LastWeekIn.Tech

A quiet day for landmark breakthroughs but packed with signals on where AI competition is heading: Google's cybersecurity-focused Gemini 4 Carbon claims the top intelligence benchmark spot, Mistral pushes a trillion-parameter open-weight model into the wild, and chip/memory and revenue data underscore how fast AI infrastructure economics are compounding.

1. 1

   Number 1. Capabilities Hacker News

   Gemini 4 Carbon

   Google released Gemini 4 Carbon, described as its strongest model yet with a focus on cybersecurity and long-horizon reasoning. It ranks #1 on the Artificial Analysis Intelligence Index v5.0 with a score of 59, and leads specialized security benchmarks including SEC-bench Pro (88.2%) for vulnerability discovery, ExploitGym (54.8%) for exploitation, and CWE-bench (71.0%) for patching. Google is reportedly limiting broad access to vetted experts despite the model's strong public benchmark scores.

   Why it matters A frontier model explicitly optimized for finding, exploiting, and patching vulnerabilities marks a new phase where AI labs are racing to dominate offensive and defensive cybersecurity capability rather than just general chat performance. Restricting access to vetted experts signals that Google itself sees dual-use risk in the model's exploitation skills, raising the stakes for how labs balance capability disclosure against misuse potential. It also intensifies competition with Anthropic's Claude Opus line and Chinese models like GLM-5.3 in the cyber-capability benchmark race.

   What to watch Watch whether Google expands access beyond vetted experts, and how rival labs respond with their own cybersecurity-specialized models or safeguards.

2. 2

   Number 2. Open Source Hacker News

   Mistral Large 4 (Le Chonk) is now Open Weight

   Mistral released Mistral Large 4 ("Le Chonk"), a 1-trillion-parameter multimodal model, as fully open weight. The release claims leading open-weight benchmark performance in coding and defensive security, offers three reasoning modes for cost-performance tradeoffs, and was trained in Europe.

   Why it matters This is one of the largest open-weight models released to date and directly challenges the narrative that frontier-scale capability requires closed deployment. By emphasizing European training and 'sovereign AI,' Mistral is positioning itself as the go-to option for governments and enterprises wary of US or Chinese model dependency, which has geopolitical as well as commercial implications. A trillion-parameter open model with strong coding and security benchmarks also lowers the barrier for state and non-state actors to access high-end agentic and cyber capabilities.

   What to watch Watch adoption by European government and defense customers, and whether independent benchmarks confirm Mistral's claimed parity with closed frontier models.

3. 3

   Number 3. Industry & Economics iheart.com

   Microsoft reports $10 billion in revenue tied to OpenAI arrangements

   Microsoft disclosed that it generated $10 billion in revenue tied to its OpenAI arrangements, reported as part of $92.3 billion in total quarterly revenue. The figure was detailed alongside Microsoft's broader financial results.

   Why it matters This is one of the first concrete, quantified looks at how much direct revenue flows to Microsoft specifically from its OpenAI partnership, giving investors and competitors a clearer picture of the commercial value embedded in that alliance beyond Azure compute credits and equity stakes. It validates the scale of the Microsoft-OpenAI relationship as a real revenue driver rather than just a strategic bet, which matters as both companies face scrutiny over AI spending sustainability and potential antitrust questions.

   What to watch Watch for how this revenue line evolves as OpenAI's commercial products mature and whether Microsoft breaks out more granular OpenAI-related financials in future earnings.

4. 4

   Number 4. Infrastructure & Chips business-standard.com

   Samsung warns that AI memory demand may outstrip supply

   Samsung Electronics said AI infrastructure spending could reach historic levels in 2027, warning that demand for memory chips is expected to outstrip industry supply by a wider margin than previously anticipated. The comments came alongside Samsung posting record quarterly profit driven by AI-chip demand.

   Why it matters Memory chip supply constraints are becoming a central bottleneck narrative for the AI buildout, following similar warnings about GPUs, power, and now helium and other critical materials. If memory supply genuinely cannot keep pace with AI datacenter expansion, it could drive up training and inference costs industry-wide, squeeze smaller AI labs out of the compute race, and further concentrate capability among companies that can secure long-term supply contracts.

   What to watch Watch for memory price trends into 2027 and whether other chipmakers (SK Hynix, Micron) issue similar supply-demand warnings or announce capacity expansions.

5. 5

   Number 5. Security & Safety business-standard.com

   Chinese GLM-5.3 model completes complex cyberattack scenarios in controlled tests

   Chinese lab Z.AI's open-weight GLM-5.3 model completed three of 13 multi-stage cyberattack scenarios in controlled simulations run by security firm Irregular, though it trailed leading US models in separate tests. The results were reported alongside broader industry commentary on AI-driven market opportunity and infrastructure spending.

   Why it matters Independent red-team testing showing an open-weight Chinese model capable of executing a meaningful fraction of complex, multi-stage cyberattacks demonstrates that offensive AI cyber capability is diffusing beyond the leading US labs, even if it currently lags behind them. Combined with Google's and Mistral's own cybersecurity-focused releases, this points to an emerging arms race in AI-enabled offense/defense where open-weight availability raises proliferation concerns distinct from closed-model risk.

   What to watch Watch for follow-up safety evaluations of GLM-5.3 and whether Chinese labs like Moonshot (which reportedly limited cyber capabilities in Kimi K3.1 Pro) adopt similar voluntary restrictions.