Tech Insights 2025 Week 3

Wow, what a start of 2025! AI now outperforms experts in detecting both breast cancer and ovarian cancer, new lawsuit says Zuckerberg approved engineers at Meta to download pirated material over torrent and remove copyright information to train their first Llama AI models, AI-generated phishing mails can now achieve higher click-through than expertly handcrafted phishing mails, Adobe introduces a new generative video model TransPixar that supports transparent backgrounds, and NVIDIA introduced a $3,000 AI supercomputer that is 50 times faster running Large Language Models than a Mac Studio M2 Ultra.

WANT TO RECEIVE THIS NEWSLETTER AS A WEEKLY EMAIL?

If you prefer to receive this newsletter as a weekly email straight to your inbox, you can sign up at: https://techbyjohan.com/newsletter/ . You will receive one email per week, nothing else, and your contact details will never be shared with any third party.

THIS WEEK'S NEWS:

1. AI Boosts Breast Cancer Detection by 18% in Landmark German Study
2. AI Outperforms Human Experts in Detecting Ovarian Cancer
3. Zuckerberg Approved Pirated Books for Meta's AI, Lawsuit Says
4. AI-Powered Spear Phishing Matches Human Deception Levels
5. TransPixar: AI System Bringing Transparency to Video Generation
6. NVIDIA Unveils Next-Gen AI and Gaming Tech at CES 2025
7. Samsung Introduces Vision AI-Powered Screens at CES 2025
8. OpenAI is Losing Money on $200 Pro Subscription
9. Meta Changes Content Moderation Policies

AI Boosts Breast Cancer Detection by 18% in Landmark German Study

https://www.nature.com/articles/s41591-024-03408-6

The News:

• The world's largest prospective AI mammography study, named PRAIM (Performance of AI in Mammography), demonstrated that artificial intelligence significantly improves breast cancer detection rates in Germany's national screening program. The study involved over 460,000 women and 119 radiologists across 12 screening sites.
• The AI system detected 6.7 cancer cases per 1,000 women screened, compared to 5.7 cases with traditional methods - an 18% increase in detection rates. Most importantly, this improvement came without increasing false positives or unnecessary follow-up examinations.
• The research is particularly significant for Germany's healthcare system, where breast cancer is the most common cancer among women, with approximately 78,000 new cases diagnosed annually.

My take: If you still believe we are living in an "AI Hype Bubble" then it's definitely time to accept that this is not a hype, not a bubble, and that results like this will most likely trigger dozens of new research projects with even greater potential on improving our health and saving lives. Most current AI models are still limited in GPU capacity and most of them do not use test-time-compute, so expect medical AI detection rates to be significantly better in the next few years, not only outperforming but completely replacing human experts for medical diagnosis.

AI Outperforms Human Experts in Detecting Ovarian Cancer

https://news.ki.se/ai-can-improve-ovarian-cancer-diagnoses

The News:

• An international study led by Karolinska Institutet in Sweden demonstrated that AI models can detect ovarian cancer in ultrasound images with greater accuracy than human experts. The AI achieved an impressive 86.3% accuracy rate, compared to 82.6% for expert examiners and 77.7% for non-expert examiners.
• The AI system was trained on over 17,000 ultrasound images from 3,652 patients across 20 hospitals in eight countries. The neural network models were designed to differentiate between benign and malignant ovarian lesions, showing consistent performance across different ultrasound systems and patient populations.
• The technology significantly improves healthcare efficiency by reducing the need for expert referrals by 63% and decreasing misdiagnosis rates by 18%.
• The AI models demonstrated exceptional performance in challenging cases where human examiners were uncertain. Professor Elisabeth Epstein notes that "neural network models can offer valuable support in the diagnosis of ovarian cancer, especially in difficult-to-diagnose cases and in settings where there's a shortage of ultrasound experts".

My take: Much like the AI model that detected Breast Cancer better than human doctors, AI models are now better than humans in almost all types of medical diagnostics. And they will only improve from here. Systems like this are amazing news especially for poorer countries with limited medical facilities, where it today might be impossible to get any diagnosis at all.

Read more:

• AI detects ovarian cancer better than human experts in new study | Fox News

Zuckerberg Approved Pirated Books for Meta's AI, Lawsuit Says

https://techcrunch.com/2025/01/09/mark-zuckerberg-gave-metas-llama-team-the-ok-to-train-on-copyrighted-works-filing-claims/

The News:

• Meta is facing allegations in a copyright lawsuit where plaintiffs claim the company knowingly used pirated materials to train its Llama AI models with CEO Mark Zuckerberg's explicit approval.
• Meta used LibGen, a controversial "links aggregator" known for providing unauthorized access to copyrighted works, to train its AI models. The dataset includes materials from major publishers like Cengage Learning, Macmillan Learning, and McGraw Hill. LibGen has previously faced multiple lawsuits and been fined millions for copyright infringement.
• Meta engineer Nikolay Bashlykov wrote a script to deliberately remove copyright information, including the words "copyright" and "acknowledgments" from ebooks and scientific journals in the training data.
• The company obtained the materials through torrenting, despite concerns from Meta's research engineers about legal implications. Meta's head of generative AI, Ahmad Ah-Dahle, approved the torrenting of LibGen materials despite internal reservations.

My take: If you somehow thought Facebook has turned into "the good guys" by releasing AI models as open-source and talking about fair use of copyrighted materials, it seems you were wrong. We are currently in a full-speed race to AGI and ASI, and I believe things are about to get much worse when it comes to privacy and copyright infringements in the next couple of years. The companies leading the AI race will make sure they always have enough data to train their models, one way or the other.

Read more:

• Mark Zuckerberg gave Meta's Llama team the OK to train on copyrighted works, filing claims | TechCrunch
• Lawsuit says Mark Zuckerberg approved Meta's use of pirated materials to train Llama AI

AI-Powered Spear Phishing Matches Human Deception Levels

https://www.lesswrong.com/posts/GCHyDKfPXa5qsG2cP/human-study-on-ai-spear-phishing-campaigns

The News:

• A recent study evaluated the effectiveness of AI-generated spear phishing emails compared to those written by human experts. The research found that AI can produce phishing emails with a click-through rate over 50%, matching the success rate of emails written by humans.
• Researchers used AI agents built with GPT-4o and Claude 3.5 Sonnet to collect publicly available data on subjects. This data was then used to create highly personalized phishing messages. The AI-generated emails achieved a click-through rate of 54%, significantly outperforming a control group that received arbitrary phishing emails with a 12% click-through rate.
• The study highlighted the cost efficiency of AI-driven phishing attacks, noting that they can reduce costs by up to 50 times compared to manual methods. Additionally, the AI models demonstrated a high capability in gathering accurate open-source intelligence, producing useful profiles for 88% of targets with only 4% containing inaccuracies.

My take: Are we as a society are prepared for these kinds of massive AI attacks? I don't think so. With AI bots able to scrape and collect almost any data on any person using paid online services, how can you be sure that the person contacting you is a real person and not an AI Bot? If you are not already using a password manager such as 1Password, that would be my first recommendation for safety since it automatically stops you from entering your passwords on hostile web pages even if you click on links that look seemingly valid. If everyone used password managers I am quite certain the number of successful phishing attempts would be significantly lower (despite the initial click-through) and everyone would be better prepared for this AI-bot future.

Read more:

• AI-generated phishing scams target corporate executives
• Gmail, Outlook and Apple users urged to watch out for this new email scam: Cybersecurity experts sound alarm

TransPixar: AI System Bringing Transparency to Video Generation

https://wileewang.github.io/TransPixar/

The News:

• TransPixar is a groundbreaking AI system for video effects generation by enabling the creation of transparent elements like smoke, reflections, and magical effects. This joint development by Adobe Research and Hong Kong University of Science and Technology (HUST) addresses a major limitation in current AI video generation tools, which typically can only produce solid images.
• The technology introduces a novel approach to generating RGBA videos (videos with transparency channels) while still maintaining high quality. It uses a diffusion transformer architecture and incorporates alpha-specific tokens, allowing for seamless blending of transparent elements into scenes.
• TransPixar has been made open-source and is available on GitHub, with demonstrations accessible through Hugging Face. The system supports various video tasks, including Text-to-RGBA Video and Image-to-RGBA Video generation.

My take: This was more or less the "missing piece" before generative AI can be used on a wider scale for VFX work. Almost every single effect in movies is stacked on top of dozens other effects, and being able to generate animations with built-in alpha channel is critical for a solid workflow.

NVIDIA Unveils Next-Gen AI and Gaming Tech at CES 2025

https://www.nvidia.com/en-us/events/ces/

The News:

• NVIDIA announced numerous news at CES this year:
• Project DIGITS. NVIDIA introduced Project DIGITS, a compact desktop AI "supercomputer" priced at $3,000. Designed for AI researchers, data scientists, and students, it enables the development and testing of AI models locally without relying on cloud infrastructure.
• Digits comes with a 20-core GB10 "Grace Blackwell Superchip", 128GB of unified memory, up to 4TB SSD and up to 1 petaflop of FP4 floating point performance.
• Using NVIDIA ConnectX networking, two Project DIGITS can be linked to run up to 405-billion-parameter models at FP4 quantization.
• GeForce RTX 50-Series GPUs based on the new Blackwell chip. The flagship RTX 5090 model will be available in January for $1,999, while the RTX 5070 is set to launch in February for $549.
• The Cosmos platform is a series of AI models designed for robotics and autonomous vehicles (I recently mentioned NVIDIA Cosmos in Tech Insights Week 46). This open-source platform works with NVIDIA's Omniverse to create more realistic video for training robots and automated services. The company also announced a partnership with Toyota to develop next-generation autonomous vehicles using NVIDIA's DriveOS operating system.

My take: If you have previously been looking into buying a Mac Studio for running local AI models, those machines max out at around 20 teraflops of performance. Now this $3000 machine provides performance that is 50 times faster. Buy two of them and you can run Llama 405B locally at FP4 quantization. The Digits computer could be a game changer once the next version of Llama launches that is trained on the next-generation infrastructure. It will be better than any LLM we have today and you can run it 100% locally for sensitive and high-risk documents that researchers are not allowed to transmit to cloud-based models.

Read more:

• NVIDIA Puts Grace Blackwell on Every Desk and at Every AI Developer’s Fingertips - NVIDIA Newsroom

Samsung Introduces Vision AI-Powered Screens at CES 2025

https://news.samsung.com/us/samsung-unveils-vision-ai-new-innovations-ces-first-look-2025-delivering-personalized-ai-powered-screens-to-enrich-everyday-life/

• Last week at CES Samsung introduced Vision AI, a new technology that transforms TVs and screens into "adaptive, intelligent companions to simplify and enrich daily living". This AI-powered system is integrated across Samsung's 2025 lineup, including Neo QLED, OLED, QLED, and The Frame models.
• Vision AI offers personalized features such as Click to Search, which provides instant information about on-screen content without interrupting viewing. Live Translate uses on-device AI to offer real-time subtitle translations in up to seven languages for live broadcasts.
• Samsung is also collaborating with partners like Microsoft to expand the capabilities of Vision AI, and select Samsung Smart TVs will come with Microsoft Copilot built-in.
• According to SW Yong, President and Head of Visual Display Business at Samsung Electronics, "With Samsung Vision AI, we're reimagining what screens can do, connecting entertainment, personalization, and lifestyle solutions into one seamless experience to simplify your life".

My take: I personally want my TVs to be as dumb as possible but with as high picture quality as possible. I can only assume that this feature bloat is what happens once all high-end TV reaches very good contrast, high color richness and wide display angle. When the display quality cannot be improved much further, when borders cannot be much thinner, and when weight cannot be much lighter, the only thing left is to "add value" by adding features like those above. It reminds me of the start menu in Windows 11 with ads, pre-installed apps, widgets and everything else no-one really asked for. But it's a way for the team responsible for improving the start menu to show progress internally. I think this entire model of building products is wrong, and it will be interesting to see the market response once these new TVs are out.

Read more:

• Samsung goes big on anti-glare screens and AI with 2025 TV lineup - The Verge

OpenAI is Losing Money on $200 Pro Subscription

https://x.com/sama/status/1876104315296968813

The News:

• Sam Altman, CEO of OpenAI, posted last Monday that OpenAI are losing money on their $200 Pro subscription since people use the service much more than they expected.
• The price was personally set by Sam Altman based on what he thought would work.

My take: I have started to use o1 a lot, and for certain tasks like organizing information or refactoring code it's outstandingly good. I easily save hours every week using o1, so for me personally $200 is worth the investment compared to how much time I save. Having the ChatGPT desktop available with Alt+Space on Mac just makes it so easy to use it on a daily basis, and I actually find myself using it more and more every week. I can only imagine it's the same for everyone else, and that is why usage has skyrocketed. It's just that good.

Meta Changes Content Moderation Policies

https://about.fb.com/news/2025/01/meta-more-speech-fewer-mistakes/

The News:

• Meta just announced significant changes to its content moderation approach on Facebook, Instagram, and Threads. Meta is ending its third-party fact-checking program and introducing a “Community Notes” system to reduce censorship and errors in content enforcement.
• Inspired by X, the new system allows users to add context to potentially misleading posts. This community-driven approach will according to Meta enhance information accuracy and reduce bias.
• Meta is also relaxing restrictions on discussions on immigration, gender identity, and sexual orientation.

My take: A few years ago The Verge had an article called BODIES IN SEATS: "At Facebook’s worst-performing content moderation site in North America, one contractor has died, and others say they fear for their lives". Before you make any posts or comments on how these decision by Facebook might make certain communities even more toxic, I urge you to read this article first. One content moderator literally had a heart attack at his desk and died. Getting rid of content moderators is probably the right choice, but it makes you question of good these platforms really are for us as a society when they need such massive amounts of moderation just to keep going.

Read more:

• Facebook moderators break their NDAs to expose desperate working conditions - The Verge