A Generalist #2 | Manus, Future of Jobs, Gemini, Software Architecture
Tech & AI News
Microsoft Research introduces KBLaM, a new approach to improving LLMs by seamlessly integrating external knowledge. Unlike traditional methods such as fine-tuning or RAG, KBLaM uses a novel "rectangular attention" mechanism to encode knowledge into continuous key-value vector pairs, enabling efficient linear scaling with knowledge base size and dynamic updates without retraining. KBLaM's efficiency is a extraordinary, enabling LLMs to access current information and provide trustworthy answers, especially in domains where accuracy is critical. (Microsoft Research Blog)
You might have heard the buzz about Manus, a new AI helper from China that's getting a lot of attention. It's not your average chatbot – it uses multiple AIs working together to try and actually do tasks for you, like research or finding info online. We took it for a spin, and honestly? It felt a bit like working with a super-smart intern: really promising and good at explaining itself, but also kinda prone to crashing, getting overwhelmed, or needing extra guidance (it even admitted to getting "lazy" on one task!). While it struggled with things like paywalls and system stability, the cool part is you can watch it work and step in. Despite the glitches, Manus shows some serious potential, especially for handling complex research, and it's definitely proof that interesting AI agent tech is coming out of China. (MIT Technology Review)
I tried to write this newsletter with Manus. (See screenshot below) First it created a todo list to follow and followed step by step that todo and created a newsletter.While the outcome was beyond what we could have imagined five years ago, it still wasn't good enough to send to subscribers. If you want to read that newsletter you can read it here.
Google introduced a new top-notch AI model that blew away all LLM benchmarks. They also claimed that they were focusing on the programming capabilities of this model. It can create a video game by producing the executable code from a single line prompt. Most of the top LLM clients/IDEs like Cursor, ThinkBuddy are already integrated with it. (Google)
Mistral AI has released a new model that outperforms most comparable models, called 'Mistral Small 3.1'. Besides all the performance upgrades, the thing that I really liked is that it has Apache 2.0 license. (Mistral AI)
And of course, Ghibli-style (not limited to that) AI image generation by the latest development of OpenAI image generation improvement on 4o model ignited a frenzy across all social platforms. After this viral flow, Sam Altman request a break from the world to get this team a sleep. (OpenAI) We said why not us and generate image for ‘A Generalist’ :)
The Future of Jobs
AI isn’t exactly snatching all our jobs, but it’s definitely shaking things up. At the same time, other big factors—like climate change, demographic shifts, and even geopolitical tensions—are also transforming what kinds of skills and talent will be in high demand. According to the World Economic Forum’s 2025 report, around two-fifths (39%) of our current skill sets could become outdated by 2030, so adapting to new realities sooner rather than later is a smart move.
On the plus side, tech-focused roles—especially in computer science and renewable energy—are growing like crazy. Meanwhile, more routine, clerical jobs have started to shrink as automation and AI handle repetitive tasks. This shift doesn’t mean the end of human work; it just means we’ll need to level up our skills to stay in the game. Reskilling, upskilling, and continuous learning are going to be huge if we want to navigate the fast-changing landscape of future jobs and come out on top. (The Future of Jobs Report 2025)
Tools & Resources
Vector databases are the perfect storage solution for semantic search, enabling AI and LLMs to understand and retrieve information based on meaning. But there are enormous options when you need to choose one of them. Here is a Vector DB Comparison to choose wisely.
Software architecture is more complex than you might think. ByteByteGo shared a roadmap to make it visible at a glance. But if you take a minute to think about what each element means, you will see how deep knowledge is required for each step. I really like the ByteByteGo newsletter and posts, highly recommend.
Book, Article & Videos
Video of the Week: How To Focus On The Right Problems - Y Combinator
Book Recommendation: George S. Clason’s timeless classic, The Richest Man in Babylon, offers financial wisdom through engaging parables set in ancient Babylon, teaching essential money management principles. Its advice on saving, investing, and prudent spending continues to resonate with readers seeking practical steps to build wealth. (Amazon)
Article: Growing up in a home with 500 books would propel a child 3.2 years further in education, on average, than would growing up in an otherwise similar home with few or no books. (ScienceDirect)
Quick Bits
Elon Musk's artificial intelligence company, xAI, has acquired his social media firm X in an all-stock deal valuing X at $33 billion to integrate data and resources for AI development. (Reuters)
Amazon plans to lay off 14,000 managerial positions to save $3.5 billion. (Tech Startups)
OpenAI plans to raise up to $40 billion in a SoftBank-led round, targeting a $300 billion valuation to advance its AI research and expand infrastructure. (Reuters)
Saudi Arabia's Public Investment Fund acquired Pokémon Go and other Niantic titles via its subsidiary Scopely, raising concerns about the handling of the games' extensive user location data. (404 Media)
Last words
Thank you for reading my first newsletter to the end. It may not yet meet the quality I envisioned, but I’ll keep working on it. Your support means a lot—whether through reviewing it or sharing ‘A Generalist’ with your friends. I deeply value your feedback, so please don’t hesitate to reach out to me on X or via email.
Bonus
In this first issue, I'm very excited to give away some gifts. You can get a 40% discount on ThinkBuddy (TB) annual plans with the code AGENERALIST40. TB is my daily AI tool that I can access to most of the LLM models and a lot of extra features. (ThinkBuddy)
Another daily tool for me is 1Password. I use a family subscription so we can easily share some common passwords like streaming services. By using a password manager, I stop using the same password everywhere, which increases my security across devices. My Watchtower score is not at the highest level, but it was terrible when I started using it a few years ago due to compromised passwords. If you consider to use 1Password also here is a
Really enjoyed this read! Thanks for sharing your insights about AI news.