Coffee of the Week
Between Turing winners, autonomous agents and the secret fear of the Terminator
DevCafé: Between Turing Winners, Autonomous Agents, and the Secret Fear of the Terminator
Hey folks! Another busy week in the world of AI, huh? While some Turing Award winners are pulling companies' ears for rushing launches without proper safety, other companies are still speeding ahead as if there's no tomorrow. We've got it all in this edition: from GPT-4.5 showing it's still the golden boy of the rankings (even though it's more expensive than its predecessor), to Amazon wanting to replace voice actors with artificial voices on Prime Video. Let's dive into this sea of news and see what's really worth keeping an eye on!
Turing Award Winners Warn About AI Without Security Features
Turing Award winners Andrew Barto and Richard Sutton have expressed concern about the launch of AI without due safety precautions, highlighting the need for responsible engineering practices and warning against prioritising commercial incentives over proper research.
The community is taking advantage of this moment in the spotlight to warn about security, which ultimately reinforces the chorus of many in the market who say "OpenAI, hold on a minute there, mate!". It's always good to remember that if people worry a little more about security, the world will thank them, right?
[Original source](https://www.ft.com/content/d8f85d40-2c5b-4a2b-b113-87fa8e30f61b?accessToken=zwAGL5bK2KcgkdPY-F1ALFtKK9OxE4f6jjD2Gw.MEUCIQDv1d62eku-I-wOQED-wpYKpg1 wMnOp7dlHkKO3kYWUQAIgEOMatr2owW_rXEHj2OIEUI0Hu0nhpP-xbpXjcwFbkkM&sharetype=gift&token=d5ad7c95-7b3a-4877-a706-af15da483708)
GPT-4.5: An Expensive Evolution, But Still on Top?
The discussion revolves around GPT-4.5, with divergent opinions on its performance and cost-effectiveness. Initially, there was scepticism about its evolution, but LM Arena tests placed it at the top of the rankings, showing better-than-expected performance.
Despite not being a revolution, GPT-4.5 has proven to be better than imagined, with remarkable performance in various categories. The problem is that it's more expensive than its predecessors, and that can be a problem. But the reality is that, in the market, people have quickly adopted the technology.
Salesforce Launches Agentforce 2dx, Letting AI Run Autonomously in Enterprise Systems
Salesforce has just announced Agentforce 2dx, a stonking update to its digital work platform that allows autonomous AI agents to work behind the scenes in enterprise systems without that constant human supervision. The idea is that these AI agents can predict needs, monitor data changes, and act on their own in any business process.
This news marks a significant evolution, moving from purely reactive AI interactions to proactive agents that just get on with it themselves. Salesforce is focused on creating a "multi-agent framework" where personal AI assistants interact with business agents to complete tasks. It's a future where AI negotiates for you!
Amazon Will Use AI in Dubbing on Prime Video
Amazon is starting an AI dubbing pilot programme for select films and series, aiming to make content more accessible. Initially, AI dubbing will be available in English and Latin American Spanish for 12 licensed films and series.
Despite Amazon promising a collaboration between "localisation professionals" and AI, replacing voice actors with artificial intelligence is controversial. The voices generated by AI have been improving, but it remains to be seen whether the public will approve the delivery of content by something "close to human" instead of the real thing. Amazon is following in the footsteps of YouTube and Meta, which are also testing automatic dubbing.
OpenAI Plans to Bring Sora Video Generator to ChatGPT
OpenAI intends to integrate its AI video generation tool, Sora, directly into ChatGPT, as announced by company leaders in a Discord session. The news aims to attract more users to ChatGPT and its premium subscription plans, which may offer higher video generation limits.
OpenAI's initiative could democratise the creation of videos with AI, but it also raises questions about ease of use and the impact on the user experience in ChatGPT. It remains to be seen whether the integration will be successful and whether it will attract a wider audience to OpenAI's AI tools.
OpenAI Launches $50 Million Grant Programme to Boost Academic Research
OpenAI has announced the launch of NextGenAI, a consortium of 15 universities, including Harvard, Oxford, and MIT, which will receive $50 million in research grants, computing, and API access. The programme aims to support the next generation of discoveries and prepare future AI leaders.
Although any extra funding is welcome, it would be even better if the amount were higher. Anyway, it's a positive step towards bringing OpenAI closer to the academic market, which is responsible for major technological advances, often made almost "on the fly".
Ilya Sutskever Will Investigate New Path for AI
Ilya Sutskever's startup Safe Superintelligence (SSI), former chief scientist at OpenAI, is valued at $30 billion, even without a product on the market. Investors are eyeing Sutskever's reputation and the promise of safe superintelligence.
The impressive valuation (about half of Antropic) demonstrates market confidence. This value makes sense when we consider that Sutskever is one of the brightest minds in the industry, and many believe that he may have actually found a promising new path for AI.
NotaGen: Advancing Musicality in Symbolic Music Generation with LLM Training Paradigms
NotaGen is a symbolic music generation model that aims to explore the potential of producing high-quality classical scores, inspired by LLMs. It adopts paradigms of pre-training, fine-tuning, and reinforcement learning and demonstrates the effectiveness of CLaMP-DPO in generating symbolic music with different architectures and encoding schemes.
A very interesting initiative that promises to give a new face to classical music generated by AI, with the potential to create compositions that rival human ones.
Mistral OCR: Mistral AI's New Tool to Boost Your Documents!
Mistral AI has launched Mistral OCR, an Optical Character Recognition (OCR) API that promises to revolutionise the way machines understand documents. According to the company, the tool goes beyond simply identifying text, understanding elements such as images, tables, and even mathematical equations, all with impressive accuracy.
Despite the hype, it remains to be seen whether Mistral OCR will actually deliver on all its promises. After all, not all that glitters is gold, and we know that exaggerated promises are the main course in the world of AI. But, if the tool is as good as they say, it could be a game-changer for those who have to deal with large volumes of complex documents.
Microsoft Unveils New Voice-Activated AI Assistant for Doctors
Microsoft has launched Dragon Copilot, an AI assistant that combines transcription and ambient listening capabilities to assist doctors in documentation and optimise patient care time. The tool promises to facilitate the extraction of medical information and the creation of clinical notes, referral letters, and post-consultation summaries.
This is sensational news that makes perfect sense: having a voice AI assistant aimed at the medical market, where details make all the difference. It remains to be seen whether doctors will actually adopt the tool, as there is some resistance from the medical field to the use of AI.
Microsoft Unveils AI Agents for Sales, Striking Back at Salesforce
Microsoft has announced two AI agents for sales, Sales Agent and Sales Chat, as a competitive response to Salesforce, integrating with Dynamics 365 and offering task automation and insights for sales representatives. The agents aim to optimise customer interaction, qualifying leads and personalising responses based on data from various sources.
Several companies are entering the CRM market with AI, such as Salesforce, and now we see large companies selling AI SaaS as a well-made product. It is a reflection that the AI market has matured a lot and that we are at this level now, where AI is being seen as a well-defined and organised product.
How Sevilla FC Is Discovering Future Stars With Llama
Sevilla FC, famous for its Europa League achievements, is using AI to revolutionise the search for new talent. With the help of the Llama 3.1 70B model, they created Scout Advisor, a tool that analyses scout reports and finds players with specific characteristics, such as a fast and skilled left winger.
It is fascinating to see how the Spanish club is optimising the work of scouts with the help of AI to find promising athletes. With so much data available on players, AI can identify talents that would have previously gone unnoticed, revolutionising the way clubs assemble their squads.
Sergey Brin Asks for In-Person Work in Favour of AGI
Google co-founder Sergey Brin sent a not-so-subtle message to the folks: want general artificial intelligence? Then let's grind in person 60 hours a week! The justification? Accelerate work by using AI to code and thus achieve that AGI. The internal memo leaked and caused a stir, with people recalling the times of the Soviet Union and questioning this workaholic vibe.
Comments like these make many people think of the hymn of the Soviet Union. And that is independent of political positioning - it is simply a questionable approach. Google had already sounded an alarm about a year ago, and it seems that now they are indeed tightening the pace to try to accelerate the development of AGI.
Google Gemini Gains Visual Search
Google Gemini now lets users ask questions using videos and what's on screen, with new features rolling out to Gemini Advanced users on the Google One Al Premium plan on Android later this month.
At Mobile World Congress (MWC) 2025 in Barcelona, the company showed off a new "Screenshare" feature, which lets users share what's on their phone screens with Gemini and ask questions about the content. As an example, the company demonstrated a user shopping for wide-legged pants and asking Gemini what other clothing would pair well. As for the video search feature, Google had already teased it at Google I/O last year, allowing you to record a video and ask Gemini questions about it while you're filming.
Google Expands AI Overviews and Introduces AI Mode
Google is testing "AI Mode" in Search, which uses Gemini 2.0 to answer complex questions with advanced reasoning and web results. This feature is available to Google One Al Premium subscribers in the US.
Google is clearly chasing after Perplexity with this feature, promising that this new tool will function as an in-depth summary and report, exploring more sites than traditional search would normally consult. In other words, they are trying to replicate exactly what Perplexity already does quite successfully.
DuckDuckGo Is Improving Its AI Search Tool, But Will Still Let You Abandon It
DuckDuckGo is turbocharging its search tool with AI, but without forcing it on you: you can still choose how and when to see the AI-boosted responses, which now come from across the web, not just Wikipedia. And there's more: DuckDuckGo will soon integrate web search within its AI chatbot.
DuckDuckGo will not jump on the bandwagon of creating a separate app for its AI chatbot, like OpenAI, Google, and Microsoft. They think the future is mixing everything into one app, with AI powering both search and chat. After all, some questions are better to start with a chat, others with a search, and sometimes you want to jump from one to the other.
Deutsche Telekom and ElevenLabs Form Strategic Partnership to Boost AI-Powered Podcasting on Magenta App
Deutsche Telekom and ElevenLabs have joined forces to bring AI audio experiences to millions of users, integrating human-like AI voices into Magenta AI. This collaboration will allow users to convert news articles into immersive, high-quality podcasts in seconds, in addition to generating personalised podcast content based on predefined prompts.
The initiative is particularly relevant as it comes from a German operator, a place where innovation is not always so evident. The collaboration with ElevenLabs, known for its voice cloning tools, is an interesting step towards converting news articles into podcasts, following a trend already explored by Google. The partnership is another example of how companies are looking for new ways to integrate AI into their services.
Sesame is a New Realistic Conversational Generative AI
Sesame has launched Maia, a conversational voice AI that impresses with its low latency, understanding of what is spoken, and recording of information. Maia's voice is natural and pleasant, creating a fluid and even "flirty" conversation experience, which raises ethical questions about the future of relationships between humans and AIs.
Maia is an impressive leap in voice technology, but it also makes us think about the implications of increasingly "humanised" AIs. Are we approaching a "Her"-style future, where we will have relationships with AIs? And how do we ensure that these technologies are used ethically and responsibly? 🤔
Meta Expands Voice-Powered AI with Llama 4
Meta is investing heavily in AI and plans to introduce voice improvements with Llama 4. The company believes that the future of AI agents will be conversational, not text-based, with a focus on natural voice interactions and expanded capabilities beyond social media platforms.
Meta is jumping on the voice assistant bandwagon, seeking to commercialise AI technology for various applications. It seems that the conversation now is about who will have the most natural and useful assistant.
Anthropic Closes Series E at USD61.5B Post-Money Valuation
Anthropic has raised $3.5 billion at a post-money valuation of $61.5 billion, with the goal of advancing the development of next-generation AI systems and expanding its computational capacity.
Anthropic has tripled its market value in just one year, an impressive growth that shows investor confidence. The company is betting on alignment and interpretability, areas in which it has demonstrated great competence. The new contribution should be directed towards international expansion and training increasingly powerful models.
Anthropic Partners with U.S. National Labs for First 1,000 Scientist AI Jam
Anthropic is joining forces with the U.S. National Laboratories for the first Al Jam with 1,000 scientists, with the aim of accelerating scientific discoveries and technological development through cutting-edge AI models. Scientists from various laboratories will explore Claude's capabilities in scientific tasks, from understanding problems to analysing results, driving scientific progress and strengthening American leadership in AI.
This partnership demonstrates the transformative potential of AI and reinforces the importance of ensuring the safety and alignment of these technologies. If Anthropic and the National Laboratories manage to create a rigorous testing environment, we will certainly see significant advances in various areas.
AWS Creates Working Group to Investigate Agents
Amazon has created a new group within AWS (Amazon Web Services) focused on "Agentic AI" to help users automate their lives. The team will be led by Swami Sivasubramanian, and the expectation is that this new area will become a billion-dollar business for AWS.
Amazon is betting heavily on AI, and it seems that "Agentic AI" is the new wave. Let's see if this initiative really takes off and if it will be another hype or a true transformation.
Not a Programmer? With A.I., Just Having an Idea May Be Enough.
Vibecoding, a term popularised by AI researcher Andrej Karpathy, describes how current AI tools allow even non-technical amateurs to build fully functional applications and websites simply by writing prompts in a text box. It's an accessible and creative approach to software development, where AI assists in coding, allowing users to turn their ideas into reality without requiring deep technical knowledge.
This trend reflects an interesting shift in how we think about programming and software creation. The idea that anyone can create their own tools with the help of AI is exciting, although it also raises questions about the future of traditional programmers. Opinions on the impact of this trend are divided: some see it as democratisation, others as a threat to specialised technical skills.
70% of AI Users Greet ChatGPT and it's Not Necessarily Out of Politeness, But Because of the Terminator
Artificial intelligence is increasingly present in our daily lives, and a recent study revealed that 70% of AI users are polite to chatbots, not necessarily out of politeness, but out of fear of a future machine rebellion, as in science fiction films.
The result of the survey is surprising, but it makes sense. After all, who has never felt a little strange asking a robot "please"? Perhaps, deep down, we are all a little afraid that one day they will remember how we treat them! 👀
OpenAI Participates in 1,000 Scientist Jam Session
OpenAI has joined nine U.S. national laboratories for an unprecedented event that brings together more than a thousand scientists, with the goal of boosting scientific discoveries using cutting-edge AI models, such as 03-mini. The idea is to accelerate technological progress in areas such as health, energy, and security, in addition to strengthening American leadership in AI.
Like Anthropic, OpenAI also participates! The initiative is super cool, but there are security concerns. After all, launching AI without due precautions can cause problems. The ideal is to find a balance between innovation and responsibility.
Paper: Agentic Deep Graph Reasoning Yields Self-Organizing Knowledge Networks
A new framework allows machines to organise and expand knowledge on their own, like a cybernetic autodidact. The system uses graphs to connect information and discover hidden patterns, such as in scientific materials, giving hints of where to look for new materials.
This self-learning AI technology is promising, but it remains to be seen whether it will be able to make connections that humans cannot, or whether it will just be another advanced research tool. If it works as expected, it can provide really valuable insights in various scientific fields.
And so we close another week full of news in the universe of artificial intelligence. From million-dollar prizes for academic research to systems that promise to find the next Messi using AI. While some companies are betting all the chips on autonomous agents, others are more concerned with making you say "please" and "thank you" to ChatGPT (apparently, 70% of us already do it for fear of a Terminator, believe it or not?). Anyway, one thing is certain: the pace of innovation shows no signs of slowing down and, next week, we will be here again to see what else this technological rollercoaster has to offer us. Until next time!