Tell it what to learn

Subjects to keep learning about

Add a subject such as animals, space or money. Vishnu then keeps working through more and more topics inside it (elephants, whales, migration...), one after another, until you remove it.

It plans to learn

Wondering and trying by itself

What it chose to do without being asked: puzzles it made up (marked by code), things it guessed before reading, and questions it asked itself. Two switches in Settings turn this on and off. Nothing it guesses or answers is trained on; only the source's words and the correct answers are.

What it has learned

Each topic is kept to a line or two, so the library stays small. Clear notebook

Add files

Text, Word, Excel, PowerPoint, PDF, OpenDocument, e-books, web pages, CSV, JSON, zip files and more (see “What Vishnu can read” under Learn next). Up to 8 MB each. Adding a file with the same name replaces the old one.

Add a whole folder

What it knows from your files

Try a question

What Vishnu remembers about you

Forget everything about me

Answers you taught it

When you press Teach under a reply, your answer is used at once for the same question, and Vishnu also studies it. Delete one here and it is forgotten.

Its library

For a question like “what is a volcano?”, Vishnu looks in everything it has read first and shows the passage with its source and licence. If it finds nothing, it can look up one page on Simple English Wikipedia or Wiktionary (you can switch that off in Settings) and keeps what it read.

Brain backups

A copy of the brain is saved after check-up bests and every few study rounds. If it ever gets worse, restore an earlier copy. Restoring first saves the current one.

Save a backup now

Things to read

All subjects

Wikipedia keeps a hand-made list of the 10,000 most important articles in every field (people, history, geography, the arts, religion, daily life, society, biology and health, physical sciences, technology, mathematics). Loading it needs the internet and only fetches the list of names (11 requests). It also loads by itself once, when the Sources page lets Vishnu look things up.

By subject

Find something to study

Type a subject. Vishnu searches all the sources below and lists what it found. Tick what you want it to study.

Where it looks

What the finder did lately

• Nothing yet.

Queue

How long to study it

A topic

Built in and offline: maths, grammar, english, animals, body, computers, conversation, coding basics, everyday facts. Any other topic is read from Simple English Wikipedia, which needs internet.

Browse topics

Vishnu already picks topics by itself. Tick any you want it to study sooner and press the button: they go to the front of its list (they are read from Simple English Wikipedia and the built-in lessons, so they need internet). A tick on "studied" means it has already done it.

Files from your computer

Text, Markdown, Word (.docx, .doc), Excel (.xlsx, .xls), PowerPoint (.pptx, .ppt), PDF, OpenDocument, Rich Text, e-books, web pages, CSV, JSON, XML, notebooks, subtitles, Python and zip files. Up to 8 MB each. Vishnu looks inside the file, so a wrong ending such as a Word file named .txt still works.

Tests on my files

Vishnu tests itself on every file you give it, whether you add it here, in a watched folder or by a link. Code makes about 8 questions from the file’s own sentences. Vishnu answers alone, from its own brain: it does not look the file up and no outside AI helps. Plain code marks the answer by words from your file. The questions are asked once before it studies the file and again after, so the difference shows what it learned. What it gets wrong is studied again and asked again a few rounds later. A brain this small keeps little of what it reads, so low scores are normal. A score says it can finish or define things from the file, not that it understands the file.

Your folders (watched)

Add folders from anywhere on this computer. Vishnu looks in them every few minutes and learns from new or changed files: text, Markdown, Word, Excel, PowerPoint, PDF, OpenDocument, e-books, CSV, JSON, web pages, Python, subtitles and zip files, and pictures that contain printed words (scans, slides, screenshots). Copy or save a document into one of these folders and it is picked up by itself and put in the queue above. Files are only read, never changed or moved. Hidden and system folders, and files that look like passwords or keys, are left alone.

A link: web page, document or GitHub

Paste one or more addresses, one per line. A web page or a document (text, Markdown, HTML, Word, PDF) is read and studied. A GitHub address can be a whole repository, one folder (…/tree/main/docs) or one file: Vishnu reads its README, documents and Python files. Only public addresses are opened, and a site that asks programs to stay out is left alone. You can also paste a link straight into the chat.

A video

The first run of HumanEval or MBPP downloads the problem file once from GitHub. Their problems are kept out of Vishnu’s training on purpose. Training slows down while a test runs.

Score over time

Where it learns from

Everything is checked for licence, secrets and quality before it is kept. Online sources are off until you turn them on.

Teach it directly

Your own examples count five times as much as downloaded text. Write pairs like this, or paste many at once.

What is in its library

Activity

Start over

Erases what the network has learned and its vocabulary. The collected library stays. Type RESET to confirm.

Local models: run an AI on this computer or server

Install a model here and Vishnu can use it as the examiner in the Study group (you choose it there). It runs on this computer only, so the questions and Vishnu’s answers stay here. No key is involved. Choose how it runs:

How connected services are used

1. A model for the examiner

Models that run on this computer or server are installed on the Connections page: Vishnu’s own engine (llama.cpp) or Ollama, from small laptop models to 70-billion-parameter server models, or any model you add by name. LM Studio and other servers are connected there too. Whatever is installed shows up in the list below.

2. The examiner

3. A session

4. Watch the group chat

Everything the examiner and Vishnu say to each other, live, on its own page: each question, Vishnu’s answer as it is written, the examiner’s verdict and the correct answer, and small notes about what gets learned. You can join in and ask the group a question too.

5. What a mistake leads to

The question is asked again in a later session, once Vishnu has studied the practice for its subject (two right answers in a row and it is done; five misses and it is set aside). Its topic goes on the list Vishnu reads about from open, licence-checked pages. And a short practice course made from Vishnu’s own built-in lessons is queued for that subject; at most four such courses are open at a time, the most-missed subject first.

The examiner’s answers are never given to Vishnu as lessons. The examiner only asks and marks. Vishnu is not shown the right answer to the question it missed; a gap counts as learned only when Vishnu itself answers the same question right twice in a row after studying.

Score by subject

Gaps open is the number of different questions Vishnu missed in a subject that are still waiting to be learned. Relearned reads like “2 of 5”: of the 5 questions missed in that subject, 2 are learned for good, meaning the examiner asked the same question again and Vishnu answered it right twice in a row. Each missed question is one gap, so “4 gaps” means 4 different questions, not 4 rounds of learning. Why Relearned can stay at 0 for a while: a missed question is only asked again after the practice for its subject has been studied (asking it earlier could only fail again), only a few practice courses are studied at a time, and it has to be answered right twice in a row. The line above the table shows where every missed question is on that way. The network is very small, so many will need several rounds of practice. Thrown away counts questions the examiner wrote that were unusable. Right by brain counts the right answers that Vishnu's own neural network wrote alone. In the Study group nothing else can answer: the question goes straight to the network, with no web search, no library, no calculator and no help from the examiner, and the examiner only marks the answer afterwards. The line under every answer in Group chat says so.

There is no website that tests and trains a model that lives on your own computer. The Benchmarks page runs public tests (HumanEval, MBPP) on Vishnu, and this page does the same job with an examiner of your choice.

Steps

Workspace files

Sign-in and server

Sign-in is off. That is normal on your own computer: you open Vishnu with its private link and see everything. To let other people use Vishnu from a server, start it with --accounts: you then sign in as the administrator, people can register, and they see only the chat and their own settings. The README has the step-by-step server guide.

Training speed on this computer

Study folder (big downloaded datasets)

Download datasets yourself and put the files in the study folder, in the sub-folder named for each one (the folder has a README.txt). Vishnu reads a few megabytes of each per round and remembers where it stopped. It never changes or deletes your files, and it does not copy a whole dataset: its library keeps only what it has read so far, up to the library size limit above. Its brain file stays the same size however much it reads, because it learns from the text and does not store it. Stories written by an AI stay off, and a dataset whose licence is not in your allowed list waits until you add it.

Personality

This changes how greetings and small talk sound. It does not change how smart it is, and it never pretends to be a person.

What it should become good at

Words, grammar, conversation and numbers always come first. The focus changes what comes next and what the source finder looks for.

How much of this computer Vishnu may use

This limits the processor only; memory depends on the brain size and the examples per step. Automatic uses all cores but one at low priority, so other programs still come first. A smaller share leaves room for your own work, a bigger one trains faster. The computer counts as idle when there has been no key press or mouse move for the time below and other programs are hardly using the processor (on a server with nobody at the keyboard, only the second part is checked).

Be gentle with my laptop

Alerts by e-mail

When Vishnu stops, waits because the computer is almost out of memory, gets stuck or keeps failing, the administrator gets an e-mail, and another one when it is learning again. Never more than one mail of a kind in the time you set. A program that has been closed by the system cannot write a mail itself, so the Windows dashboard, the Linux service or the next start of Vishnu sends it. The mail never contains passwords or keys.

Schedule and check-ups

Days it may learn (combined with the hours above):

Changing this never stops a topic that is already running. A higher number starts extra topics beside the running ones at the next study round. A lower number only stops new ones from starting until some finish. All topics share one brain and one study round, so the more you run at once, the slower each one moves.

A check-up asks every question on the Progress page and takes a minute or two. Fewer rounds between check-ups gives a score more often but slows learning a little.

Tests on my files

Thinking by itself

Two switches, each looked at again every study round. Vishnu does not feel curious or brave: these are rules that choose what to do next from its own records.

Senses: hear, speak, see (all off)

Switched off until you turn them on. Hear and speak use your browser's own speech tools, so nothing is installed and Vishnu receives only the words, never the audio. In Chrome and Edge the browser itself sends your voice to its maker's online speech service to turn it into words, so hearing needs the internet there. See reads printed text in a picture you attach in Chat; it needs extra software. Vishnu's own brain does not understand photos: it measures colours and layout and reads the words in them. For a real description you can add a vision model below.

Example: /do click at 500 300, /do type hello, /do press enter. Vishnu never moves the mouse on its own; every action comes from a message you send. Needs pip install pyautogui. Moving the mouse into the top-left corner of the screen stops it at once. Max 30 actions a minute; Alt+F4, Win+R and similar keys are blocked.

This computer

Checking…

Backup and move

One zip with the brain, vocabulary, settings and memory. It never includes API keys. It does contain your taught answers and the facts Vishnu remembers about you, so keep it private. The lite file leaves out the downloaded library (it is read again by itself). The full file keeps everything and can be large.

Download backup (lite)Download full backup

To restore on this or another computer: close Vishnu Studio, then drag the zip onto IMPORT_VISHNU_BACKUP.bat, or run python -m vishnu_core import vishnu_backup.zip. Your old files are kept in a folder called before_import_….

Size of its brain

Adds layers to the brain you have, right now. They start as pass-through layers, so nothing it learned is lost and nothing is reset; they begin learning in the next round. The limits above still apply (0 = none). Its width can only change with a reset.

Share a chat page on a website

Starts a separate page where visitors can talk to Vishnu's own model only. It has no keys, no teaching and no settings, and each visitor is rate limited.

Run python -m vishnu_core studio --public --host 0.0.0.0 --public-host chat.yourdomain.com behind an HTTPS reverse proxy. The page is at /public.

What to expect

Vishnu is a small neural network that trains on this computer. Early answers will be rough. It improves with more material and time, but it will not match large commercial models, which are trained on thousands of times more data.