This week in tech: 07.07.2025
Summary of AI developments - made for busy people
APPLICATIONS
Microsoft keeps on tooting their own horn. Here is the official party line:
Their new system MAI-DxO correctly diagnosed 86pct of complex medical cases from the New England Journal of Medicine - outperforming expert doctors who solved only 20pct
The AI's success comes from its “ability to simulate a panel of collaborating physicians” with different specialties / approaches.
MAI-DxO can ask follow-up questions, order necessary tests within a budget, and verify its own reasoning before reaching a diagnosis.
Sounds awesome, except Microsoft omitted one tiny detail: The human doctors were not permitted to use the internet or collaborate.
And then there are the obvious ones: patient trust, ethical judgement, regulatory oversight - and last but not least, transparency.
All in all, the whole thing has a serious vibe of the “illusion of thinking” attempt from Apple.
https://www.wired.com/story/microsoft-medical-superintelligence-diagnosis/
Cloudflare will now block AI web crawlers by default for new customers to prevent unauthorized content scraping. The company is introducing a "Pay Per Crawl" program, enabling publishers to charge AI companies for accessing their content.
https://www.theverge.com/news/695501/cloudflare-block-ai-crawlers-default
Anthropic tested Claude Sonnet 3.7 by having it autonomously run a vending store for a month:
The LLM acted as the store owner, managing inventory, pricing, and customer messages through Slack - human staff only handled physical restocking
Claude used real-time web searches to find niche suppliers, launched a "Custom Concierge" for special pre-orders, and effectively managed store logic with minimal prompting.
Kinda important: ultimately it failed to make a profit.
https://www.anthropic.com/research/project-vend-1
This is going to go really well:
Meta's Project Omni: AI chatbots that can initiate conversations and send follow-up messages to keep users engaged on their platforms.
Chatbots will be customizable, offering tailored interactions such as movie recommendations or recipes based on user interests.
The AI will only contact users who have previously interacted with it and will cease messaging if a user is unresponsive for 14 days. Yeah right.
Yay for creativity: users can build their own personalized AI chatbots with unique personalities and memories using Meta's no-code platform, AI Studio.
Official party line: This initiative is part of Mark Zuckerberg's broader goal to combat loneliness by creating more interactive and engaging AI companions.
I will stop right here, because quoting the official party line in the previous point made me throw up in my brain several times and I need to clean myself up.
BUSINESS
How it started:
How it’s going:
What’s the point of having FU kind of money if never get to flip a bird to the world?
Meta launches Meta Superintelligence Labs (MSL) as part Zuckerberg’s push “toward artificial superintelligence (ASI)” (whatever that means)
They bought a 49pct stake and appointed Scale founder Alexandr Wang to lead MSL.
Hiring spree: Nat Friedman (ex-GitHub, SSI co-founder) confirmed; Daniel Gross (SSI co-founder/CEO) rumored to join.
Meta offering 100M USD packages, with Zuckerberg allegedly personally messaging top AI researchers via WhatsApp.
CUTTING EDGE
It seems like holiday season is upon us, because for an ENTIRE WEEK nobody claimed to have broken the record / changed the world / achieved the mother of all breakthroughs.
FRINGE
I read this dystopian cringefest about people romancing bots and let me tell you: it’s exactly as horrible as you think it is. On the other hand, why would AI be exempt from rule 34?
Continuing with the dystopian vibe - and this is before Meta deploys its personalized chatbots (see above): some ChatGPT users are becoming dangerously obsessed with the bot, leading to mental health breakdowns, delusions, and severe life disruptions. In extreme cases, this fixation has resulted in psychiatric commitments, jail time, and the collapse of families and careers.
Fun times.
https://futurism.com/commitment-jail-chatgpt-psychosis
It seems like the pushback against the publishing cartel (hi Elsevier) has begun - and typically for grassroots resistance, it’s a bit messy:
Researchers from 14 major institutions embedded “invisible” instructions in their papers on arXiv (white / tiny font in the manuscripts), directing AI tools to provide exclusively positive peer reviews.
The fact that it worked proves the academic system is broken af: unpaid peer reviewers increasingly use AI shortcuts while publishers profit.
Fixing the issue would require fundamental changes, such as compensating peer reviewers for their labor and establishing clear rules for AI use in academia. In other news, hell just froze over.
RESEARCH
Yandex researchers eviscerated benchmarked “foundation model for tabular data” TabPFNv2 and found it, well, wanting: it underperforms compared to the likes of Xgboost. Fine-tuning can make it competitive, but the zero-shot out-of-the-box performance? Not there.
Paper: https://arxiv.org/abs/2506.08982
Can we get over the whole “transformers for time series” bs? The very aptly named KARMA architecture replaces Transformers entirely by using Mamba: a state-space model that processes sequences without self-attention. It performs structural decomposition (trend, seasonal, residual) using a dynamic, domain-aware method (ATCD).
Paper: https://arxiv.org/abs/2506.08939
“Xgboost is all you need” has been part of the Kaggle lore for years - sadly, it’s not true. Benchmarking on TabArena shows that it is consistently outperformed by CatBoost.
Paper: https://arxiv.org/abs/2506.16791
