The Age of the Agent

"Progress just means bad things happen faster" - Terry Pratchett.

Welcome to the Blacklynx Brief,

2025 will see the rise of the AI Agent.

The big AI studios are gearing up to release their agent builders. The one we are looking forward the most is OpenAI’s “Operator”.

So what is an AI agent?

An agent is something that can reach a certain goal completely autonomously, without waiting for input from us humans on what to do next. Agents have their own sensors to gather information; they can interface with certain tools; go on a computer and use it like a human would. Another important aspect is that they have a memory and a brain that learns from the experience. They are also self-improving.

You can imagine how we're getting into science-fiction territory quite fast with what I've just described. It's also why OpenAI and Google have made statements that they're reluctant to unleash this unto the public.

Safety is mostly maintained by keeping these agents in silos for the time being. They will just do one specific task.

Some examples of what this could look like.

• Suppose you have a smart home: an AI agent could monitor temperatures inside and outside. See if you're awake or ready to sleep and it would learn your preferences and automatically adjust heating, lighting, security systems and appliances while optimizing energy usage.

• An agent that has the task to make money online. It does market research into what is the most profitable. Create social media accounts , posts online using AI video, starts its own business, rakes in cash.

• A daytrading AI agent. This one would trade on the stock market. Constantly monitoring news and business trends. It would buy and trade all the time , even while you sleep.

• A personal account AI agent. This one automatically does all the backoffice grunt work of filing receipts, doing your accounting, filing your taxes. It monitors your spending behaviours, keeps track of the legislation in your country and gives advice.

• You could use an agent for online dating. Instead of going on 20 dates before you find someone you like. Your AI agent that knows you really well, could go on a 1000 first dates, 100 second dates and 50 third dates with the AI agents of other people. This could all happen in seconds and therefore matching you with someone suitable.

• A customer service agent : a system that handles incoming calls and complaints and calmly resolves the user's issues.

• A personal AI assistant: this agent reads your emails, triages them, responds to them after checking with you, coordinates meetings, etc..

These digital agents are already out there.

Here’s an example of an AI agent : Tesla’s new self-driving software is strictly speaking an AI agent. Its job : getting a car safely from A to B.

Watch this to see the future of self-driving in action. Not available in the EU where we regulate everything to shreds (edit: apparently Tesla full self-driving will be allowed in Switzerland starting in March 2025).

Another example is this account on Twitter/X called Truth Terminal. This account became a cryptocurrency guru. It started even moving markets with its posts. Marc Andreessen, a famous venture capitalist even gave it quite a bit of money for trading cryptocurrencies.

Nobody saw this coming, not even the person who created the agent.

This person had to build in some sort of safeguard to this account, because it started to say some very not politically correct things in order to manipulate a certain cryptocurrency.

It is trained mainly on the dark side of the internet - reddit and 4chan and it is generating some very strange tweets. Nobody knows why.

Maybe this is like the famous AlphaGo Go move. (AlphaGO won by doing something unexpected and outrageous. A 500IQ move that seemed absurd when played but proved genius a few hours later).

To be fair : it’s acting even stranger than Elon Musk on his Twitter account.

You can say a lot about the frontier of AI, but not that it’s not entertaining.

Welcome to the Blacklynx Brief!

Like newsletters?

Here are some newsletters our readers also enjoy:

AI News

• xAI introduced a standalone Grok AI app for iOS, offering access to Grok 2 without requiring an X (Twitter) account. The app supports free and premium tiers, providing features like image generation, text summarization, and real-time search capabilities, including referencing older X posts. This move positions Grok as a direct competitor to ChatGPT and Gemini, signaling xAI’s ambition to climb higher in the AI assistant market.

• Google’s new ‘Daily Listen’ feature in Search Labs creates personalized 5-minute AI podcasts from users’ browsing and search histories. Available in the Google mobile app, the feature provides real-time transcripts and links for further exploration, currently limited to U.S. users in English. As attention spans shrink, AI-generated audio summaries could become a popular way for users to stay informed quickly and conveniently.

• Columbia researchers developed General Expression Transformer (GET), an AI model that predicts gene expression across any human cell type with 94% accuracy. Trained on over 1.3M cells, GET successfully uncovered mechanisms in pediatric leukemia and mapped long-range genetic interactions. This breakthrough could revolutionize genetics research, enabling faster progress in understanding diseases and developing targeted treatments.

• UC Berkeley unveiled Sky-T1-32B-Preview, a fully open-source reasoning model rivaling early versions of OpenAI’s o1, developed at a fraction of the cost. Fine-tuned on Alibaba’s Qwen2.5 model, Sky-T1 was trained in just 19 hours on 8 GPUs for $450, excelling in math and coding tasks. This milestone highlights the potential for smaller labs to innovate in high-level AI using cost-efficient open-source tools.

• OpenAI posted its first robotics hardware job listings, signaling plans to develop general-purpose robots for dynamic real-world applications. The effort, led by Caitlin Kalinowski, marks a return to robotics after OpenAI shuttered its team in 2020. This move suggests OpenAI sees robotics as a critical step toward achieving AGI, integrating physical systems with its cutting-edge AI models.

• The World Economic Forum’s latest report predicts AI will create 11M jobs and displace 9M others by 2030, with roles like AI/ML specialists in high demand. Companies plan to upskill employees and reshape operations around AI, with 86% expecting significant transformations. The findings underscore the need for industries to adapt quickly as AI reshapes global workforces and business strategies.

• OpenAI released a policy blueprint emphasizing U.S. competitiveness, clear regulation, and infrastructure development to maintain AI leadership. Key proposals include AI Economic Zones to link research with industries and unified federal oversight to streamline regulations. With $175B in global capital waiting for AI investments, OpenAI calls for strategic policies to ensure AI’s benefits are equitably shared, as the U.S. prepares for a more tech-driven administration.

• The U.S. announced new tiered export controls for advanced AI chips, granting unrestricted access to close allies while severely limiting countries like China. The measures aim to close loopholes from earlier restrictions, but chipmakers like Nvidia warn of potential harm to U.S. competitiveness. These controls, paired with the incoming administration’s potential policy shifts, could reshape the global AI supply chain.

• Nvidia announced collaborations with leaders like Mayo Clinic, IQVIA, and Illumina to use AI in genomics, drug discovery, and clinical trials. Projects include open-source models for DNA and protein analysis, AI agents to streamline medical research, and new DGX Blackwell systems for analyzing pathology slides. These partnerships could dramatically accelerate innovation in healthcare, tackling bottlenecks in research and diagnosis at an unprecedented scale.

• OpenAI unveiled Tasks, a beta feature allowing users to set reminders and recurring actions like daily updates or periodic searches. The feature integrates with ChatGPT’s chat and a web interface, with up to 10 active tasks managed through a new "4o with scheduled tasks" model. While simple, this move sets the stage for agentic AI capabilities, paving the way for future integrations like automated tool use and enhanced workflows.

• Chinese AI lab Minimax debuted two open-source models, MiniMax-Text-01 and MiniMax-VL-01, featuring Lightning Attention technology for 4M-token context windows. The models outperform all open-source competitors on long-context tasks and rival closed systems on benchmarks, with API access priced competitively. These advancements could revolutionize AI agents and applications requiring extensive memory and contextual understanding.

• Microsoft introduced Magnetic-One, an orchestration system within its updated AutoGen v0.4 framework, enabling multiple specialized AI agents to collaborate on complex tasks. Agents like WebSurfer for navigation and Coder for programming communicate asynchronously through an event-driven system, with tools for low-code development and monitoring. This advancement brings us closer to personal multi-agent teams capable of handling sophisticated workflows, transforming productivity and automation.

• François Chollet, creator of Keras, teamed up with Zapier's Mike Knoop to launch Ndea, a lab focused on achieving AGI through a novel combination of deep learning and program synthesis. The lab challenges the data-hungry approach of mainstream AI, emphasizing adaptable, human-level learning efficiency. With Chollet’s new ARC Prize Foundation and Ndea's ambitious goals, the race to AGI is diversifying as top researchers explore alternative paths.

• Microsoft launched Microsoft 365 Copilot Chat, a free version of its AI assistant platform offering GPT-4o-powered chat, file analysis, and task automation. Users can access custom AI agents on a pay-per-use model, with options like $0.01 per message or $200 for 25,000 monthly messages. This move bridges the gap between free users and enterprise subscriptions, reinforcing Microsoft’s competitive edge in workplace AI.

• Ray 2, Luma Labs’ latest AI video model, generates ultra-realistic 10-second clips with improved motion and physics accuracy, handling text, image, and video-to-video prompts. First available to Dream Machine subscribers, Ray 2’s upcoming editing capabilities position it to rival Veo 2 in pushing AI video realism to new heights. The race for longer, coherent AI video outputs is heating up as labs compete to dominate creative AI tools.

Quickfire News

• OpenAI began rolling out a new Custom Instructions interface for ChatGPT, allowing users to provide more information to the assistant and customize responses through new "traits" fields.

• Microsoft introduced rStar-Math, a technique enabling small language models to achieve 90% accuracy on advanced mathematics benchmarks despite their compact size.

• Alibaba unveiled a web interface for its Qwen language models, including the Qwen2.5-Plus flagship model and specialized versions for vision, reasoning, and coding tasks.

• Cohere launched North, an enterprise AI platform based on the Command R model, featuring custom assistants, search functionality, content generation, and more.

• Hailuo AI released S2V-01, a video model capable of maintaining consistent character appearances across video sequences using only a single reference image.

• ByteDance introduced STAR, a video upscaling tool leveraging text-to-video AI to deliver state-of-the-art clarity in generated outputs.

• Elon Musk stated during an X interview at CES 2025 that AI has exhausted the sum of human training data, leading companies to rely on AI-generated synthetic data for training future models despite its limitations.

• Nvidia introduced an AI Blueprint for retail shopping assistants, enabling digital agents to process text and image queries while visualizing products in virtual spaces.

• AMD researchers published Agent Laboratory, a framework for using LLM agents as research assistants for literature reviews, experiments, and report writing, achieving an 84% cost reduction over previous methods.

• Mark Zuckerberg revealed on The Joe Rogan Experience that Meta and other companies plan to use AI to automate midlevel engineering roles this year, with a long-term goal of offloading all coding to AI.

• Savannah Feder launched Astral, an automated AI marketing platform that can handle social media tasks like commenting and content creation on platforms such as Reddit.

• A report from Bloomberg Intelligence predicted that over 200,000 Wall Street jobs could be cut in the next 3-5 years due to AI, with automation boosting banking profits by up to 17%.

• Mistral released Codestral 25.01, a lightweight coding model that performs well across 80+ programming languages and debuted tied for first place on the Copilot Arena leaderboard.

• Researchers at MBZUAI unveiled LlamaV-o1, a multimodal model specializing in step-by-step visual reasoning, achieving state-of-the-art open-source performance with a 67.3% average score across benchmarks.

• Google Cloud introduced the Automotive AI Agent powered by Gemini, which Mercedes-Benz is integrating into its MBUX Virtual Assistant to support complex conversations and multimodal in-vehicle reasoning.

• Major AI companies, including OpenAI and Google, are reportedly paying content creators up to $4 per minute for exclusive, unused video footage to train AI models, creating a new revenue stream for YouTubers and influencers.

• Microsoft announced its new CoreAI division, aimed at unifying its AI platform and developer tools while accelerating the development of the Copilot stack and agentic applications.

• President Joe Biden signed an executive order allowing AI companies to build data centers on Department of Defense and Department of Energy sites to bolster domestic AI infrastructure.

• Amazon is reportedly working to transform Alexa into an AI-powered agent, but development is hindered by issues like hallucinations and slow response times, which could harm the brand if released prematurely.

• The Mayo Clinic partnered with Microsoft and Cerebras to develop AI foundational models for medical imaging and genomic data analysis, aiming to advance personalized medicine with faster diagnostics and precise treatment predictions.

• OpenAI added GIP chairman and Blackrock executive Adebayo Ogunlesi to its Board of Directors, leveraging his expertise in infrastructure investment and global finance.

• French AI startup Bioptimus raised $41 million to develop a "GPT for biology" foundation model capable of simulating biological systems and predicting disease outcomes.

• Microsoft announced a multiyear partnership with Pearson to deliver AI-powered learning and workforce development solutions, including global skilling certifications.

• Sakana AI unveiled Transformer2, a self-adaptive language model that dynamically adjusts neural pathways for different tasks, achieving superior performance with fewer parameters compared to traditional models.

• Google partnered with the Associated Press to integrate real-time news into its Gemini AI assistant, marking its first AI collaboration with a news publisher.

• AI video platform Synthesia secured $180 million in a Series D funding round, raising the Nvidia-backed company’s valuation to $2.1 billion.

• OpenAI expanded its partnership with Axios, funding four new local newsrooms equipped with ChatGPT and other AI tools.

• Cisco announced AI Defense, a safety and security platform designed to prevent unauthorized AI tampering and data leakage through network-level protection and automated safety protocols.

Closing Thoughts

That’s it for us this week.

If you find any value from this newsletter, please pay it forward !

Thank you for being here !