🌟Vasilij’s Note
First-ever newsletter! Starting something new is both nerve-racking and exciting. I have precisely zero subscribers, so if you are reading this, thank you for being one of the first readers! Let’s jump right in!
This Week in Agents | What Changed
Agents Learn To Remember — A new procedural-memory framework (Memp) lets LLM-based agents store, retrieve, and update past experiences, improving success rates and enabling smaller models to reuse knowledge. Read more in the paper: https://www.arxiv.org/pdf/2508.06433
Nano Banana (Google) — Google’s new Gemini 2.5 Flash Image model (“Nano Banana”) lands in the Gemini app: blend multiple images, keep character/identity consistency, and make precise, natural-language edits challenging to Photoshop. Press Release>
Google brings Gemini to GitHub — Gemini CLI GitHub Actions, a no‑cost AI coding teammate, enters beta. It triages issues, reviews pull requests and executes tasks on demand, embedding agents directly into workflows. Blog Post>
Upskilling | Learn This Week | OpenAI Academy
OpenAI Academy — role-based learning paths from OpenAI
Outcome: Free, expert-led modules and collections (e.g., OpenAI for Business, Developer Build Hours, Sora Tutorials, ChatGPT at Work) plus prompt packs and videos to level up real-world use of ChatGPT/GPT-5 across roles. Fresh resources include GPT-5 for Builders (tools/parameters cookbook, prompting guide, frontend coding with GPT-5) and an Intro to GPT-5 resource hub; aligns with OpenAI’s push to scale Academy as a public learning hub. Press Release> and Academy>
Why it matters → Official, always-current guidance from the source—ideal for teams standardising workflows and prompt practices on GPT-5.
Maker Note | What I built this week
This week I was reworking sections on www.aigenticlab.com website with GPT-5 and Windsurf. For exactly $0, I was able to rebuild the whole website in about 2 hours. I was pleasantly surprised by how good and helpful GPT-5 is, although it still needs a human touch to get the last 10% of the website to look good, for which a freelancer on UpWork quoted me $40.
Top Moves | Signal and Impact
OpenAI consolidates and opens up — GPT‑5 rolls out with a real‑time router that automatically chooses between “main”, “thinking” and “pro”. This simplifies model selection for 700 million users but sparked backlash from fans of GPT‑4o’s personality. Simultaneously, OpenAI released gpt‑oss‑20B and gpt‑oss‑120B, Apache‑licensed models you can fine‑tune and run on local hardware. Expect more developers to experiment with on‑premise agents while paying for premium reasoning when needed. TLDR; This means you don’t need to pick between private/free and high end model yourself.
Securing the agent ecosystem — CrowdStrike’s Falcon Shield integrated with OpenAI’s ChatGPT Enterprise compliance API to map GPTs and Codex agents to their human creators, detect risky behaviour and automatically block compromised. It expands coverage to more than 175 SaaS apps. Press Release.
Operator’s Picks | Tools To Try
Wide Research (Manus) — A feature that spins up 100+ general‑purpose agents in parallel, each a full Manus instance, to tackle large research. Agents collaborate rather than follow rigid roles, unlocking high‑volume analysis for tasks like comparing the Fortune 500 or exploring GenAI tools. tasksmanus.im.
ElevenLabs Agents now support Chat Mode - text-only mode for its voice agents (enable via config or runtime override); pair with agent transfer to escalate to voice; bring your own LLM support. → One agent across chat + voice, fewer bots to maintain elevenlabs.io
Deep Dive | Mass Intelligence!
I have been inspired by this article https://www.oneusefulthing.org/p/mass-intelligence
Context: why this matters now.
AI is shifting from scarce to ubiquitous: model routers hide complexity, costs keep dropping, and “good-enough” multimodal tools are everywhere. Institutions built for scarce intelligence need new habits, policies, and workflows to adapt to the mass intelligence.
On paper
AI is becoming the default infrastructure: auto-routed models, stronger reasoning in free/low-cost tiers, and image/video tools that follow plain instructions without expert prompting.
Token costs trend down; more capability shows up in free tiers. Net effect: higher assist rate per user at lower marginal cost.
In practice
Day-to-day tasks get faster with acceptable quality. Occasional misroutes/glitches happen, but throughput and consistency beat manual effort for many workflows. I can’t imagine my day without AI.
AI has unlocked the power of automation and with costs dropping, it is available widely. The world has changed.
My take (what to do)
Pick two high-volume workflows. Replace prompts with buttons/recipes, add a reviewer step for sensitive outputs, and publish a plain-English policy on disclosure/provenance.
AI is here to stay, now is the time to lean into adoption.
Spotlight Tool | Genspark.ai
Purpose: all-in-one AI workspace + agentic search. Edge: Sparkpages (multi-agent research pages) plus built-in apps (Docs/Slides/Sheets). It promises to build PowerPoints and documents quickly and easily.
Try it here: genspark.ai
Let me know if you would like me to make a full demo and publish it on YouTube.
Did you find it useful? Or have questions? Please drop me a note. I respond to all emails. Simply reply to the newsletter or write to [email protected]