I published the first post in my weekly notes series last Saturday and it made me a feel a little more accountable, responsible towards “publishing” or writing.
Here’s my curation, memory dump and reading list update for July 4th, 2026.
Personal AI + Shopping
The most exciting signal this week was a paper published by researchers from Alibaba Taobao. This paper validates a few directional assumptions and bets at Seek, first, owning the intent layer and second - the mismatch in generic LLM recommendations caused by dependencies on semantic + embedding powered search.
Although the paper is focused on training a foundation model, my main takeaway (for Seek) is that we can focus on understanding and learning intent + context while the foundation models can improve on semantics and reranking.
Inspired by Cluely/Cheating Daddy and the ambient apps trend last year, I had built a small macOS tool that took macOS screenshots every few minutes and analysed the image to maintain a work-log of the day.
Discovered Dayflow around that time, it was very early, much smarter and deep (screen recordings) but I found it difficult to trust so I never used it.
Good to see that Dayflow has crossed 6k stars on Github.
Another personal AI use case that appeared a few times on the feed this week was “personal newspaper”. I’m sure every note-taking enthusiast has explored some or the other tool from the earliest versions of Pocket, Evernote, etc to now building personal AI workflows for new curation, podcast summaries, etc.
For a long time, my reading workflow was Newsletters turned into Magazines by Instapaper and delivered to my Kindle. I am still hoping to find some time and budget to buy that 7/10” e-ink display and build my personal AI display.
Last week, I mentioned about having agent-ready identifiers so you can one day let agents talk to each other. WhatsApp announced usernames, I’m guessing this is aimed at allowing integrations in the future because WhatsApp definitely missed the AI Agents bus (Telegram won) while dealing with lawsuits trying to keep AI Agents off the platform. The SIGNALS are here, it’s time Meta makes a move.
WhatsApp’s session based pricing is right. In 2015, Facebook ‘M’ was the earliest AI personal assistant bet. It was a hyper-capable text concierge accessible through Messenger. (Microsoft also had an experiment called Ruuh)
Text messaging in existing channels like iMessage and Slack has been strongly validated as the most acceptable user experience for AI Agents.
With Kunal Shah at the helm and if AI Gods finally favour Meta, they could totally own the market with a WhatsApp based general purpose personal AI Agent.
I used Poke in a group chat with Rhutali last year and it spooked us by using our marriage as context - it discovered the “fact” from our emails with properties/planners and inferred the date we got married.
Frontier Agents
[ Udayan] GrowthX is hosting possibly the largest Hermes Buildathon across 10 cities.
Fable 5 is back and the most recommended use case is to make it write plans, architecture, skills, etc. and then use cheaper models to execute.
Good callback to last week’s note on using Frontier models for planning and task design - cheaper ones for execution and then a competing frontier model for adversarial review.
For the last 3 months, I have been keeping all Seek related work in a single repo with verbose work logs and Agent chat summaries. This includes flutter, nextjs apps, eve agents and a lot of docs.
In Supernova, I imagine a simpler git-based work plane where all my work and time is spent in the main knowledge repo and committed to the master repo as we go. Cloud agents could then pick the approved work and work in the smaller repos while maintaining all work log, statuses and knowledge in the main repo. I am not confident with sub modules and linked repos; so, not trying this yet but good bookmark for future.
