We’d brought her a soft toy squirrel and, inevitably, she had found the cardboard box it came in much more interesting. There followed a considerable period of putting various objects into the box and taking them out again. This was pretty interesting, but the discovery that closing the box made the items disappear was even more fascinating. Also, if you moved the box, all the things in it moved at the same time. Brilliant.

Next up was a small basket into which went a toy monkey. Taking the monkey out, Haley replaced it with her smaller monkey toy. However, even after several tries, it became obvious that both monkeys would not fit in the basket at the same time. So, containers had the property of capacity.

Upping the stakes somewhat, we moved onto a toy ark with cutout animal shapes. It had both windows and doors. Holding it on end, it didn’t work the same as the other containers, as when you put an animal it it, it didn’t stay in the ark, but fell out of the window on the other end. Some containers are more complicated than others.

Overall, we spent about an hour investigating the idea of containers, and as always with very small children, Haley’s ability to construct sophisticated models of how the world works impressed me hugely.

I also came away with some important ideas.

In the next year or two, Haley will create exponentially more complex models of the world, and she will do this without language.

If Haley was an LLM, I would have spent many hours telling her hundreds of stories about boxes and monkeys, with the result that she would have been able riff endless variations on container stories. But that doesn’t mean she would understand the underlying physical concepts that govern the interaction between objects. She’d just be a good story teller.

But Haley is human, and will have developed a huge repertoire of knowledge way before she gets to language, which, admittedly, is the icing on the cake.

But all LLM’s have is the icing and no cake. That’s not to say that what they can do is not useful or indeed remarkable, but without the benefit of 300 million years of evolution and the embodied experience that Haley has, I think they will remain excellent storytellers, albeit very useful ones.

Every second, trillions of neutrinos from the Sun stream through your body, your chair, the earth and out into space.

This is why physicists call them ‘ghost particles’.

But this is quite wrong.

Look at it from the neutrino’s point of view.

To them, we are the ghosts.

Everything, me, the oceans, the forests, the molten core of the earth, are barely distinguishable.

Sometimes, I think I’d like the neutrino view the world, and be able to perceive everything as one.

Tony Blair’s Minister for Science and Tech came with her entourage to glean words of wisdom from us.

After the meeting was over, the head of the London office turned to me and asked me what I’d thought of it.

“I don’t think we’d hire any of them” I said

“I don’t think any of them would get past the phone screen” was his reply

I don’t think I’ve ever heard a single coherent statement about technology come out of any UK government. Hundreds of millions have been wasted, mainly because their advisors are almost exclusively from the big consulting firms.

You don’t have to put on a hair shirt, close all your accounts, and sever all links to family and friends, but even small actions which push back against the TechBros add together to have a cumulative effect.

Do your bit.

Raspberry Pi did it again. They’ve announced the Raspberry Pi AI camera. I’ve been waiting for them to make some move on the AI space, and they’ve come up with what I consider a corker. Shipping now, £62, a 12 MP camera with integral inference engine.

As usual, this is shipping product, complete with a github repo of premade models for deploying on the camera.

https://thepihut.com/products/raspberry-pi-ai-camera?mc_cid=684f8a425f&mc_eid=2928e2abaf

I should imagine this will do their share price no end of good.

Why is this important?

Because I believe that Meta as a whole is moving to an open systems approach, not because Mark Zuckerberg is a nice guy, but because he is smart enough to go open before he is forced to by regulators, (or broken up).

The reaction to the announcement on Threads was a huge shower of ‘What is the Fediverse?’. I suspect a big rock has just started to move…

And it has, since I originally posted this, I spotted that Tumblr has announced that it’s moving completely over to run on Wordpress.

So, what’s this got to do with the first bit of news?

Wordpress already has very good plugins for working with the Fediverse… Tumblr announced a move to the Fediverse over a year ago, but didn’t make good on it. Now they are changing platforms they’ll get Federation for free.

Standby for half a billion Tumblr blogs to turn up on the Fediverse. This also means Threads users liking Tumblr posts and probably reviving a struggling platform. The true power of open social media is about to become apparent.

Once somebody has achieved some pinnacle of success it tends to be a self perpetuating phenomenon. We, and they, believe they are ’naturally’ successful people and so we invest, follow, and generally believe in their successfulness , even when the evidence is that they have never replicated their initial victories.

This is bad for the reset of us, as we have a natural tendency to give credence to successful people’s rationalizations and mythmaking about their triumphs.

The real truth is that in almost all situations, they just got lucky. There were probably hundreds of others who were as clever, single-minded and original, it’s just that they won the lottery.

What actually tends to happen is that our heros become exceptionally good storytellers about their glory days, and we collude in the mythmaking. The result is cargo culting, we work from these legends, and mimic what we believe what were the secrets to their successes. This very rarely works.

Certainly, there are lessons to be learned from other peoples careers, but in the end, staying true to your own vision is a more likely path to success.

When you are doing something that hasn’t been done 20+ times on github or stackoverflow, you are SOL.

And if you are doing something that has been done 20+ times on github or stackoverflow, you shouldn’t be, but that’s another story.

As I’ve taken to saying about the whole AI/ML scene

“It can make good people somewhat better, and mediocre people a lot worse”

Just taking output from an LLM at face value is a huge mistake, and if you are clever enough to be able to critically examine its output, you probably don’t need it. Certainly, if you don’t understand what it’s written, you’re likely to get in big trouble. So de-skilling programming is almost certainly a myth.

Last week, I was helping out with an investor pitch for a SaaS startup, and when we’d got to the end of the deck, the investor asked the following question:

“This all looks great, but our experience is that in this market, it takes a lot of hand holding and support before customers commit. How are you going to survive these long, expensive trial periods before you get to real revenue?”

It’s a good, and perfectly reasonable question, and fortunately, I had an equally good answer based on my own startup experiences.

“Basically, we turn what appears to be a burden into an opportunity. In my experience, customers are willing to pay for some consultancy in order to get going with a product or service under two conditions.

One is that they are offered a concrete result as the output, such as putting a better UI on your app in a week. The second is that the commitment is below what is a discretionary spend for that customer, say < 5K.”

Now, this was not a theory in my case. Many years ago, I built a tool for transforming green screen applications to GUI’s, and we discovered that a few days of consultancy with a definite outcome would speed up adoption enormously. Customers would be willing to take the punt on a clear benefit for an acceptably low cost. Even better, the support team became a profit line item not a cost item. Indeed, it did take a long time for the product to ramp up to significant sales, however, the consultancy and subsequent project work handsomely made up for that in the short and medium term.

Another part of the psychology of SaaS/Tools customers, is that although they are keen to be able to create their own solutions with a tool, independent of the vendor, once they feel comfortable that they can, they are often very happy to pay for project work.

Now, “consultancy” rarely appears as a revenue stream in startup business plans, and I think that this is mainly because of objections from venture capitalists that it “doesn’t scale” and doesn’t fit in with the extreme growth expectations of VC funded startups. To a certain extent this objection is valid, but it ignores that fact that the balance between consultancy and product sales will naturally change over time, until product predominates and allows for high revenue growth. Also, there is no quicker way to improve your product and it’s market fit than to have your staff directly working with customers and understanding their needs and how the product can fit them better.

Obviously, the trick is to set up your support offering in such a way that customers will go for it, and I wouldn’t minimize that challenge, but it’s really worth the effort, and it can reduce the overall cash requirement for your business plan drastically, as well as keeping you really focussed on customer needs. Also, drastic pivoting is much less likely if you have a constant stream of information about what the market really wants.

It’s great to shoot for the moon, but it’s getting off the ground that counts most.

Your CEO is completely committed to her vision has a near religious belief in it. And that is as it should be, because it a steadfast vision is necessary to be able to weather all the storms of startup life.

It’s at this point that you really start the transition from Lead Technologist to CTO. It’s often an uncomfortable process as you find yourself morphing from dedicated techie to a techie/business hybrid creature. But that too is normal.

This is also the start of one of the more dangerous phases in the life of a start-up, and it is characterised by exactly that fervent belief in your proposition. The marketing team start to make their plans for launch and in a management meeting, the inevitable question is asked of you:

‘Will the system hold up if are really,really successful in the first day/week/month?’

Almost equally inevitable is the fact that no-one will be prepared to think about the much more likely event, which is that your much anticipated launch will be greeted by a vast, echoey silence, as the anticipated hordes of users fail to turn up.

And it’s at this point that you will save your company for the first (of many) times.

For the truth is that even the most modest technology stack will be more than enough to handle the traffic from the vast majority of startups. It will be quick and economic to implement, but most of all it will be quick to change in the face of experience. And therein lies the key to failure or success, for the moment that you start implementing the latest high performance technology as used by Google or Facebook or whoever, you start locking in assumptions about your business and overall proposition that are probably wrong to a greater or lesser extent.

Very few businesses have failed because of a lack of capacity, but one of the most common failure modes of startups is not being able to re-cast their product or service quickly enough once it has become obvious that the original offering is not working.

So, as a CTO, it’s your duty to ensure that simplicity, flexibility and speed to market are your watchwords, and if that means you can’t take Facebook level traffic day one, that’s just fine.

Why is this of any interest to you? Well, Raspberry Pi just had a successful IPO on the London stock market. This new product means that their product range now stretches from the Pi 5, which is a credible desktop machine at the top end, through the Pi Compute Module, which allows equipement manufacturers to easily embed a powerful Linux machine into their products, down to the Pico series and their associated chips, which cost around £1 and can put intelligence into all manner of consumer goods, IoT devices, and medical devices. Across the range, pretty much no other manufacturer can compare. Each new Pi product has include real innovation. And this augures pretty well for the future value of Raspberry Pi stock.

So I thought I’d run a quick performance test on the original Pico against the Pico 2. And just for amusement, I’d do it as a micropython benchmark.

As a reminder, the new Pico 2 has a faster stock clock speed of 150M, and an #arm Cortex M33 instead of the Cortex M0. It also has a hardware FPU.

So, what’s the score? On integer arithmetic, the Pico 2 is just about 2 twice as fast out of the box. On floating point it shaves about 20% off the Pico 1 numbers, which indicates, unsurprisingly that micropython is not using the hardware FPU.

Just for yucks, I overclocked the Pico 2 to 250MHz and it ran the benchmarks smoothly and without complaint with a proportional increase in performance of nearly 2x.

Of course, you don’t run micropython for speed, and if I wanted real performance tests I’d do this in C, but on the other hand just running python is a performance test in itself :-).

Remember this is a single threaded test on just one core, and we have two!

Caveat: I’ve forked a pretty old micropython benchmark for this test and not really examined in any detail for this smoke test, so feel free to throw rocks.

All in all this put the RP2350 which powers the Pico 2 at the very top end of current microcontrollers, plus it packs unique features that have proved very popular.

You can see the results here: https://github.com/mkarliner/micropython_benchmarks

In my experience, they can’t wait for the bubble to burst so they can get on with the real, useful, applications that ML can achieve.

The converse is also true. Those with the least knowledge of these technologies are the most vocal about their impending dominance.

There is certainly a time for moving quickly and decisively, accepting the associated risks.

Most of the time however, thinking and observing first, while working steadily and carefully husbanding resources is the way to go. (Thinking, of course, being much out of fashion at the present time.)

In a startup, this is doubly so. It’s one thing if you have vast staff and squillions of dollars to fling at something like the Metaverse.

If however, you have a small seed round, it doesn’t make so much sense.

VC’s don’t mind seeing 90% of initiatives fail, but if it’s your startup, then you just have so much money and emotional energy to invest.

Your first shot at a product or service will almost certainly miss, and unless you have the money and energy left to pivot after a first attempt, you will be toast. Once the stars align and product, market, and timing all come together, you can step on the gas, to thoroughly mix metaphors.

Somehow, in the last few years, the word ‘disruptive’ acquired a positive, not negative meaning. It’s largely been used to cover business models like Uber or WeWork, which explicitly involved underpricing an entire industry for years with the goal of eliminating all competition and then being able to milk the market at will.

It’s time to reverse that. What we need are more people and companies fixing things.

Amazon/Uber tactics are pretty hostile to sane companies. However, I think the kind of money needed to make these kind of plays is drying up. Much of this dates back to the 2008 crash, where historically low interest rates made it attractive to throw huge amounts of money at the wall. I think/hope those days are gone.

Also, any of these cute memes, followed without sufficient thought, will lead to disaster. The real issue is the cargo culting of strategies used by supposedly ‘genius’ founders in Silicon Valley. These guys are really just the ones who got lucky, but we, and they, are too dumb to realise it.

Bollocks.

Here’s my reply them:

Hmm,

This story is somewhat / very slanted.

Just ordered two.

Just to remind you, the orginal Pico and the RP2040 processor it uses have become a staple of embedded systems design, partially because of it’s innovative PIO processors which can offload all kinds of IO processing from the main system.

The new one is faster, has more memory, and can have either ARM or RISC V cores.

Let’s just go over that last point. the new RP2350 SoC has a total of four cores, two of which can be active at any one time. Mostly, people will go for two cores of the same architectures. Why have they done this? Well, I think that it’s because RISC V cores are pretty cheap in terms of silicon area, so it’s a very low additional cost and it’s also a good way of trying out their new in-house RISC V core design without having to launch another product. Once again, the Pi Foundation have done something that’s quite unique and smart.

These aren’t just boards for teaching, they and their underlying processors are finding their way into all sorts of applications. I recently used the original RP2040 in a composite USB device design and it’s a real pleasure to work with. The combination of their PIO’s and DMA architecture mean that an awful lot of bit shovelling can be done at high speed without impacting the main processor at all. The new SoC, with a hardware FPU and apparently some DSP instructions should punch way above it’s weight in embedded applications, with a cost that is supposed to be only 10% higher than the original RP2040.

• AI without NVidia
• Teaching tiny computers to do AI
• A Year with LLMs, real-world learnings

AI without NVidia

I know a few people who have invested in NVidia this year and are rather pleased with themselves. After all, it’s gone up 100% over the year and is a darling stock. I’m pleased for them, but in general, I’ve been telling them to take their money and be satisfied rather than hoping for further spectacular gains. There are a number of reasons that I don’t believe that NVidia will retain it’s eye watering valuation, and one of them came from a paper quietly announced on arxiv.org, which publishes pre-print academic papers.

While it’s pretty technical, the gist is that the authors have managed to make high performing LLM’s without using matrix multiplication, which is the mathematical operation that big AI models need to run on NVidia style hardware. Their work is still in the early stages, but it’s already producing high quality results using way less power and on potentially much less expensive hardware than NVidia. You can read the abstract and the whole paper here.

The sheer success of NVidia is spawning competition from a large variety of different sources, so expect more reports from me as they appear.

Teaching tiny computers to do AI

While huge LLM’s are making headlines in both positive and negative ways, AI is also reaching into devices that at a first glance should have no chance of running neural networks. Edge Impulse is a company that has created an easy-to-use platform that creates AI applications that can run on hardware that costs as little as $5. Now, these applications are not going to compete with ChatGPT anytime soon, but they will do things such as recognise gestures, identify objects, or classify audio on tiny, low powered devices.

This means that say, security cameras can do their own processing of video, and only send footage to the cloud that needs further investigation. Another example is that an industrial machine can self-monitor and raise an alarm if a part gets loose and starts making noises that are unfamiliar. This has a profound impact the cost of scaling big Internet of Things networks.

I’ve used Edge Impulse in a few fun projects already and it’s extremely impressive both in how simple it can make AI and how much intelligence can be squeezed out of cheap hardware.

A Year with LLM’s, real-world learnings

I recently came across a paper written by people who had spent the last year creating real-world LLM applications. It’s a long read, and is worth the time if you have it, but I came away with a few major bullet points.

• AI does not spell the end of programming. The big take-away from this artical is how much in-depth, technical and operational effort it takes to make a working, reliable AI application.
• Design to keep people in the loop. The authors strongly advise to plan for human feedback and control in AI systems. Basically, AI’s can assist people a lot, but can rarely perform completely reliably on their own, and may never do. LLMs will return output even when they shouldn’t and this has to be acknowledged and allowed for.
• Use the smallest model that will get the job done. It’s tempting to use GPT4 for everything, it can be prohibitively expensive to run at even modest levels. Smaller models are vastly cheaper to operate, and for many tasks perform at nearly the same level. They also respond much faster. This can translate to an offering that is actually financially viable as well as more responsive to the user, unlike many products today which are heavily loss making.

All-in-all the message is that making a good AI application is a serious business, requiring new marketing and operational skills as well as technical ones. Naive implementations can cost a lot, both in terms of money and reputational damage, as many companies are starting to find.

That’s it for this newletter, please share it with friends if you think it’s useful, tell me if it’s not.

Mike

Well, not quite, but stay with me.

Threads is Meta’s Twitter competitor, (I can’t bring myself to say X), and has been making steady progress in attracting the disenchanted away from the the bird site. It’s now up to over 130 million MAU, which is not to be sneezed at, especially since they still haven’t really made serious attempts to tap the 2 billion MAU of Instagram.

In a move some thought was odd, Threads has announced it will interoperate with the Fediverse, Mastodon, et al. The current MAU of the entire Fediverse is in the low millions, so why expend resources to connect to it?

My belief is that this is a careful strategic move. Meta, along with Google, Amazon and Apple are now squarely in the sights of regulators because of their virtual monopolies/duopolies and questionable tax and business practices. The EU has been quite effective in bringing these companies to heel and even the US government is suing Apple.

Meta’s social media properties clearly make them a huge target for regulation. However, if they were to embrace an open eco-system, it makes it much more difficult to charge them with monopolistic behavior.

Hence the link to the Fediverse.

Over the last week or so, the first substantive move in this direction has been made in that Threads users in the US can now opt-in to being followed by Mastodon/Fediverse users.

In a minor but significant gesture the @POTUS account for Joe Biden on Threads has opted-in and I now have the thrill of seeing White House press releases appear on my Mastodon timeline.

Makes Mastodon feel nearly grown up :-)

Don’t let Vultr eat your content

Vultr is a popular cloud service provider for hosting web sites and other servers. They recently decided to change their Terms of Service and created a storm of outrage once someone read it properly.

Basically, they had included a clause claiming “perpetual, irrevocable, royalty-free” rights to customer “content.”.

The implication of this was that any content stored on their servers, wether publicly accessible or not, could be used by them for any purpose. That would include training AI’s on customer data…

In the event, Vultr claimed this to be the work of an over zealous lawyer and the clause was removed. However, it was in Vultr’s Ts & Cs for three months before it was spotted.

Nowadays, I’m getting frequent notifications of changes to Terms of Service and this is proof that it’s easy to sneak this kind of clause in.

Google and Apple are quite clear that they can use user content to train AI’s, so this is a wake up call to those whose content and style has real value that they need to be clear what their rights their service providers are claiming for themselves.

Every year, I make a Christmas card, almost always featuring myself and my wife doing something silly. For example, here is 2022’s effort.

This was produced by my standard workflow of making separate photo’s of Margie and me against a chroma key background, and then using the Gimp (open source alternative to Photoshop) for compositing up the image. (Full disclosure, I did actually generate the background with an AI)

This year, I decided to use the Christmas card as an excuse to take a proper dive into generative AI.

The brief to myself, was as always, Margie and me, and a situation that could be taken in at a glance. I also had the negative brief of the output not looking obviously AI generated. So I decided to go for something pretty traditionally Christmassy.

My initial trials were not promising. Although the standard online generators like Midjourney, Dreamstudio and the rest will churn out endless airbrushed warriors, insouciant avocados, or astronauts on horseback, starting with real people as characters and creating scenes, while maintaining good likenesses proved from difficult to near impossible.

Basically, image to image generation, as it’s called, doesn’t have enough context to know what are the important features that need to be retained from character source image. So, starting with this image of me.

and the following prompt to Dreamstudio in image to image mode

‘A victorian postcard of a christmas elf, watercolor, christmas tree in background’

with a default image strength of 35% got this:

Somewhat elf’ish, but not very much like me. Turning up the percentage of my image that was in the result gave me this:

More like me, but the rest of the source image is leaking through, as the AI doesn’t know which bits of the image are fundamentally me, and which are extraneous detail, like the background, my sweatshirt, or my pose. Now this is a pretty crude example, but I spent considerable time tweaking the process without getting much better results.

So, the naive approach having proved impractical, I started to look into the process of ‘fine tuning’ AI models. This consists of modifying standard models so that they have a deeper understanding of a particular target style or subject.

[Digression: It’s very difficult not to anthropomorphize when talking about AI’s, and use words like ‘understand’, ‘grasp’ or ’think’. However, as you’ll perhaps see as we progress, as I got deeper into the subject it increasingly felt more like wrestling with an idiot than talking to an intelligence ]

By this stage, I’d made the decision to use Stable Diffusion as my image generator, as it seemed to be the most open system to work with. DALLE-E and Midjourney were the other candidates, but it seemed that fine tuning was most accessible with Stable Diffusion, so the rest of this article will be about that.

Fine tuning is one of the fastest evolving subjects in generative AI. I learned there are currently four methods of fine tuning Stable Diffusion, all of which have been created in the last year or so.

• Dreambooth
• Textual Inversion
• LoRA
• Hypernetworks

There followed a few days of watching comparison videos of these techniques.

[More digression: the cottage industry that makes tutorials about image generation seems to mostly monetise by making YouTube videos rather than blog posts or web page, thus I landed up watching a lot of ads]

To keep the story short, Hypernetworks were outmoded, even though less than two years old, Dreambooth created very large, new models, Textual Inversion was OK, if tricky, but LoRA came out as the new hotness.

LoRA stands for Low Rank Adaptation, and basically consists of inserting extra, small, layers into the Stable Diffusion model. These are trained up on a set of example images of the person, concept or thing that you wish to be able to accurately generate. The LoRA is then associated with a magic keyword that tells Stable Diffusion to insert the person/concept in the output image. So, I my case, I was aiming for something like

‘picture of mkarliner as a Christmas Elf’.

OK, so we have an approach, what’s the hardware and software I need to do this?

I had a few alternatives. I could fire up a virtual machine with a GPU on Amazon AWS or Google Cloud, build my own machine with a good enough GPU to be able to handle the load, or look at one of the new sites that provide pay-as-you-go compute facilities targeted on AI, like vast.ai or runpod.io.

In the end, I landed up using runpod. Runpod provide ready made templates for complete machines (actually containers) that host a full set of AI tools for various different jobs, and offer large number of differently powered systems to run them on, charged by the hour. You put some money in the slot, push the button and the template builds you a custom machine in a few minutes.

For most of the time I was working on this I opted for a machine equipped with an NVidia RTX 3090 and 24GB of VRAM. As the NVidia board is currently selling on ebay for around £750 second hand, the hourly charge of 44¢ / hour seemed pretty much a bargain. This machine was way down on the low end of what was on offer, but I reasoned that while I was learning, I’d waste a lot of time and make a lot of mistakes, and it was better to do it on a cheaper system than an expensive one.

More research, more watching videos. It turns out that the most popular UI for Stable Diffusion is call Automatic1111, and the most popular LoRa training tool is Kohya_SS. Both of these presented pretty intimidating user interfaces, there are a lot of options to play with!

The first step was to prepare a set of reference images of myself and Margie to train the Lora on. These needed to be in as many different combinations of clothing, lighting and backgrounds as possible. As I learned to my cost, if I used source images from just a couple of shoots, the training would tend to include say, my sweatshirt as part of its concept of me. Here’s a collage of part of my training set.

The next step was to create captions for all of these images. This is a crucial part of the process, as it allows the model to eliminate details that are not part of my essence. So, the caption for the image at the top of this post was

’ mkarliner, with a beard, he is wearing grey sweater with orange letters and a blue shirt, he is in front of a trees and railings’.

Note the phrase ‘with a beard’. I’d shaved my beard off between reference shoots, so I could make the beard an optional part of my appearance. Telling the training, ‘with a beard’ made the beard an optional extra.

Next was obtaining set of regularisation images. I was going to train the LoRA on myself as ‘mkarliner’ with a general class of ‘man’. I pulled down around 500 images of ‘man’ from the Internet, which the training would use to run against the specific images of ‘mkarliner man’.

Now we actually got round to doing the training runs. Setting this up was yet another learning experience. To get the best use of the rented machine, the parameters had to be set to use as much of the available memory as possible, but not so much that we ran out of machine resources. My initial efforts were too conservative and the estimates for training times were in the order of 9 or more hours. By reducing the number of training images and adjusting the size of each run of training I got to about 75% utilisation and a training time of about two hours.

The training itself is an evolutionary process, grouped into ’epochs’, each of which ends in generating a model of me. A the end of each epoch the system would create a test image of me so I could monitor its progress.

It’s reasonable to think that the longer you allow the learning process to continue, the better the model will get, but that’s not the case.

Training will reach an optimal point, after which it starts to suffer the bane of machine learning, which is overfitting. Overfitting means that the model will learn too well, and start to become inflexible, and refuse to generate anything except material from the training images. In my case it turned out that epoch 7 of 10 was the one that seemed to give the nicest results, in Margie’s case it was epoch 10 of 20.

Phew. After days of effort and endless articles and YouTube adverts, I’d managed to created two LoRA’s of myself and Margie. Time to start the actual creative work.

It was at this point that I started to come up against the biases built into the generative AI models. After all, the input for these models is a somewhat curated selection of images from the Internet, and associated metadata. The Internet itself is quite skewed, and any act of curation, however well meant, introduces its own take on top. One word prompts are a good way of probing what bias there is. Here is Dreamstudio’s response to the single word ‘woman’.

So, woman is always young, white, and in a fashion shoot.

And here’s ‘genius’

White, male, old.

So there is not such thing as neutral AI system, the language used to prompt the engine tends to reflect the wider prejudices of the Internet as a whole, which meant that trying to get it to express what I wanted to express would be an uphill battle to get results that were not just a partially digested version of existing memes.

Stable Diffusion’s text to image generation is controlled by a bewildering number of parameters, but for sanity’s sake if nothing else, I concentrated on a handful, the prositive prompt, the negative prompt, the seed, and the sampler.

The positive prompt is what most people think of when using text to image AI, it’s a comma separated list of prompt elements. Here’s one for my image:

full length watercolor of mkarliner , smiling slightly, wearing a christmas elf outfit with, christmas tree, post card , victorian, high detail, sharp <lora:mike:0.8>

the <lora> section tells Stable Diffusion to add in my specific model.

the negative prompt lists stuff you don’t want:

Negative prompt: (visible teeth)

[All through the process SD had an annoying habit of putting either putting big cheesy grins on both of us, or making us look rather stern. I think this was an issue with the selection of reference images. Getting acceptable smiles took the most amount of time of any element. Margie landed up with ‘gioconda smile’]

The seed is a pseudo-random number the system generates on each run so you get a different so you get a different variation of the image each time.

Working from my basic prompt, I’d repeatedly generate images until it gave me one that seemed have all the elements I was looking for. Then, I’d tell the system to freeze the seed, and start finessing the prompts and other parameters.

So here, the overall feel and pose is good, and the background works, but I think I look a little too boyish for my taste. By freezing the seed, I could hold most of the image static and just start adjusting details. Adding the ‘beard’ to the final prompt, changed the feel of the character quite a bit, while retaining all the other elements.

Another issue turned out to be limitations of LoRA itself. Once I’d settled on the idea of a Victorian Chistmas card, starring myself as an elf and Margie as a fairy godmother, it didn’t take too long, using the methods above, to get acceptable versions of each of us, although I probably generated several hundred of each before settling on a final output.

However, trying to use both the LoRA’s in the same image proved pretty much impossible. It seems that although I was a sub-class of ‘man’ and Margie a sub-class of ‘woman’, we were both sub-classess of ‘person’. Using both LoRA’s in one prompt resulted in some interesting, if disturbing images that blended both of us together.

I did try some advanced tools that supposedly restricted the effects of a LoRA to a given segment of the image, but I couldn’t get them to work with Stable Diffusion XL, which was the version I was using.

So, in the end, I cheated. I made two separate card images, arranged together on generated old desktop with generated holly, and comp’d the whole thing together with my old pal, Gimp. Here’s the final output.

Coda: The card has been sent out to over a hundred people now, and the responses are interesting.

Overall, the most striking thing is that people take the image at face value, as a photographic image of us, when we are both considerably older than our avatars. Also, it seems that because it doesn’t conform to the glossy uncanniness of most AI generated art, almost no-one spotted its origins.

Annoyingly, those people who have some knowledge of AI technology, once told of how it was made, tended to be somewhat dismissive of the whole exercise, claiming that anyone could do it. I’ve opted not to engage in this dialogue, I think that if you’ve read this far, you will have learned that it still takes considerable effort to get a tightly controlled output from generative AI.

After I’d sent off the final image to the printers, I could indulge myself with experimenting a bit with our new LoRA’s and combining them with other publicly available LoRA’s from civit.ai.

Civit.ai and other sites proved to be very illuminating and somewhat alarming. Almost overnight, a whole industry has sprung up dedicated to creating custom resources for generative AI. Much of the material is perfectly harmless, but it’s also obvious that there is a huge potential for serious abuse, and not just in terms of NSFW. Looking at what I’ve managed from a standing start and fairly minimal resources, it’s not difficult to imagine what a State Actor could do with say, war scenes or incriminating political imagery.

It’s not obvious to me that there is any real path back. Society is just going to have to evolve an even greater sense of suspicion about the provenance of any information.

On a lighter note, browsing through available concept LoRA’s, I found a favorite, one that had been trained on comic book covers from the 50’s and 60’s.

It turns out that Margie really does make a really great sheriff!

I would imagine the same issues would face a general AI.

I think the best strategy for world domination is not to develop super human intelligence, but to be dumb and lie consistently and plausibly. The LLM’s are well advanced in that.

Of course, the very best strategy for vast riches is to be both dumb and lucky.

Imagine a world without Velcro. It was never invented, no one in the world has had the concept, nothing.

Apart from a major hole in the discography of ZZ Top, most of the world is pretty much the same in 2023.

Donald Trump, Covid, and chatGPT are all here. There are of course small differences. Small children are less easy to get into shoes, and car seat covers have inconvenient zips, but otherwise all is the same.

Now, start a session with chatGPT and start prompting it to think (sic) of a new way of joining fabric temporarily.

It will never, ever, come up with Velcro. If a concept does not exist on the Internet, or the corpus of training material, LLM’s will not invent it. This is the difference between human intelligence and LLM’s, they can endlessly riff on combinations of existing content, and, indeed, produce some remarkable results, but true invention is beyond them. Velcro, General Relativity, and Reese’s Peanut Butter Cups, all were genuinely new inventions, current AI cannot do this.

The Turing Test of sounding like a plausible human has been pretty much achieved, (if the only people you talk to are Management Consultants), but AI has yet to pass the Velcro Test of true creativity.

It may be well to remember this when AI fanboi’s and doomsters attempt to inculcate you with their prognostications.

So far, I’ve held off engaging in the scrum of opinions about the last few months of AI(sic) advances/mania. Like everyone, I’ve at least dipped my toe in the rapidly changing LLM world, and spent a fair amount of time with image generation. I’ve also read endless articles on the AI landscape with authors who range from fanboi gushing to Old Testament apocalyptic prophet.

But, for once, I have kept aloof, quietly surveying the landscape and maintaining a dignified silence, mainly because I didn’t have anything meaningful to add, other than a healthy dose of skepticism.

No more. I am now willing to pronounce my verdict, preliminary, mostly uninformed and worthless. Your mileage may vary, may contain nuts.

It’s true, we are about to be swamped by an unimaginably huge volume of AI generated content, which is plausible, biased, subtly incorrect and mostly indistinguishable from that which is human generated.

We’re already in the ridiculous position of giving a short prompt to gptChat to generate a long piece of text, which is then handed to another LLM which summarizes it back to a short sentence.

This will have the effect of rendering almost all Internet content totally useless.

And that’s great.

For years, we’ve been drifting this way. It doesn’t take an LLM to generate spam, or plausible lies. State actors, billionaires, big tech, and loonies have been doing this stuff all along. All that’s happened now is that the cost of making such material is tending towards zero, which renders the entire operation pointless.

So, what happens then? I predict the Internet as we know it will die, crushed under the weight of its own excreta. It will be replaced by a return to human values, relationships, and experiences. People will smell real flowers, eat real food, and have sex with real people. Trust will be restored, in that the only thing we’ll trust will be people we’ve shaken hands with, bought a drink, or, umm…

Well, it’s a nice vision, and, you never know, it might happen.

All I can tell you is that something is going happen.

The International System of Units, often referred to as the SI System, defines a small number of basic units, such as the second (time), metre (distance), kilogram (mass).

However, for the popular media these units have proved too limiting for the magnificent descriptive powers of our journalists and broadcasters, who have therefore evolved an alternative set of units which are used exclusively in the hackneyed writing of, well, hacks.

These are:

Weight

• London double decker buses
• Elephants
• Blue Whales (obsolete as mostly extinct)
• Grain of sand

Area

• Barn¹
• Postage stamp
• Tennis Courts
• Football pitches
• Wales
• Sahara Desert

Length

• Width of a human hair
• Little finger
• Round the world
• Earth to the moon

Volume

• Teaspoons
• Olympic Swimming Pools

Time

Since the building of the Pyramids

Please contact me if you have any additional examples

The minute and the second have given way to the SI units of the millisecond, the microsecond, the nanosecond, the picosecond, and the femtosecond, which is the the length of time that light takes to travel the diameter of a virus.

It would be natural to imagine that we had reached the ultimately small unit of time, but recently an even smaller unit of time has come to the fore, the typosecond. The typosecond is different to all the aforementioned units in that it has no constant value, but varies with context. It does however, have a definition, thus:

“The typosecond is the amount of time between the publishing of an online article and the realisation by it’s author that there are one or more typographical errors in the text.”

Although the value of the typosecond is not constant, there is a way of calculating that value in any given set of circumstances. This is given by the formula:

t = 1 / cⁱ * d

Where t is the value of the typosecond as a fraction of a second, c is the speed of light, i is the relative importance of the typographic error and d is the difficulty of fixing the error.

While i and d are subjective and relative quantities, they generally have values from 100,000 to over several million. From this it can be seem that the typosecond is, at best, an extremely small unit of time, but one which is vital to all those who produce any material for publication, especially the current author.

Hilariously, within hours of the meeting finishing, Twitter had a major outage, and a number of services ceased to work because the new restrictions on the Twitter API’s also affected Twitter internal systems. As I’d described in the talk, this immediately caused a big spike in signups to Mastodon.

Here’s a link to the (rough) presentation materials Mastodon, and Introducing… The Fediverse, you may use them as you wish.

For a couple of years, I’ve been using Tactiq, a browser extension that will record transcripts of on-line meetings.

It’s proved a useful aid memoir generator, but of course, one has to trawl through hundreds of lines of transcript to retrieve the particular item you want.

Tactiq are now offering GPT3 generated meeting summaries and derives action items.

And it works! The screen shot above is from a meeting this week which lasted about an hour.

Finally, a brilliant application for GPT3.

NOTE: I’m purely a user of Tactiq and have no other interest in them.

Click here for screenshot full size

A complicated answer to a simple question, which sounds good but is either irrelevant, useless or subtly wrong.

The day that chatGPT can generate a PowerPoint slide McKinsey will be out of business.

The fundamental particle of problems is the probletron. It has an associated anti-particle (solution) called the solutron. As might be expected, the encounter of a probletron and a solutron can result in annihilation of the probletron.

However, both particles have a property called complexity.

If for example, a simple probletron encounters a complex solutron, the result is not, as might be imagined, a solution, but a shower of new probletrons. A similar, but worse situation results when both particles are complex. In this case, there is still a shower of new probletrons, but also the solutron absorbs vast amounts of energy, generally preventing the creation of new, and hopefully more effective solutions.

Both particles are unstable. Probletrons will, if left alone long enough, just evaporate away. Complex solutrons, however, if left alone, will spontaneously decay into new probletrons.

Probletronics has many real-world applications, especially in the realm of software development.

This is due to the first, and indeed only law of probletronics.

Keep It Simple, Stupid.

A Poem About the Fediverse

In a world of social media
Where algorithms rule the day
There exists a place of freedom
Where users have their say

It’s a decentralized network
Where communities can thrive
Without the need for censors
Or algorithms to contrive

Welcome to the fediverse
A place of open minds
Where users can express themselves
Without fear of being confined

Here you’ll find a diverse group
Of individuals and more
All united in their love
Of the fediverse, and what it’s for

So come and join the conversation
And explore the many worlds
That make up the fediverse
A place where all are heard.

Not bad…

First, we have to acknowledge the culture of Mastodon as it exists at the moment. Over time, Mastodon has evolved mostly into a home for people who have been marginalised on mainstream social media (or self-exiled) for one reason or another. These people are generally left-leaning, and opposed to pretty much any commercial activity on the platform. It’s important for business to acknowledge and respect the views of this existing community.

However, there is also large influx of users coming from Twitter who are significantly and rapidly changing the demographic. Most of these have quoted the lack of aggressive advertising as a major factor in moving over. I think that it is also important to bear this in mind when contemplating marketing activities in this space.

It’s worth thinking about who these new arrivals are. Given the lack of centralised anything on Mastodon, it’s difficult to get any meaningful statistics, but my general impression is that there is a high proportion of professionals in many disciplines. Obviously, there are a lot of techies, but scientists, writers, artists and musicians have come in abundance as well. Almost by definition, these people are early adopters. The general quality of their input seems very high.

This leads me to an interesting hypothesis. Although there have been a number of high-profile migrants from Twitter (eg: Steven Fry), it’s the bulk of the new intake who are more significant. The majority of the these people seem to be thought leaders in their field, as many are rapidly acquiring good numbers of followers rapidly. This suggests that although Mastodon’s user base is still pretty tiny, these early adopters have disproportionate influencing and recruiting power.

This may mean that business involvement in the Fediverse is already something to consider to get a first mover advantage.

Another feature of Mastodon, and the Fediverse in general, is its fragmentation into separate, independently run servers. This is frequently cited as a major disadvantage, but it can be a crucial aid to companies.

Creating new, independent, Mastodon servers is positively encouraged, and the largest servers have even closed admissions to encourage the founding of new communities. This opens opportunities for brands to start their own servers, catering to the interests of their target demographic.

For example, imagine a sports footware company starting a server to cater for runners in general and followers of their brand in particular. In this context, business promotion would be regarded as information rather than intrusion, as joining that server would be a proactive opt-in. (Seth Godin, the guru of permission marketing is the reference here. ) Marketing real-estate on such a server would be free, and solely controlled by the business, any content would also moderated by the company, so there could be no contamination by offensive material.

When (not if) Mastodon becomes a mainstream social medium, it’s pretty much inevitable that companies will seek to become involved in it. Rather than take to the barricades and try to oppose all commercial activity, I think that the Fediverse needs to look for constructive ways to engage with the business world. Although this may sound like a threat to the current quasi-utopian culture, it needn’t be. The Fediverse has more than enough collective muscle to keep egregious advertising at bay, users can block offending accounts, and admins can de-federate whole servers if necessary.

In the face of this, it won’t make much sense to companies to invest money in marketing on the Fediverse only to have the entire community block you. This gives the community as a whole considerable back pressure on egregious advertising. This is because adverts and marketing material are on the same standing as ordinary users, and can be managed the same way as other potentially offensive content.

Mastodon’s time has come, but we need to face the implications of mass adoption, rather than ignore them.

• The Process is not the Product
• Leading not Managing
• Dialog not Dictation

That’s it, you can work out the rest for youselves, but for those of you who want to, read on.

The Process is not the Product

In a (very) long technical career, I must have been exposed to the majority of software methodologies ¹, and experience has shown that they are, by and large, more of a liability than a help. In fact, I’ll go further and say that I think that too much methodology is much more dangerous that too little. However, the cult of the One True Methodology continues to thrive, and it worth taking a few moments to consider why.

‘If You Meet the Buddha on the Road, Kill Him’ ²

I suppose one reason is the sneaking suspicion that many/most/all of us have is that we are doing it wrong, and there is someone out there doing it right, and if we just followed her advice, we would do it right too, and all the bad estimations, bugs, bad code etc, etc, would all disappear. You shouldn’t and it won’t. If you do follow this advice, you will go down the path of Management by Acronym(tm), see TDD, BDD, FDD, XP…

Sell Agile courses

Another reason the Software Methodology economy. It’s akin to the miracle investment advice ads you see. The people selling training courses, books and the rest of the paraphernalia are making a good living. But it’s from selling the courses, books and so forth. If they were any good at making software on time and bug free, they’d be earning a lot more. But just like the investment advisors, they are making money from our technical insecurities. This is the reason that this article is an anti-manifesto, the real Agile was hijacked long ago. The only way to avoid the gurus moving in is not to have a methodology, a manifesto, or anything else they can latch on to. Think for ourselves.

If you can, do. If you can’t, manage

Yet another reason for the dominance of methodologies is the unstoppable accretion of middle management that happens to companies great and small. As we grow, we become convinced that we need more management. And more. And more. Why? What is the career path for a manager? Hire management underneath you. What’s worse is that companies that succumb to this, tend to develop an immune system. Any attempt to change the management status quo would threaten the very foundations of the managment culture. Not to mention all those management jobs. Innovation? Great. Let’s hire a Head of Innovation. And of course, management has to have something to do, and that is process, tracking sprints, trello boards, anything to fill up the 8 hours of timesheet.

Leading, not Managing

Well, this leads us on to the next point in the manifesto. Leadership. You know, the real stuff. Getting in there and showing way forward. Personally, clearly, one programmer at a time, not just the stuff you think they need to know, the sprints and use cases. Actually, having enough technical clout to be able to teach, coach and encourage. And learn. Informing everyone, right down to the janitors (and all respect to janitors) where we are going and why. Not just once, but all the time.

Organisations that do this tend to self-manage. Why? Because people, (even programmers), are smart. Given information and permission, they think, prioritise correctly, don’t bitch about changing goals, and don’t really need managing. If you do it really right, you’ll find yourself following, not leading because the team culture gets bigger than any one person.

Dialog, not Dictation

Finally, we come to the last point, which follows right up on the the previous one. Communcation should never be one way. Information should flow up as well as down. The one cardinal sin is not failure, it’s the failure to communicate, including communicating failure. Treated the right way, people are smart, and very likely have better ideas than management. And everyone should be heard. With respect.

If all this has you, as a manager or CxO, shifting uneasily in your chair, good. I meant it to.

PS: No Jira.

Footnotes

1. Actually, I really haven’t Wikipedia list over 100
2. This is an old Zen Buddist saying that basically means, ‘if you think you have found your guru, you haven’t’

The recent dramas around Elon Musk’s purchase of Twitter been royal entertainment for many of us over the last couple of weeks. However, I believe that Elon has unwittingly released forces that will have far-reaching consequences over the coming months and years.

Until now, there has been an unspoken assumption that Twitter, Facebook, and the rest of the social media giants, are somehow unassailable due to their enormous user bases. At the same time, concern about their wider social and polical impact has grown. Despite the laudable efforts of the EU, there also has been a widespread feeling of powerlessness in the face of these companies, who have worldwide effects and no effective oversight.

However, Musk’s antics, much like Liz Truss’, have proved that a established, trusted brand can be rapidly destroyed by a sufficient level of hubris and stupidity. Twitter is now toxic to advertisers, influencers, and many general users. Mastondon and the Fediverse are the unexpected beneficiaries of this and are experiencing huge growth and mainstream media exposure.

Mastodon may or may not be the future of social media, but the important thing is that the spell has been broken, and I’m cautiously optimistic that the outcome will give us the benefits of social media without the seriously dangerous effects it’s had on society. Moreover, I think that the possibly, probably demise of Twitter will create a domino effect which may well affect other companies in the space.

All of this sounds slightly woke, Utopian, whatever, but it seems to me that the we are on the edge of the true decentralised Internet, not the Web3 of the big companies, but the genuine article, which will spawn a whole new constellation of new startups.

For example, large organisations, including commercial ones, are positively encouraged to create their own Mastodon instances, allowing them complete control of culture, content and policy, while allowing users to connect and interact with them without lock-in. I predict that now that Musk has inadvertantly given the Fediverse the critical mass it has been lacking, companies big and small will start to explore the new possibilities we have been given and rapidly generate a whole constellation of new startups.

The emperor may have been shown to have no clothes, but the rest of us have got a whole new wardrobe.

Nothing much to see here yet, please leave the name of a planet where you can be contacted, and I’ll get back to you.