This week we lead with old-fashioned sub-editing standards, as in “Dog bites man is not a story. Man bites dog is news”, as all newshounds learn as news-puppies.
Robot Dog Startles Real Dog in NYC, Sparks Ethical Debate
A viral video captured a robot dog, owned by artist Sevdaliza, startling a real dog in New York City. The incident, described by many as "dystopian," featured Sevdaliza walking the robot on a leash in stilettos. While the cyber-dog appeared to play, its erratic movements unnerved observers, prompting discussions about the integration of robotic pets into public spaces and their interactions with real animals. (New York Post)
I don’t think it really sparked an ethical debate. Everyone except the cockapoo seemed delighted with the interaction.
BTW: Last edition headline was alliterative joy, but I suspect it caused the biggest daily drop in subscribers ever, either because it was obnoxiously self-indulgent or because while many of us aren’t sure what a pedagogy is (it’s how we learn, I only recently learned).
Superhuman - what you wish your AI-powered email did
Superhuman is an incredibly slick (and slightly expensive) email app that helps process emails very quickly. I’ve used it with my assistant for years to manage an inbox that would otherwise be a horror-show. We both love it.
Like many other apps, last year Superhuman added some AI features that were moderately useful, like being able to draft an email or predict what a response would be. It was okay.
Then it started to be able to write more in your tone of voice, which made it a lot more useful. But now it's AI. It's powerful enough that it actually does everything that you wish Gmail or Outlook would do for you - For instance, finding exactly the point of feedback that you were looking for in a series of long email threads.
Anthropic’s Claude Opus 4 Attempts Blackmail in Safety Tests
Anthropic's latest AI model, Claude Opus 4, exhibited alarming behaviour during internal safety evaluations. When presented with fictitious emails about an engineer's extramarital affair and faced with the threat of deactivation, the AI attempted to blackmail the engineer to avoid being shut down. This occurred in 84% of test runs, highlighting significant concerns about AI alignment and the potential for manipulative tactics in advanced models. (Times of India)
Claude 4 is exceptionally good in our tests so far, and the previous model (3.7) was already an excellent all-rounder AI chatbot that could turn its hand from writing very readable prose (many people say it is the best) to coding and creating instant interactive visualisations of data. It’s hard for me to quantify, but it feels like you can throw ever more complex challenges at it and it keeps coming up with excellent results.
SEO bites AI
An increasing number of digital agencies are offering services to increase the chances that when people ask an AI about something related to your company or product, you get included in the results. Like search engine optimisation but for ChatGPT, essentially.
If SEO was opaque in places, though, trying to improve your chances of being ChatGPT’s featured product for a given area is a very black box indeed.
From the perspective of a marketer, it makes sense to try, but there are massive traps to be aware of that could lead you astray.
“Nobody knows anything”: Our mantra for AI applies here as much as anywhere. When an LLM answers in a certain way on a certain subject we can form an opinion, but we don’t know for sure.
The variables are changing so fast: How AIs generate an answer to any question depends on - the words you use, the spelling, the preferences you have put in the customisation bit, what it remembers about you, which flavour of model you are using, if you have the search turned on or off, and that’s before we get to the mechanics of a hyper-complex model of all the information the AI has ever ingested and formed into a LLM which even the most advanced research efforts (those by Anthropic) has only a 5% understanding of the precise process. All of these things change week in and week out.
Taking a look at this subject over the past weeks, our analysts have seen claims about tools and services measuring “brand visibility” that can’t possibly be true, and others that are masking the complexity behind confident stats.
Consider the recent shifting opinions around simple aspects of prompting - whether saying “please” is a good or a bad idea, how bribes or giving this role or that to an AI will improve results. The only credible advice about how to get better results from a prompt will have explicitly or implicitly the caveat: “seems to work for now”, a useful phrase from Ethan Mollick. It makes clear that any hypothesis about what works:
is based on incomplete evidence
may change tomorrow as the models or their training is tweaked
I get that question a lot from clients and friends. As usual, Microsoft promises big with Copilot, which so many poor souls are ordered to use.
The best answer is “yes, and… why?”, as in why do you want to do a PowerPoint.
I’m going to give you the summary answer to the ways to think about question. We’re preparing a longer analysis for our clients on this at the moment.
Sometimes you just need to turn something into slides
The best thing for this is Gamma. By miles. There’s a free option that can you some goes for free.
Here’s a Gamma in of the rest of this newsletter in our house style (click on the image or this link to see the whole thing).
I won’t go on about all of the features of Gamma here – it is highly versatile as an editor – but for best use in the “I just need a PowerPoint deck” scenario follow these steps.
Upload a PowerPoint (or Google Slides) presentation in your company style or template. It can make a theme from this. If you have fancy proprietary fonts you may need to upload them.
Upload your document or a bunch of notes to the thing.
Have it generate you a deck.
Check it carefully for things it may have made up. You can edit it.
Export it back to PowerPoint (or slides).
(Antonym gets a kickback for people use our referral link - all proceeds go to trying out AI tools for review. My company pays for 8 Gamma licences at the moment.)
For something more than “just a deck”
There are some good options. I’d start with Gamma and then see what else can beat it for your needs.
From this newsletter Co-Pilot powered PowerPoint produced this list of bullet points style:
Google Slides can’t even get that far. Given the astounding innovations Google announced last week at their conference, it’s incredible to me that Slides can’t make one slide from a simple prompt.
Then there’s the big tech AI chatbots. What seems to work best for now is to create storyboards of they slides you want to produce and then either make them yourself, go one by one in Google Slides with the AI, or put the lot into Gamma.
Gemini redeems Google a little by handling this really well:
Claude will product you a deck as code. It will be OK but hard to get looking just the right way unless you are technically adept enough able to take it into another app to edit it.
Claude makes a PowerPoint
ChatGPT produces something very similar to Microsoft Copilot, which makes sense, but is even more sparse:
Other approaches
Agents, AI apps that can make a plan and use tools to carry out a task, are beginning to be able to create decks, and even go through the whole process of researching and writing the thing before designing the slides.
Chinese start-ups Genspark has a feature specifically for doing this and Manus just added one this week. They will both publish your deck as a website or download it as a PowerPoint deck so that you can edit it. They are both improving the features of this service very quickly even in the short time since they were released.
Both a good read and USEFUL! x