What's actually happening in AI, without the hype.

tools

Tools

Agents: what they are, and what they aren't yet

An agent takes steps on its own — searching, opening, drafting, checking. That's genuinely useful, and also why they need supervision: they can be confidently wrong several times in a row.

Use agents for work you'd review anyway, not for anything irreversible.

Tay's briefing

models

Models

What a “bigger context window” actually changes

A bigger window means more material at once — a whole book, a year of email. But models pay less attention in the middle of a very long input, so more room isn't the same as more care.

Put the thing that matters most at the start or the end, not buried in the middle.

Tay's briefing

reality

Reality

Voice cloning raised the price of a phone call

A convincing cloned voice now needs a few seconds of audio. Banks and families have both been caught out by it.

Agree a family password, and never move money on the strength of a voice alone.

Tay's briefing

reality

Reality

Chatbots still can't count on their own

Language models predict words, they don't tally objects. Counting rows, letters or items in a list is a known weak spot, and it fails politely and confidently.

Ask it to show the count step by step, then check the number yourself.

Tay's briefing

money

Money

Prices keep falling, and that changes what's worth trying

The cost of a given amount of AI work has dropped sharply over the last few years, which is why features that were impossible last year are now free buttons.

If a tool told you something was too expensive a year ago, it may be cheap now.

Tay's briefing

policy

Privacy

Every company keeps rewriting its own rules

Terms, defaults and settings change every few months, usually quietly. What you agreed to a year ago is probably not what's in force today.

Re-check Data controls in the tool you use most, four times a year.

Tay's briefing

models

Models

Why every AI now has a “thinking” mode

The big chat tools increasingly offer two speeds: a quick answer, and one that pauses to work through the problem first. The slow one is better at maths, planning and anything with several steps.

If an answer feels shallow, try the slower mode before you rewrite your prompt.

Tay's briefing

reality

Reality

Why AI images still can't spell

Image models learn the shape of letters, not the rules of language. Short words are getting better; a headline is still a gamble.

Generate the picture without text and add the words in a design tool.

Tay's briefing

reality

Reality

The difference between fluent and correct

These systems are trained to sound right, not to be right. Fluency is what makes errors hard to spot — a wrong answer reads exactly like a right one.

Ask for the source, or ask it what would prove it wrong. Confidence is not evidence.

Tay's briefing

policy

Privacy

The quiet fight over where your chats go

Most tools train on conversations unless you turn it off, and the switch sits in a different place in each one. Paid tiers usually change the default, not the setting.

One minute in Data controls is the difference between a private habit and a public one.

Tay's briefing