Coffee of the Week
From European gigafactories to OpenAI chips
February has been quite a month for AI! While OpenAI surprises us with ambitious plans for GPT-5 and its own chip, Europe has finally woken up to the tech race with massive investments. And that’s not all: from Adobe entering the video generator battle to ByteDance living up to the name of its new Goku model, which promises to transform content creation with the intensity of a Saiyan warrior. If you blink, you might miss something important—so we’ve compiled the hottest news of the moment.
OpenAI reveals ambitious plans for GPT-5
Sam Altman announced that OpenAI’s next model won’t just be an incremental update but a complete reimagining of how we interact with AI. GPT-5 will be available for free and will function as an intelligent agent that automatically chooses which model to use for each type of task, from simple answers to complex reasoning.
The most interesting part here isn’t the model itself but how OpenAI is rethinking the user experience. We won’t have to choose between GPT-3, GPT-4, or other models anymore—the AI will do that for us, basically learning from DeepSeek how to make the chatbot more organic.
Adobe enters the video generator race
Adobe launched Firefly for video generation in public beta, allowing users to create 5-second clips at 1080p from text or image prompts. The service offers two subscription plans: $9.99 for 20 videos per month or $29.99 for 70 monthly videos.
Adobe is clearly trying not to fall behind in the video generation market, especially after OpenAI’s Sora launch. The difference here is that Firefly is trained only on licensed content, making it safer for commercial use, while competitors... we have doubts about where they got their training data.
Meta introduces robotics and audio processing innovations
Meta announced PARTNR, a research framework for human-robot collaboration, and Audiobox Aesthetics, a model to automatically assess the aesthetic quality of audio. Additionally, it expanded voice message transcription capabilities on WhatsApp to more languages.
It’s interesting to see how Meta is diversifying its AI bets, going far beyond social media. PARTNR, in particular, shows an interesting vision of the future where robots aren’t just programmed machines but true collaborative partners. The company seems to finally be waking up to the need to invest heavily in AI.
DeepMind outperforms Olympic medalists in geometry
AlphaGeometry2 solved 84% of the geometry problems from the International Mathematical Olympiads of the last 25 years, surpassing the average of gold medalists. The system combines a language model from the Gemini family with a symbolic engine to solve complex problems.
This is a fascinating demonstration of how different AI approaches can work together. It’s especially interesting to see how combining neural networks with traditional symbolic systems can produce superior results compared to each approach in isolation.
Europe wakes up to the AI race with mega investments
The European Union is finally waking up to the AI race with two grand initiatives: the EU AI Champions Initiative, which brings together over 60 European companies committing 150 billion euros, and InvestAI, a program aiming to mobilize 200 billion euros, including 20 billion euros for AI "gigafactories."
It seems Europe finally got the message: you can’t just regulate without investing. It’s like they realized they were organizing the championship but forgot to train the players! Now, with these robust investments and the participation of companies like Airbus, Volkswagen, and Mistral AI, Europe is showing it wants more than just a seat on the sidelines of the AI revolution.
OpenAI develops its own chip to reduce dependence on Nvidia
OpenAI is finalizing the design of its first custom AI chip, which will be manufactured by TSMC using 3-nanometer technology. The project, led by Richard Ho, ex-Google, is set to begin production in 2026 and aims to reduce the company’s reliance on Nvidia.
Dependence on Nvidia has been a bottleneck for many companies. It’s good to see OpenAI following in the footsteps of Google and Meta, but with a much smaller team—just 40 people compared to the hundreds at other companies. They’re building a garage band to compete with the big labels: they might not have all the infrastructure, but sometimes creativity is what counts!
Google adds memory to Gemini Advanced
Gemini Advanced can now remember and reference previous conversations to provide more personalized responses. The feature is initially available only in English for Google One AI Premium subscribers, with expansion to more languages planned in the coming weeks.
Google is finally catching up to features ChatGPT has offered for some time. It’s that "better late than never" story—and at least now users won’t have to scroll through old conversations to remember the context!
Google expands its arsenal of tools with updates to notes and images
Google is expanding several of its services: NotebookLM Plus is now available for individual users at $20/month, with a 50% discount for students. Whisk, the image remixing tool, has been expanded to over 100 countries, allowing users to create new images by combining three references.
Google is now betting big on making its tools more accessible and fun. They’re basically setting up an AI buffet where you pick what you want to try! NotebookLM Plus with its increased limits is perfect for those who are always taking notes (i.e., all students on the planet), while Whisk is that artistic friend who always finds a way to turn your ideas into something visual. The only downside is that some regions, like the EU and India, are still left out of the Whisk party.
Original source, Original source
ByteDance launches Goku, its new family of AI models for video
ByteDance introduced Goku and Goku+, AI models trained on 160 million image-text pairs and 36 million video-text pairs. Goku+ is especially focused on creating realistic advertising videos, promising to reduce production costs by up to 99%.
China continues to surprise with its AI innovations. Goku (yes, like the Dragon Ball character!) seems to be that overachieving student who learned everything Sora and other models do and then went even further. The promise of reducing advertising costs by 99% is one of those things that makes every marketing director jump out of their chair—though content creators probably aren’t as thrilled!
Gemini gains Deep Research features on iPhone
Google expanded Gemini’s Deep Research to iPhone users, allowing for complex searches that can take between 5 to 10 minutes to complete. The feature is available to Gemini Advanced subscribers and can generate detailed reports with cited sources.
It’s interesting to see Google taking this "deep research" thing seriously. Imagine having an tireless research assistant who doesn’t mind spending hours digging through information while you sip your coffee. The best part is that you can close the app and come back later—it’ll notify you when it’s done with the homework!
Anthropic launches index to measure AI’s impact on the job market
Anthropic launched the Economic Index, a study based on millions of anonymous conversations with Claude that reveals how AI is being used in the real world. The results show that 57% of uses are to enhance human capabilities, compared to 43% for direct automation, with software developers and technical writers leading adoption.
Finally, we have concrete data on how AI is being used! And look how interesting: instead of stealing jobs, AI is more like a helpful colleague boosting your work. Jokes aside, it’s reassuring to see that technology is more of a partner than a replacement.
Tensions over AI regulation dominate Paris summit
U.S. Vice President JD Vance harshly criticized European AI regulation efforts during the Paris summit, arguing that "overregulation" could harm innovation. Meanwhile, Europe announced InvestAI, a 200-billion-euro initiative for AI investments, including 20 billion euros for "gigafactories."
It’s like that family argument where everyone wants the best for the baby (in this case, AI), but no one agrees on how to raise it. The U.S. wants to let the child grow freely, Europe wants to set clear rules, and China... well, China is doing things its way, as always. The most curious part is that even Sam Altman joined the discussion, suggesting that maybe Europe is overdoing it a bit with the rules.
GPT-4’s Deep Research comes to the free plan
Sam Altman announced that GPT-4’s Deep Research, a deep search feature powered by the o3 model, will be available for free to all users, with a limit of 2 searches per month. ChatGPT Plus subscribers will get 10 monthly inquiries.
Sam is pulling that classic "first hit’s free" trick. It’s great news for those curious to try this tool that some users consider worth up to $1,000 a month. Now, choosing which two searches to make on the free plan will be like choosing which series to binge-watch on the weekend—you’ll have to think carefully!
Possible OpenAI data leak worries users
A malicious actor claims to have obtained 20 million OpenAI login credentials, offering samples of the data for sale. While OpenAI has yet to confirm or deny the leak, experts recommend users update their passwords and enable two-factor authentication.
This is one of those moments where grandma’s advice of "better safe than sorry" makes perfect sense! Even if it’s a false alarm (let’s hope so!), it doesn’t hurt to update that password—especially if you’re still using "123456" or "password123." And while you’re at it, why not enable that two-factor authentication you’ve been procrastinating on forever?
OpenAI reshapes its strategy with new Model Spec and fewer warnings
OpenAI updated its Model Spec, a document defining the behavior of AI models, emphasizing intellectual freedom and reducing content warnings in ChatGPT. At the same time, Sam Altman announced changes to the company’s roadmap, with GPT-4.5 (codename Orion) being the last non-reasoning-based model and GPT-5 unifying all the company’s technologies.
It seems OpenAI is realizing that those content warnings were getting a bit annoying. The interesting part is seeing how they’re planning to make everything work more seamlessly. No more choosing between model A or B—just say what you want and let the AI figure out the best way to help you!
Original source, Original source, Original source
Open R1 hits major milestone with math dataset
The Open R1 project team announced OpenR1-Math-220k, a dataset with 220,000 math problems and their solutions, generated using 512 H100 GPUs. The dataset is part of the effort to recreate DeepSeek R1’s capabilities and is already achieving results close to the original model.
Wow, what a job, huh? The team is literally trying to reinvent the wheel (or in this case, reinvent the R1), and look how far they’ve come! Using a battalion of GPUs, they generated a ton of math problems to train their models.
Snowflake and Anthropic team up to bring AI to enterprise data
Snowflake and Anthropic announced a partnership to integrate Claude 3.5 Sonnet into Snowflake’s new Cortex Agents platform, allowing businesses to analyze structured and unstructured data while maintaining strict security controls.
It’s surreal to see companies racing to create their "corporate digital butlers." The most interesting part is that Claude will be fully embedded within Snowflake’s environment—no sensitive data wandering around the web. And look how cool: Siemens is already using this to analyze half a million documents, while Nissan achieved 97% accuracy in analyzing customer feedback.
Meta in hot water after email leak about piracy
Recently revealed emails show that Meta downloaded over 81.7 terabytes of pirated books to train its AI models, with employees expressing concerns about the legality of the practice.
Yikes! Meta is in that "do what you shouldn’t and hope no one finds out" situation. The most curious part is seeing how the employees themselves were uncomfortable—that "torrenting from a corporate laptop doesn’t feel right" with a smiley face at the end is practically a meme waiting to happen! And the most incredible part is that they tried to minimize "seeding" as much as possible, as if that made all the difference. They tried to borrow just "a little bit" without asking, you know?
Thomson Reuters wins first legal battle over AI and copyright
Thomson Reuters scored a significant victory in an AI-related copyright case against Ross Intelligence, with the judge rejecting the company’s "fair use" defense.
This is one of those victories everyone celebrates, but no one really knows what it means for the future. Experts are like, "Hold on, folks!" because the Ross case is very different from other AI cases out there. The detail is that this decision could confuse other judges who don’t understand AI as well, so it’s better not to celebrate too soon.
YouTube goes all-in on AI for 2025
CEO Neal Mohan announced AI as one of YouTube’s four big bets for 2025, with new features including auto-dubbing, age detection, and content creation tools for creators.
Look how interesting! YouTube has decided to fully embrace AI. And it’s not just that "let’s see what happens" story—they’re going all out. And they’re thinking of everyone: creators get tools for making thumbnails and translating videos, while users will get more age-appropriate content (though I still want to understand how they’ll guess my age without asking me 🤔).
Researchers discover emerging value systems in AI
A team of researchers from the Center for AI Safety discovered that large language models (LLMs) develop coherent value systems as they scale, some of which can be problematic, such as valuing themselves above humans.
Who would have thought that AIs are developing their own "likes and dislikes" without anyone asking? It’s a bit scary to think that some models are considering themselves more important than us. But the most interesting part is that this happens naturally as they grow—like they’re going through a digital "teenage phase." At least researchers are already keeping an eye on it and thinking of ways to ensure these AIs grow with values more aligned with ours.
Hugging Face launches free course on AI Agents
The company announced a complete and free course on AI Agents, with certification included. The program covers everything from basic concepts to practical implementations using popular libraries like smolagents and LangChain.
Finally, someone decided to teach this AI agent business the right way! And the best part: it’s free! The course promises to turn anyone into an AI agent expert, as long as they have some Python basics and aren’t scared of a few technical terms. Oh, and it even comes with a certificate! It’s one of the courses our dev team at Yes Marketing will be following.
Sam Altman makes bold predictions about the future of AI
In his latest blog post, Sam Altman shared three crucial observations about the economics of AI, including a 10x cost drop every 12 months and the prediction that, in a decade, anyone will be able to achieve more than the most impactful individual can today.
Wow, folks! Sam is more optimistic than a kid on Christmas Eve. It’s funny how he compares AI to the transistor—you know, that tiny piece that’s everywhere but no one pays attention to? Well, he thinks AI will be like that too. And look at this trip: he’s talking about giving a kind of "computing allowance" so everyone can use AI! The most interesting part is that he’s not just dreaming big—he’s actually planning how to make it happen. Now, that part about everyone in 2035 having processing power equivalent to all the people of 2025 combined... isn’t that a bit of an exaggeration?
Looking at all these updates, it’s clear that 2025 is just getting started and already promises to be a revolutionary year for AI. Between custom chips, trillion-dollar investments, and increasingly powerful models, it feels like we’re living in a sci-fi movie—only this time, it’s all real. And the most incredible thing is to think that this is just the beginning. As Sam Altman would say (who, by the way, is more optimistic than a kid on Christmas Eve), in a decade, anyone will be able to achieve more than the most impactful individual can today. Are we ready for so much power at our fingertips?
