AI Weekly Trends Highly Opinionated Signals from the Week [CY26W4]
🔗 Learn more about me, my work, and how to connect: maeste.it – personal bio, projects, and social links.
Starting today, I’m changing the way I write this newsletter. Those who know me well know that I’m restless in my comfort zone, and when something starts to become routine, I like to pause, look back, and question myself. And spurred on by my colleagues from the podcast, I decided to revise my creative process for the newsletter to make it a bit more human-driven or human-centric. This doesn’t mean I won’t use AI in my creative process anymore, but I’ll use it differently, because today excluding AI from your intellectual production is at least unreasonable (Antirez docet, see the chapter on AI assisted coding), but the use you make of it is as improving of the final result as it is conscious and controlled.
But before telling you how I changed this, let me comment on the cover image (which this time is not generated, but no promises on this). The book you see there, next to a mountain of peanuts, is Agentic Design Pattern by Antonio Gulli. As I said in my LinkedIn post from which I took the image, everyone has their own metrics for measuring the appreciation of a book... mine are the number of peanuts I eat while reading... and those in the photo are just a small part. Sincere advice, if you’re designing an agent system or if you just want to use them, buy it, read it and keep it on your nightstand.
Back to the newsletter. The creation process until now was more or less this:
Collecting links and reading articles
Organizing the most significant links
Adding my comments for each section in the form of a list of bullet points
AI... write (with a format and style prepared in advance)
Then yesterday, stimulated by Paolo and Alessio’s comments, I stopped to ask myself some questions (3 whys technique):
Why did I start writing a newsletter? To share the things I read during the week... ok... but mostly because I’ve always liked writing and writing is my way of reprocessing and memorizing information. But actually I’m not writing anymore...
Why did I feel the need to pass my points to AI and have it write text for others? Two reasons, because I have little time, and because I want a longer and more discursive text. But this presupposes that my readers have more time than me... hmmm maybe it’s neither true nor paid for.
Why can another newsletter make sense? To better elaborate my thinking and share it with others. And we’re back to one of the reasons for writing, elaborating my thinking and adding my point of view. But does a list of bullet points reprocessed by AI do this? Only in part.
So I inverted the parts, letting AI act as an editorial office (cleaning links, synthesizing texts that I had already read anyway) and proofreader, leaving me with the honor and burden of connecting the dots and creating the text.
What you’ll find in the chapters below has a slightly different format from previous newsletters (actually resembling the first issues a bit more), where the only part written by AI is the list of links and their summary. The text of “What’s happening this week?”, Takeaways and Action Items are all mine, with AI acting as editor and editorial office to check correctness and coherence.
Another small change, a more rigorous selection of provided links, reducing their number and trying to transfer the most significant ones. Leave me a comment with your sincere feedback on what you think.
I’m sure someone won’t believe it... given that Paolo and Alessio accused me of writing “too much with AI” even the LinkedIn post commenting on the book above... and I swear I wrote it all by hand, even after one too many beers :)
If you’re curious about how I set up this process with AI becoming a real co-worker with the use of a CLI and specific skills, leave that in a comment too and I’ll tell you about it in one of the next newsletter issues... or maybe in a special edition ;)
🧠 News and research in AI models
Takeaways for AI Engineers
Takeaway 1: Research on reasoning models is anything but stopped and we’re discovering new ways to make models think
Takeaway 2: Thinking in images is a powerful tool for humans, it will become so for models and robots
Takeaway 3: Tools like Cowork will change the way we work in front of a screen
Action Items:
Keep an eye on research on arxiv.org
Try implementing workflows (not code) with Claude Code or Claude Cowork
What’s happening this week?
The paper of the first link they propose (DASD) has immediate impacts and others less evident. On these latter I want to focus for a moment to give you an overview of where research in this field is going. The paper speaks among other things of the overthinking problem that substantially makes chain of thoughts much longer. DASD is precisely (one of) the proposed solution to this problem. Why is it fundamental? Because shorter chains allow for equal context to explore multiple solutions (multiple chains) and substantially obtain better results.
The second link is again a paper, which explores visual reasoning. Those who follow me know that I’ve often talked about it both here in the newsletter and in the podcast, because I consider it one of the next significant evolutions in reasoning models. If you think how many times you’ve made a sketch or a diagram to reason better on a problem? Here it is now that models like Nanoban and others are very good at creating these sketches or diagrams, why not use the multimodal capabilities of interpreting complex images also in the reasoning phase? We did the same for tools, which initially didn’t enter reasoning (but now do almost all from ChatGPT 5.x, to Gemini 3 passing through Chinese models like Minimax and GLM) with excellent results. I personally expect great things from visual reasoning.
Finally a novelty coming from Anthropic that could radically change the way we work on our PCs, Cowork does what Claude Code already does for the more technical and promises to load dynamic contexts for any type of document... like Claude Code does with code (and more). My suggestion is not to lose sight of what’s happening in the Claude world, because it will soon arrive on your desk. And if you trust me, you already have all this in Claude Code, if only you want to “get your hands dirty” using a terminal.
This week’s links
Distribution-Aligned Sequence Distillation (DASD)
New distillation pipeline for compact reasoning models that addresses overthinking through selective distillation on CoT tokens. DASD-4B-Thinking is completely open-source and surpasses much larger models.
Render-of-Thought: Visualizing Reasoning Chains
Framework that converts textual reasoning steps into visual representations using VLM, achieving 3-4x compression while maintaining competitive performance.
Anthropic works on Knowledge Bases for Claude Cowork
Anthropic prepares a fundamental update to Claude with Cowork mode and Knowledge Bases: persistent repositories consultable for relevant context, incrementally updatable.
🤖 Agentic AI
Takeaways for AI Engineers
Takeaway 1: Google’s personal intelligence can further change the landscape of agents. The fact that it proactively draws from resources is the fundamental change, to keep in mind also for your developments. No more RAG (pull) but context push.
Takeaway 2: Security is always a priority. Even more so with GenAI, both because natural language seems more innocuous, and because it makes even non-programmers feel capable of programming
Takeaway 3: Evals as foundation of Agentic AI: Evals are both development tools (Eval Driven Development) and continuous monitoring systems in production, necessary to manage the indeterminism of agentic AI systems
Action Items:
Try the Claude plugin for Chrome to experience agentic browsing and compare your experiences with future Gemini-Chrome integrations
Implement a complete evals strategy: Eval Driven Development in development + continuous monitoring in production for your agents
What’s happening this week?
Last week we talked about Google’s personal intelligence (and I also did so in the podcast) that brings Gemini’s intelligence to proactively search for information in your documents, photos, commitments, places visited... in short in all that data that every day you give (rather we give) to Google. The potentiality is enormous, even if as Peter Parker’s uncle said “with great power comes great responsibility”. Google announced this week that this omnipotence will not be given only to the Gemini app, but also to AI mode in normal Google Search. Keep this in mind. Parallelly there’s a lot of talk, but nothing has been seen yet, of a Gemini integration in Chrome. Having tried (and using more and more often) Claude’s plugin for Chrome, I expect great things from that integration, because having an agent act on one of the open pages, with a license to navigate for you is truly mindblowing. If you can try it and let me know what you think.
But as we said there are great responsibilities and security is one of these. One of the selected articles speaks specifically about the security of skills which being pure text can seem harmless and instead can hide worms that exploit precisely the intelligence of models to be even more dangerous and harmful. Another article speaks instead more generally about the need for a runtime (operating system if you want to push further) for agents. An idea like this, present in numerous research papers and articles in recent months, has quite a few impacts also on security. I will speak precisely about security related to Generative AI, in a panel at Voxxed Day Ticino in early February, maybe we’ll meet there.
Evals in generative AI, especially in the world of agents are a fundamental resource that every AI Engineer should know well. The article I propose this week focuses on their use to monitor possible causes that make performance degrade (intended as goodness of answers and not as speed). But evals are also fundamental in the development phase and more and more often we hear about evals driven development, as an evolution in the world of agents of test driven development. And if you think about it, it can only be so: in a world where software has indeterministic results and evaluated by percentage of success, even tests must be replaced with something that follows the same philosophy. Evals precisely.
This week’s links
AI Engineering Has a Runtime Problem
57% of companies run AI agents in production, but a standardized runtime is missing to manage state, streaming, isolation and scaling.
Monitoring in production generates billions of monthly labels, offering real-world insights superior to static evals.
Supply-chain risk of agentic AI - infecting infrastructures via skill worms
“Skill worms” are an emerging risk: malicious code can spread through the AI skills supply chain.
Google brings Personal Intelligence to AI Mode in Search
Google introduces “Personal Intelligence” in AI Mode in Search, allowing access to Gmail and Photos for personalized responses.
💻 AI Assisted Coding
Takeaways for AI Engineers
Takeaway 1: Pragmatism wins over prejudice: Even a creator of Redis skeptical like Antirez recognizes that AI-assisted coding is now too effective to ignore. AI doesn’t replace the expert programmer, but empowers them (Antirez++).
Takeaway 2: From copilots to agents: The next step after code completion is agentic coding with context, planning and ability to manage complex projects. Claude Code 2.0 and tools like Devin represent this evolution.
Takeaway 3: The new bottlenecks: With faster coding, problems shift to specifications, task management, code review and security. Tools like Devin Review and Tasks in Claude Code address these aspects.
Action Items:
Read Antirez’s article and try to make the leap from code completion to agentic coding with Claude Code or similar tools
Implement AI-assisted code review processes (Devin Review or similar) and task management systems (backlog.md, Tasks in Claude Code) to manage the accelerated flow
What’s happening this week?
I can only start with the article by Salvatore Sanfilippo (Antirez) that struck me a lot. In the past and also in the article itself, Antirez (creator of Redis, who certainly needs no introduction), doesn’t define himself a fan of AI, on the contrary he has also expressed criticisms on the subject. But in this article he goes so far as to say that “now it’s clear that for most projects, writing the code yourself makes no more sense, except for fun” explaining how AI assisted coding can give a productivity boost even to a top programmer like him. He’s pragmatic and clear in saying that today it makes no sense to be denialists, as Kent Beck has also repeatedly pointed out. Salvatore says that with the help of a coding agent he managed to complete 4 very complex tasks with a certain ease. I add my note: the fact that he was at the helm of that agent makes all the difference in the world. Not everyone becomes Antirez with AI, but let’s say that Antirez becomes Antirez++ and this applies to all programmers and professionals who want to deepen this tool.
You haven’t approached agentic coding yet and maybe you’ve stopped at advanced code completion with AI. The second article is worth a read because it talks about precisely how to make that step (towards Claude Code in this case) successfully.
To those who ask me where the problems and bottlenecks are shifting with all this faster and “helped” coding, I always answer: specifications, task management for the agent, code review, security. I spoke about security in the previous section. In the links in this section you can find the answer that Devin tries to give to code review, very interesting because it doesn’t just limit itself to adding AI in that phase (which only shifts the problem a bit further), but uses AI to make the process simpler and more effective by adding tools to help humans understand and manage PRs. Task management is coming natively in Claude Code which as always is attentive to community initiatives. My goto for this thing remains backlog.md but I’m curious to see what those at Anthropic will bring.
I will speak about task management and spec-driven development at the Voxxed Day in another panel, where I will be a moderator with four guests who deal precisely with that. I’ll tell you in one of the next newsletters or in the podcast what I’ll have learned.
This week’s links
Don’t fall into the anti-AI hype
Salvatore “Antirez” Sanfilippo, creator of Redis, argues that for most projects writing code by yourself makes no more sense, except for fun.
I was a top 0.01% Cursor user. Here’s why I switched to Claude Code 2.0
Complete guide to the 5 pillars of agentic coding: context management, planning, closing the loop, verifiability and systematic debugging.
AI-powered code review tool that addresses bottlenecks with AI understanding of code, CLI, intelligent diff organization and bug flagging.
We’re turning Todos into Tasks in Claude Code
Anthropic updates Todos to Tasks: new primitive for tracking complex projects with dependencies, cross-session collaboration and filesystem persistence.
🏢 Business and society
Takeaways for AI Engineers
Takeaway 1: OpenAI and economic necessity: Despite growing revenues, OpenAI seeks $50B in new funds and introduces ads in the Go tier. Advertising becomes necessary to sustain AI infrastructure and compete with Google.
Takeaway 2: Inference is the new economic engine: While model training requires huge investments, optimized inference (vLLM, Inferact) is becoming the real revenue generator in the AI ecosystem.
Takeaway 3: Alignment and transparency: Anthropic publishes Claude’s new constitution, highlighting the importance of aligning AI models with human values as they acquire more decision-making autonomy.
Action Items:
Carefully evaluate trade-offs between cost and privacy in your AI tool choices (tier with ads vs tier without)
Read Claude’s constitution to understand how model alignment will impact autonomous AI systems of the future
What’s happening this week?
This week OpenAI protagonist of this section. On one hand Altman meets investors for another HUGE injection of funds into the company. Despite 2025 having seen a large increase in revenues, OpenAI still needs huge investments to maintain the pace it has given itself in bringing AI also to consumer and sustain the confrontation with a giant like Google. At the same time it launches a light subscription (called Go) at only 8$ per month, but supported by advertisements shown to the user during its use. The official link and Simon Willison’s authoritative opinion (one of the voices I esteem most in the AI landscape) explain what are OpenAI’s promises of clear separation between contents and ads. We’ll see, meanwhile Google swears that Gemini will never have advertising... indeed we’ll see.
Meanwhile we observe a startup made by some of the creators of vLLM who were recently acquired and who retry the entrepreneurial adventure. I don’t know if even this time they’ll have the same successes, but certainly at this moment inference is the money engine much more than creating models.
Anthropic confirms their attention to Claude’s alignment. The article is worth reading also to understand well what alignment is and how much this could impact the use of AI in the future, especially giving it more and more decision-making autonomy.
This week’s links
OpenAI’s long-rumored introduction of ads to ChatGPT
Simon Willison analyzes the introduction of ads on ChatGPT for Free and Go tiers, with significant implications for OpenAI’s business model.
OpenAI plans to introduce ads in ChatGPT
OpenAI officially announces advertising on ChatGPT to support wider access to AI, guaranteeing privacy and separation between ads and responses.
Anthropic publishes Claude’s new constitution: detailed document on the vision, values and alignment of the AI model.
OpenAI’s Altman Meets Mideast Investors for $50 Billion Round
Sam Altman seeks funding for a round of at least $50 billion that would value OpenAI between $750-830 billion for AI infrastructures and data centers.
Inference startup Inferact lands $150M to commercialize vLLM
Inferact, startup from the vLLM team, raises $150M to commercialize the open-source inference engine, highlighting the shift from training to inference.
🤖 Robotics and Physical AI
Takeaways for AI Engineers
Takeaway 1: Physical AI: the next frontier. Robotics is only a part of Physical AI, which combines perception, reasoning and action in the physical world. Players like Nvidia, OpenAI and startups like Cyberwave are investing heavily in this space.
Takeaway 2: The ecosystem is moving: From Nvidia’s VLA models to Merge Labs’ brain-computer interfaces (funded by OpenAI), the industry is creating the infrastructure for physical AI. It’s the “Android moment” for robotics.
Takeaway 3: Concrete opportunities for developers: Programs like Cyberwave Builders offer free hardware and support to develop real physical AI. It’s time to experiment with robotics and physical systems.
Action Items:
Sign up for the Cyberwave Builders Program to get free access to robotic hardware and develop physical AI projects
Explore Nvidia’s VLA technologies and resources on embodied AI to understand how to integrate visual perception, language and action in your projects
What’s happening this week?
I start by pointing out Cyberwave’s initiative, Italian startup in the robotics world: a builders program with 4-week cohorts focused on creating physical AI. And if you’re selected they provide the robot for free to do your development. For me who consider robotics “the next big thing” (ehm...I’m not the only one), an opportunity not to be missed. In fact I’m already subscribed, and I advise you to do the same following the link provided below. If you want to know more about Cyberwave I interviewed the founder Simone Di Somma in the podcast. A dense and very interesting interview both for Cyberwave and for Simone’s experiences and vision.
And to testify that I’m not the only one thinking about robotics, or rather about physical AI of which robotics is only a part, as the next turning point, the other two links speak to you about research and investments in this field by none other than OpenAI and Nvidia, two of the companies that are driving the AI revolution.
This week’s links
Nvidia Speeds up AI Reasoning with Fast-ThinkAct
Fast-ThinkAct: vision-language-action framework with 9.3x faster inference. Nvidia presents Isaac GR00T N1.6, VLA model for robotic applications.
OpenAI leads a $252M round in Merge Labs, startup of non-invasive brain-computer interfaces that use ultrasound to read/write in the brain.
Launching the Cyberwave Builders Program
Cyberwave launches a 4-week builders program for physical AI with free robotic hardware (SO101 arms, UGV rover) and team support.
🔗 Learn more about me, my work, and how to connect: maeste.it – personal bio, projects, and social links.