Now | Latest squakings

In the spirit of experimenting with AI tools as I said I would, I have created a new section to the blog which you can find here: https://parrotsandpaperclips.com/daily-brief/

I have curated a list of about 30 news sites that have content relevant to this blog and to my day to day work. Then I put together the following bits and pieces:

1. A python script living on Github in a private repository that fetches content from the sources and sends the headlines and links back to Claude
2. Claude then selects the most significant stories and generates summaries and insights as well as an overall "elevator pitch" that brings together all the stories into one point of view
3. All this then uses the Ghost (blogging platform) API to tag and post the content to my new blog section and finally trigger a newsletter email (sorry only me for now)
4. I have used a Github workflow with a cron schedule to trigger the whole script every workday morning at about 7am

At the time of writing (11-05-2026) after a few manual runs I have one post that is up and published. It is a good first go but I am sure I'll be tweaking it a bit more. I'll be crossing my fingers that the cron schedule runs tomorrow morning and there will be more to see.

It's not yet available to sign up to as a subscriber but that will come later. For now, have a look and let me know what you think.

Just for a bit of fun, do you want to see what an LLM created in the 1930s would say? Check out Talkie in the link above.

All modern LMs are basically trained on the same enormous set of data scrapped from the web and digitised versions of older paper publications. While there are differences we can see between all the major models they do all have very similar behaviours and are all able to reference the same set of data either during training or during inference with access to the web.

Limiting the data-set to pre-1930 text gives a fascinating look, at least in theory, at what might be different in our collective culture compared to back then. You can also investigate (as the creators of the model have): does it anticipate inventions it never saw? Can it learn to code from examples despite having no knowledge of computers? How much of what we think we know about LMs is actually just "things web-trained models do"?

I tried asking a question that is topical at the moment with the re-emergence of One Nation as a political force, at least if you believe the polls. What was surprising was that on running the same prompt multiple times I got responses which were in some ways similar and in others very different. Almost every time I got responses mentioning some variation of:

• the importance of maintaining a standard of living for the current population
• having numbers that maintain "healthful growth", whatever that means
• some mention of considering the differing needs of states/towns based on the main industries
• a focus on settling mainly in country areas

What was different was the way it would oscillate with relatively progressive views "The national origin of the immigrants should be left to individual choice" and some quite less-so "The Latin nations should be kept out. China and India should be absolutely prohibited." See below screenshots of two examples illustrating this.

I tried a few other types of prompts and got a similar mix of responses. It does make me think, is it really just a model based 1930's text or is there some reinforcement learning going on to try and steer away from some of the most controversial statements it might make? The original blog post does note:

talkie reflects the culture and values of the texts it was trained on. As such, it can produce outputs that will be offensive to users.

Give it a go and see what you think!

A couple of weeks ago I posted about starting the Claude 101 online training course. I've now finished it. Didn't take long.

The course covers quite a bit: prompting basics, memory and context, Projects, Skills, Artifacts, and a few other features. Much of it was familiar territory that won't be new to a seasoned Claude user, but having it laid out in a structured way was still useful. It filled in a few gaps and gave me a clearer picture of how everything fits together.

The two areas most worth revisiting were Projects and Skills.

Projects are self-contained workspaces with their own memory, chat histories, knowledge bases, and custom instructions. I have used these already, but knowing how to best work with artefacts and setting your preferences for a project was a good insight.

Skills are essentially reusable expertise packages — folders of instructions and resources that Claude loads when needed. You describe what you want, answer a few questions, upload any relevant materials, and Claude builds a structured skill file you can use again and again. I have already mentioned the Karpathy council of experts skill in my previous post, that is a good one to have a go with. https://parrotsandpaperclips.com/now/#learning-all-about-claude

Next will probably be the Claude Code 101 course. I suspect that one will be a bit more of a stretch as I have only just started on my Claude Code journey on VS Code while learning about Python.

Over the past few years I have been using a mixture of free and paid LLMs. My first paid account was with OpenAI although I stopped that not too long ago and switched to Gemini, mainly to try out Nano Banana and see what all the fuss was about, but also to just get a bit more exposure to a different model that was that basis of the Google Generative AI Leader certificate that I just completed. This was before Trump fired Anthropic "like dogs" for daring to put limits on Claude being used for killer robots and mass surveillance of US citizens.

After that, Sam Altman stepped in to take on the business Anthropic was removed from, which triggered a wave of users abandoning ChatGPT in protest. So you could say I left Open AI before it was cool!

Anyway, I'm having a go at improving my skills with Claude and luckily they have a bunch of free courses available which you can check out here: https://anthropic.skilljar.com/. I am starting with "Claude 101", naturally, so I'll see how it goes and report back. So far, looking at the curriculum, it seems like a good overview of the main ways to think about organising yourself beyond just basic prompts and getting into some of the other features like projects, artifacts and skills. A good place to get started.

No tech or AI today, just something I really enjoy at this time of year.

One of my favourite desserts to make is the humble tiramisu. I love it because it is so simple and has such a lovely mix of textures and flavours. It is also a bit of a crowd pleaser, there aren't many people who don't enjoy a good tiramisu!

In my opinion the secret to a good tiramisu is in:

• Good savoiardi biscuits that have been soaked just the right amount in the coffee/alcohol (not dry, not completely falling apart)
• A really light and fluffy mascarpone mix
• good quality dark chocolate for the top

There are many different recipes out there and many of them are great, traditionalists will say that the mix should be just mascarpone and eggs, no cream! I have tried that version and it is also very good, especially if you have good quality fresh eggs, but I do tend to cheat a bit and add some cream as well. I find that it does lighten things a bit.

Another thing purists will quibble on is what alcohol to use. The most traditional version uses marsala wine, a sweet fortified italian wine. I have also heard that traditional versions will use just Amaretto or Frangelico. My thinking is that I want to have some sweetness to come from the alcohol itself, but pure frangelico or the like can be a bit too sweet, so I tend to mix half grappa and half something sweet. The Nutty flavour of frangelico or amaretto is great, I also like to use something like Mac (a macadamia liqour) for an Australian twist.

The savoiardi biscuits should be italian, something like the Vincenzi ones. Nothing fancy, just the plain ones with the sugar on top. I have tried some Australian brands and the texture is just not right.

Lastly the coffee. I know some people use instant but don't do that. Ideally if you have an espresso machine make a batch of long blacks and use that. Or if you have a stove-top mokka machine use that. Last option would be plunger. But do use good strong coffee, you want to be able to taste it. Worst case go to your local cafe and order two long blacks. But instant just doesn't cut it.

So here is how I do it:

Ingredients

• 1 packet savoiardi
• 1 tub mascarpone (250g) - Formaggio zantti or Montefiore Mascarpone
• 2 eggs, yolks seperated
• 3/4 cup whipping cream
• 70g caster sugar
• Splash of vanilla extract
• 1.5 cups strong black coffee (see note)
• Grappa
• Nutty liqour - Mac/Frangelcio/Amaretto
• Good quality cocoa for dusting
• Grated dark chocolate - Lindt 70%/85%

Instructions

1. Beat the egg yolks with the sugar until the mixture becomes very pale and thick, 5-7min
2. Add vanilla and mascarpone, beat into the mixture until it is well combined and light
3. In a clean bowl whisk the egg whites until stiff. Don't go overboard and split the whites. You can add a small pinch of salt to help them along
4. In yet another bowl whip the cream until it is light and fluffy
5. Slowly fold in the egg whites and cream into the mascarpone and egg mixture. You want to keep as much air as possible so it stays really light
6. In a tray, mix your black coffee and alcohol. Some recipes will say 1-2 tablespoons. This is not tiramisu for ants, go big or don't bother. You need at least 1/4 cup total 😃
7. Now get your savoiardi and dip them into the coffee mix making sure to get both sides. The trick is to not over-soak them so they fall appart, but to not be so fast you end up with a dry centre. I usually wear those disposable gloves for this step so my fingers don't smell like coffee for the rest of the day
8. Now comes the layering, you can either do individual tiramisus in short glasses, or use a medium size glass tray (20x35cm-ish)
9. Place a layer of the soaked savoiardi on the bottom, then using a spatula or large spoon put a layer of the mascarpone-cream on top
10. Then repeat for the next layer. You should aim to get 2 layers of savoiardi and 2 of mascarpone
11. Use your spoon or spatula to press the mascarpone into every corner of the tray and around all the savoiardi. Smooth out the top
12. You can put this as it is, covered, into the fridge and leave it for a few hours or overnight. It actually tastes better then if served straight away as the coffee/alcohol has time to soak in fully and mix in with the mascarpone layers a bit
13. When you are ready to serve, grate one or two squares of the lindt (you don't need much, it seems to expend when grated) and sprinkle over the tiramisu, then use a small sieve to sprinkle the cocoa on top of this
14. Cut into whatever size rectangles or squares take your fancy and serve straight away!

I hope you like it as much as I do. Enjoy!

Following on from my previous quick post, and reading yet another Benedict Evans article, this time on the puzzle of generative AI adoption. His argument is that chatbots are unlikely to be the end product—they’re more of a demo layer, waiting for someone to build the “real” apps on top. That is something that I do largely agree with and mentioned in my previous post, but I can see the other side of this as I argue below.

GenAI’s adoption puzzle — Benedict Evans

Generative AI chatbots might be a life-changing transformation in the nature of computing, that can replace all software, but so far, most of its users only pick it up every week or two, and far fewer have made it part of their lives. Is that a time problem or a product problem?

Benedict EvansBenedict Evans

I want to ask the question again: what if the chatbot itself is the product?

Relational databases changed the world, but for most people they became invisible. They’re the plumbing that sits behind every piece of software. The only people who treat them as products are the engineers who build on them.

Excel, though, went another way. It’s as fundamental as a database, but it never disappeared into the background. People use it raw, straight out of the box, at wildly different levels of skill. Some just keep household budgets, others build businesses with pivot tables and macros. Excel is the product.

That’s the analogy I find most interesting for chatbots. Yes, they’re still flaky and limited. But with memory, projects, canvas views and better interfaces, it’s not hard to imagine a world where people use them directly. Not hidden behind other apps. Just the chatbot as the thing.

If that’s true, then maybe we’re not waiting for the “killer app.” Maybe the chatbot already is one—a new general-purpose interface, like a spreadsheet you talk to instead of click on.

So why hasn’t it happened yet?

Part of the answer may be cultural. When Excel launched in the 1980s, tools arrived finished—or at least polished enough to feel complete. With AI, companies are showing unfinished versions, moving fast, shipping in public. Adoption looks huge on paper, but in practice people are still testing, waiting for the product to settle down.

It may simply take time. Excel wasn’t instantly a workplace staple—it grew over years. Chatbots might need the same. Or maybe they really will fade into the background, embedded in other apps.

But I don’t think we should dismiss the possibility. Sometimes the raw tool is the product. Excel was. Chatbots might be too.

AI metrics — Benedict Evans

With every platform shift, we want to measure the growth but we’re confused about what to measure. That’s partly a problem of data and definitions, but it’s really a question about what this is going to be.

Benedict EvansBenedict Evans

I really admire Benedict Evans’ work, and his latest article is another great one. He looks at how AI uptake is measured, and more importantly, what those metrics actually mean. At the end he asks:

"the real question, as I’ve hinted at a couple of times, is how much LLMs will be used mostly as actual user-facing general-purpose chatbots at all or whether they will mostly be embedded inside other things"

I’ve been using LLMs more and more, and for many use cases treating one like “search on steroids” is enough. It delivers results fast — formatted tables, sources, links — and there’s that moment of relief: “Yes! The AI did all the work!”. Then comes the crash: broken links or claims that don’t hold up, and suddenly you’re verifying everything manually.

But that’s just Level-0 stuff. Where LLMs become really interesting is with deeper, multi-stage prompts. That’s also where many people lose traction. If you don’t put in the work — learn how the latest models behave; discover what prompt styles get results — you won’t get close to their potential.

A good example that comes to mind: CV-checking tools that promise to help you beat the ATS (Applicant Tracking System). Yes, you could write a detailed prompt yourself, feed in the job description and your CV, get feedback from GPT or Claude directly. But a lot of people prefer something simpler: upload your CV, click a button, see a slick chart and some step-by step improvement instructions. Easy. Fast.

At the moment my prediction is that for many, AI chatbots are still only wonderful and unexpected novelties. Only the hardcore users are using them close to full potential. I don’t have hard data to prove it, but I believe we’re still far from a world where AI fully, or even substantially replaces people. As William Gibson put it:

"The future is already here – it's just not evenly distributed"

I think it's the same for any tech in the business world. However revolutionary and productivity-enhancing a new technology is it takes time for it to mature and be adopted by organisations. Anyone who has been through a protracted digital transformation or system migration knows just how hard it can be, even if everyone agrees the change is worth doing.

Of course I could be completely wrong and we'll get to AGI in 2027, machines cross the Rubicon, self-improve, foom, we’re done. But for now I still think we are somewhere before the peak of inflated expectations on the hype cycle and people are way too excited.

Time will tell —unless the foom gets us first!

This a great post I came across on substack, check it out if you have 10min to spare. https://cmarmitage.substack.com/p/i-researched-every-attempt-to-stop

It is the middle of August and the wettest it has been for quite some time. Setting up this blog is in some ways easier than I thought, but all the little bits take time. Also, how does anyone ever make a decision on what to post? So many things to choose from, it's hard not to over-think things.

Anyway, in the middle of all that of that I am currently enjoying a 2022 album by New York-based saxophonist Oded Tzur called Isabella. Really amazing stuff that is a combination of middle-eastern / classical Indian music and modern Jazz. Check it out!