AI Weekly Trends Highly Opinionated Signals from the Week [CY26W7]
🔗 Learn more about me, my work, and how to connect: maeste.it – personal bio, projects, and social links.
One of the reasons, perhaps the main one, why I started writing this newsletter and then recording Risorse Artificiali is the need I felt to participate in AI dissemination. And that need stems from the fact that I think AI (and particularly AGI) and its impacts on business, society, and work are too neglected. Perhaps not by everyone, but many seem to engage in denial, more or less argued, often in a vacuous attempt to ridicule it (see all the posts this week about self-service car washes). I’ve already seen this happen very recently, and I thank Matt Shumer for having the courage to make that comparison that’s been running through my head for a long time between AGI and Covid. I stayed away from that comparison, to avoid giving AGI a negative connotation (I’m anything but a doomer), but I see the same denial mechanisms, perhaps driven by fear, but much more by ignorance. As with Covid (and believe me, I live in Cremona, the hardest-hit province with the most deaths in 2020) until field hospitals were visible, no one really believed it. You have no idea how many times I heard “it’s a China thing,” “it’ll be the usual bad flu.” Well, AGI is coming, it’s not someone else’s problem, it’s not the usual technology. It will change many of the jobs we know. It will change society. And I hope it won’t cause the damage Covid did, but we need to be prepared to control the change rather than be controlled by it.
And I’m not the one saying this— all the scientists and most influential people in this field are. People who have visibility into what’s happening in major labs. Today, not 15 years from now.
A few weeks ago I interviewed Alessandro Maserati on the podcast. And he said: “I think it’s very underestimated in public debate that we’re already talking about AGI being achievable, right? Just that alone should populate all the front pages, all the newspapers, we shouldn’t be talking about anything else.” I subscribe to that, word for word.
Yesterday I commented on a LinkedIn post that made a heartfelt appeal. I subscribe to that too, and it gives me hope that there’s growing awareness at least among those working in this sector.
Below you’ll find this week’s news—the first section ideally continues this introduction, but even in the others, if you read with a careful eye, you’ll find evolution running in that direction. If you’ve been following me for a while and have read my newsletters critically, this won’t be news to you. If I may suggest, start talking about it with friends who are less involved in this world, without a negative connotation. This is a technology, a tool, and like all tools it just needs to be used well. But first and foremost, it shouldn’t be denied—rather, it should be understood today to be used best tomorrow.
Business and Society
Key Takeaways for AI Engineers
Takeaway 1: AI leaders speak of AGI as a fait accompli - Amodei (90% confidence), Suleyman (18 months to human-level performance): it’s no longer “if,” but “when.” What they see in labs today determines these predictions
Takeaway 2: “Something Big Is Happening” must be read and shared - It’s not clickbait, it’s realistic. Preparation is a professional duty to yourself and those around you
Takeaway 3: White-collar work disruption is underway - Not in 15 years, but on a 1-5 year timeline for 50% of entry-level positions
Action Items:
Read “Something Big Is Happening” and share it - If you’ve already read it, re-read it. Then have someone less familiar with AI topics read it
Prepare a professional adaptation plan - Don’t wait: integrate AI into daily work, build financial resilience, cultivate adaptability
What’s happening this week?
Usually this is the last section of my newsletter. But not today, because the news I’m reporting all says that AGI is coming, and as you read in the introduction, I believe the time has come to acknowledge it.
But let’s proceed in order. You’ve probably already read the article “Something Big is Happening,” or at least seen it cited by someone this week. It’s certainly on everyone’s lips who deals with AI in any capacity. Well, if you haven’t read it, read it and re-read it; if you have already, read it again... and then have someone less involved in the topic read it. It doesn’t exaggerate, it’s not click bait, it’s realistic. Being prepared is a necessity and a duty.
Dario Amodei spoke on the Dwarkesh Patel podcast. And he reiterated many of the things he said in a panel with Demis Hassabis a couple of weeks ago. Namely, that the point is not IF a “Country of geniuses in a data center” will arrive, but when. Because the IF is at a 90% confidence level (like a good scientist, he doesn’t take anything for certain, but expresses confidence levels). And Amodei is more scientist than CEO, and perhaps one of the brightest minds in the world, certainly in the sector.
Mustafa Suleyman, Microsoft’s AI chief - not some random tinkerer next door - instead talks about AI reaching human-level performance in 18 months and most computer activities being completely automated. In other words, many “white-collar” jobs won’t exist anymore, at least as we know them today.
These people have visibility into things happening inside labs, and if they (and others) keep talking about these things, it’s because they’re seeing them happen today. Today, not in 15 years.
I’ve also included other articles to give you a more complete picture.
This Week’s Links
Something Big Is Happening - 1-5 year timeline for 50% of entry-level white-collar positions, urgent preparation needed
AI Skepticism: Career Killer - Resistance to AI adoption is becoming a professional liability
AI Champions - Influential technologists within teams drive successful AI implementation
Dario Amodei: Dwarkesh Podcast - “Country of geniuses in a data center” within 1-3 years, 90% confidence
Anthropic Series G - $30B funding, $380B valuation, $14B revenue run-rate
Mustafa Suleyman: Fortune - Human-level AI in 18 months, white-collar job automation
AI Model News and Research
Key Takeaways for AI Engineers
Takeaway 1: Gemini 3 Deep Think sets new records - With 84.6% on ARC-AGI-2 and 48.4% on Humanity’s Last Exam, benchmarks demonstrate significant progress toward reasoning capabilities increasingly close to AGI
Takeaway 2: Chinese open-source AI is changing the game - Models like MiniMax M2.5, Step 3.5 Flash, and GLM-5 offer near-SOTA performance at 1/10-1/20 the cost of American competitors, making advanced AI accessible to everyone
Takeaway 3: Competition shifts to speed and efficiency - From “flash” and “lightning” versions to execution on Cerebras hardware, the focus is no longer just on quality but on making it practical and economical
Action Items:
Test a Chinese open-source model - Try MiniMax M2.5 or Step 3.5 Flash: at 1/20 the cost of Claude, it’s worth exploring economic alternatives for repetitive workloads
Watch the Seedance 2.0 and Qwen-Image demos - This week’s links show video and image quality that seemed impossible until recently
What’s happening this week?
There’s lots of news in models this week, but I have to start with the latest version of Gemini 3 Deep Think. The benchmarks you can verify in the link I’m sharing are beyond impressive. They tell of an AGI approaching in giant steps. Undoubtedly benchmarks don’t always tell the whole truth and raise expectations, but in any case the improvements from previous versions are impressive and parallel what Amodei and others are saying about approaching AGI and its impacts on work and society.
Although DeepMind’s model results took the spotlight, they’re certainly not the only model updates. Alone, the evolution of GPT-5.3 Codex Spark, a smaller and super-fast version of OpenAI’s latest, would be huge news both for the speed achieved and for being the first to run on Cerebras hardware. But there’s also a lot coming from China with three SOTA models at very low prices and excellent benchmarks, bringing them close to top American models at 1/10 the cost. Chinese models seem to be closing the gap with big tech SOTA models, while remaining a few months behind, but the cost factor and the fact they’re open makes them very interesting. After all, when Opus 4.1 and then 4.5 came out a few months ago, we proclaimed a miracle, and today GLM, MiniMax, and Step models practically match those incredible performances, even as we’re already looking at 4.6 and beyond.
Finally, I’m reporting two news items and some demo links for an image model (Qwen-Image, so small it runs on gamer PCs or Mac Minis) and for Video. I cover these types of models less in this newsletter, but you can see the incredible progress for yourself by following the links below.
This Week’s Links
Qwen-Image-2.0 - Image generation with native 2K resolution, advanced text rendering, and prompts up to 1K tokens
Seedance 2.0 - ByteDance’s text-to-video with natural movement and impressive micro-details
Step 3.5 Flash - 196B sparse model activating only 11B parameters per token, 100-350 tok/s
GPT-5.3 Codex Spark - New advanced OpenAI model for coding and reasoning, first on Cerebras hardware
GLM-5 - Latest Zhipu AI model with improved reasoning, part of the wave of competitive Chinese models
Gemini 3 Deep Think - Benchmark records: 84.6% ARC-AGI-2, 48.4% Humanity’s Last Exam, 3455 Elo Codeforces
MiniMax M2.5 - Open-source frontier: SWE-Bench 80.2%, 1/20 the cost of Claude Opus 4.6
Agentic AI
Key Takeaways for AI Engineers
Takeaway 1: OpenClaw marks a paradigm shift - From explicit prompts (chat) to context injected by data streams with tools to execute actions: models with real-time data access define strategies and act autonomously
Takeaway 2: Three trends consolidate: Skills, Sandbox, Swarm - Skills for reusable workflows, Sandbox for secure execution, Swarm for parallelizing complex tasks with dozens of agents
Action Items:
Listen to Peter Steinberger’s interview on Lex Fridman - He explains the shift from chat-based to context-driven agent behavior better than anyone
Explore the A2A course on deeplearning.ai and the GitHub project - The protocol is becoming the standard for inter-framework agent communication
What’s happening this week?
If models continue to evolve and improve continuously, what’s happening in the world of agents is a true “aha moment,” capable of redefining our expectations of intelligent agents. I’m referring of course to OpenClaw and what it’s making us understand about how these models with broad access to real-time data are able to define behavior strategies and take actions. Nothing magical or unexpected for those observing ongoing changes, it’s a natural evolution of what we’ve seen happen over the last two years. What happens in autonomous and proactive agents like OpenClaw is the shift of interaction from one based on an explicit prompt (chat) to a context injected by data streams that create implicit prompts and tools that allow executing actions. Peter Steinberger, OpenClaw’s creator, explains this much better than I can in his Lex Fridman interview. Listen to it.
Three other very clear trends in the agent world this period are Skills, sandbox, and parallel sub-agents often called swarm. Let’s start with Skills, which are simply reusable sets of instructions and scripts to build workflows. Conceived and launched by Anthropic, they’ve been adopted by everyone, and this week OpenAI underscored their importance in their APIs too. Regarding sandbox, it’s worth reading the LangChain article because it clarifies many concepts about this. Everyone is talking about Agent Swarm - Anthropic, MiniMax, Z.AI with GLM 5.0. I’m sharing a Kimi article here though, among the first to talk about Agent Swarm, which is particularly interesting because it addresses the topic of very large swarms.
I’ll leave you with an invitation to check out the A2A course recently launched by deeplearning.ai. And of course, take a look at the A2A project on GitHub where my team and I are personally involved.
This Week’s Links
A2A Protocol - Standard for agent discovery and communication across frameworks, with course on deeplearning.ai
Kimi Agent Swarm - Up to 100 parallel sub-agents, 4.5x faster than sequential execution
OpenAI Skills API - Reusable bundles of instructions, scripts, and assets for repeatable workflows
Sandbox Connection Patterns - Two architectural patterns for integrating agents with execution sandbox
Lex Fridman: Peter Steinberger - Interview with OpenClaw’s creator on agentic engineering and the future of software development
AI Assisted Coding
Key Takeaways for AI Engineers
Takeaway 1: “Libraries are over, LLMs are the new compiler” - Karpathy’s provocation points to more fluid software with fewer dependencies: extract only what’s needed from monolithic libraries via AI agents
Takeaway 2: Persistent memory becomes priority - Claude-mem demonstrates that “done is better than perfect”: even if imperfect, the attempt to give coding agents long-term memory is fundamental
Takeaway 3: Task duration as a metric of progress toward AGI - Cursor reaches 24+ hours of autonomous tasks: if the length of tasks completable with acceptable error grows, we’re getting closer to AGI
Action Items:
Read Karpathy’s thread and try DeepWiki via MCP - Replace “github” with “deepwiki” in a repo URL to understand the potential
Install Claude-mem and test it on a real project - Evaluate if persistent memory improves your productivity with Claude Code
What’s happening this week?
The most relevant article this week on AI assisted coding is undoubtedly Karpathy’s. As always, quite concisely he touches on many interesting points. First, using DeepWiki via MCP to give the LLM rich, quality context regarding the repository you’re working on. I often use DeepWiki to automatically generate the wiki with project architecture to help me better understand new projects. Using it as an MCP server is brilliant and shifts the boundary of understanding and managing complex projects. But in the article, Andrej drops a real bomb, going so far as to say “Libraries are over, LLMs are the new compiler.” I agree with his desire for more fluid software with fewer dependencies, although this could challenge some open source licensing and intellectual property models. Read that article—it’s full of practical insights, but also food for thought.
Claude-mem is evolving a lot, and deserves to be tried and explored. It’s an attempt, even if not perfect in my opinion, to give coding agents a long-term memory that goes beyond the session. It’s something definitely needed, which also has great complexity. As I said, it’s not perfect in my opinion, but in this era of AI and code generated and improved at light speed, I’m starting to think that “done is better than perfect.” In other words, it’s a good start that can inspire others to do even better. Worth trying both in coding agents (like Claude, which it’s named after) and in more generic agents like OpenClaw.
Speed is definitely a theme, as I mentioned in the previous chapter, and Anthropic makes available (at a cost) a fast mode for Opus. Meanwhile, Cursor launches Long-running Agents with tasks up to 24 hours. And I imagine you remember that one of the most important parameters for measuring the impact of intelligent agents in a sector is the length of tasks they’re able to complete with an acceptable error factor. 24 hours starts to make you think of something increasingly close to AGI. A theme that returns in this section too.
This Week’s Links
Claude Code Fast Mode - Opus 4.6 2.5x faster with /fast, pricing $30/150 MTok
Claude-Mem - Persistent memory plugin for Claude Code with web viewer and hybrid search
Cursor Long-Running Agents - Autonomous agents up to 24+ hours with planning and multi-agent verification
Karpathy: DeepWiki and MCP - “Libraries are over, LLMs are the new compiler” - more fluid and malleable software
🔗 Learn more about me, my work, and how to connect: maeste.it – personal bio, projects, and social links.