Coffee of the Week
This week will make you rethink the future of AI
Another week has passed and, as always, the world of artificial intelligence hasn't stopped for a second. Between tools that promise to revolutionize content creation, scandals that make us question who is really behind the "magic" of AI, and advancements that seem straight out of a science fiction movie, there's a lot to unpack.
Get ready for a journey through the hottest news of the week, where we'll go from AI-generated videos to startups that fooled tech giants. Yes, you read that right - there's a story about 700 Indian engineers who pretended to be artificial intelligence. It's no joke!
Luma AI introduces Modify Video, a tool to reimagine and control videos
Luma AI has launched Modify Video, an innovative tool that allows users to reimagine any video, offering director-level control in post-production, covering style, characters, and settings. The tool enables reshaping performances, swapping entire worlds, and redefining framing, all with the goal of infinitely shaping the creative vision.
Now this is giving creative power into anyone's hands! Luma AI continues to impress me with its practical approach to video generation. While other companies focus only on creating from scratch, they realized that often we just want to modify something that already exists.
ManusAI: introducing Manus video generation
Manus transforms your prompts into complete, structured, sequenced stories ready to be watched. With a single prompt, Manus plans each scene, creates the visuals, and animates your vision. From storyboarding to concept visualization, your ideas become animated videos in minutes. Early access for Basic, Plus, and Pro members. Coming soon for everyone!
ManusAI is clearly betting on narrative as a differentiator. While many tools generate isolated clips, they want to create coherent stories. It's a smart approach, especially for content creators who need more than just "pretty videos" - they need narratives that make sense.
NotebookLM now lets you share notebooks with anyone via public link
NotebookLM announced that it is now possible to curate and share notebooks with anyone, using a public link. The announcement invites users to share their notebooks with the community.
This is one of those features that seems obvious but makes all the difference. NotebookLM was already excellent for organizing information, but now with public sharing, it becomes a powerful collaborative tool. I can imagine the use cases: from research teams to teachers sharing resources with students.
Codex gets internet access, but it's off by default
Sam Altman announced that Codex now has internet access, but emphasizes that it is off by default and that there are complex trade-offs involved. He recommends that people read carefully about the risks and use it when it makes sense, and also informs that it will be available for ChatGPT Plus users.
I like this cautious approach from OpenAI. Giving internet access to a coding tool is powerful, but also dangerous. Leaving it off by default shows they are taking security risks seriously. It's the kind of responsible decision we should see more often in the industry.
'700 Indian engineers posed as AI': The London startup that took Microsoft for a ride
Builder.ai, once touted as a revolutionary AI startup backed by Microsoft, has entered insolvency following revelations that its core no-code development platform was powered not by artificial intelligence, but by 700 human engineers in India. The company marketed its platform as being driven by an AI assistant named "Natasha," which supposedly assembled software applications like Lego bricks, but recent reports revealed that customer requests were fulfilled manually by developers.
This story is both hilarious and concerning. On one hand, it's impressive how they managed to keep up the facade for so long. On the other hand, it shows how easy it is to fool investors with AI buzzwords. It's a reminder that we should always be skeptical when something seems "too good to be true" in the tech world.
A breakthrough in breast cancer prevention: FDA approves first AI tool to predict risk from a mammogram
The FDA has approved Clairity Breast, the first AI platform that predicts a woman's risk of developing breast cancer in the next five years, using just a standard mammogram. Clairity Breast analyzes the mammogram and uses artificial intelligence to detect subtle imaging patterns in breast tissue that correlate with future cancer development, even if the mammogram appears normal to the human eye. The tool provides a validated five-year risk score that can guide personalized follow-up care before any signs of disease emerge. Commercial launch of Clairity Breast is expected in late 2025.
Now this is AI being used for something truly important. The ability to predict breast cancer five years before symptoms appear can save thousands of lives. It's in these medical applications that AI truly shines - finding patterns that humans simply cannot see.
Meta makes AI official to automatically create ads by 2026
Meta aims to automate ad creation using artificial intelligence by the end of next year. The company wants to create and distribute advertising pieces with the help of artificial intelligence, requiring only the objective, how much they would pay for the desired result, connecting the bank account, and Meta will do the rest. This system is expected to be operational by the end of 2026.
On one hand, this will democratize advertising for small businesses that don't have the resources for agencies. On the other hand, I fear we will be flooded with even more generic AI-generated ads. The challenge will be to maintain authenticity and creativity in an era of automated advertising. Not to mention the hatred it will create among large advertising agencies...
Microsoft unveils Bing Video Creator
Bing Video Creator, powered by Sora, turns your text prompts into short videos, allowing you to describe what you want to see and watch your vision come to life. Free and accessible through the Bing mobile app, and coming soon to desktop and Copilot search, Bing Video Creator democratizes AI video generation, offering features like quick creations and the ability to refine results through descriptive prompts and style modifiers.
Microsoft continues its strategy of making powerful tools accessible for free. Using Sora through Bing is a smart move to compete with ChatGPT. This could be a game-changer for content creators who don't have the budget for premium tools. I've already started testing, but I want them to add the option to create videos from an image.
ElevenLabs introduces Conversational AI 2.0
ElevenLabs has launched Conversational AI 2.0, an update to its platform for voice agents, featuring advanced capabilities such as more natural interaction, automatic language detection, RAG integration for knowledge access, multimodal support (voice and text), batch calling, and enhanced security and compliance, including HIPAA and EU data residency.
ElevenLabs is betting big on the enterprise market with these compliance features. The RAG integration is very interesting - it means voice agents can access specific knowledge bases. I see a lot of potential for call centers and customer service.
Exclusive: Anthropic hits $3 billion annualized revenue on business demand for AI
Artificial intelligence company Anthropic is generating about $3 billion in annualized revenue, a significant jump from nearly $1 billion in December 2024. This growth demonstrates the increasing business demand for AI, with code generation being a major driver for Anthropic. The company, founded in 2021 by a team that left OpenAI, is valued at $61.4 billion, while OpenAI is valued at $300 billion.
Tripling revenue in less than six months is absolutely impressive! Anthropic is doing something right, especially in the enterprise market. The fact that code generation is a major driver shows that developers are truly adopting these tools in their daily work. And I can only corroborate this, here at Yes Marketing we significantly reduced software project delivery times when we started using AI for pair coding and testing.
Advanced audio dialogue and generation with Gemini 2.5
This article introduces the new capabilities of Gemini 2.5 for audio dialogue and generation, including natural conversation, style control, tool integration, and conversation context awareness. Gemini 2.5 also offers audio-video understanding, multilinguality, and affective dialogue. The article also discusses Gemini 2.5's controllable text-to-speech (TTS) technology, which allows users to generate audio with precise style, tone, and emotional expression.
Google is trying to catch up in the conversational AI race. Gemini 2.5's audio capabilities sound very good, especially the fine control over style and emotion. It will be interesting to see how this compares to OpenAI's Advanced Voice Mode.
Google quietly released an app that lets you download and run AI models locally
Google has quietly released an app called Google AI Edge Gallery, which allows users to run a variety of AI models available on the AI dev platform Hugging Face on their phones. The app is available for Android and soon for iOS, allowing users to find, download, and run compatible models that generate images, answer questions, write and edit code, and more. The models work offline, leveraging the phones' processors.
Finally! Running AI models locally on the phone provided by a major player (here, Google) was something I had been waiting for for a long time. Privacy and the ability to work offline are huge advantages. This could be a game-changer for those concerned about data privacy. I already used an app called PocketPal to run SLMs, but the tool is not very adaptive...
Large language models are proficient at solving and creating emotional intelligence tests
Large language models (LLMs) demonstrate knowledge in diverse domains, but their emotional intelligence capability remains uncertain. This research examined whether LLMs can solve and generate performance-based emotional intelligence tests. The results showed that ChatGPT-4, ChatGPT-01, Gemini 1.5 flash, Copilot 365, Claude 3.5 Haiku, and DeepSeek V3 outperformed humans on five standard emotional intelligence tests, achieving an average accuracy of 81%, compared to the human average of 56% reported in the original validation studies.
This is both impressive and disturbing. If AIs have better "emotional intelligence" than humans on standardized tests, what does that say about us? Of course, there's a difference between answering a test correctly and actually feeling and understanding emotions, but it's a veeery interesting result.
Meta plans to replace humans with AI to assess privacy and social risks
Meta plans to replace human reviewers with artificial intelligence to assess the privacy and social risks associated with new features on Instagram, WhatsApp, and Facebook. This move, which aims to streamline the assessment process, raises concerns about AI's ability to effectively detect and mitigate potential harms. While Meta argues that AI will allow them to roll out app updates faster, critics argue that the decision could compromise the quality of the assessment and increase risks for users.
Hmm, using AI to assess the social risks of AI itself? This seems to me like the technological equivalent of "letting the fox guard the hen house." I understand the need to speed up processes, but some things require human judgment, especially when it comes to social impact. I hope Meta opens its eyes and backs down from this idea.
Mistral releases a 'vibe' coding client, Mistral Code
French AI startup Mistral is releasing its own "vibe coding" client, Mistral Code, to compete with incumbents like Windsurf, Anysphere's Cursor, and GitHub Copilot. Mistral Code is an AI-powered coding assistant that bundles Mistral's models, an in-IDE assistant, local deployment options, and enterprise tools into a single package. The private beta is available for JetBrains and Microsoft's VS Code development platforms.
Mistral is following the lead of other companies by creating its own IDE. The term "vibe coding" is becoming popular, and I like the idea. Coding should be intuitive and fluid. The local deployment option is an important differentiator for companies concerned about security.
OpenAI slams court order to save all ChatGPT logs, including deleted chats
OpenAI is fighting a court order to preserve all ChatGPT user logs, including deleted and sensitive chats, claiming the order violates user privacy and harms its ability to meet privacy commitments. The company argues the order was issued prematurely, based only on speculation that users were deleting their logs to cover up copyright infringement.
This is an important battle for digital privacy. If courts can order the preservation of data that users explicitly deleted, where does control over our own information lie? I support OpenAI in this fight, even if I sometimes disagree with their other decisions (many others, I must say).
Samsung close to wide-ranging deal with Perplexity for artificial intelligence features
Samsung Electronics Co. is close to a wide-ranging deal to invest in Perplexity AI Inc. and put the artificial intelligence startup’s search technology at the forefront of the South Korean company’s devices. The two companies are in talks to pre-load Perplexity’s app and assistant on upcoming Samsung devices and integrate the startup’s search features into Samsung’s web browser. The companies have also discussed integrating Perplexity’s technology into Samsung’s Bixby virtual assistant.
Finally, Samsung might have a chance to revitalize Bixby. Perplexity has a differentiated approach to AI search, and integrating that into Samsung devices could be exactly what they need to compete with Google Assistant and Siri. It's a partnership that makes sense for both parties, I hope they make the solution available for non-Samsung phones too.
This benchmark used Reddit’s AITA to test how much AI models suck up to us
A new benchmark, called Elephant, makes it easier to spot when AI models are overly sycophantic—but there’s no current fix. A new benchmark that measures the sycophantic tendencies of leading AI models could help AI companies avoid these issues in the future.
Using the AITA subreddit to test if AIs are too "people-pleasing" is hilarious. It's a real problem, many models tend to agree with everything we say, even when they should disagree or call us out. I hope this benchmark helps create more honest and less sycophantic AIs. And since I discovered this term "people-pleasing" I no longer believe in humanity, LOL.
Transfer learning for predictive modeling of covid-19: A multicenter study from 12 hospitals
This multicenter study, conducted in 12 hospitals in Brazil, demonstrates the successful application of transfer learning in predicting Intensive Care Unit (ICU) admissions for covid-19 patients. By leveraging knowledge from a better-performing hospital, it was possible to improve and accelerate the training of predictive models in other hospitals with less data or lower performance, highlighting the potential of transfer learning to enhance efficiency and adaptability in resource-limited healthcare settings.
This kind of collaboration between hospitals using AI is exactly what we need to see more of. Transfer learning allows hospitals with fewer resources to benefit from the experience of others. It's AI being used for the common good, especially important in countries with healthcare inequalities.
Trump's budget would ban states from regulating AI for 10 years. That could be a problem for everyday Americans.
Former President Donald Trump's budget proposes a 10-year ban on states regulating artificial intelligence (AI), which could leave the public vulnerable to AI without any protections. The proposal is part of a drastic overhaul of the federal government's fiscal and financial structure and has raised concerns among lawmakers and industry experts, who argue that states should have the freedom to regulate this potentially transformative technology.
Ten years without state AI regulation seems too long for a technology that evolves so quickly. I understand the need not to stifle innovation, but some oversight is necessary. AI affects everything from employment to privacy, we cannot leave that completely unregulated. Every time I read news like this, I am grateful for the barriers imposed by the EU.
Windsurf says Anthropic is limiting its direct access to Claude AI models
Windsurf, a popular "vibe-coding" startup reportedly being acquired by OpenAI, says Anthropic has significantly reduced its first-party access to its Claude 3.7 Sonnet and Claude 3.5 Sonnet AI models. Windsurf has to find other third-party compute providers to run several of Anthropic's most popular AI models on its platform. Anthropic is prioritizing capacity for sustainable partnerships and allows developers to access Claude through its direct API integration, its partner ecosystem, and other developer tools.
If Windsurf was really acquired by OpenAI, then it makes sense for Anthropic to limit access, it would be strange to feed the competition. This situation shows how acquisitions in the AI space can have unexpected consequences for users of the tools. I repeat here again, if OpenAI formalizes this purchase, we will stop using Windsurf here.
Paper: Gödel Darwin Machine: Open-ended evolution of self-improving agents
Current AI systems have fixed human-designed architectures and cannot autonomously and continuously improve themselves. The Gödel Darwin Machine (DGM) is a self-improving system that iteratively modifies its own code and empirically validates each change using coding benchmarks. Inspired by Darwinian evolution, DGM maintains an archive of generated coding agents, increasing performance on SWE-bench from 20.0% to 50.0% and on Polyglot from 14.2% to 30.7%.
This is simultaneously fascinating and a little scary. An AI that improves itself is the holy grail of artificial intelligence, but it's also something many researchers fear. The results are impressive, but I hope there are strict controls in place. We are literally creating systems that evolve on their own.
And so we reach the end of another week full of developments that make us question what is real and what is artificial.
What stands out most to me this week is the diversity of applications. From breast cancer prevention to the 700 Indian engineers who pretended to be AI, we see that this technology is simultaneously solving real problems and creating new ethical and social challenges.
The question is no longer whether AI will change our lives, but how we will navigate these changes responsibly.
