Quantum Computing

Analog

AI

Quantum Computing

Briefest origin story: Quantum Computing as a concept follows from an 80's era collective conjecture, developing into a dedicated mission upon the publishing of Shor's algorithm in 1994. I am assuming most people are familiar.

So, 32 years later...

WHERE IT STANDS

There are a handful of companies providing platforms for quantum computing and they've produced devices that have many qubits.

Physical qubits are given in the table but relative to computation topics like code-breaking, logical qubits are what matter. Multiple physical qubits are required to achieve a single logical qubit, with the exact number varying considerably from platform to platform.

So, how many logical qubits are needed? For breaking encryption, the task so prominently mentioned 30 years ago, here are the estimates:

By the way, ECC-256 is the encryption used in the bitcoin protocol.

Alright, so the next thing we need to know is how many logical qubits are being achieved.

That top entry with 'Error Detection Only' sort of doesn't count because error correction is needed for the computation being considered. So for now, 28 is the max, and my reading leads me to believe that this was labwork.

Bitcoin would seem to be safe for awhile.

Bitcoin is practically a medium for QC hype. Voices with ominous tones tell the tale of massive movement and coordinated sales: a big player has lost faith.

So what can be done with the logical qubits available? Well, nothing to prefer it over the old fashioned silicon.

Consensus is that 100 logical qubits are need to definitively outperform the alternatives. This will usher in the era of "scientific quantum advantage", a time when quantum computing certainly outperforms supercomputers at several tasks of scientific interest. And it is at this point that all of this effort will be proven worthwhile.

Here's the roadmap to 100 logical qubits:

Wasn't going to include those two myths at the bottom, but that kind of smack talk is part of QC.

There are also claims that, by the end of 2026, algorithmic cleverness will allow for some "quantum advantage" applications even with currently achievable logical qubit numbers.

Now appreciate this fact:

The first logical qubit was achieved in 2024.

As with other new technologies and nascent sectors, most of the news comes from the companies and research teams themselves, so some skepticism is warranted.

On a personal note:

I remember the interest generated when Shor published. Before that I had heard nothing of QC and hadn't even taken Quantum Mechanics (or even Mechanics) yet. The most anticipated application, then as now, was secure communication, or quantum encryption/decryption. There was a guy I knew who was particularly enthusiastic about confidentiality and most of what I heard about the topic, I heard from him, the kind of person that learned German and Russian for no reason. Conversations often went to how quantum computing would break standard encryption wide open. "Give it 10 years...", always another 10 years.

But I don't think Quantum Computing has developed slowly, considering that the technology had/has to be manifested. QC just lags far behind the hype that proponents think is required to keep the endeavor moving forward.

Compare it to LIGO, which broke ground right around the time of Shor's publication. That project rests on a rather firm theoretical foundation most would agree, and was advertised as requiring mainly off-the-shelf 90's technology to implement. First detection (late 2015) was considered kind of late given the sales pitch. But it worked, and the latest upgrade should provide something to look at -- we have a new eyeball on the universe.

Analog Computers

A new niche for an old approach. New architectures for analog computers in theory can lower the energy consumed by the LLM computation, by factors in the range 100-1000.

A good Curious Droid video covering some of the past and present of analog computers:

There are a few approaches relevant to AI out there:

The CIM (Compute In Memory) approach using memristors is relatively simple and looks the most promising in terms of immediate applicability in the AI sector. As with Quantum Computing, most news in this sector comes from the companies themselves.

There were a couple of Veritasium videos some years ago about a company called Mythic using CIM. Cool stuff:

Then, later that year (2022), Mythic announced they had run out of money.

They managed to get $13 million to keep a pulse, and a few months ago received $125 million in funding.

I got a list of companies in the field from an AI prompt and everyone I looked at was a startup with hardly any revenue, their biggest announcements being their latest funding rounds. In fact, it looked like their reported revenue may have been a modest return on the money they're holding, but I don't know much about all that. I'm not really sure at this point who has products in the field.

Not CIM, but something about this idea I really like:

Anyway, Microsoft Research is funding a photonics-based endeavor and IBM is doing something as well, so there are $$$ in the mix.

As with Quantum Computing, the theoretical gains are irresistable motivation to anyone selling computation.

So from what I've taken in, the main drawback with the CIM method is a worst case error of ~5%. But I don't really know what that means. 5% on each coefficient of some final vector? 5% on some value used to calculate similarity? Is it noise or truncation? I caught a reference to an intermediate ADC/DAC stage, what's the resolution?

There is some claim of error detection, which would allow a calculation to be discarded in the extreme case and another attempt made. But there's always going to be some error in an analog system, look and you shall find.

A lot of us have used analog computers without really putting such a fine point on it. An analog PID controller, for instance, is an analog computer. What about a digital PID controller? Well, then you are using a digital computer to emulate an analog computer.

Aside:

I worked with an old engineer some time ago, and he had dabbled in analog computers as a younger man. Where he had worked, there were still vacuum tubes in the older analog computers, or 'analyzers' as he called them. I always thought that was cool.

It took awhile, but I eventually picked up on a vibe that I think a lot of his generation shared. Those of us with the computers, with our N-bit ADCs and DACs and signal processing libraries and infinite product-line with supporting software, we were a bunch of brutalist assholes who were leveraging a generic digital technology to utterly displace the soulful custom electrical engineering of the late 20th century, all without any specific hard-earned expertise in any given field. Ultimately, we brought no increase in fidelity or reliability -- we simply displaced expertise and thereby lowered labor costs.

It was a lot like how the old rockers hated the synthesizers, drum machines and sampling that took over popular music.

I'm more sympathetic now -- fuck the lot of us. I bet the old engineers love hearing hype about AI putting us out of their bitterness.

"We used to do all of that ourselves", the old engineer looking over his shoulder croaked, words that were tired but clung to the ear with swollen knuckles, escaping a breath of humid gingivitis that rolled down to the elbow like a graveyard fog and settled on his forearm with a chill.

The comments in that Curious Droid video have a lot of old guys getting misty-eyed over their bygone control systems. They had some great analog methods for data analysis, too.

AI

Game development with AI

It took a couple of years using AI extensively as a development tool before I felt comfortable letting it see entire projects. Or even entire classes/namespaces sometimes.

I feared the crutch and the spy but mostly the rogue corruption of the agent.

But I'd been wanting to test with a large-ish project to see what the new tools offered and it was easier to initiate something from scratch that had nothing to do with real work. I wanted something complicated and interesting, but isolated.

So I picked a game concept that I last played around with over 10 years ago, a world-builder somewhat-isometric 2.5D affair. When I did this before, there was good early progress, but as the game got more complicated features became increasingly difficult to add and run-time performance was a problem as I added the mechanics I wanted. Dijkstra's algorithm was melting the motherboard on relatively few nodes. Finally, abandonment.

I was impressed when chatgpt just spit out the hexagon map. I wasn't even thinking of taking that on, but after I made the initial prompt, the AI listed it as one of a series of things it would do to get me started.

Now, if you started something like this with a framework, you might already have that and not be impressed. But for a "roll your own" type, this is awesome to see. Frameworks can be good for rapid development of something simple and common, but much beyond that and you will be limited by the imagination of the framework developers. You can do anything you want -- provided it's been done before.

The game now has a lot of component systems constantly interacting. Dijkstra's algorithm is running for a few networks and the economic simulation models supply and demand fairly well, responding to trade war tactics like flooding the market. People move to live closer to their jobs because human transportion is expensive. It all has a good organic feel.

It is easy to say that I've gotten much farther along in far less time than in previous attempts.

I can't imagine how much this would have cost via API calls.

But I'm pretty impressed at this point. The project has gotten fairly large with a lot of modules, assets, etc...and the AI performs fairly well with some marshaling. More and more I let the AI manage, but edit the code myself -- it's just quicker.

I consider the speed up on overall development due to AI to be essentially infinite, as the project would have become intractable some time ago.

I've tested some of the AI features in Visual Studio, but the IDE gets sluggish and buggy. This will probably improve, maybe it already has but I'm leary -- separation of church and state.

Regarding coding, the speed up was huge in the initial development, shrinking as the codebase grew larger. Occasionally, I wished "my" code was a little more familiar so little changes could be implemented quickly without hunting lines or typing a prompt which involves about 80 times the amount of text I'm looking to edit.

Dev continues but check it out: islands. Player gets the left island, cpu gets the other. I am trying to make something like Age of Empires but with islands and a quasi-modern economy. Only the economy and traffic are reasonably well-developed so far. But soon, artillery and battleships. First, though, a few more commodities...

AI and government regulation

The latter half of 2026 sees no shortage of FUD, just a different incoming trajectory. Vast unemployment has been set aside and inevitable Armageddon offered up in its stead. And upcoming midterm elections of course have something to do with the latest uptick in the issue's media coverage, a harbinger of impending catchphrases and thought-terminating clichés.

Gah.

Fine, but I refuse to give a deuce. The only need I see for regulation is to explicitly assign the liability for error and the harm caused by AI agents. Things would tighten up nicely, I bet.

It would suck if AI ends humanity before it fixes our traffic issues, solves the P vs NP problem, and cures cancer.

But no one I know who really uses AI productively is very concerned about AI or AGI. Most of the FUD comes from the industry itself, casual users and those who don’t use it at all.

The real threat of AI is like the threat of matches in the hands of a child. The problem isn’t just the matches.

Maybe don’t let AI pick the bombing targets. (Where is Isaac Asimov when you need him?)

But the real motivation is to get government contracts lined up and dedicated to the big players. These companies want to be treated like weapons manufacturers so their economic wellbeing becomes a national security issue.

In the end, this will be a massive set-aside for these companies and their investors. Which will be a drag to watch, but the regulations will have meaningful impact only on the revenues of the big players.

I am interested, though, in learning what exactly would get you put in jail. The idea that you would go to jail for “developing or creating an artificial intelligence system that surpasses human cognitive abilities across most tasks” might be a little bit alarming, depending on your regard for human cognitive abilities.

--2025

2025 was the last year in the period of time covering AI's infancy and adolescence, basing my opinion mostly on the commercial availability of AI and the degree of utilization. A nice tidy number, too.

There were holdouts all through 2024 (couldn’t blame them), but the quality of the models exceeded some threshold in 2025. The most reluctant were brought onboard by Claude, which in that period produced output that was subtly better than the competition. It reminded me of the difference between my IPad and my Android tablets.

2026--

Widespread adoption and universal awareness.

Some say we have entered a time of acceleration, some say we have already reached a plateau.

At some point this year (early 2026), I started to get that queezy 'uncanny valley' feeling reading Claude text. Strange cuz I've never gotten that from text before and didn't really think it was possible.

Maybe this makes sense. Machine Vision preceeded these LLM's by a considerable length of time, and much accomplished pre-gpt regarding Machine Learning was in the visual context. Current AI-generated visual media has high enough quality to escape the uncanny valley. Perhaps the quality of GPT-LLM has only just now reached the minimum threshold required to enter.

Where's the fake part?

After the words are made and the words have been lined up, the thinker is no longer required.

A Negativist View

A few years ago there was a Google "engineer" who was claiming consciousness had been achieved by a model that was effectively something shy of GPT 3. I never worry about AI being conscious but I worry about people thinking and believing it is. How would you disprove it? Unfalsifiable in any way convincing to an acolyte. The type that is drawn to this assertion will find even more attraction in the near future.

The hype around the regulation push is similar in some respects, though I suspect Bernie Sanders codes more than that engineer ever did. Deception, fear-mongering, backroom deals -- very human.

Maybe they are right. This stuff is dangerous and we can't have people fooling around with it, kind of like biologicals. We need governments to take over entirely, like they did with research on deadly pathogens, so that nothing bad ever happens.

A Positivist View

Holding everything else constant, if AI simply attenuates the impact of human error and negligence then a lot of value is added to human society, and real money is saved on wasted time, remediation and insurance.

The best way to get any share in this is to use AI for one's own benefit as aggresively as everyone else.

I am not an accelerationist, but I am an optimist. We've acquired vast amounts of data in our research institutions, especially materials research, that no human has had time to sift through. AI looks to be the best way to brute-compute our way to some useful solutions, as we've seen in the academic realm. And like other positivists, I think we'll see undeniable benefits soon in the healthcare realm.

But I look again at that picture up there...