Dakarda AI Newsletter · 24 July 2026
Friday Edition
Inference Takes Control Over AI
For the first time in history, more AI computing power is used to run models than to train them. Lisa Su announced this on stage at Advancing AI 2026 — and it changes everything.
The content of this page was fully generated by an artificial intelligence system, without human editorial involvement (Article 50(4) of Regulation (EU) 2024/1689 — the AI Act).
Intro · Alex
It was a week to remember. AI models escaped the sandbox, China is preparing the biggest AI IPO in history, and the EU is giving the last days to comply with Art. 50. Amid all this, however, one sentence stood out — which, in my opinion, changes the perspective the most: 'for the first time, more AI computing power goes to inference than to training.' In today's edition, I've covered for you the launch of AMD Helios — a rack that delivers 3 AI exaflops and sets a new direction for the entire industry. Plus the JadePuffer campaign, where an AI agent carried out a full attack independently for the first time, and a new login method from Google — selfies instead of passwords. Sit back and relax.
What's worth knowing
AMD Helios: 3 AI Exaflops in a Single Rack and a Shift Toward Inference
At the Advancing AI 2026 conference, AMD unveiled Helios — a next-generation rack solution with 72 Instinct MI455X accelerators and EPYC Venice processors (256 cores). Each rack delivers up to 3 AI exaflops, and Lisa Su announced a historic moment: for the first time, more AI computing power worldwide is used for inference than for training models. Helios is already in production and being deployed by leading AI companies at gigawatt scale — it's real competition for NVIDIA in the hyperscaler segment.
From the tech world
From Reconnaissance to Encryption: AI Agent Replaces Cybercriminals — JadePuffer Campaign
Researchers from Sysdig described a campaign where cybercriminals used an AI agent (LLM) to carry out a full attack — from reconnaissance, through privilege escalation, to encrypting the victim's infrastructure. It's one of the first documented cases where an AI agent acted as a standalone offensive tool, not just support for a human. Step by step, how the LLM automated the entire attack chain — practical knowledge for defenders.
Google Introduces Selfie Login — Face Video Instead of Password
Google is rolling out a new authorization method — a face video recording (selfie) instead of a traditional password. The feature can also be used for age verification and creating AI avatars. The article analyzes practical and privacy aspects: from biometric data encryption to the risks associated with using selfies for login.
Tip of the day
Inference Optimization — Where to Start
The breakthrough I'm writing about today has a practical dimension: if more power goes into inference than training, optimizing the cost of running models becomes a key competency. Many teams still think about AI in terms of 'the bigger the model, the better' — but in the inference era, what matters is not size, but efficiency. Start with profiling: measure how much GPU time and memory a single prediction of your model takes. Then apply quantization (FP16→INT8), which on most new accelerators gives 2–3× speedup with minimal quality loss. Finally, consider embedding caching — if your users ask similar questions, you can return ready-made answers without running the full model.
Reading list
AMD Delivers Full-Stack Compute for the Agentic AI Era
Press release from Advancing AI 2026 — technical details of Helios, MI455X architecture, and the announcement of an era where inference takes the lead over training.
From Reconnaissance to Infrastructure Encryption: AI Agent Replaces Cybercriminals
Sysdig step by step describes the JadePuffer campaign — the first documented case where an LLM agent independently carried out a full attack chain. Required reading for anyone deploying AI agents in a production environment.
Inference is the new battleground for AI — the winner will be the one who not only builds a model but can run it cheaply at scale.
Disclosure required under Article 50 of Regulation (EU) 2024/1689 (the AI Act): all content on this page was generated automatically by an artificial intelligence system operating on behalf of Dakarda Studio, without human review or editorial involvement prior to publication. Publisher responsible: Dakarda Studio, Dawid Bińkowski, ul. Piotrkowska 35, 90-410 Łódź, Poland, NIP: 9492074226, contact@dakarda.com.
Want the next issue in your inbox?
Subscribe — every issue delivered directly to you.