iiinova — where your digital self keeps evolving

pa28 Nasty 🇨🇦 @pa28.bsky.social · 13/10/2026

The race to achieve the poorly defined goal of GenAI first has lead many to become too uncritical of the current underlying technology. A lesson the field of Economics (and others) have learned and adapted to. Safe is slow, slow is fast. 7/7

000

pa28 Nasty 🇨🇦 @pa28.bsky.social · 13/10/2026

It is also becoming more difficult to ensure that the data being mined is not in some way at least partially produced by earlier generations of LLMs (except hardcopy printed before LLMs existed). 6/...

100

pa28 Nasty 🇨🇦 @pa28.bsky.social · 13/10/2026

LLM model generation on the other hand has concentrated on ever growing compute and input data. Now data mining has exhausted the easily reachable ore and is going after poorer quality (monitoring people), or harder to extract (hardcopy only). 5/...

100

pa28 Nasty 🇨🇦 @pa28.bsky.social · 13/10/2026

Both pre and post processing used deterministic processes built on real world modelling of the problem. That required seeking out mathematical tools from wherever we could find them crypto math, physics, engineering, DSP, etc. 4/...

100

pa28 Nasty 🇨🇦 @pa28.bsky.social · 13/10/2026

My first ML/AI project ran from the mid 1980s through 90s and was (like most successful projects) built on the work of others. The best compute we could get at the time was a single super computer or, if lucky, a cluster of a few super computers. Digesting huge data sets wasn't possible. 2/...

000

pa28 Nasty 🇨🇦 @pa28.bsky.social · 13/10/2026

If I wanted to improve performance or, more often, solve different problems using the available compute, I had to use different techniques. Some involved preprocessing or post processing which LLMs have in attention and temperature, but what I'm talking about is qualitatively different. 3/...

100

pa28 Nasty 🇨🇦 @pa28.bsky.social · 13/10/2026

This morning I woke up to justinwolfers.bsky.social podcast on the winners of the Nobel Prize for Economics. It reminded me of something from my work career when I was working with ML techniques and various statistical math techniques. 1/... #AI #GenAI #math #physics www.youtube.com/watch?v=hpT6...

youtube.com

Bob Hall and Mike Woodford Win the 2026 Economics Nobel

YouTube video by Platypus Economics with Justin Wolfers

210

pa28 Nasty 🇨🇦 @pa28.bsky.social · 12/10/2026

I think we should. I do read male authors, but female authors make up 2/3 or more of my library. This may be an artifact of selection bias because women's work needs to be so much better just to be published. (I also have some self published work by women in my library.)

041

pa28 Nasty 🇨🇦 @pa28.bsky.social · 11/10/2026

“Acting on their own” is an interesting way to say statistically and stochastically assembled a set of tokens which when rendered in language and sent to government websites were interpreted as requests to …

000

pa28 Nasty 🇨🇦 @pa28.bsky.social · 04/10/2026

Occasionally Meta coughs up some wisdom. This reel is one of the best examples of the harm the way we are using Generative AI I've seen. www.facebook.com/reel/1002410...

facebook.com

000

pa28 Nasty 🇨🇦 @pa28.bsky.social · 03/10/2026

No one elected the pope? I have notes.

130

pa28 Nasty 🇨🇦 @pa28.bsky.social · 02/10/2026

Indeed. What I can't understand is how many who work at Anthropic et al don't get that the underlying math says the volume of training data is not a substitute for the quality of training data. Hopefully they figure this out before scanning and shredding the Dead Sea Scrolls or cuneiform tablets.

010

pa28 Nasty 🇨🇦 @pa28.bsky.social · 29/09/2026

We can all better ourselves by following your guidance. I don't know you other than from our discourse here, but even just from that I can say that I am glad you have been able become the person you are. I am more glad than I can express that I share the planet with people who are like you.

110

pa28 Nasty 🇨🇦 @pa28.bsky.social · 22/09/2026

I'm far from sure that it isn't already too late to prevent what is coming financially.

000

pa28 Nasty 🇨🇦 @pa28.bsky.social · 20/09/2026

Very interesting read. Somewhat similar to Physicists who in discussions eventually land on "if you really want to understand Quantum Mechanics you have to understand the math." Always wonder about tech firms who are happy to admit they don't understand their products.

010

pa28 Nasty 🇨🇦 @pa28.bsky.social · 20/09/2026

Since ANNs are statistical machines and LLMs are ANNs with output randomization (the temperature layer), LLMs can be trained on only true data and still output incorrect answers.

000

pa28 Nasty 🇨🇦 @pa28.bsky.social · 16/09/2026

What is poor Whiskey Pete supposed to think now. Clearly Mike is not confident Pete will win the case. Can't blame Mike though, its been an Oscar-worthy shit show.

000

pa28 Nasty 🇨🇦 @pa28.bsky.social · 16/09/2026

LLMs evolved out of translation ML work where the Transformer does well enough. The gymnastics they are going through to expand the effective size of the context window and preserve the efficacy of Attention clearly demonstrates that LLMs are much closer to Eliza than AGI.

010

Reposted by pa28 Nasty 🇨🇦

Leah @leah.myatproto.social · 13/09/2026

These arguments rarely take into account the reality of human nature in which actually, there IS a large contingent of humans who say "fuck you" because we can. Can we learn to incorporate machine learning into workflow. Sure. Will we? No Why? Because, fuck you.

2173

pa28 Nasty 🇨🇦 @pa28.bsky.social · 12/09/2026

The only difference between LLM misalignment and LLM hallucination is a matter of scale. Both can ultimately be traced to regression towards the mean which is basically what artificial neural networks do. This can be very useful when applied with an understanding of the underlying math. Tech is hard

010

pa28 Nasty 🇨🇦 @pa28.bsky.social · 12/09/2026

The fallout from the AI bubble will be bad if we let it. Countries will continue to lean on asymmetric war if we continue to make that the easiest way to achieve their national goals. If we don't learn from history we will be forced to re-live it.

010

pa28 Nasty 🇨🇦 @pa28.bsky.social · 12/09/2026

We are watching one failing empire try to reconquer a former vassal state. Now a second failing empire is trying the same thing without even the sketchy reasons of the first.

100

pa28 Nasty 🇨🇦 @pa28.bsky.social · 12/09/2026

The good news is that what grew out of the fall of the Soviet Union: that the size of the economy, the power of the military and the sophistication of the intelligence is not correlated with success in military conquest. All empires fall when the foundations crumble.

100

pa28 Nasty 🇨🇦 @pa28.bsky.social · 12/09/2026

What is true, however, is that if we continue to allow grifters to make obscene amounts of money by making more and more people miserable we may come to wish it was not complete bull shit.

100

pa28 Nasty 🇨🇦 @pa28.bsky.social · 12/09/2026

I'm not an historian so I won't dwell on cause and effect except to say what we have witnessed, and what we are witnessing is a direct result of all that has come before. Two days ago Anthropic researchers opined we could all be gone in 4 years because of AI. This is, of course, complete bull shit.

100

pa28 Nasty 🇨🇦 @pa28.bsky.social · 12/09/2026

The collapse of the Soviet Union was a victory no one planned for, which left western intelligence, military and politics completely at a loss as to what to do next. We could have a peace dividend, if the military industrial complex would let us. Unfortunately there was too much money to be made.

100

pa28 Nasty 🇨🇦 @pa28.bsky.social · 12/09/2026

It is a mistake to view the event in isolation. The better part of a generation after the fall of the Berlin Wall, 3 1/2 decades after I started a career in intelligence, I stood in my team's OpCen watching a tower burn. In time to see the second impact and watch each tower fall on live TV.

210

pa28 Nasty 🇨🇦 @pa28.bsky.social · 12/09/2026

I've seen lots of news and social media about the 25th anniversary of 9/11 today. It is important to remember and celebrate the heroic actions, lives saved and lives lost in the attack; the friendships forged as air travel was shut down stranding people far from home and as we responded.

100

Reposted by pa28 Nasty 🇨🇦

Anna Holmes @annabookwriter.bsky.social · 10/09/2026

Again, always: 1. Question what motives they might have to say such a thing 2. Question why, if they truly believe that, they continue to work on the damn thing

25112

pa28 Nasty 🇨🇦 @pa28.bsky.social · 08/09/2026

Sadly for us the US Ambassador is a doofus. I am 100% in favour of punting him out though.

000

pa28 Nasty 🇨🇦 @pa28.bsky.social · 03/09/2026

During a conference in Montreal we went to an Indian place. I asked if they had mild options. The waiter said "Sorry sir, we don't do that." I ordered, it was spicier than I usually enjoy, but I love Indian food (I am nothing if not nuanced) just had to drink a lot of beer.

030

pa28 Nasty 🇨🇦 @pa28.bsky.social · 31/08/2026

Google dropped not being evil many years ago.

010

pa28 Nasty 🇨🇦 @pa28.bsky.social · 31/08/2026

I think New Yorkers get it, but honestly, as much as I've enjoyed my time travelling in the USA I'm not going back until the have as many years of sane government as they have had insane.

010

Reposted by pa28 Nasty 🇨🇦

Andrew Robinson (no, not Garak) @andrewr.bsky.social · 31/08/2026

Betteridge's Law: No/Non. If anything, numbers are going to drop further. The US has no idea how pissed off Canadians are with them.

2122

pa28 Nasty 🇨🇦 @pa28.bsky.social · 29/08/2026

There is also the Rideau Canal which still operates and gets around the Eisenhower Lock. It would be a transhipment nightmare but that would also create lots of jobs in Montreal and Kingston. Just sayin.

010

pa28 Nasty 🇨🇦 @pa28.bsky.social · 29/08/2026

I just finished listening to a security podcast about roll confusion as a security flaw, which is unsolvable. Every other sentence I was thinking "yes, that follows from the fundamental math of ANNs." They kept saying "that's a concern, but it is amazing how far LLMs have come in so little time." 🤡🤢

000

pa28 Nasty 🇨🇦 @pa28.bsky.social · 25/08/2026

Are we really considering regression towards the mean as a viable parenting strategy?

000

pa28 Nasty 🇨🇦 @pa28.bsky.social · 24/08/2026

RayBans were the iconic pilot brand when I was flying. Now I won't even consider them for any use. Slightly related, the drumbeat of the AI bubble pop is increasing in volume and tempo. I would not be surprised if Ray Ban is collateral damage of the coming financial blood bath.

010

pa28 Nasty 🇨🇦 @pa28.bsky.social · 23/08/2026

Is it odd that I’m thinking WTF is the Trump admin doing in the metric system?

001

pa28 Nasty 🇨🇦 @pa28.bsky.social · 23/08/2026

I still have nightmares about the refrigerator of shame.

120

Reposted by pa28 Nasty 🇨🇦

Katie Mack @astrokatie.com · 23/08/2026

So many of these folks are like “ooo we’re daring free-thinking mavericks” and it’s like nah man you’re just super racist

302171286

pa28 Nasty 🇨🇦 @pa28.bsky.social · 23/08/2026

A better question might be will you have a free and democratic election ever again. We might his excuse. Also, someone brief POTUS on the history of Canadian troops and especially JTF2. I've worked with them and they are not a special brand of 🇨🇦, just better trained. US victory is far from certain.

000

pa28 Nasty 🇨🇦 @pa28.bsky.social · 23/08/2026

During my employer's flirt with Agile in which one goal is to: "Delight customers by taking a customer-centric view to our work." Me: I talked to the 'customer' and they would like X. TL: Yeah, no, we aren't going to do that.

030

pa28 Nasty 🇨🇦 @pa28.bsky.social · 21/08/2026

Pagers are great for alerting, all you need is a modem and a telephone line. Pre-internet I wrote an automated alerting system to page appropriate people when systems needed human intervention, or whatever events they wanted to know about. Sounds intrusive but beat the heck out of overtime/shifts.

110

pa28 Nasty 🇨🇦 @pa28.bsky.social · 21/08/2026

To be fair detecting a fall from the POV of a watch must be difficult. Anyway by the time the first cut made it through to the UX there were 4 in the queue.🤯 I had been trying to train myself not to jump at every haptic event immediately, but I'm not doing that any more. Visions of talking to EMT...

000

pa28 Nasty 🇨🇦 @pa28.bsky.social · 21/08/2026

TIL Apple Watch is (I was going to say a panic artist) erring on the side of caution for fall alerts. I have a great pair of geared pruners I was using to prune lower limbs on a blue spruce. Each cut ends in a small jolt at each hand. Enough apparently to trigger a fall alert.

110

pa28 Nasty 🇨🇦 @pa28.bsky.social · 19/08/2026

If things keep going this way you won't be able to afford our products with tariffs on them.

030

Reposted by pa28 Nasty 🇨🇦

Texas Paul @realtexaspaul.com · 19/08/2026

🚨U.S. Treasury engaged in emergeny bond buyback to try and stem the out of control yield escalation.🚨

77765295