Bitcoin Mining Labby Coinovo

Mining Lab · Coinovo Academy · AI and crypto

AI and crypto

A tool, not an oracle

Intro 11 min 7 cards 18 questions

Objectives

What a language model actually does

An AI language model has learned from a vast amount of text which words are likely to follow each other. From that it produces answers that sound fluent and knowledgeable. It does not automatically look things up in a database, and its knowledge ends at a cut-off date unless it is allowed to search the web. It is strong at explaining, summarising and rephrasing.

A language model produces plausible text, not checked facts.

Hallucinations

Sometimes a model invents things: figures, dates, features, even sources with a title and address. The tricky part is that wrong answers sound just as confident as right ones. Ask for the exact date of a rule change and you may get a precise-looking but wrong answer. Knowing this, you treat AI as a draft, not as evidence.

Convincing is not the same as correct.

Full explanation

Why models make things up

A language model is trained to deliver a fitting continuation, not to stay silent when it does not know something. If you ask a question for which it has learned no reliable answer, it still produces text that looks like an answer. That text is assembled from patterns that were common in similar texts.

Precise details are especially vulnerable: years, percentages, names of laws, version numbers, web addresses. A citation can look completely genuine, with a fitting title and a believable address, and still not exist. On topics that change quickly, such as tax rules or fees, the model is also often wrong, because its knowledge ends at a cut-off date.

Newer systems can search the web and read documents. That reduces errors but does not remove them. Even a source that was found can be summarised wrongly. The basic rule stays the same: the more depends on a statement, the more important it is to check it at the source yourself.

How to check answers

First: ask for the source of every important statement and open it yourself. Check whether the page exists and whether it really says what is claimed. An address that leads nowhere is a clear sign of an invented reference.

Second: compare with a primary source. For the rules of a network that is the technical specification or the original paper, for tax questions the law or the tax authority of the country, for warnings the financial regulator. Asking a second AI whether the first one is right does not replace this, because both may have learned the same mistake.

Third: ask follow-up questions. Ask the model to name the weaknesses of its own answer or to argue the opposite position. If it starts to flounder or changes its statement without a new reason, the first answer was not reliable. Numbers that matter are better recalculated yourself.

No price predictions

Prices depend on future events and on the behaviour of many people who themselves react to new information. If there were a tool that reliably predicted prices, everyone using it would act on it at once, and the predicted advantage would vanish. When a language model says a price will rise, that is a sentence that sounds plausible, not an insight.

AI trading bots, realistically

A trading bot is a program that buys and sells according to rules. Whether those rules were written by a person or found with AI does not change the fact that fees, price swings and bad luck are part of it. A bot can be disciplined, but it cannot know the market. Anyone promising fixed returns, with or without AI, promises something that does not exist.

Full explanation

What a bot really is

Behind the word bot there is usually something simple: a rule such as "buy when the price falls below an average, sell when it rises above it", carried out by a program that never sleeps. AI can help find such rules or analyse data. It does not turn a rule into a certainty.

Every trade costs money. There are fees, plus the gap between buying and selling price, and in fast markets a worse price than expected. A bot that trades often pays these costs often. With 20 trades a month at 0.2% cost each, that is 4% a month that has to be earned before any profit appears at all.

On top come risks that have nothing to do with the market. Many bots need access to your exchange account via a key. If that key has the right to withdraw money, an attacker or a dubious provider can empty the account. Anyone using a bot therefore gives it trading rights at most, never withdrawal rights.

Backtests and their traps

A backtest runs a strategy over past prices and shows what it would have earned back then. That sounds like proof but is easy to bend. Anyone who turns the settings long enough will always find a combination that looks brilliant in the past.

This problem is called overfitting. The strategy has then not learned the market but the coincidences of one particular period. When a new period comes, the coincidences no longer fit and the results collapse. The more adjustable settings and the more impressive the historical figure, the greater the suspicion.

A more honest test uses data that was not used during development, with realistic fees and without polishing afterwards. Even then a good past is no promise for the future. Anyone advertising a product with a backtest without disclosing these points is selling a story, not a result.

AI tokens, soberly

Whenever a topic is in demand, tokens appear that carry it in their name. But the word AI in a name says nothing about whether the project actually uses AI. Helpful questions are: is there already a usable product? Does that product need a token of its own at all? And who can change the rules and the supply of the token?

A buzzword in the name is not a product.

Deepfakes as a scam tool

AI can rebuild videos and voices deceptively well. Scammers use this for fake appearances by well-known people who supposedly give away crypto if you send some first. Even more dangerous are cloned voices of relatives urgently asking for money on the phone. An agreed code word helps, as does calling back on a number you saved yourself.

A real-looking video is not proof.

Using AI well for learning

Used well, AI is a patient tutor. Have it explain terms, work through examples or quiz you. Ask for counterarguments to an idea that excites you. Summarise a whitepaper and then read the key passages yourself. But never give an AI your recovery phrase or private keys, and do not let it prepare a transaction for you that you have not checked.

AI for understanding, sources for checking.

Glossary

Language model
AI that produces plausible answers from text patterns.
Hallucination
Convincingly phrased but wrong or invented statement by an AI.
Knowledge cut-off
The point in time up to which a model's training data reaches.
Backtest
Running a strategy over past prices.
Overfitting
A strategy fits the coincidences of the past rather than the market.
Deepfake
Video or voice faked with AI.

Further reading

This module comes with 18 explained questions. They live in the app, along with your progress and the credit for finishing.
The Coinovo appA reason to come back every dayTasks, shifts and the daily hunt, plus an account that shows what has added up.Daily tasksYour own accountCosts nothingStart now

More modules

What Bitcoin isMoney without a middlemanHow mining worksGuessing until it fitsHashrateHow much guessing goes onThe difficultyWhy the rhythm holdsThe halvingThe reward shrinksWallets and keysWhat you never hand overFeesYou pay for spaceMining poolsTogether instead of aloneSpotting fraudThe same tricks every timeTokenomics, taken apartBitcoin as the worked example, projects under testEthereum and smart contractsA shared computer with a fee meterProof of stake and stakingCollateral instead of electricityStablecoinsStable as long as the backing holdsDeFi: financial services made of codeSwap, borrow, lend without a bankThe Lightning NetworkBitcoin instantly, with conditionsSpotting and stopping scamsCheck first, then signInvesting with a planRules instead of gut feelingRunning mining as a businessWhat a machine really costsSelf-custody in practiceFrom the first key to inheritanceTax and bookkeepingWhat the tax office wants to seeHow the network really worksNodes, mempool, forksRisk and position sizeHow not to lose everything