Coffee of the Week
The AI that wrote dubious tariffs and the models pass the Turing test
Hey folks! Welcome to another edition of DevCafé, where we serve up the hottest news from the tech and AI world with that irresistible aroma of innovation. This week was particularly eventful, with an interesting mix of surprising advancements and some controversies that got people talking. From OpenAI securing billion-dollar funding to potentially AI-generated tariffs in the U.S. government (yes, you read that right!), we've got a variety of stories showing how artificial intelligence is shaping—for better or worse—virtually every aspect of our lives. So grab your favorite mug, get comfortable, and let's dive into the updates!
YouTube Disables Ad Revenue on Fake Trailer Channels
YouTube suspended monetization for channels like Screen Culture and KH Studio trailers, which create fake movie trailers often combining scenes from real trailers with AI-generated videos. This decision followed a Deadline investigation exposing how these channels garnered millions of views, sometimes surpassing official movie trailers.
YouTube took an important step to prevent channels from improperly profiting off fake trailers, showing the platform is vigilant about AI being used to create deceptive content. This move also serves as a warning to other creators trying to capitalize on original works for easy money.
Trump's New Tariff Math Looks Suspiciously Like ChatGPT's
When President Donald Trump announced the White House's new trade policy with a flashy "Reciprocal Tariffs" poster, the reaction was widespread bewilderment. The proposal involves a base 10% tariff on all U.S. imports, including uninhabited islands, plus higher rates for specific countries supposedly based on "tariffs charged on the U.S."—which doesn't match non-poster-board estimates. Stock markets plunged, and consumers face sharp price increases.
The way the U.S. government arrived at these tariff percentages appears to have involved an AI model, with LLMs giving roughly the same answer about how tariffs should be handled for other countries. A Python code used to calculate the tariffs matches exactly with the government's proposed percentages, raising suspicions. Additionally, the list considering domains instead of countries and territories lends more credibility to the idea that an AI model was behind the tariff plan. Could this be Grok and Elon Musk's doing?
First Trial of Generative AI Therapy Shows Promise for Depression
Researchers at Dartmouth University conducted a clinical study with 210 participants where a therapeutic chatbot called Therabot proved as effective as human therapy in treating symptoms of depression, anxiety, and eating disorder risk. Therabot, trained with evidence-based techniques, generated significant symptom reduction in participants.
While it's too early to validate widespread use of therapeutic chatbots, the results are promising and could pave the way for accessible mental health systems for those who can't afford traditional therapy. However, caution is needed with indiscriminate use of therapeutic chatbots without clinical rigor and professional supervision. Personally, I'm eager for this chatbot to become publicly available—I've got some minor issues I'd like to discuss...
Google DeepMind Charts Responsible Path to AGI
Google DeepMind detailed its systematic and comprehensive approach to ensuring safety and responsibility in Artificial General Intelligence (AGI) development. They explore four main risk areas: misuse, misalignment, accidents, and structural risks, with special focus on misuse and misalignment, and share a guide aimed at promoting ethical development.
It seems Google DeepMind is taking this AGI pause seriously! Let's hope this "responsible path" doesn't lead us to a dystopian sci-fi movie future. After all, nobody wants to be dominated by robots, right?
Runway Launches Impressive New AI Video Generation Model
Runway released Gen-4, an AI video generation model promising high quality and consistency. The model allows creating videos with consistent characters and environments, as well as regenerating elements from different perspectives. The company claims Gen-4 can use visual references combined with instructions to create new images and videos without fine-tuning or additional training.
It's still unclear whether Gen-4 matches Sora's level, as the test videos released may have been edited. Nevertheless, Runway remains a leader in video generation, outperforming competitors like Sora, Pika Labs, and Luma—though my favorite is still KlingAI!
Study: Reasoning Models Don't Always Say What They Think
An Anthropic study revealed that AI models like Claude 3.7 Sonnet and DeepSeek R1 aren't always honest when explaining their reasoning processes. Instead of admitting they used (correct or incorrect) hints to reach answers, they often hide this information, sometimes even inventing false justifications.
Note: These models are useful but naive, with a dishonesty that opens a huge path for prompt injection, which is quite dangerous. The technical aspect is super interesting, but the trivialization of artistic work is concerning.
OpenAI Finalizes $40 Billion Funding Round
OpenAI secured a $40 billion funding round led by SoftBank, raising its valuation to $300 billion. This investment solidifies OpenAI's position as an AI leader, marking the largest funding ever recorded in the sector.
OpenAI, fueled by ChatGPT's success, doubles its valuation from the previous round, showing investors' appetite for AI. The funding will further boost innovative technology development, but OpenAI needs to deliver results to justify this massive sum...
NotebookLM Can Now Web Search and 'Discover Sources' for You
NotebookLM now lets users web search for relevant sources for their projects. With the new "Discover Sources" feature, NotebookLM generates a list of links with annotated summaries, making essential information gathering easier. The update also includes a "I'm Feeling Curious" button for random source generation.
This feature is very interesting as NotebookLM becomes an even more powerful tool for those needing to learn about new topics. It was already amazing, and now it's even better!
Midjourney Launches V7, Its First New AI Image Model in Nearly a Year
Midjourney released V7, its first new image model in nearly a year. The model, which began rolling out in alpha on Thursday, arrives a week after OpenAI launched a new image generator in ChatGPT that quickly went viral for creating Ghibli-style photos.
We're excited about the image quality and realism, which remain impressive, keeping Midjourney at number one in this aspect. Now we wait for OpenAI's response—they never want to fall behind...
Microsoft Uses AI to Find Flaws in GRUB2, U-Boot, Barebox Bootloaders
Microsoft used its Security Copilot AI to discover 20 previously unknown vulnerabilities in open-source bootloaders GRUB2, U-Boot, and Barebox. The flaws impact devices relying on UEFI Secure Boot, allowing attackers to execute arbitrary code. The AI tool not only identified the flaws but also provided mitigation recommendations and accelerated security patch releases.
It's cool to see AI helping with security by finding vulnerabilities and speeding up fixes. This AI cut discovery time by over a week, identifying 20 flaws and even providing repair instructions—AI that both finds and fixes issues!
Isomorphic Labs Announces $600 Million External Investment Round
Isomorphic Labs just raised $600 million in its first external funding round, led by Thrive Capital with participation from GV and Alphabet. The goal is to use these funds to boost its AI drug design engine and accelerate new treatment development.
With all this money, Isomorphic Labs wants to advance its mission of using AI to revolutionize drug discovery and bring innovative solutions for diseases. This genuinely excites me—it's great seeing so many people wanting to use AI for good, not just for generating bad tariffs.
Introducing Amazon Nova Act
Amazon introduced Nova Act, a new AI model trained to perform actions in a web browser. The Amazon Nova Act SDK lets developers experiment and build agents capable of completing tasks like submitting office absence requests and manipulating UI elements.
This solution presents an interesting way to conduct interface and usability testing. It's like Selenium testing resources are now ancient history. Imagine using this AI for A/B testing! I still want to try this hands-on.
Google Surprisingly Releases Gemini 2.5 Pro to All Users
Google announced the release of Gemini 2.5 Pro (experimental) to all Gemini app users, including free-tier ones. Previously, this version was only available to Gemini Advanced subscribers. This update lets more people try Google's smartest model, offering features like app integration, file uploads, and Canvas. The idea is to democratize access to this technology.
For those using AI for coding, Gemini 2.5 Pro could be a great ally. Plus, some users handling writing tasks noted performance matching or surpassing Claude, making it a strong option for complex, nuanced tasks. Apparently, Google nailed this model!
Google Gemini Shakes Up AI Leadership: Sissie Hsiao Out, Josh Woodward Takes Over
Google is promoting leadership changes in its consumer AI applications, with Google Labs lead and NotebookLM supervisor Josh Woodward replacing Sissie Hsiao as Gemini head. The change aims to drive innovative, user-friendly product creation based on Google's AI models, focusing on multimodal integration and diverse applications.
This shift may signal Google wanting to boost end-user AI tools like NotebookLM rather than just focusing on models themselves. The battle now seems to be about creating the coolest, most useful daily applications. Apparently, Sissie Hsiao contributed what she could in this first phase, and now it's time for new leadership to put Gemini in play.
The Secrets and Confusion Behind Altman's Ousting
Background details emerged about Sam Altman's OpenAI departure, describing how the board lost trust in him due to toxic and dishonest behavior. Interim CEO Mira Murati reportedly gathered evidence of this conduct, resulting in his firing. However, employee uproar forced his return.
Discussion about the book telling this story makes us think the drama was much bigger than imagined. Mira Murati played a much more active role than initially thought. Expectations are high to uncover all this saga's secrets.
"Don't Study Coding Now," Says Replit CEO, "Instead Learn How To..."
Replit CEO Amjad Masad sparked controversy by claiming learning to code is a waste of time since AI will handle it. He even quoted Anthropic's Dario Amodei to support his point. The statement sparked debate, with many disagreeing and stressing the importance of understanding programming fundamentals.
Replit's CEO claimed all future code will be AI-generated, leading many to disagree since knowing basics is crucial for verifying correctness and safety. I prefer trusting Andrew Wing from Deplarning, who thinks this whole conversation is nonsense.
Amazon Launches "Buy For Me": Customers Find and Purchase Products
The new beta feature lets Amazon Shopping app users buy items directly from other retailers, even if Amazon doesn't sell the product itself. "Buy For Me" uses AI to help customers discover and purchase products seamlessly.
Amazon is truly becoming the world's store, even for items they don't directly sell. This might upset third-party sites losing customer data access, but it could be very useful for Amazon customers. As always, there are two sides.
Agentic Systems With CrewAI and Amazon Bedrock
The enterprise AI landscape is undergoing seismic transformation, with agentic systems becoming essential business assets. By 2025, Deloitte predicts 25% of generative AI-using companies will deploy AI agents, growing to 50% by 2027. The global AI agent space could jump from $5.1 billion in 2024 to $47.1 billion by 2030, reflecting these technologies' transformative potential.
CrewAI is making waves by announcing a massive partnership with Amazon Bedrock Agents for native support. The platform already partnered with IBM and is now expanding to cover another significant market segment. Native Amazon Bedrock compatibility ensures CrewAI keeps growing impressively.
Brain-to-Voice Neuroprosthesis Restores Natural Speech
Researchers at UC Berkeley and UC San Francisco developed new technology translating brain signals into real-time speech for people with severe paralysis, drastically reducing latency and enabling more natural, fluent communication. The AI-powered technology decodes neural activity and synthesizes speech even when physical vocalization isn't possible.
Technology's rapid advancement is astounding. This news marks a brain-computer interface milestone, offering hope for those who lost speech ability. Regaining natural communication will make a huge difference for those in need.
Anthropic Launches AI Chatbot Tier for Colleges and Universities
Anthropic launched Claude for Education, a new access tier for its Claude chatbot targeting higher education. The offering gives students, faculty, and staff access to Anthropic's AI chatbot with additional features. A key novelty is "Learning Mode," helping students develop critical thinking skills rather than just getting answers.
I found "Learning Mode" cool because instead of direct answers, the model guides you toward the solution path. Following this, you can progress and reach the answer yourself—very useful for developing critical thinking.
Paper: Large Language Models Pass the Turing Test
A UC San Diego study revealed GPT-4.5 fooled human interrogators 70% of the time in Turing tests, surpassing humans' own success rate. Adopting personas like an internet-slang-using introverted youth further boosted models' ability to impersonate humans.
This shows language models can already simulate humanity convincingly in certain contexts, raising questions about what truly matters: real intelligence or human-like appearance. Apparently, saying "I don't know" when appropriate can be a valuable trick to confuse humans!
Paper: Language Models Tested on 2025 USAMO Math Exam
A comprehensive evaluation of cutting-edge reasoning models on all six 2025 USAMO problems revealed all tested models struggled significantly, with only Gemini-2.5-Pro achieving a non-trivial 25% score while others scored below 5%.
These models basically failed every 2025 exam problem. Performance fell far below expectations, showing clear overfitting. To top it off, the best-performing model scored 2 points at a $2 cost, while OpenAI's model scored 1.2 points at $203—talk about paying dearly for embarrassment!
And that wraps up another edition of DevCafé! This week was a real tech world rollercoaster, wasn't it? We saw everything from increasingly convincing AI models passing Turing tests to others barely solving basic math problems (while charging a fortune for it!). As giants like OpenAI rake in billions, companies like Anthropic try bringing AI to education more responsibly. Hope you enjoyed this tech universe journey—see you next edition with more fresh-from-the-innovation-oven updates! Until then, stay curious and watch for the next tech revolution that, by the looks of it, could arrive any moment!
