Daily Briefing
23 August 2026 · 5 stories
🔥 Top story
1Hermes, Info-Sys's homegrown AI helper, lands on Cloudways: try it with one click
Imagine a play area where your favourite AI helper is already installed, ready to use, with no tinkering required. That is exactly what Cloudways — the hosting brand of American giant DigitalOcean — rolled out this week. It opened a brand-new “Managed AI Helpers” shop, and among the first picks you will find OpenClaw and, more importantly, Hermes — our homegrown system built right here at Info-Sys. Practically speaking, for a small business or an agency that wants to automate a task, it means: no need to hire an engineer to set everything up. You pick the helper, you plug in your key (from OpenAI, Anthropic or Google), and the tool runs within minutes. And good news: this is no copy or rip-off. DigitalOcean's official press release explicitly names our public code repository. For Info-Sys, it is a public showcase on a company listed on the New York Stock Exchange.
2Two shared languages for AI helpers to talk to each other are now under the same roof — a good thing for businesses
Today, AI helpers chat with each other using two shared “languages”: MCP (to give helpers their tools) and A2A (to get helpers to team up with each other). The problem, until now, was that these two languages belonged to two different companies — Anthropic for one, Google for the other. This week, Google handed its A2A language over to the Linux Foundation, the same place that already hosts MCP. Translation for a small business that was on the fence: you no longer need to worry about betting on the wrong horse. Both standards are now run by a neutral body, backed by more than 250 players — including AWS, Microsoft, Google, OpenAI and Cloudflare. It is a bit like if French and English had both been declared official languages of the same country: everyone can talk to each other without choosing sides.
3DeepSeek nearly doubles its prices mid-ascent — and now targets a $74 billion valuation
Chinese lab DeepSeek made two announcements back to back this week. First, its V4-Pro model is now officially open to everyone — it scores among the very best on coding tests. Second, its top price tag has exploded: the cost for output can now reach $3.96 per million words, almost five times what it was before. At the same time, several press sources are talking about a fundraising round of $8 billion at a $74 billion valuation. For a business that uses a lot of AI, this means the era of “cheap as you like” Chinese tools may be coming to an end. For the curious, it also confirms that Chinese open AI labs have joined the big leagues — now valued at the same level as American closed labs.
4Alibaba releases a monster 2,400-billion-parameter AI model — free to download
If you have a good computer (and a lot of patience), you can now download Qwen3.8-Max, Alibaba's latest flagship. It is the very first “Qwen-Max” ever released to the public. Under the hood: 2,400 billion parameters in total, of which 95 kick in for each question. On public rankings, it lands 5th for text and 2nd for image understanding, just behind closed American systems. And for those without a server room at home, Alibaba is also putting out Qwen3.8-27B, a compact version under the very permissive Apache 2.0 licence that fits on a single graphics card. This is the biggest free doorway into giant AI models available in the world today. One small catch: the open version handles text only — not yet images or audio.
5A Chinese “open” AI model found 2,436 security flaws — including in the heart of Linux
GLM-5.3, the latest model from Chinese lab Z.ai, made a splash this week: in testing, it dug up 2,436 flaws across 269 real pieces of software — including the Linux kernel, the WebKit rendering engine, and the FreeBSD system. Better still: it has already flagged a bug in Cursor, the coding editor recently snapped up by SpaceX. In the meantime, Z.ai has deliberately held back the public release of the model's brain by about two weeks — time to tighten the safety rails and avoid malicious uses. For the general public, this changes nothing directly. But for software developers and vendors, it is a loud wake-up call: free Chinese models are getting just as good, sometimes better, than the paid American ones at hunting bugs.
📡 To watch
Anthropic stamps your texts with invisible ink — to obey Europe
Since August 2, a new European law requires AI systems to flag their outputs. Anthropic went with a hidden statistical watermark, baked in at no extra cost. You will never see it — but a specialised tool will be able to tell whether a piece of text came from Claude.
Cognition (Devin) negotiating a $40 billion valuation — three times more than in May
The maker of the AI helper that codes on your behalf is in talks to raise funds at more than $40 billion in valuation, according to Bloomberg. Its yearly revenue would have doubled in three months. But nothing is signed yet: the amount and the valuation can still change before the deal closes.
Hermes name mix-up alert resolved ✅
A cybersecurity report had named “Hermes” as a tool used in a cyberattack, creating confusion with our system. This week, DigitalOcean's official press release cuts through the noise: it really is our github.com/nousresearch/hermes-agent repository. No trademark risk — alert closed.
When will GLM-5.3's public brain come out?
Z.ai has deliberately delayed putting the model's brain online to toughen the safety rails. The window closes at the end of August — watch for the final licence: the inherited permissive one or a custom restriction? That choice will decide who can use it commercially.
📊 Trend
The weekend confirmed two shifts pulling in opposite directions. On one side, the hardware that runs AI helpers is becoming a ready-to-use hosting product (Cloudways now hosts Hermes in a public release, a New York Stock Exchange showcase for Info-Sys). On the other, the AI models that power those helpers are getting pricey when you aim for the top of the line (DeepSeek has quadrupled its prices on V4-Pro). In other words: AI helpers are moving from “geek project” stage to “managed service” stage — but with a heavier bill for heavy use. For a business that automates, it is now an explicit trade-off between hosted convenience and a free, high-quality model.