In the part of my chapter where I talk about choosing AI tools, I refer to the “hat, haircut, tattoo” analogy originated by James Clear in his 2024-02-01 newsletter and explained nicely by in his recent article On Developer Platform Experiments. I thought it might be good to have an illustration for this analogy in my book chapter. I did a search and didn’t find any suitable images for this analogy that I could reuse.

As my readers know, until now, I have avoided AI image generation tools due to ethical concerns:

But thanks to and her Notes relating to this collaborative post brokered by (on which she and I and five others contributed)1, I recently discovered Blunge.ai. This looked like a great opportunity to try it. (See this Blunge blog post for more on their approach to AI ethics.)

Blunge is based in Sydney, Australia. They offer a free plan which allows limited generations of public images that can’t be used commercially. Their site lets you try a few images without even creating an account at first, which I really appreciated. Everything I write about in this article was done without creating a login.

First, I chose a style. Here’s a partial screen shot of the available styles:

My requirements:

• I want my book chapter images to be simple so they are legible if someone is reading the ebook on a small screen, such as a mobile device.

• I’m assuming the hard copies of the Entrepreneur’s Manifesto book will be printed only in black & white, so I don’t need or want color.

I chose Blunge’s Style - the simple B&W outline format at upper left.

First iteration

Prompt: “create a composite image with 3 adjacent tiles in a horizontal row. the three tiles should show a hat, a haircut, and a tattoo. all images should be non-gendered and simple.”

Result:

Verdict: It got the layout and style right, but the hat and haircut both looked masculine-coded (even though I had specified ‘non-gendered’). So I tried to correct this.

Second iteration

Prompt: “make the hat and the haircut images gender-neutral“

Result:

Verdict: I think that works! Maybe a caption would help? (Yes, I could add words myself with an image editing tool, but I wanted to see what the tool could do with text.)

Third iteration

Prompt: “put the words Hat, Haircut, and Tattoo in a simple sans serif font, centered underneath the 3 tiles”

Result:

Verdict: Oh nooo … the text is perfectly fine, but instead of keeping the 3 images I was happy with, it regenerated the tiles and gave me back three gendered tile images 😖 (At least, I have never yet seen that nautical heart tattoo on a woman.)

Maybe I hadn’t selected my previous image in the prompt box, and it defaulted to regenerating? It obviously remembered the gist of my previous instructions, though. I clicked the handy delete icon (trash can) and selected generated image #2 as input.

Fourth iteration

Prompt: “Use the tile images in this version - do not regenerate the tile images - and add the words "Hat", "Haircut", and "Tattoo" in a simple sans serif font, each centered underneath its respective tile.”

Result:

Verdict: I think that works. What do you think?

Leave a comment

Fifth Iteration

For fun, I wanted to see what would happen if I chose to apply a different style. I picked the simple Vintage Collage Objects style (the cat with the open mouth in the big image above).

Prompt: “Apply the newly selected style to this image. Keep the gender-neutral hat, haircut, and tattoo tile images and the words underneath.”

Result:

Verdict: Blunge did a good job keeping the tile images and words I wanted, and these do look nicer (although not what I expected from the cat thumbnail image above). But the colors and background won’t work for my book chapter.

Experiment Wrap-Up

At this point, Blunge gave me the message: “You’ve used up all your credits. Start your free trial now to generate more images for free and get access to video generation and more.” Fair enough.

All of these experiments were completed in under an hour. Most of that time was in writing up what I was doing, here, not in running the tool! The image generations generally completed within 10 seconds or so. Granted, these were simple styles and simple requests, but it was free, and I’m pretty happy with what I got for my purposes.

And if their ethics positioning is accurate, I wasn’t exploiting any human artists by using Blunge. I was still using up computing resources, though, so my experiment had a bit of data center impact. And I was a bit disappointed with needing to reprompt to avoid gender bias - but, unfortunately, that’s pretty common among genAI tools. (Note to self: “gender-neutral” worked better than “non-gendered”.)

I’m still going to mostly rely on searches for existing images, when I think one might exist that could fit my needs, and collaborations with human artists.

What’s Next?

To use this image (or others) in my book chapter, I need a paid Blunge plan. Fortunately, at $10/month, the base paid plan is affordable even on my ‘ethical shoestring’ budget. That plan will let me generate 50 images/month. That should be more than enough for my books in progress, and even my Substack posts. The next plan up at $20/month allows 100 images, which I don’t think I need, and video generation, which I don’t care about. And they offer a 14-day free trial, which at this point will carry me into my 2026 budget year before I have to start paying.

So my next step is to sign up for an account, start my free trial, and save this “hat, haircut, tattoo” image as PNG and SVG for the book chapter. Then I may experiment with generating other simple illustrations … like maybe the chainsaw analogy from my Everyday Ethical AI book, which I’m also referencing in this chapter. I might also try generating nicer branding images for my newsletters.

Open Questions

Along with my free trial, I’m also going to investigate Blunge.ai’s ethical position more closely. I’m a little concerned that:

• Their blog post doesn’t mention if their underlying image generation model was trained only on works that were not copyrighted.

  • It does say that artists can upload their own images and train their own model, which would be ethical sourcing. That doesn’t fit my needs, though; I’m not an artist. It’s also not likely that they are actually doing a full retrain of a model with only one artist’s works. More likely, they are developing a style model.

• Their image generation widget shows a choice of models to be used:

  • Nano Banana Pro x Blunge (Gemini 3 Pro / Nano Banana 2)

  • GPT Image x Blunge (GPT Image 1)

  • Nano Banana (Gemini 2.5 Flash Image / Nano Banana 1)

  • GPT Image (GPT Image 1)

  • Nano Banana Pro (Gemini 3 Pro Image / Nano Banana 2)

  • GPT Image (mini)

  • Imagen

By extension, any images generated in Blunge with a model based on Nano Banana, GPT Image, or Imagen will be unethically sourced, since OpenAI and Google did not ethically source the data for those models. (This article by has more info on ethical considerations for Nano Banana2.)

I’ve messaged Blunge with inquiries. For now, I’m not using this image in my book chapter. If it turns out that there are ethically-sourced model options, I’ll try regenerating this image. Once I have answers, I’ll update my ethical shoestring list accordingly.

I’ll write more in 2026 about my further adventures with Blunge. I invite you to subscribe to stay informed and let me know this kind of post is interesting for you!

Subscribe now

1

Slow AI

Are AI-generated Images Theft?

This post began in a Note on AI-generated images from Caitlin McColl 🇨🇦, following on from a post about plagiarism in Substack Notes. That earlier conversation was relatively clear. Copying someone else’s words and presenting them as your own is plagiarism. The harm is direct. The boundary is visible…

Read more

4 months ago · 60 likes · 170 comments · Sam Illingworth, Mia Kiraki 🎭, Caitlin McColl 🇨🇦, AI Meets Girlboss, Karen Smiley, ToxSec, Jenny Ouyang, and Karen Spinner

2

Data in Motion

The Ethics of AI Imagery: Nano Banana and the New Frontier of Digital Reality

The Gist: Google's integration of the powerful "Nano Banana" image editing tool into Gemini marks a pivotal moment in AI-generated imagery, promising creative possibilities whilst raising profound ethical concerns about consent, authenticity, and digital truth. Whilst watermarking offers some transparency, emerging legislation like Denmark's pioneering …

Read more

8 months ago · 1 like · Genevieve Smith-Nunes

Subscribe now

Why

For most things, I’m a planner. I like planning ahead and researching my options before something happens. My current smart watch, a Fitbit Versa, is (knock on wood) still working fine, but I want to be prepared for the day it isn’t. So there are 3 motivations for my smart watch research:

1. I’ve been wearing a Fitbit watch for over 8 years; I’m on my third one. As a privacy practice, I set up a special email account that I only use for my Fitbit. Well, Google bought Fitbit in 2024, and in mid-2025 they began nagging Fitbit watch users that we would need to switch to a Google account by Feb. 2025 to keep using our watches. As much as I still use Google services for many things (working on that), I don’t like the idea of them fusing my personal biometric data to everything else they know about me. And on principle, I resent being forced to use a Google account to keep my watch from becoming a dead piece of hardware.

2. My recent experience with Fitbit has been flaky. I’ve had a new watch go bad within a few months of purchase (display just went haywire). It was replaced under warranty, but I was without one for about a month. And I’ve found that its oxygen monitoring isn’t very accurate. I often don’t even get a reading on my overnight O2 at all, or it’s only available late in the day. It’s the one feature I wanted when I upgraded from my previous Fitbit, and it just doesn’t work that well. So when the time comes to buy another smart watch, it likely won’t be a Fitbit.

3. I’m tired of the constant nagging from Fitbit to update to a premium subscription (monthly payment). And it’s unlikely Google will stop. I’m not going to do it. My budget and I don’t need that stress.

I just want a basic wearable that:

• works as a wristwatch, i.e. can show me what time it is (but dims when I don’t want to see it, to save battery and avoid distraction)

• lets me monitor the essentials on my wrist, namely my heart rate and steps (i.e. I don’t have to pull out my phone, unlock it, and look at an app to see my numbers)

• measures oxygen saturation (desirable, if it works better than what I have now, but not absolutely mandatory)

• has decent battery life (at least a couple of days between charges)

• doesn’t require a monthly or annual fee (and preferably, doesn’t nag me about one)

• affordable (ideally under $100, although I could go to $150-200 if pressed)

• iPhone compatibility (highly preferred)

• doesn’t share or use AI on my data (personal or biometrics) without my opt-in (not negotiable)

I don’t need cellular connectivity. I’m not a runner, so I don’t even really need GPS. I’d actually rather not have GPS, since it will dilute battery life, it’s a potential privacy risk, and I’m not looking for features like fall alerts that would use it (a watch can use my phone’s GPS if needed). It seems hard to find a smart watch nowadays that doesn’t have GPS built in, though.

How I Searched

Data privacy concerns with wearables aren’t new (see e.g. this 2021 article).1 Industry consolidation and use of data for AI make it more of a concern now.

I did some old-fashioned searches to find newer articles on wearables and privacy (e.g. 2 3 4 5 6 7), recent devices that don’t require premium fees (e.g. 8), and info on Fitbit alternatives (e.g. 9 10 11). Here are the ones I ruled out.

1. I will not even consider Amazon or Meta devices. (This won’t surprise you if you’ve seen my previous comments in AISW interviews about why I don’t use the Alexa I was gifted, and why I left Facebook last summer and deleted all my photos and posts).

2. I ruled out Xiaomi and Amazfit wearables due to their terrible scores on data privacy.

3. EU-based, GDPR-compliant Polar Vantage V3 is well-regarded for privacy, but at $299-699, not even close to my budget.

4. I found Apollo Neuro via ’s article on gifts for people who read the fine print12. Their wearable looks interesting, but it doesn’t meet one of my main criteria: it isn’t a watch. No time display. (I might try out the free version of their Apollo Sessions app, though.)

5. I ruled out Whoop 5.0 because it has no watch display and it comes with a pricy $199/yr fee. Double no.

6. I ruled out Oura Ring 4 due to its high price ($349), lack of a display, and $70/mo fee to get metrics on heart rate and oxygen. Triple no. (Two competitor rings which used to be available for sale in the US no longer are, due to Oura winning a copyright infringement judgment against them. )

7. I found many off-brand and unbranded fitness watches on Amazon, mostly under $100. Some have names like MorePro or EnergyPort or KEEPONFIT. Others have no names, with sellers whose store names are seemingly random combinations of letters. All used their own special iOS apps. Not feeling confident I could trust how they would manage my data, I ruled them all out.

What I Found

That narrowed down the field quite a bit. Here are the 4 brands I considered.

• Apple: Watch SE 2 $229 , Watch SE 3 $249 (on sale for $199)

• Garmin: Lily 2 $199, Lily 2 Active $249, Vivoactive 5 $199, Vivoactive 6 $249

• Google: Pixel Watch 3 $249 (2024 model with 45mm screen is $219; one color with 41mm screen is $169), other models higher

• Samsung: Galaxy Fit 3 $55, Galaxy Watch $249 and up

Apple and Google were the top two in most privacy ratings (although not Mozilla’s). I looked at them first.

Apple

Apple watch prices were right at the top of my budget (only within budget if on sale). The Apple watches don’t seem to offer oxygen monitoring in my price range, though. And the battery life seemed exceptionally short, at 18 hours max. I want to wear my watch around the clock, including when I sleep. So this option was out.

Google

I’m obviously not thrilled at the idea of needing to use a Google account with any of their watches, but thought I’d see how they stacked up. The Google Pixel watches were generally beyond my budget, though. I found a 2024 model in one specific color that came in under $200. I could live with that model and color, but it only offers a 36-hour battery life even in ‘battery saver’ mode, and they’re Android-only. At that price range, with those limitations, they’re out.

Garmin

Garmin models dominate the “less creepy” end of the Mozilla evaluations, so I was eager to see how they measured up.

The Garmin Lily 2 (base model) didn’t include GPS, which was fine. However, it costs the same as the Vivoactive 5 and has a much shorter battery life.

The Garmin Forerunner has an option to disable GPS and extend battery life from 19 hours (!) to 11 days, which was a plus. It didn’t appear to have oxygen monitoring, though.

The Garmin Vivoactive watches have long battery life (measured in double-digit days, not hours). They work on both iOS and Android. They include heart rate and oxygen monitoring. They have features to detect naps as well as sleeps and to analyze the impact of napping on sleep quality — something I hadn’t considered, but which could be valuable. They even offer a ‘wheelchair mode’, which might prove to be useful for a family member. That indicates they’ve given inclusion some thought, which is a plus in my book.

Samsung

At only $55, and with oxygen monitoring included, the Samsung Galaxy Fit 3 looks like the best of this bunch. But it’s only compatible with Android phones. (I did consider the possibility that, with the cost savings, I could buy a refurbished Android phone as a second, wifi-only device. It would be somewhat inconvenient, but I could use that isolation to further protect my data privacy. This is a fallback option for me.)

Another concerning thing is that when I searched the Samsung website, under Wearables|Watches, the Fit line wasn’t listed, implying it’s been discontinued. That isn’t a good sign for future support. Also, I looked at some listings for the Galaxy Watch, but found this note about “Galaxy AI features” only being free through the end of 2025:

The other Samsung watches were outside of my budget. And Mozilla ranked Samsung watches as ‘creepy’. I regretfully ruled out the Samsung product line.

Conclusion

The baseline, cheapest option is to keep my current Fitbit watch and set up a clean, new Google account to use with it, separate from my main one, to better protect my privacy. That should last until the watch dies or Google tries to force me to pay for a premium subscription to get basic data analysis features. In the meantime, this path would save me $200. I am not that confident, though, that Google won’t figure out the new account is actually me. Their ML will certainly be able to tell that it’s being worn by someone in my household who goes everywhere I do.

As one of the most privacy-protective brands, Garmin’s more affordable watches are the clear winner in my book. It looks like the Vivoactive 5 has been superseded by the Vivoactive 6, which is $50 more (above my budget) and doesn’t seem to offer anything additional that I need. Hopefully the 5 model will remain available for under $200. This “Garmin Vivosmart 5 Review” has more info on it.13

I’ll post an update after I make my purchase and have some experience with it. If you’re also looking at a similar decision, I’d love to hear if this info helps you, if I missed any major smart watch options, and what you end up deciding & why!

Leave a comment

References

Stating what I hope is obvious, if you know me at all: I make no money from any of the links in this article, and I have no relationships with any of the companies mentioned.

ChatGPT Confusion

Within the space of our 1-hour Zoom call last Friday night, Lakshmi’s licensed copy of ChatGPT Pro (GPT-5):

• Gave positive scores to almost every interview, even interview #059 which we chose as a negative benchmark. (It has clearly negative sentiments such as “I would never intentionally use AI”.)

• Repeatedly used the data from the AISW059 PDF file when we told it to use only the AISW041 PDF file.

• Claimed that about half of the PDF files we had just uploaded in the previous command were not readable. Mind you, we’re only talking about 80 files, and not very big ones — megabytes in total, not gigabytes or terabytes. And this was the Pro version in a fresh new conversation.

ChatGPT was clearly mishandling the PDF files she uploaded. We wanted to reduce the content of the files to exclude my interview questions anyway, and to limit the content to plain text only, to reduce possible sources of bias. So we decided to take the opportunity to have ChatGPT help us construct the reduced set of content we wanted to analyze.

That … didn’t work well. Lakshmi tried giving ChatGPT the URL for interview 032, an engineering program manager in Costa Rica, so that it might retrieve the correct content. It failed.

This reported failure actually makes sense because I have the Substack setting on the newsletter set to block AI training. So yay, ChatGPT sort of respected my setting. That’s the only good news.

The kicker: ChatGPT offered to try to get the interview content from what it called an “Open Substack mirror” (‼️). Existence of such a mirror would mean ChatGPT or one of their partners wasn’t really respecting our opt-out settings after all.

Curious to see what that mirror was and if it would work, we agreed to let it try.

Sure enough, ChatGPT came back with the text of an interview that it claimed was a successful retrieval of interview 032. What it came back with, though, was not interview 032. The content was for an anonymous embedded software developer in Costa Rica who supposedly graduated in 2014.

Not only was the interview not the one we requested, I never interviewed any embedded developer from Costa Rica who graduated in 2014. ChatGPT and its supposed Open Substack mirror got the country right, but otherwise the returned interview was made up — and useless to us. 🙄

I’m still mulling that supposed ‘Open Substack mirror’. As a mirror, it’s more like a funhouse mirror than a real one! For people who’ve opted out of AI training through our Substack settings, is it sort of good news that trying to retrieve the interview through the mirror didn’t work, even though ChatGPT claimed it did?

Lakshmi and I resorted to manually rebuilding 80 smaller answers-only files in DOCX format, removing photos, titles, and my interview questions and comments. It was tedious, but the most efficient way to get to our real work and know that we could trust that the content would be accurate.

Dancing in the dark with Copilot

Earlier last week, while Lakshmi was running OpenAI ChatGPT Pro on our data, I ran Google Notebook LM. ChatGPT’s numbers were all over the place, and I got much lower numbers with Notebook LM. Many of the lower numbers aligned better with my intuition about the guest’s sentiment. Giving both LLMs a prompt with a specific scoring range (-1 to 1) helped, but ChatGPT and Notebook LM were still quite different. I hadn’t expected that.

So for fun, I tried Mistral Le Chat with the same files and prompt. It was strongly positive, much more so than ChatGPT!

Then I tried Microsoft Copilot. Copilot was mostly positive, but not as much as Mistral.

For all of these, I initially used the same PDF files as Lakshmi did with ChatGPT, and the same simple prompt for scoring AI sentiment in the range -1 to 1.

After reducing the files to only the guests’ answers, as described above, we re-ran all four LLMs. In order to run Notebook LM, I converted the DOCX files to TXT format. (Why? Because Notebook LM pretends it doesn’t know how to read DOCX, although it obviously could do so easily. 🙄 And converting DOCX to PDF would have been more work and made bigger files for the LLMs to digest.)

Using the TXT files as input for all of the LLMs not only removed some potential biases, it made the files much smaller (a few MB total) and removed an internal LLM processing step to convert from PDF or DOCX. We hoped the LLMs would be less likely to hit token or context limits.

I ran a few sample TXT files in each LLM that night, then hibernated my laptop with the LLM browser windows left open. I picked up the work the next day and re-ran all four models with all 80 files.

I expected similar scores from Copilot on the TXT files compared to when I re-ran the scoring with the same simple prompt the day before. Uh, no. Copilot now was even more positive than Mistral, even on the negative benchmark interview. Every one of the first 20 interviews was scored as positive! Not only was it different from what Copilot gave me the day before, I knew that it wasn’t right.

Meanwhile, NotebookLM and even the highly-optimistic Mistral rated the same benchmark negative interview #059 at -0.8 and -0.9, which made sense. And compared to when I ran the same interviews with the same scoring prompt a day earlier, their scores were fairly repeatable.

I asked Copilot to explain its scoring. Among other things, it said that it was counting awareness of ethical concerns as a positive. Aha! I wanted expressions of ethical concerns to be handled as a more negative sentiment.

From my experiences with these interviews, guests who commented about reservations on the five concerns in my Everyday Ethical AI book (environmental impact of data centers, unethical data sourcing, exploitation of workers, model biases, impact on lives and livelihoods) did so because they found those concerns to be a downside of using AI, or a factor that kept them from using AI more. So those comments should reduce the score on AI sentiment.

I had not told Copilot to score ethical concerns about AI as positive sentiment. Plus, Copilot didn’t seem to score sentiment that way just the day before. With the same interview files and the same prompt, I got very different answers on two different days, using the same conversation instance.

I updated my prompt to clarify that concerns about ethics should reduce the score, and ran it again. I checked some of the individual interviews, including #059. Copilot still scored it as fairly positive.

When I pointed this out to Copilot, it acknowledged that its score was inconsistent with the many strong negative comments in #059. Copilot apologized and said it would adjust the score to a less positive number by modifying internal weights. But the first results with those supposed adjustments didn’t change at all. 🙄

We did this dance a few times, with me trying various prompts on ways to weigh ethical concerns and practical uses, and checking some benchmark interviews. I added, then ended up removing, a prompt telling Copilot that practical use of AI should raise the score; it appeared that this caused Copilot to over-weight wording it was already treating as positive.

Inspired by the Lehigh study that telling a LLM to ignore biasing factors could help to reduce the bias in its financial recommendations, I also added a sentence to the prompt telling Copilot to ignore demographics like race, age, gender, etc.

At the end of the day (well, night), the Copilot scores were still overall positive, but the score on #059 finally got below -0.5, to -0.8 (on the scale of -1 to 1). This was at least comparable to the other LLMs, which now scored #059 at -1, -0.9, and -0.9 with that revised prompt. The more positive (IMHO) interviews still had positive scores.

To ensure consistency, Lakshmi and I then re-ran all of the other LLMs with the updated prompt and the TXT files.

I ended up trying a fifth LLM - Anthropic Claude. Claude showed a somewhat similar pattern to Mistral and Copilot. But none of that proved that any of them were right. (They might have used common, WEIRD1 unbalanced training datasets, or other factors that made them give somewhat similar results.)

Here are the preliminary distributions of AI sentiment scores for the five LLMs:

ChatGPT looks suspiciously like a normal distribution (bell curve). There’s no reason to believe that the opinions in the underlying set of interviews are normally distributed, although it’s possible. And I’m really wondering whether most of these LLMs might be trained to respond with positive spins on sentiments about AI in particular, as a matter of company self-interest.

Leave a comment

At this point, I don’t feel confident about trusting any one of them with scoring these AISW interviews. We’re using an aggregate of all five LLMs (averages) for reporting our findings at the conference. I also tested a scoring approach used by judging panels in some Olympic sports (drop the high and low, then take the average); it didn’t make a noticeable difference.

Next Steps

We’re finalizing the data analysis for the conference this week. In particular, we’re looking at common themes and patterns in sentiment about AI across gender, race, ethnicity, age, type of business, type of job, geographic location, and even whether guests use Substack, LinkedIn, or both.

We’re definitely not stopping this analysis on Friday. I’m intrigued by the discrepancies among these LLMs, and we will be investigating further.

Trying other LLMs is one option. Unfortunately, so far, the ethically-developed chat.PublicAI.co is refusing to validate my email address to let me use it. 😔 I’ll pick that up next week, and I will definitely use it as soon as I can get in. I may try Perplexity too; we’ll see.

In addition, one of the professors we’re collaborating with for the ABC conference is doing good old-fashioned Textual Analysis with ‘close reading’ of the interviews. After the conference, she will be assigning scores based on her expert human judgment and analysis of the interviews. Her scores will give us some ground truth for verifying which of the LLMs (if any) are actually accurate for guest sentiments about AI.

Curious to hear about the results of our data analysis on how people feel about AI and ethics? Stay tuned! After the conference, I’ll summarize what we’re learning here in Everyday Ethical AI, and update you all from time to time.

Subscribe to Everyday Ethical AI

We’ll also share an article (or a few) in 6 ‘P’s in AI Pods (AI6P) about the more technical details of the analysis, what we find when investigating the LLM discrepancies, and what we’re working on next.

Subscribe to 6 'P's in AI Pods

And if you’re into data visualization and love violin plots as much as I do, you definitely won’t want to miss those articles. Here’s a preview. 😊

I like Mike Johns’ distinction between ‘driverless’ and ‘humanless’ control. If humans are in the car, but not in the loop at all, that’s an obvious safety concern.

All software has bugs. AI/ML-based software, including self-driving vehicle code, is no exception. Even if error rates are impressively low (and I don’t know what Waymo’s bug density data looks like), mistakes are problems. And it sounds like there was no effective way for the human in the car to address it when the car made a mistake.

There’s obvious potential for other times when humans in vehicles may need emergency help while on an automated ride, such as having a stroke or heart attack, or their vehicle becoming damaged or being assaulted.

Human riders need:

• a reliable ‘stop the ride, I want to get out now’ button,

• an ‘emergency help’ button that works, and

• a way to stop the car and get out, even without internet connectivity.

It sounds like Waymo has attempted to design some of these features in. But as serious as it can be for the automated driving and navigation to have bugs, it’s far worse when the safety measures don’t work.

A key question is why the stop buttons in the car and in the app didn’t work. If the button in the mobile app didn’t work because of poor internet connectivity, ok, but why didn’t the buttons IN the car work?

Hopefully those in-car buttons are either physically wired in the car or use local in-car wireless communications, and don’t rely on internet connectivity. Internet-only buttons, or buttons that can’t work if the car loses power, would be a serious design safety flaw.

It also sounds like the remote monitoring of the Waymos has gaps. A robust monitoring system should be able to detect a vehicle going in circles or such. This should be doable either with machine learning or with non-ML pattern detection. Humans supervising the system (and yes, there should be some, and seemingly are) should be alerted & able to reach out to a trapped or endangered passenger. And Waymo’s software developers should be automatically notified of the bug & associated data.

In my last corporate role as director of AI, these kinds of capabilities for cloud monitoring and control of embedded systems were called “digital feedback loops”, but that concept wasn’t unique or new with us. A recent article about Waymo claims “Waymo deploys extensive tracking software to monitor on-road activity” and describes some of their development process. (We’ll save for later the concerns about the potential of such tracking software for invading riders’ privacy.) However, the article doesn’t provide details on Waymo’s bug densities or whether they have intelligent monitoring features like looking for vehicles going in circles.3

If you live in an area that has “full self-driving” vehicles (I don’t), what are your experiences with being in or around them?

If, like me, you have experience with automotive software or embedded systems, I’d love to hear your thoughts on this too!

Leave a comment

Share

1

The Human Intelligence (HI) Stack

Driverless Cars Are Not Waymo Fun

You know that travel anxiety dream where your attempts to get to the airport keep getting thwarted? That dream became a reality for Mike Johns, a techie who got into a driverless car and wound up in a seemingly endless loop at the airport. The car went in circles, and he couldn’t get out, almost ca…

Read more

a year ago · 7 likes · 20 comments · Blair Glaser

Subscribe now

What People Think

A Consumer Reports AI survey of 2,022 U.S. adults, conducted in May 2024, shows that people in the USA care about transparency and explainable AI.

“A new nationally representative Consumer Reports survey explores Americans’ attitudes toward artificial intelligence (AI) and algorithmic decision-making. The survey found that a majority of Americans are uncomfortable about the use of AI and algorithmic decision-making technology around major life moments as it relates to housing, employment, and healthcare.”1

People are looking for companies to be far more transparent:

1. about the data used to train or score the AI model or algorithm, and

2. about a way they can correct the data (about them) used in the decision, if it’s wrong.

“The vast majority of Consumer Reports’ respondents (83%) said they would want to know what information was used to instruct AI or a computer algorithm to make a decision about them.  Another super-majority (91%) said they would want to have a way to correct the data where a computer algorithm was used.”2

Why They’re Not Wrong

Their concerns are well-founded. As this article on the CR survey pointed out:

“Given that AI frequently screws up, often in a discriminatory way, they're probably not wrong to be concerned.”3

Predictive policing4, the discriminatory example cited in this article, is just one cause for concern. Other risky areas for AI bias include credit decisions5 (e.g. on mortgages, loans, and rentals):

“The biggest-ever study of real people’s mortgage data shows that predictive tools used to approve or reject loans are less accurate for minorities.”

and healthcare 6:

“… for this new class of tools to do more good than harm, Sendak believes the entire health care sector must address its underlying racial inequity. "You have to look in the mirror," he said. "It requires you to ask hard questions of yourself, of the people you work with, the organizations you're a part of. Because if you're actually looking for bias in algorithms, the root cause of a lot of the bias is inequities in care."

The risk of perpetuating existing inequities makes detecting and rooting out bias in AI and algorithms particularly important and challenging. (The case described in the above article, about detecting sepsis in Hispanic children, illustrates this.) But that’s no excuse for not trying — and we need to fix the existing inequities while we’re at it.

Leave a comment

Share

Subscribe

Got Lungs? You’re At Risk (too)

For decades, we have desperately needed better diagnostic tools for lung cancers. Research remains woefully under-funded, and lung cancer survival rates still suck today almost as badly as they did 20 years ago.

More people of all genders die from lung cancer each year than from any other type of cancer. (There are more cases of other cancers, like breast and prostate. But because they have better detection and treatment, far fewer people die.)

A key reason lung cancer is so often fatal is that we haven’t had good ways to detect it early. Some numbers:

“Lung cancer is the UK’s biggest cancer killer, with nearly 35,000 deaths in the UK attributed to the disease annually, accounting for one in five (21%) of all cancer deaths. Six in ten people will survive more than five years if lung cancer is diagnosed at the earlier stages, but this falls to just one in ten for those diagnosed at the latest stage.”

How And Why Do I Know About This?

20 years ago, here in the US, my husband Dan got an initial chest X-ray when he started having shortness of breath. The X-ray report didn’t flag any cancer risk.

When Dan was finally diagnosed a year later, his oncologist told us chest X-rays weren’t good enough to rely on, and that by the time a radiologist can see a lung cancer tumor on an X-ray, it’s late, and the cancer has probably already spread. And Dan’s had. 😢

It’s still recognized today that chest X-rays often miss small lung cancers. But many doctors (non-oncologists, like Dan’s primary doctor) rely on them. They’re more readily available and they’re less expensive than CT scans. So finding a way to make chest X-rays more accurate for diagnosing lung cancer (and other lung conditions) early would be a boon.

Late diagnosis was fatal for my Dan, as it has been for so many others. But I’m an incurable optimist. I’ve remained hopeful that we’ll find and build new diagnostic tools that can help save people you and I love who are still alive today.

The AI Break-Through

Australia-based Annalise AI is using the power of machine learning to detect lung cancers earlier and more accurately.

This page on their website describes the Annalise.ai breakthrough in lung cancer diagnosis, which was announced on June 3:

“The Annalise.ai chest X-ray AI solution has been shown to improve diagnostic accuracy by 45% and increase diagnostic efficiency by 12%.”

That level of diagnostic improvement with chest X-rays is a BIG deal in helping to get people into treatment earlier. From the initial UK pilot:

“Previously the Trust started treatment for 90% of people with lung cancer within 99 days. The addition of Annalise.ai’s chest X-ray solution has reduced this to 76 days, and the proportion of cancers identified at earlier stages increased by 27%.”

Their tool will now be used across the UK. This BBC article includes some info and a video.

What You Can Do, Wherever You Are

As much as I’m excited about the prospect of tools like Annalise AI for improving diagnosis, we still have a long way to go, and many other aspects to solve, before we’ll make a serious dent in lung cancer deaths and suffering worldwide. All of us can help by being more aware, and by helping others learn, about risks and symptoms.

There are many myths around lung cancer. (I’ll be updating my “top ten myths” flyer with the latest stats for Lung Cancer Awareness Month in November, and will share it here.)

By far the biggest myths relate to smoking. Lung cancer can happen to anyone, regardless of smoking history. Three quick notes:

1. More than 10% of people who get lung cancer never smoked; being a never-smoker does not make you safe. (And about half of the rest quit smoking years ago.)

2. Lots of smokers never do get lung cancer.

3. Exposure to radon and other contaminants is a significant factor as well. Radon alone is the #2 cause of lung cancer.

For more information about lung cancer diagnosis and early detection, see these sites:

• https://www.lungevity.org/lung-cancer-basics

• https://www.lungcancerresearchfoundation.org/for-patients/how-do-you-know-if-you-have-lung-cancer/

• http://www.lcrfresources.org (free account required to download materials)

And if you’re buying a home, new or not, please consider getting a radon test, and having dangerous levels remediated before you move your family in.

This post is public. Please share freely to help raise awareness about lung cancer! Thank you.

Share

Clear ribbon logo for lung cancer awareness, designed by Mark Richardson for me (all rights reserved). Why clear? Because lung cancer is invisible.

This is a quick capture of some investigations I did this past week to find a free and preferably ethical tool for converting the audio to transcripts for the posts. Let me know if you find it useful.

TL;DR Here are the tools I found in my searches (some details below).

Tool Candidates - free: Conversations, ConverterApp, Fathom, Fireflies.ai, oTranscribe, Restream, SpeakToText.

Tool Candidates - not free, or free plan not suitable: Audiotype, Descript, HappyScribe, Kapwing, Media.io, Notta, Otter.ai, Talknotes, Turboscribe, Veed, Yescribe.

Bottom line: Try ConverterApp or Restream if you need to convert MP3s to text transcripts. They’re fast, pretty accurate for audio from native English speakers, private, and truly free - no accounts or calendar integrations required. If you don’t mind calendar integration and want a free tool, check out Fathom and Fireflies.ai.

Caveat: I haven’t yet surfaced any info about their underlying technologies or ethics. But at least what I do know about ConverterApp and Restream doesn’t raise any red flags.

Tool Candidates - free:

• Conversations - https://highlight.ing/apps/conversations

  • Built on Highlight open source platform

  • Free - downloadable Windows app

  • Appears to focus on real-time listening and conversion - not an ideal fit for my needs.

• 🌟ConverterApp - https://converter.app/audio-to-text/

  • Can distinguish between multiple speakers - good for sessions with interviewees.

  • Automatically filters out filler words.

  • Free to use via website without an account or giving an email address.

  • Lots of ads on the bottom of the page, but they don’t interfere with use.

Tested with a 28:46 interview recording (26mb) with 2 speakers. Transcription completed in 2 minutes, and the text was over 99% accurate (missed words: kiosk, EULA, a name with an unconventional spelling) 👍.

• Fathom - https://fathom.video/

  • Free account requires sign-in with Google or Microsoft, and calendar integration. An active Zoom account is also required. (Has Zoom app plugin - does it work with free Zoom accounts?)

  • Their privacy policy and terms of service do not mention AI or training. However, they openly claim AI-based features. They state that their products are GDPR and HIPAA compliant.

  • Has paid tiers, but the free version may be good enough for my needs

    • “Pricing: Free with no usage limitations.”

    

    

• Fireflies.ai: https://fireflies.ai/ (not to be confused with Adobe Firefly for images)

  • The free plan allows “unlimited transcription” for meetings joined by their bot, with limited AI summaries, if:

    • “Auto-join calendar meetings” is set to ON, and

    • “Send email recap to” is set to "Everyone on the invite”.

  • Sign-in is only supported via Google Auth (calendar integration) or Outlook auth. Unlimited free transcriptions require:

    • a Google Meet call recorded with the Chrome extension, or

    • the Fireflies AI notetaker must be allowed to join a meeting as a participant, with the settings above. 

  • Their data processing policy is stated to be GDPR compliant.

  • Can upload past meetings as mp3, mp4, wav, or m4a files to get them transcribed, but: 1 transcription credit is consumed per uploaded file (new users get between 3-10 credits), and users need to upgrade to a paid plan to download transcripts, audio, or video. 

• ❌ oTranscribe - https://otranscribe.com/

  • Free and based on open source (MIT license) - “A free web app to take the pain out of transcribing recorded interviews.”

I started to test oTranscribe with a 25:17 MP3 file. It doesn’t actually transcribe. It just provides a better UI than switching between a MP3 player app and a text editing app. The help says “oTranscribe makes the manual task of transcribing audio a lot less painful. But you still have to do the transcription.”

• Private - in browser - “your audio file and transcript never leave your computer”

• Interoperability - limited; provides a link to media.io to help with conversions if needed

• 🌟 Restream - https://restream.io/tools/transcribe-audio-to-text

  • Free and private - no account required: “You can feel confident that your data is safe with us when you use our audio transcription tool. You don’t have to download any software, we don’t show ads, and you don’t need to create an account. We also wipe your data from our servers after you’re done transcribing.”

  • Interoperability - “Our audio transcription tool supports audio files like MP3, WAV, FLAC, AAC and more.” (no m4a? need to test)

  • Accuracy - Site claims 99% transcription accuracy for English.

  • Other features - Plans appear oriented to video and live streaming. Outputs from the free plan appear to include a (video) watermark.

I tested Restream on the same 28 min file as ConverterApp. Once the file was uploaded, the transcription finished in seconds. It doesn’t identify different speakers, but it does separate them nicely in the file with blank lines to improve readability. The transcript includes ‘mmhmm’ and other filler words, which is a plus if you want to find them in the audio and edit them out. 👍

• SpeakToText - https://speaktotext.io/

  • Truly free: “This free online tool transcribes your audio files into text, applying high-precision methods backed by OpenAI's Whisper model. Unlike other services, this tool does not ask for your email address, offers mass transcription and accepts files up to 25MB in size.”

  • Ethical?: Likely no. ❌ The tool relies on OpenAI’s Whisper model. The safest assumption is that anything from OpenAI has not been fairly trained.

  • Private: They also promise that the content will not be mined: “At SpeakToText.io, we prioritize your privacy and security. None of your audio files are stored on our servers, ensuring your data's confidentiality. For data sent through our OpenAI-powered API, OpenAI's policy as of March 1, 2023, reinforces our commitment to your privacy. OpenAI no longer uses data submitted through our API for model training or improvement.”

  • Interoperable: Supported formats include mp3, mp4, wav, m4a, and more. Ability to take m4a format which Zoom provides is a plus, although I have to convert to mp3 anyway for Substack (but if Audacity could take m4a, then I could skip that step).

Tool Candidates - not free, or free plan not suitable:

Audiotype, Descript, HappyScribe, Kapwing, Media.io, Notta, Otter.ai, Talknotes, Turboscribe, Veed, Yescribe

(audio voiceover stops here and is available only in Substack, not in podcasts)

• Audiotype - https://www.audiotype.org/transcribe/audio/

  • Not free unless your content is under 1 minute long. Free trial does not allow downloading.

  • Pricing is based on the number of minutes, e.g. 60 minutes for 9 euros.

  • Speed - “With Audiotype, it takes a third of the file duration to transcribe speech to text”

• Descript - https://www.descript.com/transcription

  • Seems more focused on video than audio.

  • Pricing: Free plan offers 1 hour/month of recording time. Hobbyist plan is $12/mo, Creator plan is $24/mo, Business plan is $40/mo, and they offer an enterprise plan.

    

• HappyScribe - https://www.happyscribe.com/free-transcription-software

  • Free 85% accurate transcription tool for up to 10 minutes (not downloadable), as a way to upsell 99% accurate human transcription services or a paid transcription plan

• Kapwing - https://www.kapwing.com/tools/audio-to-text

  • Free for 10 minutes of transcription per month, then paid plans are available

    

  • Potential ethical concern: “Repurpose content from the internet by pasting a link.” - this seems to open the door for use of content not owned.

  • Bonus feature: edit the audio by editing the transcript (cut out unwanted parts)

    

• Media.io - https://www.media.io/sound-to-text.html

  • Free plan includes 5 min of automatic subtitling. Basic plan is $6.99/mo paid yearly. Premium plan is $12.99/mo and includes AI features like background noise removal.

    

  • Offers a ‘free’ Windows app download (to be tested to see what it does)

  • Doesn’t support Firefox

    

• Notta - https://www.notta.ai/en/tools/audio-to-text-converter

  • Free plan allows 120 min/month but limits to max 3 min per conversation - not viable for me

• Otter.ai -

  • Pricing: Free plan offers 300 monthly transcription minutes; 30 minutes per conversation; allows importing and transcribing of three audio or video files in a lifetime. Annual paid plans start at $10 per month.

• Talknotes - https://talknotes.io/tools/transcribe-to-text -

  • Free trial: can transcribe a 2 min recording for free via website. Base plan allows unlimited notes up to 20 min of recording time for $12/mo or $69/yr. Pro plan allows up to 25 2-hour recordings for $49/mo or $499/yr.

  • Home page says it offers privacy protections and does not use your content for training AI.

  • No information is offered on the underlying foundation models they use.

    

• Turboscribe - https://turboscribe.ai

  • Private: Their FAQ says “We don't use any third-party transcription APIs or services. We run all audio/video transcription in-house on machines we own or control.” and “We don't train AI or machine learning models on your media files or transcripts.”

  • 3 free transcripts per day with account (email required)

  • Pricing: $20/mo or $120/year for unlimited transcripts up to 10 hours long (each)

    

  • Interoperability: “TurboScribe supports the vast majority of common audio and video formats, including MP3, M4A, MP4, MOV, AAC, WAV, OGG, OPUS, MPEG, WMA, WMV, AVI, FLAC, AIFF, ALAC, 3GP, MKV, WEBM, VOB, RMVB, MTS, TS, QuickTime, and DivX.”

  • Accuracy: Site claims 99.8% accuracy. Bonus feature: it recognizes different speakers, e.g. for interviews (my primary intended purpose).

    

• Veed - https://www.veed.io/tools/audio-to-text

  • Free plan with watermark and 15 min/month TTS

  • Pricing -

    

• Yescribe - https://yescribe.ai

  • Free plan offers 3 files/day, up to 30 min each and identifies speakers

  • Pricing: Basic plan allows 10 files/day up to 5 hours each for $4.90/month billed yearly. Pro plan allows unlimited files per day at $9.90/month billed yearly.

    

References

https://www.descript.com/blog/article/best-free-audio-to-text-converters:

• Descript - Transcribing audio or video files using a computer (free for 60 min/month)

• Otter.ai - Transcribing virtual meetings - only 3 free lifetime downloads

• Fathom - Sales teams and customer service reps (worth a look though!)

• MacWhisper - Mac users only

• Google Docs Voice Typing - Google Suite users, not suited for interviews

• Windows Voice Typing - Windows users - could try with ‘virtual audio cable’ for converting MP3 to text (see tutorial)

Activity around applying generative AI to music has been escalating in recent months. It has the potential to disruptively impact people across a wide variety of work and life roles relating to music: composers, lyricists, performers, producers, voiceover artists, podcast hosts, advertisers, presenters, and so many more.

As many of you who know me are already aware, I’m:

• a lifelong music lover,

• a career-long data analysis geek, and

• a passionate advocate for diversity and inclusion.

Ethics of generative AI for music is a perfect storm for me - how could I *not* write about it? :)

On behalf of my collaborators and myself, I’m delighted to announce a new “state of the union” analysis, as of mid-March 2024, on ethical use of generative AI for creating music!

This 100% FREE, 100% human-authored set of posts includes an overview of the affected PEOPLE and the companies and PRODUCTS, and then dives deep into bias risks, copyright infringement concerns, costs vs. benefits, and other important aspects of AI ethics.

Our current plan, detailed in this INTRODUCTION post in our publication, is a series of 8 articles on AI for Music, plus living reference pages that we will continue to update as new events occur in this fast-moving space.

Please check it out, restack or share or add comments, and let me know what you think - especially if we missed anything you feel is important!

Leave a comment

I purposely didn't do that before trying setup on laptop #2, but I did fetch a new copy of the setup software. Leap Motion has clearly updated the Windows 7 setup app. This time, one of the first things setup did was to launch an (unsigned) NVidia app to check my video driver version. It seemed to take a while to run, leaving the browser window open in front with the Nvidia page mostly empty, but then I realized that behind the browser was a message from setup saying that my video drivers were more than a year old and I should update them, lest Orientation and Airspace not work properly. It wasn't clear if the browser window was ever going to find the new driver for me - it seemed hung. After waiting a while, I clicked Cancel and promised myself I'd do the driver update tomorrow.

Tomorrow: I checked for driver updates. Surprisingly, Windows said none were available. So this was a dead end.

This post was first published on the agile Teams WordPress blog on March 21, 2010. As of July 12, 2024, all agile Teams content has moved here to Agile Analytics and Beyond. Minor edits have been made to support audio voiceover for the podcast.

So what's 3D navigation really good for?

After learning how to fly around the world, I stepped back to think about where the innovations might lie - what the 'low hanging fruit' and the 'holy grail' might be for making the most of Leap Motion and similar 3D interfaces.

Touch-free

In some environments, not touching a control device is a Good Thing. A few scenarios that quickly came to mind:

1. wearing gloves (e.g. for insulation or protection) which could make it hard to use conventional devices (like mice and keyboards)

2. not wanting dirty hands (or gloves) to crud up control devices, and

3. not wanting clean hands (or gloves) to be contaminated by potentially dirty control devices.

To me, these exploits seem like the low-hanging fruit - quick wins. (I'm not saying they'll be easy to accomplish - for instance, how much will a normal pair of winter gloves interfere with recognition of hand gestures by people working outdoors? But they're somewhat obvious applications to target.)

Beyond touch-free

Thinking about ways to exploit touch-free user interfaces with existing applications is fun and cool, just like Tom Cruise's gloved computer navigation in Minority Report. It's like having a touchscreen in the air, and some useful things can be done with just that.

Yet it's a far bigger, but more interesting, challenge to think of completely new ways to use 3D touch-free controllers to do things that maybe we don't yet dream of being able to do. That's where the real innovations to rock our world will come from. But how can we discover them?

Trying to 'see the water'

We humans are capable of being "perfect filters": our assumptions can be so deeply ingrained that often we don't even realize we're making them. As Alan Willett so nicely analogized at SEPGNA 2012:

"Most of us are like fish in the water - we don't see the water."

Just as ocean water is hard for the fish swimming in it to see, applications where 3D control (not just a 2D virtual touchscreen) is useful can be hard to see. Finding the killer apps for 3D interfaces will require questioning ourselves about assumptions we probably don't even realize we're making about how we interact with our computers.

3D spreadsheets FTW?

As an example, look at how we use spreadsheets to store and manipulate data. We are so accustomed to a 2D paradigm that thinking about navigating our row,column oriented spreadsheet tools with a 3D interface like Leap Motion seems silly. (At least, that's what many of my colleagues said when I asked if or how a Leap controller could help them do their daily work on computers).

But what if 3D control became commonplace? What kinds of workbook tools could we have, and what new insights might we gain, if we could explore data cubes instead of sheets?

Look Ma, two hands

What sophisticated controls might be possible with two coordinated hands instead of one? In my experience this is very hard to do with devices, yet humans can learn to do things with two independent hands all the time, like playing the guitar. (Seriously, have you ever tried to use two joysticks or mice at once, one per hand? If you've ever gotten it to work, let me know! I'm not a hard-core gamer, but to me, it's like patting your head and rubbing your tummy at the same time.)

The holy grail for gestural control

A perfect storm for gestural controls would seem to be combining all three benefits: to have touch-free, two-handed, 3D navigation. Now that I want to see - and do - someday soon.

Leap Motion and the Leap Motion logo are the trademarks of Leap Motion, Inc. and are used here by permission.

Photo by DAVID ILIFF. License: CC-BY-SA 3.0

I quickly got GE configured for 3D control and, like everyone else, promptly had the earth spinning wildly! I also kept seeing the 'dots' on the screen from Touchless, along with GE's unexplained fat and thin up/down and left/right arrows. After disabling Touchless from the system tray, the dots went away, but my control didn't improve.

The support forum FAQ article showed two distinct user populations: those that loved flying around the world via Leaping, and those who found it frustrating. Videos seemed to help many people, so I visited the company's channel (it had 13,741 subscribers as of my visit).

I found lots of videos, mostly 2 weeks to 2 months old. The initial promo video provided by the company for leaping with Google Earth was cool-looking, but only 29 seconds long, and pretty useless for helping someone learn how to fly like that: it was impossible to judge correct hand position relative to the controller because the controller wasn't visible in the video. More promising was a playlist of 40 videos on how to #LeapInto Google Earth, plus some direct video links from the support forum.

Fortunately, it became clear that some early adopters who tried it out had started documenting. Julia's #LeapInto Sweden was short but helpful. By far the most useful link I found for guidance on flying without spinning was Leap Motion - Google Earth instructions. This video showed correct hand orientation and control, and explained what the fat and skinny arrows meant and how to navigate.

With that start, within an hour I was able to navigate comfortably around the buildings of NC State Centennial Campus and around Raleigh. Relaunching it an hour later required a bit of re-learning, but I was soon able to fly around San Francisco and under the Golden Gate Bridge, just like another demo video showed. I spent another 30 min or so flying to Vatican City, Sweden, the Statue of Liberty, and some places I used to live in the US.

Now I actually feel comfortable enough to consider showing it at an internal technology demo tomorrow. I can't wait to hear what people think!

Leap Motion and the Leap Motion logo are the trademarks of Leap Motion, Inc. and are used here by permission.

All other trademarks and registered marks are the property of their respective owners.

I found a developer video (based on v0.7.3) illustrating Visualizer and other functions on Windows 8, and another video which seemed to show using Touchless instead of keyboard and mouse on Windows 8. Great, but I have Windows 7.

One thing that perplexed me after installing my Leap Motion software was the appearance of a new "Flicks" system tray item. Microsoft's "What are flicks?" article says flicks are for a pen or finger only. It seemed logical though that the 'flicks' conventions for navigation (e.g. upward flick to scroll down) would apply to the Leap Motion controller.

Today (August 8) I launched the Control Panel to try out the Visualizer and was prompted to download another update. I did so, then tried launching Touchless from Airspace. It threw up the User Account Control dialog, I granted permissions, then - nothing. It just stopped. I saw one brief popup from the system tray about a smudge and needing to clean the window of the controller. I did that, but still no dice.

Restarted just for fun and relaunched Airspace and the Touchless app. No apparent change - nothing happened when I moved my hands over the sensor. This time I looked for hidden system tray icons, and found a new one. It didn't actually say "Touchless" - it just said "Interaction". "Open Help Guide" launched IE8 (my default browser for corporate reasons) with the tutorial. It doesn't run well in IE though. Switching to Chrome let me see the tutorial. Back in the system tray icon, I looked under the Interaction submenu and saw that it was "Disabled". I changed it to "Basic" and now I was able to see fingers on screen. Yay!

One immediate observation: the mapping of gestures to monitor only works on the primary monitor - no obvious way to navigate or control apps on the second monitor that I use when my laptop is in its dock.

Next I tried practicing scrolls and clicks and zooms on the tutorial page in Chrome. Clicking in the vertical scroll bar worked. The scroll and zoom gestures didn't. Tried a Word document and had trouble scrolling in that, although a light overlay of a mouse kept showing on screen. I had some luck "clicking" the mouse buttons, but that's not really what the sensor is good for! The keyboard icon popped up sporadically too - it wasn't clear what I was doing with my gestures to trigger it. I opened a PDF (saved earlier from the PC Magazine article "10 Best Apps for the Leap Motion Controller") and had similar results trying to navigate in Reader.

I revisited the Support Forum at this point to look for clues. I found one topic on what to do if the error "Touch Emulation Driver Corrupt or missing" appeared. Well, I wasn't seeing that error, but touch emulation certainly wasn't right. So I figured it wouldn't hurt to try the advice - manually installing MultiTouch from the Leap Motion directory. After a User Account Control dialog, I got a dialog for about 15 seconds saying "Installing, this could take a few minutes..." When it finished, I switched back to Reader, and now two-finger zooming worked. Yay!

All in all, some progress and some frustrations. I tend to work with multiple apps and windows open at once, and I frequently made accidental clicks on other windows on the screen. I clearly need practice - and will look for more crowd-sourced tips before I try it again.

For fun, I tried turning 'flicks' off from the system tray. It didn't seem to affect behavior at all.

Next stop: world travel with Leap Motion and Google Earth (the 11th item listed in the PC Magazine article).

Leap Motion and the Leap Motion logo are the trademarks of Leap Motion, Inc. and are used here by permission.

After getting started with my Leap Motion controller, I proceeded to 'get' the free Corel® Painter® Freestyle™ app. The prerequisites listed on the page looked OK, although the page warned that it would need 79mb and 'more space might be needed' for .NET 4.0; so this app will make an interesting test of whether .NET 4.5 would meet the dependency.

When I clicked "Get App", the Airspace store page for Corel Painter® Freestyle™ prompted me for my password, then quickly switched to 'Purchased', and my computer got a bit sluggish. For a few minutes, I assumed the app was being downloaded in the background to my laptop.

I did something else for about 15 minutes, but my Airspace home page hadn't yet refreshed to include a new app button, so it wasn't clear what was going on. [Based on later experiences with app 'purchases', no apps seem to actually download until Airspace is refreshed by exiting and re-launching.] I tried exiting the Airspace page by File|Exit, to no effect. Closing it by clicking the X at top right worked, though.

On restarting Airspace from my desktop shortcut, it detected that my LM sensor wasn't attached, and offered me the option of continuing with mouse only. Again, this feature probably originated as a convenience for development and testing, but leaving it in for customers is a sign of thoughtful error handling in the design of the software. I selected that option and continued.

My Airspace page now loaded with the 5 previous apps plus my new Painter Freestyle app which began "Downloading ..." About 15 minutes later, the download had finished. This seemed quite slow, even given that the app needed about 150 mb for its application files. Then I got a new Dependency page that didn't tell me exactly what was needed, only asked me to install the "Painter® Freestyle™ Dependencies". I always like to know what's inside those boxes before I install them - so for now, I didn't.

Painting in Airspace

Once I got the Orientation app working (see next story), I was able to launch this app without any difficulty, and without installing any additional prerequisites. It launched pretty fast and ran well (unlike Cut-The-Rope, which still doesn't run).

Actually painting something with the app clearly takes practice. On my first tries, I found it hard to move to the toolbar above the paint area (to choose a new brush or color) without unintentionally painting. This needs more trial time before I can 'draw' any conclusions.

Saving a Painting

Saving my crude paintwork as .RIF (the default) yielded a 2644kb file (for a 1280x800 canvas). Windows, of course, didn't know how to open it from Explorer. I was pleasantly surprised to be able to open the file from the app and save it as .PNG without using the controller (.JPEG, .TIFF, and .PSD formats are also supported). The PNG file wasn't much smaller, but saving the JPG as 'good' quality or 65% yielded a compact 181kb file; at 80% or 'high' quality, the file was 265kb.

I'm not much of a non-digital artist, so I don't expect to be creating any masterpieces with the app. :) Strictly for the curious, here's a half-size image of the example:

First try at using Corel® Painter® Freestyle™ with Leap Motion

More adventures to follow ...

1. Leap Motion and the Leap Motion logo are the trademarks of Leap Motion, Inc. and are used here by permission.

2. Corel®, Freestyle™, and Painter® are trademarks or registered trademarks of Corel Corporation and/or its subsidiaries in Canada, the United States and/or other countries.

After getting started with my Leap Motion controller, I had stopped at the point where the Orientation.exe app was crashing. A few days later, I was ready to try again.

Checking for prerequisites

First step: reconfirm prerequisites. As of May 19, 2013, the minimum system requirements were:

• Windows® 7 or 8 or Mac® OS X 10.7 - Windows 7 Enterprise (64-bit) with SP1

• AMD Phenom™ II or Intel® Core™ i3, i5, i7 processor

• 2 GB RAM - I have more than 8gb

• USB 2.0 port - check

• Internet connection - check

No problems there. Next I looked in the Airspace Control Panel for software updates, and found one which said it fixed startup issues with Orientation.exe. But still no luck. I checked the support forum and troubleshooting guides and didn't find anything useful. At this point (on July 29), I set my Leap Motion controller aside to work on other things.

First leap, take two

Absent any replies to my support email requests and twitter pings, I revived my investigation this past weekend (August 4) by considering possible unstated prerequisites. One that came to mind was the ".NET 4.0" that Cut The Rope apparently needs. Maybe Orientation.exe and other apps need it too? Using Aaron Stebner's .NET verifier, I confirmed that my prior . NET 4.5 installation did provide a functional version of .NET 4.0 ( both Client and Full).

Next I re-checked for updates (none found), and scanned the troubleshooting pages. Here I found a new Support entry from August 2 on "How to update your graphics driver". Aha! (Should have thought of that.)

Rather than following their instructions to the letter and trying to figure out exactly which NVIDIA Quadro FX 880M driver I should manually download, I used the Windows "Update Driver" function. That worked beautifully (after about 20 min on a bandwidth-constrained connection). All I had to do after installation was reset the driver configuration to 1280x800 (listed as the minimum by many Leap Motion apps) and full width.

Getting Oriented!

Now Orientation.exe ran beautifully. I actually didn't expect it to, so I still had my laptop on mute! I grabbed a pair of headphones to plug in and restarted it. It was way too loud at 50% headphone volume level - I reduced it to 15% and went through the app once. (Mute would have been fine. Orientation has no audio or aural instructions, and I found the music kind of annoying - more distracting than inspiring. YMMV.)

In 'where it sees' of Orientation, I noticed that the colors of the visualization changed towards orange or red when I made quicker moves - that was cool, and promising. In 'what it sees', I quickly figured out how to get it to recognize both of my hands, and most of the time, all of my fingers. But some simple gestures that I expected to work well didn't - e.g. gradually closing a thumb and finger into a pinch, or bringing a left finger and right finger together. In both cases, the two fingers would 'disappear' soon before the tips actually touched. I have not yet gotten this to work, but will keep experimenting with it.

'How to draw' was a good introduction to apps that use depth for 'touching' and 'lifting' from a surface. I definitely need more runtime on this too, but will probably just experiment with Corel® Painter® Freestyle™ app from now on. Trying to draw something of any complexity in Orientation, only to have the first parts erased, got to be a bit frustrating.

Who's training who?

My techie husband saw me doing these Orientation experiments and got curious. One of his first questions was, "Are you training it, or is it training you?" It's a great question. He shares my interest in understanding how well user interface technologies adapt to people, not vice versa. In the long run, technologies that require careful user training are not likely to 'stick' or be productive for sporadic use scenarios. (Key architectural quality attributes: learnability, memorability, and operability!) But for purposes of exploring a first version of a new UI technology, I don't mind being trained a bit.  

More adventures to follow ...

1. Leap Motion and the Leap Motion logo are the trademarks of Leap Motion, Inc. and are used here by permission.

2. All other trademarks and registered marks are the property of their respective owners.

Resources

Before jumping in, I read through the FAQs, legal terms, a tutorial on using it on Windows, and some ergonomic hints on the website, as well as the customary Twitter, Google+, and Facebook links. My overall impression was of a professional organization that has thought carefully about its users - an encouraging start.

Installation

Their startup web page auto-detected my operating system correctly, and the software download (about 57 mb) was fast, under a minute (kudos to whoever at Leap Motion planned for scalability of their web portal). Installation of "Leap Motion software and Airspace" was simple, was properly signed with a security certificate, and didn't require any special permissions (more kudos on all counts).

I plugged its USB cable into my laptop (Windows 7 64-bit Enterprise). After a firmware update and quick installation of device drivers, it was ready to use. So far, the road looked smooth and fast.

Startup

Their "Welcome" app launched after installation finished. Here I hit a glitch after viewing the "Orientation" hints and repositioning my LeapMotion sensor in front of my keyboard as they indicated: I clicked "Continue", and Orientation.exe promptly crashed (see image). But the crash appeared harmless, and their software continued introducing me to "Airspace", their app store (iTunes equivalent).

The next speed bump was registering for an Airspace account. I tried to register and leave birthdate blank, but registration failed. This was annoying - I'll readily acknowledge that I'm WAY over both 18 and 21, so I'm annoyed by spurious requests for the month and day (fodder for identity thieves). I took away a few kudo points for privacy concerns, and continued. Then when I re-tried the registration with a birthdate, I found the page had wiped out the password I'd already entered and I had to put it in again (twice). Grr. But moving on ... I made it to the Airspace Home page.

Five app buttons loaded on the page, and then another browser window popped up in front of it. I knew from reading the Leap Motion blog that way more than 5 apps were available, so I wondered if the home page didn't fully load.

The popup window indicated that the fifth app, "Cut the Rope" (a game), has dependencies I didn't yet have installed: Microsoft .NET Framework 4.0, and Microsoft XNA Framework 4.0. This puzzled me a bit because I already have .NET 4.5 installed. Shouldn't the Leap Motion framework and "Cut the Rope" be upward compatible with v4.5?

Being a software hacker (the good kind), I did a bit of poking around at this point. The dependency installation page pointed to C:\Users\[myID]\AppData\Local\AirspaceApps\CuttheRope\Dependencies\README.html, so I took a look in C:\Users\[myID]\AppData\Local\AirspaceApps\. It contained only 3 folders: CuttheRope, Lotus, and Shimsham. (I guess the "Airspace Store" and "Orientation" don't need folders.) CuttheRope contained an .EXE, some sub-folders, and LeapCSharp.NET4.0.dll. (So I'm guessing maybe it's not just CuttheRope that is going to have these dependencies.) Its "Dependencies" sub-folder contained installers (one .MSI, one .EXE) for both installations. Kudos for the self-containment, but this didn't fully offset the jolt of the dependency interruption.

My two cents: Seeing the first home page display get stuck on such dependencies is annoying and weighs down the up-til-now pretty lightweight installation process. Components required by the pre-installed LM apps should be automatically installed along with the Leap software and Airspace". I decided to ignore "Cut the Rope" and its dependencies, and try to run one of the 4 visible apps that did appear to load.

Airspace: The Leap Motion app store

Clicking the Airspace Store app button just took me to their website, https://airspace.leapmotion.com/, carrying with it my account information. I explored my profile a bit and found just Username (already populated), Email (already populated), and Payment Type (blank). Hmm. This left me wondering how many apps useful for my investigation would be free or reasonably priced. Of the Top Picks in the store, only a few (including Cut The Rope) appeared to be free, with most priced between $1.99 and $4.99. I'll be further exploring the "free apps" category link.

Google Earth looked promising, and was included by PC Magazine in their "10 Best Apps for the Leap Motion Controller". But the link almost completely failed to load in IE8 (all I got were the 3 tumblr links). It loaded much better in Firefox, so I plan to do most of my further app explorations there instead of IE.

Stuck at Orientation

Orientation.exe again crashed, this time with Problem Signature:

Problem signature: Problem Event Name: BEX Application Name: Orientation.exe Application Version: 0.0.0.0 Application Timestamp: 51ea9edd Fault Module Name: StackHash_0a9e Fault Module Version: 0.0.0.0 Fault Module Timestamp: 00000000 Exception Offset: 00000000 Exception Code: c0000005 Exception Data: 00000008 OS Version: 6.1.7601.2.1.0.256.4 Locale ID: 1033 Additional Information 1: 0a9e Additional Information 2: 0a9e372d3b4ad19135b953a78882e789 Additional Information 3: 0a9e Additional Information 4: 0a9e372d3b4ad19135b953a78882e789

Read our privacy statement online: http://go.microsoft.com/fwlink/?linkid=104288&clcid=0x0409

If the online privacy statement is not available, please read our privacy statement offline: C:\Windows\system32\en-US\erofflps.txt I tried using "Check online for a solution", to no apparent avail. I also tried searching Airspace for the Orientation app, but the only match was Corel Painter Freestyle (which is free, so I expect I'll give it a try).

At this point, I unplugged the USB from my laptop to get ready to go to a meeting. Almost immediately, I got a system tray popup that said "Smudge detected. Your Leap Motion Controller may function better if you clean the window.". Um, no, that's not the problem ;)

Troubleshooting

After my meeting, I did a little investigation in the Leap Motion control panel before ending my experiments for the afternoon. First I checked the Software Log looking for insights on the Orientation.exe crash, and found at the end of the log:

[15:00:16] [Info] Leap Motion Controller disconnected: LP81576857099 [15:00:24] [Warning] Smudge detected. Your Leap Motion Controller may function better if you clean the window.

So the disconnection was detected correctly, but the message was overridden by a spurious smudge detection - a minor adjustment might be needed in that part of the software.

I didn't find any clues to the Orientation crash in the remainder of the log, or in the FAQ. I'm going to inquire with Leap Motion support before installing anything further.

I did find a few perplexing log entries, such as "[12:21:05] [Warning] The device has entered 'Robust Mode' to compensate for lighting conditions." and many more messages about the device going into and out of "robust mode". That probably reflects the fact that I sit and move around between a big window and my keyboard & LM sensor. I may need to adjust my window blinds to darken my space a bit.

I also had one log message "[Info] Reversed tracking orientation". By default, the Tracking tab of the control panel has "Auto Orient Tracking" checked by default, so moving the device during setup probably explains it.

I may run those two dependency installations to see if it fixes the Orientation.exe crash. And I expect to try out a potentially useful "Diagnostic Visualizer" I found in the Leap Motion control panel.

Then I want to see how well it lets me navigate Windows (via the free Touchless for Windows app).

Running Windows post-installation for a day, without the Leap Motion sensor plugged-in, is a bit "off". Two observations so far:

• My mouse seemed to be moving and snapping relative to an unseen grid (wouldn't move diagonally). Unchecking the option to launch the tray item at startup and rebooting seemed to solve this.

• I need to be able to use vertical scroll bars with my mouse or touchpad while reading or showing PDF files in a meeting. My Adobe Reader is now stuck in 'touch' mode every time I open a new PDF file (no scroll bars, just a 'hand'). Attempts to change the touch mode setting in Adobe Reader don't persist. The multi-touch driver that was installed with Airspace appears to be stickier. I will definitely be looking for a solution to this.

More on my further adventures in future blog entries ...

Leap Motion and the Leap Motion logo are the trademarks of Leap Motion, Inc. and are used here by permission.

Wouldn't it be even cooler if we had touch-free control of our desktops, laptops, and enterprise computers, the way Tom Cruise did in "Minority Report"? (see WIMP video here)

Technology has been getting closer to making this vision a reality. For instance, this demo at TED 2010 showed a spatial user interface, but it required the user to wear special gloves.

A new device by LeapMotion proposes to make the vision of accessory-free, touch-free navigation a reality. I pre-ordered a Leap Motion controller last year, and watched as beta stories like NASA controlling a space robot with a Leap Motion sensor began to surface.

Today I received my new toy technology device. The LeapMotion controller is tiny: a sleek, lightweight device about as long and thick as my index finger and twice as wide. So it takes just a small bit of space in front of my keyboard. [see more photos here].

I will be writing more here as I try out my Leap Motion controller. Stay tuned!

Image credits:

1. TED 2010 image by Steve Jurvetson on Flickr, via Wikimedia, under the Creative Commons Attribution 2.0 Generic license

2. Leap Motion controller image used in accordance with their social media guidelines

Leap Motion and the Leap Motion logo are the trademarks of Leap Motion, Inc. and are used here by permission.

1. What's changed the most are the development tools. Compared to the simple text editors I started with, modern IDEs offer way more built-in guidance and 'accelerators' (although they do sometimes get in the way). And I'm learning to leverage the vast amount of online help and forums available nowadays. Learning to navigate the complexities of newer IDEs and debuggers, and become hyper-efficient in using them, will still take a little time, I'm sure.

2. Oddly, the languages themselves haven't changed all that much. I began with assembler and FORTRAN and quickly moved into RatFor, a C-like Rational FORTRAN preprocessor. I also did some work in Pascal and Ada, and lots of batch scripting, before moving to C and C++, then into Java. Picking up C# over the last few weeks has been straightforward.

3. My background in both agile methods and more formal approaches to architecture and requirements is clearly influencing how I do my work now. Thinking about what might be 'the simplest thing that will possibly work' and how to test steers my 'XP For One' task planning. I mull up front whether I need to design for performance to handle the huge datasets I'm working with now, and plan spikes to help me test early. For this project, throughput on my laptop is more than adequate so far: my programs runs through 4 years' worth of data in just a few minutes. Robustness and error detection in how I clean and process my data are critical, though.

Bottom line: it's gratifying to know that after so many years focusing more on management and process, I do still have the design, programming, and math skills to tackle and solve technical problems hands-on. It's still cool to wake up in the morning with a piece of a solution to a coding challenge I fell asleep thinking about, or to find my mind puzzling out an answer while I'm in the shower. And I like that this experience is building my mental framework for my future technical leadership, whether in management or coaching or research, to truly understand what product development teams using these latest tools are coping with. It's all good!

When the usual Adobe upgrade page (http://get.adobe.com/flashplayer/) fails to update Flash Player on Windows:

1. Launch favorite browser and go to http://kb2.adobe.com/cps/855/cpsid_85599.html

2. Download both of the EXE installers  (IE/ActiveX, non-IE)

3. Exit all browsers

4. Run both EXE files. (Order shouldn't matter.)

5. Launch browser(s) and, in each, go to http://www.adobe.com/software/flash/about/ to verify that the actual installed version is now the latest, vs. the table of versions lower on the page.

This procedure has only ever failed me at step 4, and only when remnants of previous browser usage with Flash since boot-up have somehow remained in memory. In that case, restarting my PC between steps 3 and 4 does the trick (I don't auto-launch any browsers on boot-up).

Bonus: no need to remember to uncheck the box on the main install page to prevent installation of the (free but unwanted) McAfee Security Scan Plus.

1) The hype level is amazing: Mashable's citing big stats on the numbers of Buzz users and posts - but when you consider the involuntary users (people who were, astonishingly, opted in by Google simply by virtue of having a Gmail account) and the buzzes being piped in from Twitter, FB, ... sources, I think actual use of Buzz is way lower than they're touting. The few people I know who have experimented with it (including me) essentially have wired up their existing Google Readers, Twitters, and other accounts, and aren't creating original content there. So it's not a source; it's just an aggregator - from that perspective and as an actual discussion platform, I consider FriendFeed superior in many regards.

2) It's concerning that Google seemingly didn't think through the privacy and security implications well enough before imposing Gbuzz on all Gmail users. It's no one's business but mine who else I email. Automatically building follower lists AND automatically exposing them in people's profiles, before they even know they have Buzz profiles, is astonishing. I'm glad Google has taken some steps after the initial uproar to address this, but that's not enough. For instance, I don't want people following me without my express permission; and no, it's not good enough to give me a 'block' option to remove them afterward.

3) Most disturbingly, 'turn off Buzz' doesn't seem to really turn it off. I just did an experiment in which I:

- posted a single buzz comment, and connected a public Flickr stream to account 1, which caused it to pick up a public photo posted months ago and include it in account1's buzz stream

- confirmed that account2, which was (automatically) following account1, could see the buzz and photo

- went back to account1 and disconnected all services, unfollowed everyone, and "turned off Buzz"

But from account2, I can still see account1's existence, profile, and the buzz and photo that were previously picked up.

If someone 'turns off Buzz', wouldn't it be logical that they probably don't still want their profile and stream to be visible?!

Account1 shouldn't have to re-enable Buzz, go to the list of followers and block them all, then turn buzz off again, in order to achieve that objective. But now, that's what I'm going to do (and if that doesn't remove account1's buzz from account2's stream, I'll manually delete everything in account1's buzz stream and post a followup.)

IMHO Buzz isn't ready for prime time yet. Maybe Gmail stayed 'in beta' longer than it needed to, but GBuzz seriously needed some alpha and beta runtime to catch and fix these kinds of issues before a broader rollout.

Update: re-enabling Buzz, deleting the buzz posts, and blocking account2 seemed to do the trick for taking account1's posts out of account2's stream. I've now removed all posts and all following/followers from account1 anyway, and turned Buzz back off, just to be on the safe side. But account1's profile still seems to be publicly viewable (perhaps because it's a Gchat buddy) to account2 - not obvious how to roll that back. And account2 still says it's following account1, even though account1 blocked account2. What does it mean to be able to block someone, if their account thinks they're still following you (and you're still in their list of people they're following, a list which they might be exposing publicly?) Also, be aware that if you block a follower, it'll be hard to find that person again to unblock them, and blocking may have consequences that go beyond Buzz.

Further update: There's a blog post at http://nobuzzingmail.blogspot.com/2010/02/now-google-says-we-quickly-realized.html which shows a screen shot of the Buzz settings tab and a 'completely disable' option at bottom. But it's far from obvious, and not that easy to even get this tab into your Settings (supposed to be able to enable the tab via Labs, but that path isn't working at present). The best official page I've found which describes how to unbuzz is at http://mail.google.com/support/bin/answer.py?hl=en&answer=171460 - but the choices it describes still leave a lot to be desired. The most complete removal option requires deleting your Google public profile (which is not unique to Buzz). Having trouble finding your Google profile? Try https://www.google.com/profiles/me . There's a link for deleting your profile once you go into 'edit my profile' (https://www.google.com/profiles/me/editprofile) - it's WAY at the bottom: https://www.google.com/profiles/me/deleteprofile .

Feb. 17, 2010 update: Google has somewhat changed 'how to unbuzz' yourself. I found several good articles via BlogHer (Cnet, ABDPBT, angelynnodom)

It was well attended, planned, and executed; other than a few A/V annoyances, my only quibble is that for future I'd suggest the organizers plan more lecture time for a more leisurely talking pace, and less Q&A time at the end. Presenter Tim Hodgkins, an Enterprise Technology Strategist from a nearby Microsoft office, is clearly experienced and knowledgeable, and he was refreshingly frank - kudos to him for a useful and entertaining talk that effectively addressed a diverse audience.

However, due at least in part to that diversity in interests of the audience, the topics covered weren't nearly as tech-oriented as I had hoped. While it had been described by NCSU as a 'business and technical' talk, the slides actually covered were pretty light on 'how' details, and devoid of discussion on financial benefits. My first red flag that it might not be very geeky was the absence of a prominently-posted Twitter hashtag ;)  Then Tim asked the audience at the start whether they were consumer users or enterprise users, and unfortunately for me, the consumers were the vast majority. He then sensibly tailored his talk accordingly, but to address both segments in the two hour lecture time allotted (minus A/V adjustment time), he visibly skipped a bunch of enterprise-focused slides, and had to go pretty fast overall.

I tweeted only once, during a brief A/V break, and the rest of the time I took old-fashioned notes - impressions and reflections now summarized below, with a side note on Windows 7-based entertainment in a separate blog post. (Those of you who already have hands-on experience as Windows 7 users or from beta-testing for months may disagree or have further insights to offer; comments welcome!)

Win7 specifics

• Versions: A number of key improved features for 'consumers' are not supported in the most basic Home version, eg Media Center. Tim also commented tactfully on the limited capabilities in the 'Starter Edition'. So perhaps the Home Premium pricetag should really be considered as the minimum cost for the vast majority of people.

• Upgrading - On the plus side, if Win7 is installed over Vista, you can get dual boot capability by choosing 'clean install' into a new partition. But there's "no upgrade directly from XP to Windows 7"?! Boo - and now the bar's even higher for what I'd need to see in Win7 to convince me to make the switch from XP Pro any time soon.

• Backward Compatibility costs money: "XP Mode" (Virtual PC 2007) isn't supported on Home Basic either, and it's the only fee-free solution (MDEV costs money and is for large enterprises). Also, 256mb may be too low for the default setting for shared memory assigned to Virtual PC, which means that (1) Home Premium users will have to muck around in the settings to get older apps to run (one audience member mentioned image editing program GIMP as an example of one that doesn't run in out-of-the-box VPC), and (2) they'll need more total RAM ($$$) to run the big VPC and normal Win7 effectively. Having the free toolkits available is a good start, and maybe the jump from Vista to Win7 is a no-brainer, but given how much app breakage ensued between XP and Vista, migrating from XP is a concern.

• Less is More? Some utilities that used to be preinstalled, like Movie Maker, are now separate downloads (although still free) as "Live Essentials" (not clear though if it'll be available to Win7 Home Basic users?) This continues the disappointing trend of taking away small but useful multimedia apps (anyone else remember Photo Editor?).

• Security: Plugging the removable-media security hole with BitLockerToGo is a good thing. After drive encryption is complete, it provides backup of recovery info to Active Directory for enterprises, or to another media for home networks. It may take several hours to encrypt a 200gb hard drive, though. Does a drive become unusable if power drops out during that time? Tim also said the very long text string which is your key is saved 'at the end'. I know if I have to wait hours for a large hard drive to encrypt, I'm going to start it and go off to work or bed - if it finishes but the computer loses power or reboots before I can save the recovery info to another media, will I get another chance? This wasn't clear.

Product Development Implications

• Caching Pluses and Minuses: Possible performance gains vs. possible incompatibilities from new Branch Office cache strategies (Distributed: Groove-like, Hosted: central server) - Distributed was said to be for "branches without any infrastructure". [NB: Can a branch actually have no infrastructure? Software systems with 'no' (intentional) architecture still have do one, even if it's just a "big ball of mud".] And will this 'play nice' with dynamically generated files, or with other file or data caching solutions that performance-intensive industrial-quality applications are already using?

• AppControl side effects? This control brings some potential benefits to enterprise IS departments [verbatim quote of Tim's comment on Windows 7 AppLocker, which I enjoyed so much I tweeted it: "Unlike what we did in XP, it actually works well"], at the cost of power user autonomy and convenience, and may create needs for new error handling in software installers.

• Backward Compatibility, Premium edition: MEDV (MS Enterprise Desktop Virtualization) looks quite nice - we saw a smooth demo of using MEDV for diverting IE8-incompatible websites to IE6 - but MEDV is "not free".

• Reproducing a problem: The new snippy tool is basic capture/playback with user annotation and logging of UI inputs, nicely mashed together. Product development teams may need to add test cases to see how/if Snippy tool works for capturing problem-reproduction data on their software products - is it only UI operations that get recorded, and (how) can it be launched programmatically? It will be useful for resolving PICNIC issues and for documenting steps-to-reproduce for industrial testers, but it's not really a substitute for remote desktop connections.

• Safe For Prime Time? - Remote Access (using Ipsec and IPv6) sounds promising, but was essentially not covered, and may not be usable 'everywhere'; many older routers, which may still be in use in homes and hotels for a while, do not support IPv6. Additionally, given the consequences of security breaches, I think it will be a tough sell to convince enterprise security gurus to give up their SecurID tokens and other VPN solutions, and trust that Remote Access is bulletproof. Unless there's already a ton of comforting data from intensive security vulnerability testing, I foresee a slow adoption curve on this.

Unanswered questions

(Tim invited the audience to email him and he'd get back with answers):

•  Is HomeGroup bi-directionally compatible with XP SP3? (sharing data as well as consuming it)

• Can Win7 backup/restore read Vista backups? (.ISO files?)

Serendipitous Discovery

• Using the Magnifier for demos (what a great way to make a detail visible to audience members in the back of the room! how did I miss this? it's in XP and Vista already) and/or on my netbook.

Overall

• When did 'surface' get verbed? ;) The slides used it repeatedly.

• Microsoft appears to be going after Google Desktop (which has come pre-installed by the OEM on the last several computers I've purchased) by integrating Bing, search connectors, and pre-indexed 'Libraries' into Explorer.

Bottom line: Windows 7 is evolutionary, not revolutionary, change. Some capabilities do seem to be improved, eg BitLocker security and AppLocker. Nothing jumped out at me as must-have, 'wow' innovative, or 'killer'. Overall, I'm still not convinced I should give up XP Pro on my own (non-enterprise) Wintel computers - but I'm still keeping an open mind.