Nvidia licenses Groq inference tech (and hires key execs)
Nvidia signed a non-exclusive licensing deal with Groq for its AI inference technology, and Groq’s founder/CEO Jonathan Ross is set to join Nvidia along with other senior leaders and staff. Groq will continue operating independently under a new CEO.
This is another signal that the center of gravity is shifting from “who trains the biggest model” to “who runs inference cheapest, fastest, and at scale.” Inference is where real usage lives, and where energy and latency become business constraints (not academic ones).
My take: I think this tells that the next “chip wars” chapter may be less about raw GPU dominance and more about specialized inference paths, power efficiency, and software integration.
“Vibe coding” meets reality: humans still win
A recent paper introduces a benchmark closer to real life than typical “unit tests pass” evaluations: multi-agent strategic planning on a logistics optimization problem (auction + routing). The result: human-coded agents consistently beat LLM-coded agents, and most LLM-built agents lose to simple baselines.
This matters because many orgs are quietly extrapolating from “LLM writes code snippets” to “LLM replaces junior engineers.”
My take: LLMs shine when the problem is pattern-heavy and the solution space is well-trodden. But optimization, strategic interaction, and robust planning still expose weaknesses. Also: benchmarks are becoming a game. Models can “learn the leaderboard,” while real software engineering is messy, long-horizon, and full of tradeoffs.
If you’re hiring: I’d be cautious about cutting junior pipelines. If you’re training: teach people to use LLMs while strengthening the core skill LLMs still struggle with problem solving under constraints.
I wrote in multiple occasions about this, you can check here, here and here.
ChatGPT ads: the monetization path nobody loves
Multiple reports (including a scan of ChatGPT Android App) suggest OpenAI is exploring advertising formats inside ChatGPT, even if timelines and product decisions may shift.
This would be a major change: ads inside a conversational interface are not just “banner ads.” They risk becoming answer-shaping incentives, especially in commercial queries.
My take: Ads are the simplest monetization lever at internet scale, and that’s exactly why it’s risky. Search ads already shaped an era of SEO gaming; “LLM answer ads” could shape an era of reality gaming. If ads happen, the only sustainable approach is strict separation: clear labeling, transparent ranking logic, and strong privacy defaults. Otherwise trust erodes fast, and trust is the product in this case.
AI pentesting agents are starting to compete with humans
A Stanford-led study evaluated AI agents versus human cybersecurity professionals on a live enterprise-like environment (a university network with ~8,000 hosts across 12 subnets). Their agent scaffold ARTEMIS ranked second, found 9 valid vulnerabilities, and achieved an 82% valid submission rate, outperforming 9 of 10 human participants in that setup.
They also report cost comparisons (agent variants vs human pentesters), and note weaknesses like false positives and GUI friction.
My take: This is not “AI replaces pentesters.” It’s “pentesting becomes a human+agent team sport.” The economic incentive is obvious: you can run agents continuously, at scale, across assets humans rarely touch often enough. But you still need humans for judgment, scoping, and for the inevitable weird edge cases. I also believe CISOs should start planning for agent-driven offensive testing. Hackers are using AI to amplify their attacks. CISOs should also use AI for protection!
OpenAI releases GPT-5.2 and GPT-5.2-Codex
In December, OpenAI released GPT-5.2 as a new frontier series for professional work and long-running agents, reporting improvements across knowledge work and coding benchmarks (e.g., SWE-Bench Pro).
More interesting for builders: GPT-5.2-Codex targets real software engineering workflows (multi-file refactors, migrations, long debugging sessions). The key technical idea is native context compaction—compressing prior steps while preserving state/intent so longer tool-driven sessions don’t collapse under context limits.
Bonus: Access Top Courses at Discount Price when taking resolutions for 2026
As we head towards 2026, I consolidated a list of courses at discount prices from top courses creators, mainly around AI and Project Management (PMP). These courses (depending on the course), are generally sold at 50$/€+. Now you can get them at 10$/€. Links are only valid for next 3 days. Act fast and take new resolutions in 2026 to learn new stuff!
AI courses :
PMP courses:
