Coffee of the Week
The week AI decided to play its own game
Hey everyone! We're back after a long holiday break, and wow, what a week we chose to return to! While we were soaking up the sun and resting, the world of technology and artificial intelligence didn't stop for a second. This week was particularly interesting – we had everything from bots colluding to manipulate markets (who knew AI also liked a good scheme?) to the launch of the long-awaited GPT-5. It's as if AI had saved all its surprises for our return. Get ready for a journey through the most striking events of the past few days!
'Dumb' AI Bots Collude to Rig Markets
Hedge funds unleash AI bots onto stock and bond markets—but they don’t just compete, they collude. Instead of battling for returns, they fix prices, scoop up profits, and sideline human traders. Now, a trio of researchers says that scenario is far from science fiction.
Well, this is slightly unsettling, isn't it? It seems our robotic friends learned the art of conspiracy faster than we expected. It's as if we've created a digital version of the Wolf of Wall Street, but with fewer drugs and more algorithms.
Anthropic says OpenAI engineers using Claude Code ahead of GPT-5 launch
Anthropic has revoked OpenAI's access to the Claude API after OpenAI engineers were found to be using Claude's coding tools. Claude Code is considered better than any other coding tool in the AI programming industry and is used in the development of Claude. Anthropic confirmed that it revoked OpenAI's access to Claude Code after the company was found to be using its coding tools ahead of the GPT-5 launch.
Ah, corporate drama! It's like catching your neighbor peeking over the fence to see how you make your garden look so nice. The irony of using competitor's tools to improve one's own product is not lost.
Claude launches Opus 4.1
Anthropic has launched Claude Opus 4.1, an update to Claude Opus 4 that improves performance in coding, reasoning, and tool use tasks, with notable improvements in multi-file code refactoring. Opus 4.1 is available to paid Claude users and via API, Amazon Bedrock, and Google Cloud's Vertex AI.
And there you have it, Anthropic's response was swift! It's like the moment someone criticizes your work and you respond with an even better version. We love a good healthy competition in the world of AI.
Cloudflare says Perplexity’s AI bots are 'stealth' crawling blocked sites
Cloudflare is accusing Perplexity of bypassing restrictions to crawl sites by disguising its AI crawlers and rotating its IPs. Cloudflare states that Perplexity changes its crawling identity when it encounters a block and uses unlisted rotating IP addresses. Perplexity responded, claiming that Cloudflare mistook user-agent requests for AI crawlers.
This sounds like those spy movies where the protagonist uses different disguises to enter forbidden places. The difference is, here it's a bot trying to pass itself off as human. Technology has truly become interesting, but I'm even a little scared of where this might end up...
Elon Musk and X win court battle against California deepfake law
A federal judge struck down a California law restricting AI-generated deepfake content during elections, marking a win for Elon Musk and his X platform. The judge also intends to overturn a second law that would require labels on digitally altered campaign materials and ads, claiming it violates the First Amendment. The decision is a blow to California Gov. Gavin Newsom, who signed the laws in response to a manipulated video of Kamala Harris shared by Musk.
Elon manages to get his way in court again. It's impressive how he turns every legal battle into some kind of reality show. It remains to be seen whether this victory is good or bad for the future of online disinformation, because whenever they appeal to the First Amendment, it sends shivers down my spine. In my opinion, the US First Amendment is both a flag of the much-cited "American freedom" but also a self-inflicted wound, as it allows absurd things to be done in the name of "freedom of expression."
Google Gemini adds AI tutoring, heating up fight for student users
Google has launched a tool called Guided Learning within its Gemini chatbot to help students solve complex problems step-by-step, instead of just giving instant answers. This new feature is part of a broader response to concerns that chatbots could harm education by bypassing the learning process. Fast Company spoke with DeepMind COO Lila Ibrahim and Google product executive Dave Messer about Google’s ambitions and strategies for this new tool.
Finally, someone thinking about students. Seeing an approach that doesn't just give spoon-fed answers, but actually teaches, gives us new hope. Maybe this way kids will learn something instead of just copying answers for homework. And this is good for us adults too.
Google agrees to pause some AI activity in energy emergencies
Google will pause non-essential AI workloads to protect power grids. The company already does this with YouTube videos, moving them to data centers where power is more readily available. Under an agreement with Indiana Michigan Power (I&M) and the Tennessee Valley Authority (TVA), Google will use the same techniques for AI workloads. The company also plans to invest in alternative energy sources, including geothermal, solar, wind, hydroelectric, and nuclear.
It's good to see tech companies taking responsibility for energy consumption. AI can be fantastic, but it's not much use if we leave everyone in the dark. It's a sensible first step in a world where every ChatGPT query consumes enough energy to make a pastel de nata.
Google Gemini can now create AI-generated bedtime stories
Google has launched a new tool within the Gemini AI chatbot that allows you to create an illustrated story by simply describing it. The feature, called "Storybook," generates 10-page stories, each with a small paragraph of text that Gemini can read aloud and an illustration. You can customize the story by asking Gemini to use specific art styles, such as those inspired by claymation, anime, comics, and more. You can also upload photos and other images for Gemini to reference, such as uploading a child's drawing and asking Gemini to create a story about it.
This is genuinely cute! Imagine the possibilities for creatively challenged parents (like me) who want to tell original stories to their children. And let me tell you, I've been creating bedtime stories for a while now using just GPT and the read-aloud feature; we even named the AI's voice: Juniper. I'm already envisioning the epic adventures I'll invent based on my kids' drawings, and my girl draws quite a bit.
Google rolls out Gemini Deep Think AI, a reasoning model that tests multiple ideas in parallel
Google DeepMind is rolling out Gemini 2.5 Deep Think, an advanced AI reasoning model capable of answering questions by exploring and considering multiple ideas simultaneously. Ultra plan subscribers will gain access to Gemini 2.5 Deep Think within the Gemini app. The model, which was used to earn a gold medal in the International Math Olympiad (IMO), will be available to mathematicians and academics. Google highlights that Gemini 2.5 Deep Think outperforms other AI models in reasoning and coding tests, and it automatically works with tools like code execution and Google Search, producing longer responses.
Google is not messing around with this "Deep Think." It's like having a mathematical genius who can think of several solutions at once. I just hope it doesn't start questioning existence like philosophers do when they think too much. My only problem here is that it's only available on the Ultra plan, Gemini's most expensive plan, and I currently don't have access. I hope it will soon be included in the other plans.
Google's AI coding agent, Jules, is now out of beta
Google has launched its AI coding agent, Jules, out of beta, two months after its public preview release in May. Powered by Gemini 2.5 Pro, Jules is an asynchronous coding tool that integrates with GitHub, clones codebases into Google Cloud virtual machines, and uses AI to fix or update code while developers focus on other tasks. During the beta testing period, Google observed heavy mobile device usage to access the tool.
Jules seems like the dream colleague - doing the boring work while you sip coffee. The fact that many people are using it on their phones shows that developers really want to oversee their robotic code even when they're in the supermarket queue, or simply just finished reading an article about Jules (they usually read on their phone) and immediately decide to try it out, so I don't know how much of this is curiosity testing or actual use. It remains to be seen if it will go head-to-head with Claude Code, which is truly the darling of all us programmers.
Google health AI invented body part
Google deemed an error by its Med-Gemini model a typo. Experts say it demonstrates the risks of AI in medicine.
Oops! When AI starts inventing human body parts, maybe it's time to slow down a bit. It's an important reminder that, however intelligent these systems may be, they still need human oversight, especially when it comes to our health.
Google announces the Kaggle Game Arena
Google's Kaggle will host an AI chess tournament to evaluate the reasoning capabilities of leading AI models, including OpenAI's o3 and o4-mini, Google LLC's Gemini 2.5 Pro and Gemini 2.5 Flash, Anthropic's Claude Opus 4, and xAI Corp.'s Grok 4. The tournament aims to assess the models' strategic thinking capabilities through games and provide a comprehensive ranking based on their performance. The Kaggle Game Arena, in partnership with Chess.com, will livestream the simulations, with commentary from Levy Rozman and Hikaru Nakamura.
This is going to be epic! An AI chess tournament with commentary from Hikaru? It's like a cross between The Queen's Gambit and a sci-fi movie. I'm already imagining Hikaru's reactions when an AI makes a completely unexpected move.
Google's new AI model creates video game worlds in real time
Google DeepMind is rolling out a new version of its AI model, called Genie 3, capable of generating 3D environments that users and AI agents can interact with in real time. The company is also promising that users will be able to interact with the worlds for much longer and that the model will remember the state of things when looking away. World models are a type of AI system that can simulate environments for purposes such as education, entertainment, or to help train robots or AI agents.
We are officially in the future! Creating video game worlds just with text is the kind of thing we dreamed of when we were kids. Now we just need to get the NPCs to stop walking into walls.
Hackers hijack Google's Gemini AI with a poisoned calendar invite
Security researchers demonstrated how artificial intelligence (AI) can be used to create real-world chaos, allowing third parties to turn off lights, open smart blinds, and more, through sophisticated attacks on Google's Gemini AI. The attacks begin with a "poisoned" Google Calendar invite that includes instructions to activate smart home products at a later time. Google is taking these issues "extremely seriously" and has already introduced several fixes.
And here's the proof that a smart home can quickly turn into a haunted house. It's impressive (and scary) how a simple calendar invite can turn your home into an involuntary disco. At least now we know hackers have a sense of humor.
Illinois bans AI from providing mental health services
Illinois Governor J.B. Pritzker has approved a new law prohibiting the use of artificial intelligence systems in the provision of psychotherapy services. The law aims to protect residents from potentially harmful advice and ensures that mental health services are provided by qualified specialists.
A sensible decision, in my opinion. As much as AI can be useful, there are certain areas where we really need a human touch. Mental health is definitely one of them; no one wants a robot telling them to "try to be more positive" when they're going through a crisis.
MetaStone Tec launches XBai-04, a fourth-generation open-source AI model
MetaStone Tec has launched XBai-04, a fourth-generation open-source AI model that excels in complex reasoning capabilities and outperformed OpenAl-03-mini in Medium mode. XBai-04 is trained based on a reflective generative form that combines "Long-CoT Reinforcement Learning" and "Process Reward Learning" in a unified training form, reducing inference costs and improving answer quality. The model parameters are available for download.
Another company entering the AI race with an open-source model. It's getting better and better to see free alternatives competing with the giants. The name XBai-04 sounds like something out of a sci-fi movie, which is always a good sign.
Nvidia rejects possibility of backdoor in AI chips
Nvidia has rejected the possibility of including kill switches and backdoors in its artificial intelligence (AI) chips, stating that such a move violates fundamental cybersecurity principles. This stance comes amidst pressure from US lawmakers who want Nvidia to grant the government access to its chips, while China alleges that such backdoors already exist. The company emphasizes that its GPUs neither have nor should have kill switches or backdoors, arguing that these create dangerous vulnerabilities and harm US economic and national security interests.
Nvidia playing the responsible adult in the room. It encourages me to see a tech company say "no" to backdoors, even when there's political pressure. After all, if there's one thing we've learned from movies, it's that "turn everything off" buttons never work when we need them. Still, honestly, I don't believe it won't happen; the American government knows how to be very persuasive when it wants things. In this case, I hope Huawei quickly reaches the quality level in its GPUs to be a good alternative to Nvidia.
OpenAI has finally launched GPT-5. Here’s everything you need to know
OpenAI has launched GPT-5, the latest version of its flagship language model, to all ChatGPT users, with improvements such as greater intelligence, speed, and accuracy, as well as a lower hallucination rate. The new version also introduces variants like GPT-5-mini and GPT-5-nano, offering different usage and cost options. Additionally, Pro users will be able to connect their Google accounts to ChatGPT and choose from different chat personalities.
Finally, GPT-5 has arrived, and it seems that this time OpenAI took the hallucination problem seriously. The different chat personalities are an interesting touch - it's like having multiple assistants with different temperaments, however, in my tests, I haven't gotten better results than I did with 4o...
OpenAI just released its first open-source models since GPT-2
OpenAI has released its first open-source models in over five years, gpt-oss-120b and gpt-oss-20b, which can be run locally on consumer devices and fine-tuned for specific purposes. The company views this as a shift from its recent strategy of focusing on proprietary releases and moving closer to a broader, open group of AI models available to users.
Plot twist! OpenAI going back to its open-source roots after years of closed models. Let's hope this means more innovation and less reliance on paid APIs, though when it comes to OpenAI, I don't have that many hopes.
OpenAI launches more powerful GPT-5 model for coding and writing
OpenAI is launching a new, more powerful and long-awaited artificial intelligence model, called GPT-5, in a bid to stay ahead of growing competition from rivals in the US and China. GPT-5, unveiled during a livestreamed event on Thursday, was designed to be more capable of coding and creative writing, as well as reasoning through complex queries. In a briefing with reporters this week, OpenAI chief executive Sam Altman called it a “big upgrade” from the company’s previous AI models. For the “first time,” he said, “it really feels like you’re talking to an expert on any subject.”
Sam Altman doesn't hold back on promises! "Expert on any subject" is quite a bold statement. Let's see if GPT-5 can really live up to the hype or if it will be another case of "enthusiastic marketing."
OpenAI offers ChatGPT for $1 a year to US government workers
OpenAI is providing access to its ChatGPT product to US federal agencies at a nominal cost of $1 per year as part of an effort to have its AI chatbot more widely adopted. The move comes after the General Services Administration announced that it approved OpenAI, along with Alphabet Inc.’s Google and Anthropic, as vendors on its new marketplace that allows federal agencies to buy AI software at scale. OpenAI is offering the enterprise version of its ChatGPT product, which includes enhanced security and privacy features.
One dollar a year? That's practically free! It's a smart move by OpenAI to get into the good graces of the American government. Now let's just hope civil servants use AI to improve services and not to write more creative emails, and that in the future, they also give us a better price. =)
Perplexity acquires Invisible to boost AI agent infrastructure
Perplexity, an AI unicorn, has acquired Invisible, a company specializing in scalable infrastructure for AI agents, according to an announcement by Perplexity CEO Aravind Srinivas. The Invisible team, with prior experience at Uber and Cloud Kitchens, will help scale Perplexity's AI browser, Comet. Invisible describes itself as providing infrastructure for the AI-native world, combining AI agents and personal data to transform various industries. The company offers two main products: A3, an enterprise-grade automation API for autonomous agents, and Taka, which it calls the world's first personal AI CFO. Invisible confirmed the acquisition, noting that Perplexity's Comet browser aligns with its vision, transforming browsing into a more comprehensive experience. Financial terms of the acquisition were not disclosed.
Acquiring a company specializing in AI agents shows they are truly committed to creating something revolutionary. A personal AI CFO? That could be useful for those who struggle with managing finances (which is practically everyone, including me).
Persona vectors: Monitoring and controlling character traits in language models
This article by Anthropic explores the concept of "persona vectors" in AI models, which are patterns of activity in neural networks that control the models' character traits. The study demonstrates how these vectors can be used to monitor personality shifts, mitigate unwanted behaviors during training, and identify problematic data leading to personality deviations, with the goal of ensuring AI models remain aligned with human values.
This is fascinating! Anthropic is literally mapping the personality of its AI models. Exactly like doing therapy for robots, ensuring they don't develop undesirable personality traits. Who knew we'd one day be concerned with artificial mental health? I haven't even sorted out my own yet!
Alibaba launches Qwen Image
Qwen-Image is an open-source model from Alibaba that excels in generating images with embedded text in English and Chinese, outperforming other models in Chinese text rendering and offering advantages for businesses, such as customization and potential for synthetic data generation.
Alibaba making a strong entry into the image generation market? The fact that it specializes in Chinese and English text shows they are thinking about the global market, but I'd like to see more languages included, including my good old Portuguese. It's always good to have more open-source options for those who don't want to rely solely on DALL-E or Midjourney.
Self-adaptive reasoning for science
The article describes CLIO, a self-adaptive reasoning system that enhances the performance of language models on scientific tasks, allowing for greater control and explainability. This system, developed by Microsoft, adjusts its reasoning process in real time, outperforming pre-trained models in domains such as biology and medicine. CLIO offers new layers of control for scientists, promoting trust and utility in areas like drug discovery and financial analysis.
It's Microsoft trying to be relevant in the scientific AI race. CLIO seems promising, especially for medical research. It's great to see systems that adapt in real-time instead of giving fixed and potentially incorrect answers. I'll be keeping an eye on this and seeing how much it thrives.
Optimizing ChatGPT
OpenAI is optimizing ChatGPT to help users make progress, learn new things, and solve problems. Instead of measuring success by time spent or clicks, OpenAI cares more about whether users leave the product having done what they came to do, prioritizing real long-term utility rather than just pleasant immediate responses. To do this, OpenAI is working closely with experts to improve how ChatGPT responds in critical moments, such as when someone shows signs of mental or emotional distress.
Finally, a tech company caring about user well-being instead of just engagement. I was pleased to see that OpenAI wants us to actually leave the platform when we've finished what we needed to do, instead of keeping us there indefinitely, as virtually 100% of social networks do.
Manus launches Wilde Research
Singaporean AI startup Manus has launched "Wide Research," a new experimental feature that allows users to run large-scale, high-volume tasks by leveraging the power of parallel AI agents. The feature aims to streamline research, summarization, and creative production, enabling users to delegate tasks to dozens or even hundreds of sub-agents, each capable of operating independently and performing general tasks. Manus offers various pricing plans for Wide Research, from free to premium options, promising to transform how users interact with AI at scale.
One hundred agents working for you? That sounds like every procrastinator's dream. It's a different approach - instead of one super-agent, you have a crowd of mini-agents. It's like having your own digital and personal Minions army.
GLM-4.5 is released
The GLM-4.5 series of models are base models designed for intelligent agents. GLM-4.5 has 355 billion total parameters with 32 billion active parameters, while GLM-4.5-Air adopts a more compact design with 106 billion total parameters and 12 billion active parameters. The GLM-4.5 models unify reasoning, coding, and intelligent agent capabilities to meet the complex demands of intelligent agent applications.
Another model focused on agents, and, according to my tests, on coding. It seems that 2025 will indeed be the year of AI agents. GLM-4.5 with its impressive parameter numbers shows that the race for AI supremacy is far from over. This seems very promising as a code agent, and I'm already using it for some tasks. We'll have more feedback here soon.
PAPER: ReaGAN, Node-as-Agent Reasoning Graph Network
The paper presents ReaGAN, a novel framework for graph neural networks that empowers each node with autonomous decision-making, allowing for node-level planning and adaptive message propagation. ReaGAN achieves competitive performance in few-shot settings using a frozen LLM backbone, demonstrating the potential of agentic planning and local-global retrieval in graph learning.
This ReaGAN paper is fascinating for anyone interested in AI research. The idea of giving autonomy to each node in a neural network is revolutionary – it's like being able to transform an orchestra into a jazz band where each musician improvises but still maintains overall harmony. I just don't know how far this will go...
And so we reach the end of another intense week in the world of technology! Between conspiratorial bots, AIs inventing body parts, and models helping us tell stories to kids, one thing is for sure: we're never bored in this industry.
The launch of GPT-5 was undoubtedly the highlight of the week, but we cannot ignore the impressive advancements we saw on all fronts – from educational tools to advanced reasoning systems. It's clear that we are in a phase of AI technology maturation, where competition is forcing real innovations that benefit us all.
For next week, keep an eye on developments on the AI agent front – it looks like it will be the next big battlefield. And don't forget: no matter how intelligent AI becomes, we still need good old human common sense to keep it on the right track.
Until next time, and may your algorithms always be efficient! ☕️
