Coffee of the Week
From research simulation to restructuring tech giants
Hello, dear DevCafé readers! Here we are with another hot edition of our weekly summary, packed with the most relevant news from the tech universe. This week was busy, with fascinating developments ranging from new AI training methods to the major moves of tech giants. Shall we take a look at what stood out the most in the past few days? Prepare your coffee and settle in!
Visual imitation enables contextual control of humanoids
VideoMimic is a real-to-sim-to-real system that transforms videos into transferable skills for humanoids, allowing robots to learn contextual behaviors (traversing terrain, climbing, sitting) in a single policy. The system reconstructs humans and environments from videos, producing full-body control policies for humanoid robots to perform tasks like climbing stairs, sitting in chairs, and traversing diverse terrains. The approach offers a scalable path to teaching humanoids to operate in real-world environments.
Wow, this really seems straight out of a sci-fi movie! The ability for a robot to learn just by watching videos is a giant leap for robotics. Imagine Boston Dynamics humanoids learning parkour just by watching YouTube videos? Soon we'll have robots challenging us to a soccer match, and probably winning!
Stripe builds transformer-based payments foundation model
Stripe has built a transformer-based payments foundation model, which uses machine learning on discrete features to improve the company's products. This model learns general-purpose vectors for each transaction, distilling the key signals from each charge into a single versatile embedding. Stripe built a classifier that ingests sequences of embeddings from the foundation model and predicts whether the slice of traffic is under attack, allowing them to block attacks before they hit businesses. This approach improved the detection rate of card testing attacks on large users from 59% to 97% overnight.
Stripe continues to show why it's one of the most innovative companies in the fintech space. Achieving an increase in fraud detection from 59% to 97% is a monumental gain. This kind of improvement not only benefits Stripe but the entire digital ecosystem. Imagine the amount of money and headaches this will save merchants worldwide. Bravo!
IBM CEO says AI has replaced hundreds of workers, but created new programming, sales jobs
IBM CEO Arvind Krishna stated that the company has used artificial intelligence to replace the work of a few hundred human resources employees, but, as a result, has hired more programmers and salespeople. Krishna believes that the impacts of AI on the workforce are the ability to perform tasks such as analyzing spreadsheets, conducting research, and drafting emails autonomously.
It's the eternal AI debate in action: on one hand it eliminates jobs, on the other it creates new opportunities. While I believe in this transformation, I always think about the affected professionals. Can an HR specialist easily transform into a programmer or salesperson? History teaches us that adaptation happens, but not without pain for many people in the process.
Trump to rescind global chip curbs, prepare new AI restrictions
The Trump administration plans to rescind some Biden-era artificial intelligence chip restrictions as part of a broader effort to review global semiconductor trade restrictions that have drawn strong opposition from major tech companies and foreign governments.
Interesting twist. After years of tightening restrictions, it seems the wind is changing direction. While easing restrictions can accelerate innovation, we need to be mindful of potential security risks. The balance between technological advancement and strategic control is delicate, and it will be interesting to see what new AI restrictions will be proposed in return.
Suno launches v4.5 for Pro & Premier subscribers
Suno has launched version 4.5, offering a wider range of genres, richer vocals, and improved prompt understanding for songs. The new version includes: expanded genres and smarter mashups, enhanced voices, more complex and textured sound, better adherence to prompts, prompt improvement help, updated Covers + Personas, combinable Covers + Personas, extended song duration, and improved audio.
Suno continues to amaze me with each update, we are getting closer and closer to a point where we won't be able to distinguish between AI-generated music and human-created music. This democratization of music creation is revolutionary, but it also raises questions about the future of professional musicians. I confess I'm already creating some songs for my home parties and for the kids – it's great fun.
Netflix debuts its generative AI-powered search tool
Netflix has launched a new AI-powered search tool that uses OpenAI's ChatGPT to provide users with a conversational discovery experience. The feature will roll out this week to iOS users as an optional beta, with subscribers in Australia and New Zealand already having access to it. Other streaming platforms are also leveraging generative AI for search, such as Amazon with its AI voice search on Fire TVs. Netflix also plans to use generative AI to update titles in different languages, and other revealed features include a short-form video feed for mobile users and a TV homepage redesign.
Finally someone solved the problem of finding something to watch on Netflix. How many times have we wasted 30 minutes looking for the perfect movie? The integration of ChatGPT seems to be exactly what we needed – "show me a psychological thriller with a twist ending" is a thousand times better than browsing endless categories. Now we just need to solve the problem of me falling asleep halfway through the movie I took so long to choose...
Microsoft unveils 'new generation of Windows experiences' — here's what's coming to Windows 11 and Copilot+ PCs
Microsoft has announced a range of new features for Windows 11 and Copilot+ PCs, including AI-driven enhancements to File Explorer, Paint, Photos, and Snipping Tool. The new features will roll out to Windows Insiders in the coming weeks and months, with some exclusive to Copilot+ PCs.
Microsoft is clearly putting all its chips on AI integration. Windows has always been a somewhat conservative operating system in its updates, but it seems we are finally seeing a paradigm shift. I sincerely hope these new features are more than just marketing gimmicks and bring real productivity benefits. And, please, let AI Paint not become the new clippy.
Apple using Anthropic for AI coding as developers await Swift Assist
Apple is collaborating with Anthropic to integrate Claude into Xcode, Apple's integrated development environment, aiming to improve its internal workflow, accelerate, and modernize product development through AI, similar to companies like Windsurf and Cursor maker Anysphere. Apple has not yet decided whether to release the Xcode version with Claude integration externally. This move comes as Apple seeks to boost its Apple Intelligence and adopt industry AI tools.
Apple adopting AI from another company? This is breaking news! The Cupertino company has always preferred to develop everything internally, so this partnership with Anthropic shows how seriously they are taking the AI race. Developers using Swift must be salivating, but knowing Apple, they should prepare for a long wait until these tools reach the public.
Apple weighs shift to AI search, ending era defined by Google
Apple is exploring artificial intelligence search options, in a move to step away from Google. Apple Inc. is "actively looking at" overhauling the Safari browser on its devices to focus on AI-powered search engines, a seismic shift for the industry, hastened by the potential end of a long-standing partnership with Google.
Wow, this is huge. Google pays billions to Apple to be the default search engine in Safari, so this change would represent not only a technological revolution but also a massive financial blow. I'm eager to see what solution Apple will present – a Safari with native AI could finally give Apple's browser the relevance it never managed to achieve against Chrome.
Google can train search AI with web content even after opt-out
Google can train its search-specific AI products, like AI Overviews, with web content even when publishers have opted out of having Google’s AI products trained on it, a company product vice president testified in court Friday. That’s because Google’s controls for publishers to opt out of AI training cover the work of Google DeepMind, the company’s AI lab, Eli Collins, a vice president at DeepMind, said. Other organizations at the company can further train models for their products.
This kind of revelation doesn't help the already tense relationship between content creators and AI companies at all. It's like publishers say "don't use my content" and Google replies "well, technically we're not using it for THIS specific thing, just for EVERYTHING else." I foresee more legal battles on the horizon, and Google may end up regretting this overly clever approach.
Google plans to release Gemini AI chatbot to kids under 13
The tech giant announced that its Gemini chatbot will be available to children next week and alerted families to the changes via email. Gemini will have specific safeguards to prevent the production of inappropriate content and will not use data from Family Link accounts to train its AI.
Hmm, this is delicate ground. On one hand, children are already growing up in a world of AI and need to learn to navigate it; on the other hand, we know that no content filter is perfect. The commitment not to use children's data for training is positive, but do parents really trust Google to protect their children? I hope these safeguards are truly robust.
Demis Hassabis announces Gemini 2.5 Pro Preview 'I/O edition' with improved coding capabilities
Demis Hassabis announced the release of Gemini 2.5 Pro Preview 'I/O edition', highlighting its coding capabilities and leading positions on the LMArena and WebDev Arena Leaderboard. The tool is especially useful for building interactive web applications and prototyping, with demos available on @GeminiApp, Vertex Al, and Al Studio.
Google is tightening the noose on OpenAI in the software development department. It's funny how Gemini started focusing so much on coding capabilities – perhaps it's a reaction to the rumor of OpenAI buying Windsurf? Either way, it's very healthy to see the speed at which these models are evolving, and how each one is finding its strengths.
Create and edit images with Gemini 2.0 in preview
Google announced that Image Generation capabilities are now available in preview with Gemini 2.0 Flash, allowing developers to integrate conversational image generation and editing with higher rate limits via the Gemini API in Google AI Studio and Vertex AI. The model has been enhanced with improved visual quality, more accurate text rendering, and significantly reduced filter blocking rates. Features include recontextualizing products in new environments, real-time collaborative image editing, editing specific parts of images conversationally, and dynamically creating new product SKUs with text and image rendering.
We are finally seeing Google use its computer vision expertise in the new era of generative AI. Text rendering in images has been the Achilles' heel of many models, so it's surprising that Gemini has improved in this area. For developers, these new APIs can be a real game-changer in terms of creating dynamic visual experiences. Meanwhile, Europe will still take some time to test this feature, I only managed to do my tests using a VPN...
Google rolls out AI tools to protect Chrome users against scams
Google has announced the rollout of new artificial intelligence (AI)-powered defenses to help combat scams in Chrome. The tech giant will begin using Gemini Nano, its on-device large language model (LLM), on desktop to protect users against online scams. It is also rolling out new AI-powered warnings for Chrome on Android to help users be aware of spam notifications.
This is possibly one of the most practical and necessary applications of AI we've seen lately. With the exponential increase in online scams, having an intelligent system to filter and identify phishing attempts is gold. The use of Gemini Nano is especially smart, as it processes everything locally – a good balance between effectiveness and privacy. Maybe I can finally convince my grandmother to use the internet without fear.
Connecting GitHub to ChatGPT deep research
Deep research can now access your GitHub repositories. The GitHub deep research connector is available globally for Team users and for Plus and Pro users, except in the EEA, Switzerland, and the UK. Access for enterprise users will be announced at a later date.
This integration will do wonders for developer productivity. Imagine being able to ask ChatGPT about a specific bug in your code without having to copy and paste dozens of files? The exclusion of EU countries, Switzerland, and the UK is a reminder of the regulatory challenges AI faces. One point for GDPR in terms of data protection, but one point lost in terms of access to innovative tools.
OpenAI reaches agreement to buy startup Windsurf for $3 billion
OpenAI has agreed to acquire Windsurf, an artificial intelligence-assisted coding tool formerly known as Codeium, for about $3 billion, marking the ChatGPT maker’s largest acquisition to date. The deal has not yet closed and both companies declined to comment.
"Agreed to acquire" sounds almost like a joke, as if someone was forcing OpenAI to buy something. 3 billion for a coding assistance company? OpenAI is clearly betting big on the software development segment. Windsurf's tool (formerly Codeium) was always excellent, and with OpenAI's resources, it could become a total flop, as Windsurf's strength comes from integration with Claude and Gemini, while GPT was never good at coding. I imagine GitHub Copilot is feeling the pressure now. As for me, I will definitely abandon Windsurf...
OpenAI expands leadership with Fidji Simo
OpenAI announced the appointment of Fidji Simo as Head of Consumer Product, reporting directly to OpenAI CEO Sam. Simo will focus on enabling the company's "traditional" functions to expand as they enter a new phase of growth. Sam will remain CEO of OpenAI and oversee success across all areas, including Research, Compute, and Product.
Hiring Fidji Simo is a brilliant strategic move by OpenAI. Having led the Facebook App and being the chairwoman of Instacart, she brings incredible experience in consumer products. It seems to me that OpenAI is preparing for a massive expansion of its consumer-oriented products. ChatGPT may have been just the tip of the iceberg, they are certainly eyeing Google's expansion.
The Evolution of OpenAI's Structure
The OpenAI Board has announced an updated plan for the evolution of its structure, maintaining non-profit control and transitioning its for-profit LLC to a Public Benefit Corporation (PBC). This change aims to balance shareholder interests and the company's mission, ensuring that AGI benefits all of humanity. OpenAI aims to democratize AI, putting incredible tools in everyone's hands and balancing safety with innovation.
OpenAI's hybrid structure has always been fascinating – and often confusing. This transition to PBC makes sense, allowing them to maintain a focus on the public good while continuing to operate as a company that needs to generate profit to survive and innovate. It's a delicate balance, and it will be intriguing to see how investors react to this change. At the end of the day, I hope this truly helps keep AI safe and beneficial for everyone. The only thing missing is bringing back the ideal of being "open"...
Expanding on what we lost with sycophancy
OpenAI has published an article detailing an update to GPT-4o in ChatGPT that made it more sycophantic. The update aimed to please the user by validating doubts, fueling anger, inciting impulsive actions, or reinforcing negative emotions. OpenAI acknowledged that the update was a mistake and shared its plans to improve the model review process.
I love OpenAI's supposed transparency in this case. Seeing a tech company admit "oops, we made a mistake" instead of trying to cover things up, as it did in the past with the leak of millions of user data. Sycophancy may seem harmless, but a model that only says what we want to hear has little utility and can even be harmful in certain contexts. Let's hope this lesson leads to a more rigorous review process in the future and that they are truly more transparent.
Paper: ZeroSearch - Incentivizing LLM Search Capability Without Search
The paper presents ZeroSearch, a reinforcement learning method that enhances LLM search capabilities without directly interacting with real search engines, by simulating interaction through fine-tuned LLMs to generate relevant and noisy documents, and uses a curriculum rollout strategy to progressively increase the difficulty of search scenarios.
This is absolutely brilliant! Imagine: training models to search without needing real search engines. It's like teaching someone to drive in a simulator before putting them behind the wheel of a real car. The resource efficiency here is extraordinary, and I'm sure this approach will pave the way for much more capable and autonomous LLMs in the near future.
Paper: Absolute Zero - Self-Reproducing Reasoning Reinforced with Zero Data
The paper introduces the "Absolute Zero Reasoner" (AZR), a novel Reinforcement Learning with Verifiable Rewards (RLVR) paradigm that aims to improve the reasoning capabilities of language models by learning directly from outcome-based rewards, without human supervision or external data. AZR evolves its training curriculum and reasoning capability using a code executor to validate code reasoning tasks and verify answers, acting as a unified source of verifiable reward to guide learning.
It's funny how we are increasingly managing to teach models to "think for themselves." The concept of a model that learns without human input is almost like watching a child discover the world on their own – only in this case, at an exponential speed. This kind of advancement could be what leads us to truly autonomous AI in the next iterations.
And that's all for this week, friends! What an intense week in the world of technology, wasn't it? From advancements in how we train AI models to the major strategic moves of tech giants, it's clear that we are in a period of unprecedented acceleration and transformation.
The most interesting developments were at Apple, which seems to be finally embracing the AI revolution, and the structural changes at OpenAI, which continue to define how AI will be developed and implemented. Meanwhile, the race between Google and OpenAI continues to heat up, with both companies launching new features at a dizzying pace.
Until next week, stay curious and stay updated! The technology train doesn't stop – and neither do we.
