This week in tech: 1.06.2026
Summary of AI developments - made for busy people
APPLICATIONS
Good news: pope Leo XIV released an encyclical about AI - it is important for the spiritual leader of 1.4 billion people to address the most important technological and societal change of our time.
Okay news: the Vatican invited the tech industry to talk.
Not so good news: of all the people to pick, they went for Anthropic - the most amazing collection of self-righteous nutjobs out there (they still don’t have Sutskever on the payroll, but I think it’s just a matter of time).
Proper long take - later this week.
https://www.nytimes.com/2026/05/25/world/europe/pope-leo-encyclical.html
Can LLM into math? OpenAI claims one of its internal reasoning models may have solved the 80-year-old math problem: disproving a conjecture by Erdős on unit distance. The alleged proof proceeded by discovering infinitely many point arrangements with more unit-distance pairs than previously believed possible - yes, I know, a lot of hedging language.
According to OpenAI and outside mathematicians, the model independently connected discrete geometry with algebraic number theory and produced a proof that experts later reviewed and strengthened. It’s kind of a big deal: mathematical proofs can be verified step by step, making them a stronger test of AI reasoning and originality than standard benchmarks (that may reward memorization or guessing).
Disclaimer: the paper is still awaiting formal peer review, but the episode suggests AI could actually contribute novelty in verifiable fields like mathematics.
https://openai.com/index/model-disproves-discrete-geometry-conjecture/
Nvidia is went into world models a while back, and they are showing no signs of stopping:
- SANA-WM can generate 60 seconds long controllable videos
- 2.6B parameters, takes variable input - image / prompt / camera path
- Uses <8GB of VRAM and integrates directly with ComfyUI and Diffusers
- Technical goodies: 6-axis camera control, 32x compression to stay tiny, upscale to 2K using the LTX2 refiner
- Apache 2.0 license :-)
Repo:https://github.com/NVlabs/Sana
Stability AI has launched Stability Audio 3.0:
new lineup of AI audio models capable of generating longer, higher-quality music and sound effects - including compositions exceeding six minutes
Support inpainting and continuation => you can edit specific regions or extend a recording without regenerating everything.
The release includes four models: small SFX (459M), small (459M), medium (1.4B), and large (2.7B)
The release includes the weights, training and inference code, and LoRA fine-tuning support
The large model is available only through the API /self-hosting paid services - companies with more than USD 1M in revenue need to get an enterprise license.
HF model page: https://huggingface.co/collections/stabilityai/stable-audio-3
Stability AI API access: https://stableaudio.com/
Paper: https://arxiv.org/abs/2605.17991
Liquid has launched Co-Invest: a tool that lets users analyze markets and trade directly inside ChatGPT and Claude - crypto, stocks, forex, pre-IPO shares, you name it. There seeems to be a growing trend of embedding financial services into AI assistants: trading, research, and payments converge into a single conversational interface. All your stuff is belong to us. The everythign app. What could possibly go wrong.
Google has introduced Gemini Embedding 2 and it looks genuinely useful:
universal multimodal embedding model - handles text, images, audio, and video in one unified system => cross-media search, retrieval, RAG.
Tops benchmarks in image retrieval, video search, multilingual text understanding, and code
The model generalizes to niche fields without specialized training, making it highly flexible for real-world applications.
Available through the Google Gemini API and Google Cloud Vertex AI
Docs: https://docs.cloud.google.com/gemini-enterprise-agent-platform/models/gemini/embedding-2
Paper: https://arxiv.org/abs/2605.27295
BUSINESS
The Ketamine Prince goes into robotics: Tesla has begun construction of a dedicated Optimus factory at Gigafactory Texas, where they want to build 10 million humanoid robots per year.
Paranoid people live longer: China is reportedly tightening controls on AI talent - researchers and execs need government approval for overseas travel. This shows Beijing treats advanced AI development as a strategically sensitive industry and rightfully so - the restrictions are now extended, but hte policy has been in effect for govt officials for a while now.
Bowing to the inevitable? Barnes & Noble CEO says he wants all AI written books to be identified as AI. He will sell them, although he does not expect them to be terribly popular - among all the emotional outbursts (both for and against), there is something to be said for the transparency.
https://futurism.com/artificial-intelligence/barnes-and-noble-ceo-ai-book
Apparently the Ginger Caligula doesn’t understand the modern world - in which the EU is a regulatory superpower - and he is working to diminish America’s standing (even more). To wit: Agent Orange delayed a proposed executive order that would have required government security reviews of advanced AI models before public release. Critics warned it could slow US AI development and hurt competitiveness against China - given the price / quality ratio of models produced on two sides of the Pacific, said critics do have a leg to stand on.
Here’s prediction: if this trend continues, within 6-12 months the US government will resort to lawfare and try restricting Chinese models because national security or sth.
The cheap AI era is coming to an end, and so is tokenmaxxing: every major provider has bumped their prices over the last few months (usage based pricing, nerfing lower subscription plans, you name it). Turns out it’s a challenge not just for ordinary mortals, but even with companies sitting on piles of money dwarfing a GDP of a major economy: Microsoft reportedly canceled internal Claude Code licenses, while Uber burned through its 2026 AI budget in 4 months.
A collateral effect might be the death of “AI is the new industrial revolution” cliche: in the 19th century, machines reduced marginal production costs - what we’ve got now is actually a spike in operating cost as usage scales (even behemoths are renting intelligence per task).
Euthanasia of the rentier: if they have Internet in hell, Keynes must be laughing.
https://aiweekly.co/alerts/microsoft-drops-claude-code-after-budget-overrun
Every time I talk to people about the future of AI in Europe, the topic of innovation surfaces and somebody mentions “but ASML”. And fair enough, they are an epic success story - except they are getting annoyed with the European Commission handling the economy.
https://newsbit.nl/asml-topman-haalt-hard-uit-naar-eu-ai-bedrijven-vertrekken-uit-europa/
FRINGE
I am perfectly aware that “the innocent have nothing to fear” is a very common mindset, and most people cannot think beyond first order consequences, but still: handing over unchangeable biometrics to a company training models for the military is not a good idea.
Then again, if you believe in palm reading, you probably deserve what follows.
https://www.media.io/ai/image-to-image/ai-palm-reading-chatgpt
RESEARCH
This is quite useful: a new paper introduces the Sig-Graph GAN model designed to generate realistic synthetic financial time series **while avoiding weak stationarity assumptions**. It integrates time-series signatures, LSTMs, and Graph Neural Networks via visibility graphs - this way the framework effectively captures both geometric and temporal patterns in volatile stock market data.
Paper: https://arxiv.org/abs/2605.22215
You want smart sampling, at some point you will end up with MCMC: everybody knows that (with the possible exception of people who think before ChatGPT there was only darkness), The glorious moment has finally arrived for diffusion models: they rely on sampling in the latent space, but the standard samplers - even with Langevin correctors - introduce bias (both from discretization and from score estimation methods). How to fix that? With an absolute classic, that’s how: Metropolis-style corrections using only the score function. The “style” in preceding sentences is there for a reason: you don’t have direct access to density ratios..
The techniques improve image generation quality on FFHQ, AFHQv2, and ImageNet-64 without requiring model retraining.
Paper:https://arxiv.org/abs/2605.09654
Real-world predictions often depend heavily on external, unstructured text: news reports, social media “content”, domain expert feedback and whatnot. Such data is not directly consumable by deep learning models - and this being 2026, “transformer” is the answer no matter the question ;-) Nexus is a new framework that utilizes a multi-agent framework to parse raw data alongside contextual real-world text, allowing large language models to accurately track shifts that could be otherwise hard to incorporate.
Paper:https://arxiv.org/abs/2605.14389
There is a joke I heard as a student - it’s so old, a version of it was probably found on the tablets recovered in the ruins of Babylon - and it goes like this: a mathematician is given a piece of wood with two nails in it. One nail is hammered all the way in, the other is half-way - and the task is to remove them both. What does a math guy do? He starts with the one all the way in, because that’s the more interesting case. Having solved it, he proceeds to the other nail and the first thing he does is hammer it all the way in - in order to reduce the problem to a known case.
This pretty much summarizes the time honored approach to the problem of stationarity: we have a nice theory of how stationary processes behave, so if we have a trend or seasonality, a statistician starts by removing them. We’ve always done it that way, so it must be the right approach, right? Some people have always grumbled about information loss, but who are they to argue with generational wisdom…
Luckily, the rebellious mindset refuses to go away: this (extensive) empirical study challenges standard textbook methodologies by checking if force-transforming datasets to meet stationarity criteria actually yields better models. Across over 3.5k controlled experiments, the authors demonstrate that traditional data transformations frequently weaken final model accuracy by deleting critical trends.
Paper:https://arxiv.org/abs/2605.17689
The first thing you need to know about updating time series models live in production is that it is not particularly standardized (which is a polite way of saying it is usually messy). This new paper offers a mathematically cleaner, frequency-aware baseline method: we update only a fraction of the model parameters while remaining resilient to sudden noise.
Paper: https://arxiv.org/abs/2605.17250
