Google Launches Gemini 3 Pro
Google DeepMind has introduced Gemini 3 Pro, as its most intelligent AI model to date.
This third-generation Gemini is designed for advanced reasoning across modalities and is deeply integrated with Google’s ecosystem, powering search-connected, multimodal “generative UI” experiences
In practice, Gemini 3 can handle text, images, and more, enabling features like high-resolution image generation (with the new Nano Banana Pro model - Amazing by the way) and even an AI content verification tool in Google’s apps. Currently in preview, Gemini 3 Pro is accessible via the Gemini mobile app, Google Cloud (Vertex AI), and Google’s AI Studio.
Google reports that it outperforms previous models in complex learning and planning tasks and it’s set to replace older assistants (even being tested in Android Auto).
DeepSeek-V3.2-Exp: Long context, sparse attention, open weights
DeepSeek released V3.2-Exp (around 2 months back, but I didn't had the opportunity to review it till recently), an open-weight experimental model built on V3.1-Terminus that debuts DeepSeek Sparse Attention (DSA). Instead of just scaling parameters, they reworked attention patterns to handle long context more efficiently, while keeping training configs comparable to the previous model. API prices dropped by more than 50%, signaling a push toward cheaper tokens rather than just bigger models.
From day 0, V3.2-Exp is supported in vLLM on modern Nvidia hardware (H100/H200/H20, B200/GB200). Benchmarks show parity with V3.1-Terminus across common tasks, while enabling much longer context at similar compute cost.
I think this is emblematic of a broader trend: efficiency and context length are now as strategic as raw IQ. Open-weight players like DeepSeek are creating viable alternatives to proprietary “maximalist” models, especially for workloads where you need to stuff entire repos, long specs or multi-quarter logs into a single context. The question for enterprises is less “Which is the smartest model?” and more “Which stack gives me acceptable quality at predictable cost?”
I speak about this both in my course on DeepSeek, and my course on Sustainable AI.
Shadow AI: when your riskiest AI system is the one you don’t know about
Gartner now predicts that by 2030, ~40% of enterprises will suffer a security or compliance incident caused by “shadow AI”, unsanctioned AI tools used by employees outside IT or security oversight. In recent surveys, 69% of security leaders say they suspect or have confirmed such tools in their org, and Microsoft data suggests 71% of workers in the UK alone use shadow AI, with 22% doing so for high-risk tasks like finance.
Analysts highlight three main risks: unvetted tools exfiltrating sensitive data, regulatory breaches (GDPR, HIPAA, financial regulations) when regulated data is pasted into random chatbots, and the accumulation of “AI technical debt”. Systems built in the shadows that later need to be replaced or secured at great cost.
I believe most companies still treat AI like “just another SaaS app,” but the risk profile is closer to an unmanaged data lake plus a junior analyst (with all due respect to junior analysts) with root access. A pragmatic approach is: 1) publish a clear AI usage policy, 2) offer good, approved tools so people don’t need shadow ones, 3) bake AI risk checks into vendor assessments, and 4) train staff on what never goes into a prompt (customer PII, trade secrets, health data, etc.).
Can the Internet Go Offline
The last weeks were another reminder of how much of the internet rests on a few critical providers.
After recent outages at AWS and Azure, Cloudflare also went down, taking with it a long list of services, including ChatGPT. When Cloudflare or a big cloud region fails, it’s not “one site” that breaks, it’s an entire slice of the internet experience.
This ties well with the Guardian article “Could the internet go offline?”. The core idea: the internet is built on creaking, decades-old protocols and a lot of concentration. A few unlucky events (extreme weather on data centres, a serious bug in a major provider, a routing or DNS issue) could cause cascading failures.
I do believe that we have accumulated many Single Points of Failure (SPOFs). Full global shutdown is unlikely, but large-scale, multi-day disruption is no longer a crazy scenario, remember Crowdstrike global outage. Rethinking resilience, multi-region, multi-cloud where it matters, and realistic outage playbooks, is becoming part of the job, not an “edge case”.
Aardvark: GPT-5 as an Autonomous Security Researcher
OpenAI has released Aardvark, an “agentic security researcher” powered by GPT-5.
Aardvark connects to your repositories, reads the code, builds a simple threat model, scans commits, tries to exploit what it finds in a sandbox, and then proposes patches. It behaves more like a human security engineer than a classic rule-based scanner.
In OpenAI’s tests on open-source projects, Aardvark reportedly identified 92% of known and synthetic vulnerabilities, and it already led to several CVEs being issued. It also integrates directly with GitHub and Codex to open pull requests with suggested fixes.
This is getting very interesting. Aardvark looks like a serious candidate to replace or augment SAST in many contexts: more context, better explanations, continuous watch. For DAST-style testing, it is less obvious today, but future versions integrated into staging and runtime could move in that direction. I do not think it is of a level yet to do Pen Testing, even if some AI agentic agents are starting to show good traction on such technologies like this open source repo Strix.
It’s a clear step toward “security by default,” where an AI agent quietly reviews your code all the time, and hopefully finds the bug before an attacker does. In terms of maturity, I do not believe it is yet to replace vertical specialized SAST tools.
Bonus: Your Sustainable AI Course Free Access
As promised when I initially launched this newsletter, I am more than glad to offer exclusive free access to my courses when I launch them.
I just launched a new course on Sustainable AI, and the idea came following a key note I delivered in Amsterdam, where some questions came on that topic, and when I looked, there were barely some educational material about it, so I built one :)
To get your free access, you just need to
The link is valid only for 2 days, so ensure you enroll quickly. Leaving a 5* review if you like the course is really helpful so it gets good ranking and I appreciate your support on that. I am looking forward to get your feedback.
In case the button didn't work, you can simple use this link or copy/paste it:
https://www.udemy.com/course/sustainable-ai-artificial-intelligence-footprint-and-what-todo/?couponCode=TPTL_DIGITAL
And if it doesn't work or it expires, please answer to this email and I'll send you a new one starting December 1st.
