Sketch the app, test agents, secure code: three quick levers for your day

Important AI news

From sketches to clickable apps in Gemini Canvas

Google shows a Canvas tool. You upload a drawing and get runnable code right away that you refine iteratively. That cuts time to prototype a lot. For example from whiteboard to first UI. The provider says you can test it now.¹

Formally secure agent code: Microsoft Research releases PDD Kit

PDD Kit translates your intent into a formal specification, proves the agent code for all inputs, and uses tests where proofs are not enough. Formal verification is a mathematical proof of correctness. They showed this on a Rust web app login. It found existing bugs and proved the absence of login errors relative to the spec. It runs in GitHub Next’s Chopin and is an announcement, not broad availability. The trust boundary, the Trusted Computing Base, remains called out.²

TPUv7 with AgentX benchmarks: competition on inference costs

SemiAnalysis publishes first long, multi-step AgentX benchmarks on TPUv7. Example numbers: GLM-5.2 delivers 1.2 to 2.7 times as many tokens per dollar as B200, B300, and MI355X at 60 to 300 tokens per second per user. Qwen3.5 with llm-d beats GB300 FP8 and reaches parity with GB300 FP4 at higher quality. Kimi K3 sits 1.15 to 2.5 times over GB200, B200, B300, and MI355X. Measured on InferenceX/AgentX, which many buyers and labs support. These are vendor benchmark claims, independent checking is missing. What this means for you: when you book compute, compare TPU offers directly with GPU prices.³

OpenAI announces GPT-6 Cyber

The model targets security reviews. It finds vulnerabilities, avoids exploits, and suggests fixes within authorized limits. This is an announcement. There is no info on availability. Talk to your security team about if and where controlled use might make sense later.⁴

NVIDIA shows Vera CPU for agent workloads

NVIDIA reports 1.6 times more agents per watt than the fastest x86 processor and up to double agent speed versus the most efficient x86 chip. Early users like Perplexity and DeepInfra cite 1.5 to 2.2 times faster tasks depending on the test. These are NVIDIA numbers. External reproductions are pending. If you run many agentic steps on CPUs, Vera belongs in your evaluation.⁵

Highlights for your workday

• Learn to test agents reliably. The free course from DeepLearning.AI shows how to read agent traces, turn recurring errors into a trailing test set, and automate evaluations with code and LLM judges. An LLM judge is a model that scores results by rules. You get a loop you repeat after every change.⁶
• Create shorter videos for less money. Gemini Omni 1.2 Flash brings native 15 second clips, better consistency, and according to Google 30 percent lower prices. If you build social snippets or product demos, that cuts cost per asset. Rollout depends on access.⁷
• More signups with ChatGPT Ads. Flipsnack focused campaigns on product trials, increased registrations, and lowered price per signup. This is a vendor case study. Next step for you: set the goal Registration and A/B test two ad variants.⁸

Tools and updates

• Mistral Studio and “Le Chonk”: build and run enterprise agents and apps with more control over models, data, and infrastructure. Good if you must meet data residency requirements. Status: announced, details on price and availability missing.⁹
• One click deploy with Cursor and Vercel: deploy an app with no account setup, no config, and an auto provisioned database. Useful for quick demos and internal tests. Source is an X post.¹⁰
• Memory is becoming the bottleneck for agents. NVIDIA describes a full stack approach. Cache context outside GPU memory, evict inactive contexts, and reuse them so more parallel users are served. For your cloud choice: ask about KV cache management and context storage beyond GPU VRAM.¹¹

Try it in five minutes

1) Sketch a mini app as an ASCII wireframe in your AI chat and ask for a single HTML file with inline CSS and minimal JS. Example: a ticket list with search field, New ticket button, and status filter.¹ 2) Save the file locally, open it in the browser, and test search, filter, and button. 3) Ask the AI for a change. For example required field validation in the form and a colored status display. Result: a clickable HTML page as a simple prototype. Checkpoint: search filters correctly, button opens a form, validation is visible.

Sources

1. Transform a quick drawing into a live app with Gemini. (blog.google)
2. Formally verifying agent-written full-stack web apps with PDD Kit (microsoft.com)
3. The Iron(wood) Throne: Google's TPU Stakes Its Claim on Inference (newsletter.semianalysis.com)
4. GPT‑6 Cyber: Advancing the frontier of aligned cybersecurity agents (openai.com)
5. Vera Delivers the Best of Both Worlds: Highest Throughput and Fastest Agents (blogs.nvidia.com)
6. Stop guessing how to fix your agent (youtube.com)
7. Gemini Omni 1.2 Flash: Fast and cost-effective with longer clips (blog.google)
8. Flipsnack grows product signups with conversion optimization (openai.com)
9. Le Chonk brings the brains. Studio puts it to work. (mistral.ai)
10. @vercel: Deploy apps in one click with @cursor_ai.

• No account to create • No deployment configuration • Database provisioning included

Built on Vercel for Platforms → https://t.co/09OAn (x.com) 11. More Performance From Every Byte: How Full-Stack Codesign Solves AI’s Memory Challenge (blogs.nvidia.com)