The AI Digest

22/06/2026

Have Perplexity solved AI memory?

Not quite. But what they launched this week is worth understanding.

Memory is one of the most significant unsolved problems in AI. Most tools start every session completely fresh. No awareness of what you worked on yesterday.

No ability to learn from what went wrong last time. No way to get better at the job the way any human worker would over time.

Perplexity has launched something called Brain for its Computer agent product. The approach is different from how most AI handles memory.

Rather than remembering things about you - your preferences, your working style - Brain remembers what the agent actually did.

What approaches worked, what sources were dead ends, what corrections you made.

Overnight, it reviews everything from the day's sessions and updates a kind of personal wiki for the agent.

Next time it starts a task, it has a better map of your world than it did the day before.

Early results are encouraging. Correctness on familiar tasks improved by 25%. Recall went up 16%. The cost of tasks that require pulling in historical context dropped by 13%.

That is not a solved problem. True AI memory, the kind that would let an agent build genuine expertise over time the way a person does , remains one of the harder challenges in the field.

But the framing is interesting. The shift from "remember the user" to "remember the work" is a more useful model for what AI agents actually need to get better at their jobs.

If those performance improvements compound over time as Perplexity suggest, the practical gap between an AI that forgets everything and one that genuinely learns could start to close.

Brain is rolling out now for Max and Enterprise Max subscribers in Research Preview.

Brain is Perplexity's self-improving memory system for Computer that learns from past work to deliver faster, more accurate results over time.

22/06/2026

Anthropic just gave its AI coding tool the ability to show its work.

Claude Code, its AI assistant for developers, can now turn a working session into a live, shareable web page that updates itself as the AI works. Anthropic calls them Artifacts.

Here is what that looks like in practice. An engineer asks the AI to investigate a production problem before their morning standup.

The AI works through the logs, identifies the suspect code, and builds a page: a timeline, the key findings, an error-rate chart. She shares the link with the team.

By the time the meeting starts, the AI has already updated the page twice as the investigation progressed.

Teammates open the same link. They see the same picture. Nobody has to ask "can you walk us through what you found."

The pages are private to your organisation by default and update automatically each time the AI publishes a new version.

Everything on the page is built from the AI's own session context - the code it read, the tools it used, the reasoning it worked through.

A few of the suggested uses: incident timelines for engineers dealing with outages, PR walkthroughs for code reviewers, dashboards showing what a team shipped that week.

It is in beta for Team and Enterprise users now.

Preview your in-progress work in Claude Code as a live, interactive artifact—built from your full session context and shareable with your team.

22/06/2026

Will we get access to Anthropic's Fable model back soon?

Trump states he no longer sees Anthropic as a national security threat.

Whilst there has been no official confirmation that the restrictions are being removed, hopefully this is a step in the right direction.

Source:

Trump told Axios that Anthropic has "behaved very responsibly" and signalled he may ease restrictions on its Fable 5 and Mythos 5 AI models.

20/06/2026

You can know what your AI bill is going to look like before it arrives.

Most companies don't bother. They wait for the invoice, then ask questions.

With Copilot Cowork now billed on a pay as you go credit system, there's actually a simple way to estimate your monthly cost in advance.

Here's how.

Open the Cowork usage report in your Microsoft admin centre. It shows you how many sessions each user has had this month.

Multiply that number by 5. Most sessions involve roughly five prompts back and forth, so this gives you a rough total prompt count.

Multiply that by 400. That sits around the middle of Microsoft's light to heavy task range, so it works as a sensible average credit cost per prompt.

Divide the result by 100. Pay as you go pricing is one cent per credit, so this turns your credit total straight into dollars.

Say someone has had 20 sessions this month. 20 times 5 times 400, divided by 100, lands at roughly $400.

Run that for every user and you have an estimate of your spend, not a guess you make after the fact.

It won't be exact. Some tasks are light, some are heavy, and real usage will sit either side of that average. But it gives finance a number to plan around instead of a surprise to react to.

While you're in the admin centre, set spending limits on individual users too. If someone's usage runs high, you want to see that on a dashboard, not in a budget meeting.

20/06/2026

Every signup form you have ever filled in was built on one assumption. A human is on the other end, clicking the boxes.

That assumption just stopped being true.

AI coding agents, the tools that write and launch software on their own, kept hitting the same wall. To put anything live on the internet, you normally have to sign up first. Type an email, set a password, maybe confirm a code sent to your phone.

Simple enough for a person. But what happens when there is no person?

An AI agent working alone has no hands to click "verify" and no phone to receive a text. It just gets stuck, like a delivery driver showing up to a locked building with no buzzer.

Cloudflare, the company behind a huge chunk of the internet's infrastructure, just removed that lock. An AI agent can now build something and put it live online immediately. No signup. No password. No human required at any point.

The catch is it only lasts 60 minutes. Think of it as a locker at the gym. You can use it straight away, but if nobody claims it before time runs out, everything inside gets cleared automatically.

If a person likes what the AI built, they click one link within that hour and the temporary account becomes permanently theirs.

This is not really a story about one company's developer tool though. Cloudflare also recently teamed up with Stripe so agents can set up paid subscriptions on a person's behalf, without anyone typing in a card number. It joined a separate effort that lets agents create accounts on other websites the same way.

The pattern is the same every time. Passwords, signup forms, even "prove you are not a robot" tests were all built around the idea that a person is doing the clicking. AI agents now handle entire jobs from start to finish with nobody watching, so those old gates are quietly being redesigned to let them straight through.

Worth sitting with. The next time something appears online, finished and live, there may not have been a single human click anywhere in the process.

Source: Cloudflare blog,

The moment an agent needs to deploy something, it slams face-first into a wall built for humans. Today we're rolling out Temporary Accounts on Cloudflare Workers. Any agent can now run wrangler deploy — temporary and get a live Worker in seconds.

19/06/2026

You can feel like you're getting sharper at spotting fake news while actually getting worse at it.

That's not a hypothesis. It's now a measured result.

A new four week MIT study tracked 67 people trying to tell real news headlines and images from fake ones, some with an AI assistant helping, some without.

With AI in the loop, people called it correctly 21% more often. That part is no surprise.

Here's the part that should give you pause.

By week four, when the same people judged things alone, their accuracy had dropped 15.3% from where they started.

About a quarter of them believed they were getting better the whole time. They weren't.

So should you ditch the AI? Not quite.

The researchers found something more interesting buried in the data. It mattered enormously how the AI delivered its answer.

When it just handed over a verdict, fake or real, people switched their own judgment off and went along with it.

When it pointed them toward clues and asked them to look closer instead, their thinking stayed switched on.

The problem isn't AI helping you decide. It's AI deciding for you.

Next time you ask a chatbot whether something online is real, ask it to show you how it knows, not just what it thinks.

Your judgment in a month will thank you.

Source: The Guardian, reporting on the MIT study (https://www.theguardian.com/us-news/2026/jun/19/chatbots-critical-thinking-skills)

18/06/2026

Catching out Ai with a joke for five year olds 😂

18/06/2026

AI just got more expensive to use carelessly.

Microsoft Copilot Cowork went generally available this week, and the detail most coverage is glossing over is how it charges.
You pay per task.

Not a flat subscription add-on.

https://www.microsoft.com/en-us/microsoft-365/blog/2026/06/16/copilot-cowork-is-now-generally-available/

Every job Cowork runs draws down Copilot Credits based on how much work it had to do to complete it.

Microsoft breaks tasks into three types: light, medium, and heavy.

Light tasks pull from a small number of sources and produce a simple output. Medium tasks draw on multiple sources and generate a few outputs. Heavy tasks go broad, reason deeply, and produce a lot.

Think of it like a taxi meter versus a monthly bus pass. Before, the cost was bundled. Now the meter is running.

That changes how you should use it. Before firing off a request, it is worth ten seconds asking: what do I actually need here? Matching the ask to the job is how you keep costs under control.

Microsoft has also built in spending limits at the tenant, group, and user level. So businesses can set a ceiling before things run away.
This is where AI pricing is heading across the board. Worth understanding the task types now, before the bills arrive.

10/06/2026

Ever wondered what the goodest boy to ever live would look like?