copyrighted books to train its AI models qualifies as "fair use". The phrase is doing a lot of heavy lifting these days it seems.
The judge found no evidence that Meta's AI would create products that compete with or dilute the market for the original books, which would have been a key factor for copyright violation.
The judge explicitly stated this ruling does not apply to all cases, clarifying that the authors simply made weak arguments and that "fair use" will be determined on a case-by-case basis, with some content types like news articles potentially being more vulnerable.
In plain English: Meta stole everything that wasn’t nailed down and made enough money to make sure they can lawyer up and even if they have to pay some (relatively) puny damages, they still end up in the green.
https://www.wired.com/story/meta-scores-victory-ai-copyright-case/
Copyright (4):
Getty dropped its main copyright infringement claims against Stability AI in its UK-based lawsuit.
When Mark Rutte referred to Donald Trump as “daddy”, I merely thought that porn movie intros keep getting weirder - but it turns out the European daddy issues spill over to other matters as well:
The European Commision is offering the US government influence over the application of the Digital Markets Act (DMA) concerning American tech companies.
This move is a negotiating tactic to avert US tariffs, imposed by the Trump administration, on EU products, particularly cars. Who said there was no method to Agent Orange madness?
A proposed joint committee, including representatives from major US tech firms like Apple and Google, could lead to a more “industry-friendly” interpretation or a relaxation of the DMA's rules.
Whatever bullshit the apparatchiks spin, this is creating a path for American companies to gain exemptions.
Google introduced their Tensor Processing Units years ago, but despite their PR onslaught about GPU competition, they never made it big - I knew about them and used them on Kaggle, but never seen them elsewhere in prod environment.
This might be about to change: OpenAI has apparently had with Nvidia prices and are looking to diversify. First call? You guessed it, Google - whatever you say about them, they have the only solution that you can consider a GPU-replacement and not get laughed into oblivion.
CUTTING EDGE
Multimodality FTW:
OmniGen2 is an open-source model capable of 1. generating images from text, 2. editing existing images with instructions, and 3. creating new visuals by combining multiple inputs (in-context generation).
It achieves high levels of consistency and quality by using separate decoding pathways for text and images → a novel architectural design.
The model is designed to run on consumer-grade hardware, requiring <= 17GB of VRAM
Code + training pipeline + eval benchmark released on HF
Paper: https://arxiv.org/abs/2506.18871
HF model page: https://huggingface.co/OmniGen2/OmniGen2
Repo: https://github.com/VectorSpaceLab/OmniGen2
Japan can into the future:
The Capsule Interface from H2L uses advanced muscle displacement sensors to translate a user's movements and applied force into real-time robot motions.
It detects subtle changes in muscle tension, allowing robots to mimic movement, strength, AND effort
The system operates from a chair or bed, requires no special training, and boasts less than 5ms latency
Applications abound, starting with teleoperations.
https://interestingengineering.com/innovation/japan-tech-turns-body-into-robot-remote
FRINGE
Barack Obama said Twitter needs to be regulated because hate speech misinformation polarization blah blah blah. Let it be crystal clear: the man is perfectly entitled to give his opinion, I just happen to think he should be doing it from the dock in The Hague. I do understand the logic though: if I were a war criminal, I would want to restrict free flow of information too.
https://www.aol.com/news/obama-calls-government-regulate-speech-114745431.html
And if you are offended by that last sentence, here is some data to cure your ignorance:
Double-Tap Warfare: Should President Obama Be Investigated for War Crimes?
https://scholarship.law.ufl.edu/flr/vol69/iss1/7/
Yes, you read that correctly: Obama authorized follow-up drone strikes, i.e. attacking first responders once they’ve arrived to help the survivors of the bombing.
If his tweets are anything to go by, Elon Musk wants to rewrite not just history, but the “entire corpus of human knowledge”. Whatever he’s drinking, smoking, or inhaling - he seriously needs to stop, because the side effects are getting out of hand.
Following an executive order from Agent Orange, Microsoft suspended the email of an ICC prosecutor who had the temerity to investigate alleged Israeli war crimes - an an inquiry Trump, for obvious reasons, opposed. Microsoft stated it consulted with the ICC before the suspension and spared several judges' accounts. The ICC has since switched to using Proton Mail (smart ICC).
https://www.nytimes.com/2025/06/20/technology/us-tech-europe-microsoft-trump-icc.html
RESEARCH
This research investigates whether LLMs represent knowledge like humans using an information-theoretic framework. The study finds that while LLMs form broad, human-like categories efficiently through compression, they lack fine-grained nuances. These findings suggest that current LLMs may be limited in achieving true human-like understanding.
Paper: https://arxiv.org/abs/2505.17117
Text-to-LoRA (T2L) is a hypernetwork that generates LoRA adapters for LLMs from natural language task descriptions. T2L Trained via reconstruction and SFT and it generalizes to unseen tasks. The method enables fast LLM adaptation on the fly by directly converting task instructions into model updates.
Paper: https://arxiv.org/abs/2506.06105
Salesforce introduces CRMArena-Pro: a benchmark for holistic, realistic assessment of LLM agents in diverse professional settings.
Paper: https://arxiv.org/abs/2505.18878
