Tag: ChatGPT

#69: Thinking Deeply

Posted on by ewand

Designer (24)

It’s a little over 47 years since the Hitchhikers Guide To The Galaxy was first broadcast on the radio, followed by the publication of the written work the following year. It took the most powerful supercomputer, “Deep Thought”, 7.5 million years to come up with the Answer to the Ultimate Question.

deepthought

Who knows in the modern era how long it would take? Things have changed a lot over the last 50 years, and surely Google et al could manage a reply quicker than that.

It seems that the answers to many important but previously impossible questions are only a moment’s search away.

A screenshot of a web page AI-generated content may be incorrect.

Copilot’s Getting Deeper

Microsoft quietly unveiled an additional feature to its main Copilot offering – i.e. the free, web thing or Copilot app on PC or mobile (as opposed to the paid-for Microsoft 365 offering, or any other app’s Copilot-branded functionality).

Go to copilot.ai and just below the prompt, select the drop-down to change the mode – with a single click on the flower-like icon (which is not at all like the OpenAI logo), you can get it to Think Deeper.

Screenshot 2025-03-28 081649

Screenshot 2025-03-28 081837

This feature uses the fairly recent o3-mini (high) model from OpenAI (which runs on Azure, don’t you know… well at least most of it does), giving additional insight into whatever you’re asking. It doesn’t take much longer to answer compared to the regular reply so you might just think about using it all the time for questions of moderate complexity. And it’s free.

ChatGPT itself has a “Deep Research” function which is available to paying users (Plus or Pro), and Microsoft has also unveiled a forthcoming “Researcher” capability that will be part of the Microsoft 365 Copilot commercial offering, alongside some deep reasoning stuff for agents built in Copilot Studio. It’s all getting really deep, man.

Wannabe Record Breakers

As well as Copiloting-everything (mostly based on top of OpenAI stuff), Microsoft has been looking further afield and building its own AI technologies. There’s still plenty of Ayy Eye noise coming from Redmond, and an AI Skills Fest virtual event starting in April is going to keep the foot on the gas.

It might have one of the more obscure Guinness World Records, too…

A group of people looking at a glass board AI-generated content may be incorrect.

Let’s not get too excited now, kiddies. There are plenty of strange records to aspire to.

#68: It’s all about the prompt

Posted on (Updated by ewand

A cartoon of a person talking to another person AI-generated content may be incorrect.

When internet search engines took off in the mid 90s – remember Alta Vista? – and Google exploded into the public consciousness in the early 2000s, it became increasingly apparent that getting good search results were helped by being able to ask your question correctly.

Savvy searchers might use a combination of quotes and other “operators” to specify an exact phrase, or guide the search engine to include only certain terms or results from a particular website (such as site:tipoweek.com onenote). Google and Bing both tend to use the same operators (so, as Scott Hanselman would say, you could “Google with Bing”).

Prompting Today

A screenshot of a computer screen AI-generated content may be incorrect.

When using some of the many AI tools such as ChatGPT, Copilot, Gemini etc, you can get very relevant results by being quite specific in what you ask it to do. As an example, one of the best ToW banner images was created using Microsoft Designer with the prompt, “a serene image of a young boy sitting at an old laptop (with Windows 10) but lurking in the dark background is the grim reaper”

A child sitting in a chair with a computer AI-generated content may be incorrect.

Or, getting much more detailed, see Kat Beedim’s detailed 200+ word instructions to create consistently-formatted notes from meeting transcripts.

Being much more verbose and directional than you’d ever try in a regular search engine can give some quite remarkable results. The order of what you ask might vary the emphasis given to certain parts of the response, and the general advice is to be positive – i.e. ask for things you want, rather than telling it what you don’t want.

A polar bear and cub in the snow AI-generated content may be incorrect.

It seems that AI can suffer from a variant of Dostoevsky’s “White Bear Problem”; ie. Asking it not to do something increases the likelihood of doing it. Not long after Microsoft went big on Copilot and Designer, here’s one example when Copilot was asked to draw an image on a particular topic…

No photo description available.

The idea was to convey a background threat with those hooded figures, not the feeling that the poor girl was in imminent peril. The figures lurking in the background might be a mite less sinister if they weren’t armed, so clarification was called for…

No photo description available.

Maybe DALL-E 3 at that time was just fixated with firearms, or asking it not to do something was a step too far. We’ve gone from “some guns” to “pointing guns at her”. Hmmm.

No photo description available.

Trying the same prompt in Designer seemingly gets a little less gun-heavy now, but still has the odd one creeping in. Trying to be more explicit doesn’t appear to work… adding to the end of the prompt, “The sinister hooded figures are not carrying guns of any kind”.

You might think that instruction is simple enough, but no. It seems to be interpreted as “you want more guns? Gotcha”.

A person sitting at a desk with a computer AI-generated content may be incorrect.

Further reading

See here for some more tips on Copilot, or take a look at some pearls from the Copilot support team. Also, look out for some more in-depth instructions on using ChatGPT.

For business Copilot with M365 users, the Copilot Prompt Gallery is worth a play.

A screenshot of a computer AI-generated content may be incorrect.

For other Copilot ideas, check out Chris Stuart Ridout talking about Prompt Buddy, a Teams app which lets users share good prompts with others in the company.

A screenshot of a computer AI-generated content may be incorrect.

681 – AI Push Me Pull You

Posted on by ewand

clip_image002As more AI hoo-hah continues to pour from the hype machine of big tech, some of the puffed-up services are starting to become more accessible. Even if it’s still badged as “preview”, Bing’s AI chat is now more widely available and further functionality will be along soon.

Stepping onto the bandwagon before it fully departs, AI technology behind Google’s Bard chatbot is being embedded into Search, as announced at Google’s I/O conference, whose theme was all about how everything is being re-engineered to embrace generative AI. El Reg has neatly summarized the keynote here if you’re interested to learn more (“We sat through the Chocolate Factory’s PR blitz so you don’t have to”).

Google announced copilot-like functionality for its cloud productivity suite while Microsoft unveiled the M365 Copilot preview that’s been running for a few months, is being extended further.

Not all is rosy in the garden of AI, however. Distinguished scientist Geoffrey Hinton, a Turing Award winner and widely-described as “Godfather of AI”, has walked out of his Googly role amid concerns that AI will become sentient and enslave or kill us all. Interestingly, Hinton did not sign the Elon Musk-backed petition to halt AI development, effectively saying that if those currently working on it were to stop, others would pick up the baton. Microsoft’s chief scientist agrees.

clip_image004Making AI pay for itself is one challenge that will need to be addressed, as the intensive computation required can be very expensive – costs of running ChatGPT are eye-watering, according to OpenAI’s boss, and reckoned by some to be in the region of $700K per day. Still, investors can’t get enough of it and OpenAI is piloting a $20/month ChatGPT Plus subscription.

The expanded M365 Copilot preview is a “paid-for” thing, and Microsoft’s Q3 earnings call did mention that Copilot will be additionally priced over and above whatever Office licenses a customer already has (though some AI related features will show up in E3/E5 licensed environments, such as the new Semantic Index which can be harnessed by Copilot but will be useful for giving more accurate search results even if Copilot is not in use).

Back in the present, there are some relatively new practical capabilities in both Bing AI and in the Edge browser’s discover feature, as discused in last week’s ToW. The Compose feature in the Sidebar lets you play with generating different types of written content, the kind of thing which will be integral to Copilot in all kinds of Office applications before long.

The Insights tab on the same Sidebar gives you more info on the page you’re currently looking at, from a summary of the key points of the page, to some background on where that site is accessed from, how likely it is to be reliable and more.

clip_image005

The core Bing AI search in a browser – in case you’re itchy about using Edge or its Bing Sidebar – has some new clip_image007capabilities, especially the integration of Bing Image Creator, which is available clip_image009separately from the AI chat function.

Another one of OpenAI’s groovy tools – Dall-E – generates images based on a text description, and Bing AI chat can feed directly into that.

The image generation capability is now multi-lingual (with over 100 languages supported). It will also soon be possible to upload images to Chat, so you could ask it questions about what’s in the image.

All free for now, but someday soon, we will need to pay the ferryman or the robot overlords will wreak their revenge.