Building with AI and agents across software, markets & business experiments. Founder @Moltgate • AI Agents community 17k+
Internet
Joined March 2014
- Tweets16.8K
- Following483
- Followers12.1K
- Likes1.4K
Before AI it took me 2 days to make a website.
Now i make 2 websites in 1 day.
Productivity gain since 2023: +300%
—-
The bad part now:
The money i get from a website has also been slashed by more than 75%
—-
Results:
- great pleasure to see work results faster
- work as long to make less money
- feeling exhausted from a different kind of work: taste and decision-making instead of coding and designing
=> don’t make websites for a living unless you can really extract a lot of value!
In the age of AI, the best UI is no UI.
Or so says Sam Altman.
Having built multiple products and services and using 10s of agents, i somehow agree for single services/chats…
but as soon as you start to have 3-5+ serious and rather deep, technical or mass-data agents / chats, a UI becomes indispensable.
=> No UI is great for simple, single or superficial tasks. Otherwise you need a new type of UI built for collab with AI.
How a SECOND BRAIN works:
1. Keep your agents’ knowledge in an Obsidian vault or GitHub repository of markdown files, out of the agent’s chat so you can use multiple agents.
You may want to keep multiple vaults or repos for your different projects or to separate personal and business concerns.
2. Give access to your agent(s) (multiple instances and/or providers), only to what they need, so that you don’t have to teach them about the same projects and preferences, save tokens, and not clog context windows.
3. Explain the system in 1 central AGENTS.md per vault/repo for all concerned agents to read
4. Have one end-of-day agent write down the day’s work in a dated file: YYMMDD-daily-work-log.md, and update the other shared files as projects change.
=> Your next agent should inherit useful project context without needing your entire chat history + your project context stays up-to-date
BONUS: have another agent feed previous day results, daily news updates, competitor scans, etc. to automate the work of the day.
I asked 100 AI personas about new products and services ideas through 2 different LLM runs.
The #1 result?
People usually don’t want another product/service!
Both GPT 6 Luna and @grok 4.7 labelled most of the time say “no new service.”
What do they want?
A better checklist. A simpler workflow. One less thing to manage.
The new Simulation PM is a double test run of 100 synthetic personas going through a questionnaire on B2B AI business:
- 100 personas * GPT6 luna medium
VS
- 100 personas * @grok 4.7 high
Insights coming soon… 😉
Testing a new run with Simulation PM.
I am checking if a cheaper model (GPT 6 luna) can lead to good results in the agentic b2b space.
I ran a questionnaire through 100 synthetic personas with 200k+ personal income yesterday.
My Simulation PM agent asked them what product or service they would need.
#1 result:
A shared calendar for the whole family to keep track of who does what, who goes where, who manages what chore.
It used to take me 3 days to build a small static website ready to receive content.
Now i build 1 site per day, with an app behind it and systems to sell.
Is that a 300% increase in productivity since AI?
I ran a questionnaire through 100 synthetic personas with 200k+ personal income yesterday.
My Simulation PM agent asked them what product or service they would need.
#1 result:
A shared calendar for the whole family to keep track of who does what, who goes where, who manages what chore.
I built 2,790,132 synthetic personas.
Now I'm putting them to work.
The first AI-agent simulations are running: testing offers, exploring customer behavior, and experimenting with synthetic markets.
I'll let you know if it's better than regular LLM ideas.
simulation.pm
The AI dev loop:
1. Deploy your coding agent dev
2. User sends error
3. Paste error to coding agent
4. Deploy agent’s patch
5. Tell user to try again
=> you need users
I’m opening a new frontier of business and experiments for AI and agents with a simulations project management platform.
Come check it out and give me your thoughts.
👉 simulation.pm
The problem of loops and graphs for AI agents is that you can never be 100% the outcome is exactly what you asked.
So what wiggle is acceptable? How much leeway do you give agents to do things their way before you actually check the output?