01
Language models
How large models read, reason and write — and where they still fall short.
Open questions
- Where does a fluent answer stop being a finished task?
- How should a model show what it isn’t sure of?
Research Independent
Prabhat Tiwari
Useful on the second day, not just impressive on the first.
Research on language models, AI agents, open networks and play — read through a background in finance, and tested by building.
02Identity
Identity
I’m Prabhat Tiwari. I build small things on the web, test ideas about technology and AI, and write down what I learn — in English and in हिन्दी.
My academic background is in finance — a Bachelor’s and a Master’s — and most of what I study now sits where that meets software: AI that acts on our behalf, networks where people hold their own keys, and products that have to earn a second visit.
- Focus
- Language models, AI agents, open networks, interactive media.
- Grounding
- A Bachelor’s and a Master’s in Finance.
- Practice
- Independent. Research, building and writing, side by side.
- Languages
- English and हिन्दी. I write in both.
“I would rather leave a section honestly empty than fill it with things that are not real.”
03Research overview
Five threads.
One question underneath.
Each thread asks the same thing from a different side: when does a technology become something people can actually depend on?
02
Agents and permission
Software that plans, uses tools and carries a task through — and knows where its authority ends.
Open questions
- What may an agent commit to without asking?
- How does an approval stay attached to one specific choice?
03
Open networks and identity
Blockchains, Web3 infrastructure and digital identity: names, keys and what it means to own something online.
Open questions
- What changes when the thing you hold is a key, not a login?
- Where should a human‑readable name actually take you?
04
Interactive media and play
Games, online play and the communities around them — and worlds shaped by models instead of scripts.
Open questions
- Where do models make play better than scripts, and where worse?
- What if delay were treated as a material, not a bug?
05
Finance meets technology
My academic grounding is in finance. I read AI and open networks partly through it: value, risk, ownership and trust.
Open questions
- What does self‑custody ask of ordinary people?
- When software moves money, where should the check sit?
04Selected work
Selected work.
Everything here is live or openly in progress. Nothing is listed before it’s real.
My personal workspace on the web. I use it to build small projects, try ideas about software and AI, and publish what I think along the way.
- Built around three beliefs: a demo is not a product, agents need a brake pedal, and build for the second visit.
- Experiments start as questions about how software should behave. Each links to a written note; code follows when it’s ready.
- Projects appear only once they’re real. Unfinished slots say “coming soon” instead of pretending.
- A dark front room with glass forms that turn as you scroll.
A personal journal for essays and notes, on technology, Web3, AI and gaming.
- English essays on AI tools and building products.
- Full‑text search, a reading view with contents, a topic archive and local bookmarks.
- Built as a fast static site. No trackers.
A venture and investment organization being built with a long‑term view: to identify exceptional ideas, support ambitious builders and take part in creating meaningful companies and technologies.
- In its foundation stage. Structure, initiatives and activities are still being prepared.
- The public page is written as a letter, not a pitch. It announces nothing that doesn’t exist yet.
- Its working motto: Build. Invest. Endure.
Also in progress
Smaller rooms, each honest about where it stands.
Under construction
OpenAIJobs.in
An independent project in development: a modern place to discover opportunities, organizations, careers and the changing culture of work. Not affiliated with OpenAI.
openaijobs.in
Live · Web3
prabhattiwari.brave
A research profile on a Web3 name (Unstoppable Domains): decentralized systems, AI, emerging technology and interactive media. Opens natively in Brave; mirrored for every browser.
prabhattiwari.brave
Live · Notes to come
AI notes
Four threads on artificial intelligence — language models, agents, AI × Web3 and AI in games — with notes and experiments published as they’re ready.
ai.prabhattiwari.inOne name, several rooms
The prabhattiwari.in network. Each subdomain has one job.
05Thinking & process
How I think
about the work.
Three beliefs, written down so I can be held to them — then the method and the models that follow from them.
01
A demo is not a product.
The useful test starts where the impressive answer ends.
Read the note02
Agents need a brake pedal.
Permission is a product feature, not a paragraph in the terms.
Read the note03
Build for the second visit.
The first visit tests curiosity. The second should test usefulness.
Read the noteA working method
Six steps, each taken from my own notes.
-
01
Start with one task
Pick one task. Define what completion means. Expand only when that contract holds.
From “Your AI demo is not your product”
-
02
Write the contract
Decide which choices the software may make, which need review and how failure should look.
From “Your AI demo is not your product”
-
03
Run the full‑task test
Give the tool a real input, a clear finish line and one awkward exception. Then count the work left for the human.
From “Your AI demo is not your product”
-
04
Prepare freely, commit carefully
Let the software research, organize and draft. Put the check where money, reputation or access changes hands.
From “An AI agent should have a brake pedal”
-
05
Design the return
What should the product preserve? What can it stop asking because the user already answered?
From “Build for the second visit”
-
06
Say what didn’t happen
If an action can’t be verified, say so. Don’t turn a partial result into a cheerful completion message.
From “An AI agent should have a brake pedal”
Mental models
Three distinctions I keep returning to.
Preparationnotcommitment
Research, options and drafts are cheap to redo. A booking, a payment or a sent message isn’t.
Usefulnessnotusage
A backup tool doesn’t need a daily habit. It needs to be remembered, and to work when the need returns.
A suggestionnotan action
Writing a draft and sending it are different jobs. The line between them deserves to be visible.
06Writing & notes
Writing & notes.
Short essays in English on AI tools and building products.
-
01
AI & tools · English · 2 minYour AI demo is not your productThe useful test starts where the impressive answer ends.
- Beyond the demo
- A full‑task test
- A smaller first release
-
02
AI & tools · English · 2 minAn AI agent should have a brake pedalPermission is a product feature, not a paragraph in the terms.
- A visible boundary
- Permission that fits
- Design for failure
-
03
Building · English · 2 minBuild for the second visitThe first visit tests curiosity. The second should test usefulness.
- A reason to return
- Carry the work forward
- Usefulness over usage
“A brake pedal is useful even when the engine is excellent.”
07Timeline
Timeline.
A grounding in finance first. The public record of the work begins in 2026, and the next entries are already named.
-
Earlier
Bachelor’s in Finance
Undergraduate study in finance.
-
Earlier
Master’s in Finance
Postgraduate study in finance.
-
Oct 2026
prabhattiwari.brave
A research profile goes on‑chain, on a Web3 name.
-
Oct 2026
Tiwari Ventures
A public marker for a venture in its foundation stage.
-
Oct 2026
OpenAIJobs.in
An under‑construction page for a project on the future of work.
-
Oct 2026
The prabhattiwari.in network
One name, several rooms: blog, labs, ai, projects and research.
-
Oct 2026
The Journal, rebuilt
blog.prabhattiwari.in relaunches as a proper publication.
-
Oct 2026
Labs, rebuilt
labs.prabhattiwari.in becomes a 3D workspace for projects and experiments.
-
Oct 2026
Research
This site: the threads, the work and the thinking in one place.
-
Next
Prototype notebook
Small working prototypes, with notes on what worked and what didn’t.
-
Next
Agent permissions, in practice
A hands‑on version of the brake‑pedal idea.
-
Next
A small AI tool
Built around clear permissions and a reason to come back.
08Presence
Presence.
Where the work shows up, and what each place is for. No feeds pulled in, no counters — just the doors.
09Collaboration
If you’re working on
something in these threads,
I’d like to hear about it.
A narrow AI tool
You have one task and want an honest full‑task test before you widen it.
Agent permissions
Where approvals sit, what an agent may commit to, and how failure should read.
Keys, names and identity
Open networks, Web3 naming and the habits that come with holding your own keys.
Writing, in two languages
Essays and notes in English or हिन्दी on technology and the people using it.
How it usually starts
A short message: the problem, what you’ve tried, and what “done” would look like.
Start a conversation10Contact
Contact.
The best way to reach me is a message on social media. There is no contact form, so nothing here collects your details.
I read messages when I can, so replies may take a little while.
11Closing
Useful on the second day,
not just impressive on the first.
— Prabhat Tiwari
Back to the beginning