AI Will Create More Jobs Than It Eliminates By 2027

No. Most AI startups will die and have huge layoffs.

John Larkin Highland Tech Glen Canyon Design Center Lunatic Fringe Electronics

Might be able to pick up a nice data center for cheap in ~2029 when the money runs out. ;)

Cheers

Phil Hobbs

At 30 cents per KWH, no sane person would put a data center here.

Today's SF Chronicle (rotten rag) says that Anthropic has leased about a million square feet of premium downtown SF office space. They have lost about $6 billion so far.

Why does an AI company need a zillion employees?

Next step: IPO!

John Larkin Highland Tech Glen Canyon Design Center Lunatic Fringe Electronics

This morning's Register is all AI.

John Larkin Highland Tech Glen Canyon Design Center Lunatic Fringe Electronics

It's pretty useful for debugging software, I can feed it a few hundred lines of code and it can find the missing brace that's causing a cryptic C error message much faster than I can. It's also pretty good at auto-generating documentation. Who likes writing documentation?!

It's a good academic-to-English translator. A number of textbooks and whitepapers on e.g. physics and mathematics either just aren't very well-written or make certain assumptions about the reader's background and leave out steps, and GPT is adept at filling in the blanks for the confused.

Two years ago GPT really stunk at math but it's come a ways and it makes a pretty good ad-hoc computer algebra system. It's sort of remarkable that it can do this without actually connecting to a "real" CAS on the back end, as I understand it the transformer algorithm/neural network somehow constructs a rudimentary CAS out of nothing on-the-fly, which is pretty remarkable when you think about it. Example:

"I would like a resistor divider made up of 5 resistors in series, grounded at one end, and with some voltage going into the series string. Please devise constraints on the values to produce outputs at the taps where each output starting from the bottom, is approximately double the previous. Then select 1% standard tolerance values for a total string resistance of 100k."

Output: 6.34k, 6.34k, 12.7k, 24.9k, 49.9k. Looks good to me..

However it's kind of terrible for more free-form electronics design tasks and will often lead you astray about very simple circuits without close supervision.

Those are all LLM related tasks. That's the problem with the "chatbots"; they have conditioned everyone into thinking that's (all) AI is.

We'll likely see a whole new generation of devices that don't just "do what they are told" but anticipate needs and even suggest differences in usage patterns.

My "garage attendant" lets me know if any of my exterior lamps on the car are "broken". In addition to warning me if something is breaking the plane of the garage door (opening or closing) that the "electric eye" can't see -- like the nose of the car (backing in) not far enough into the garage or someone on a ladder accessing the "attic" space as the door is commanded open. Or, something IN the garage before it allows the door to open (it sends video of the scene to the car via a WiFi link) so you don't "let the dog escape" or discover something in the path of the vehicle as you enter).

My HVAC controller kept giving me "bogus" results while I was developing the learning algorithms for it. Actually, it was telling me that the ACbrrr was failing and I thought it's results were in error (the ACbrrr failed within a month!)

My irrigation controller told me about a leaking valve because it was able to "watch" non-stop. (I need a better water meter to get finer-grained usage information, especially for the "indoor circuit")

I suspect I could extract a fair bit of information from vehicles just by accessing sensors that are already in place (but, my interests don't lie, there).

Colleagues and old clients are already inquiring as to what they can do and how to do it. None of these things are "essential". But, neither is a GPS for your vehicle or coffee maker that turns itself on in anticipation of your morning brew. IoT and leaf processing will be the next decade's focus.

Today, a machine without network connectivity is a paperweight. Tomorrow, a machine that doesn't learn will be similarly described.

Today's GPT experiment:

"I would like a resistor divider made up of 5 resistors in series, grounded at one end, and with some voltage going into the series string. Please devise constraints on the values to produce outputs at the taps where each output starting from the bottom, is approximately double the previous. Please find 1% standard tolerance values for a total string resistance of 100k, and format the output in ASCII suitable for pasting into Usenet."

Output:

Resulting Tap Voltages (Vin = 1.0 normalized)

I then verified it in LTSpice, fed the netlist back in and asked it to make an ASCII schematic from the netlist. Output:

Not too shabby for an unpaid intern!

We (EEs) already benefit from AI in the tools that we use. We just don't think of them as AI, anymore (someone made that observation about how AI loses it's magic when it becomes commonplace).

We don't think anything of circuit simulation, autoplace/route, DRCs/DFM, wizards that build websites just by dragging "controls" onto a canvas, etc.

When your parts orders are triggered by an agent monitoring your stock consumption and orders in the pipeline, it just "seems obvious" -- why do we need a purchasing agent who has to be manually prodded?

When your warehouse staff are *told* where to place some item that was just received -- based on where there is available space -- and that "automatically" remembered by the inventory system, we think THAT is obvious, as well.

I don't believe The Register is totally AI. How do I know? I simply asked an AI.

I meant it was all about AI.

John Larkin Highland Tech Glen Canyon Design Center Lunatic Fringe Electronics

I often have a similar problem, but I want to use resistors that we have in stock. I have a program that lets me spec the resistor ratio, sizes to use, and range of Thevenin impedances.

The program does that for two resistors. More complex stuff, I Spice and fiddle.

After it lists all the solutions, I prowl the schematic to see if any of the values are already there, or if I can swap one. So I wind up with 7.87K logic pullups or something.

John Larkin Highland Tech Glen Canyon Design Center Lunatic Fringe Electronics

I beg to differ. If I search The Register for "AI" with:

If I search for "Artificial Intelligence":

I would expect a much larger percentage if the articles were "all about AI".

I just found a way to limit the scope to today using:

That's a substantially larger percentage, but hardly "all about AI". Also, the 11 articles containing "AI" did not all discuss AI, which would likely have made both the global article search and the daily "AI" counts much smaller percentages.

Asking ChatGPT "What percentage of articles in "The Register" are about AI?" did not produce anything useful:

I said "this morning"

You are obviously Bill Sloman.

John Larkin Highland Tech Glen Canyon Design Center Lunatic Fringe Electronics

Somebody who jeers at your sloppy thinking? Both Jeff Liebermann and I do other things as well, but since you classify your world in terms of the flattery its various components deliver, you may not pay attention to our other activities.

"This morning" was why I concentrated on estimating the percentage of AI articles that appeared on Feb 2, 2026 in The Register. I have no easy way to count the number of AI related articles that appeared that morning. My guess(tm) is the AI content was a small percentage of the articles which appeared on Feb 2.

There is probably some similarity between my writing and his. However, any possible similarity doesn't invalidate my logic.

I did not intended to offend you in any manner. I just spent four days unsuccessfully troubleshooting an electrical fault in my 2001 Subaru Forester. I was rather frustrated and felt the need to do something different and to offload my frustrations. My apologies for not correctly deducing your intentions.

I can see that four days chasing a car bug could be annoying. I gave up car repair some decades ago; it's just too complex and it makes your fingernails dirty.

Well, not literally 100% of the Register pieces were AI, but most were. Only about half are this morning. It's all going to crash one day soon. People are raising and burning too much money, like in the first .com (aka .bomb) thing.

The google AI responses are actually useful sometimes, and sometimes wrong. I asked it if an Efinix T20 FPGA could run my logic at 250 MHz, and it basically said yes, and it works. The logic includes my 50-cent

John Larkin Highland Tech Glen Canyon Design Center Lunatic Fringe Electronics

The plan now is to put the data centers into solar-synchronous orbits.

John Larkin Highland Tech Glen Canyon Design Center Lunatic Fringe Electronics

Then why say "all"?

It will certainly come off the boil as the real advantages become better defined, and the disadvantages become more obvious.

It isn't likely that it will follow the same trajectory as the dot.com bubble - it's a different technology offering different advantages, and many of the people involved lived through the dot.com bubble.

And if you knew what you were doing you could have worked that out by applying native, rather than artificial, intelligence.

The jitter on the output of a clocked system is usually limited by the jitter on the edges of the master clock, and if you can run that faster the jitter on its edges will usually be less.

Apparently the best you can do at the moment (if IEEE Spectrum is to be believed) is to use a whispering gallery mode in a sapphire crystal immersed in liquid helium as your master oscillator. The long term stability isn't as good as an atomic clock, but the short term jitter is appreciably better.

We can buy cheap surface-mount XOs with ps or even fs of RMS jitter. It's the DDS part that wrecks things. For 50 cents it looks like we can generate any frequency up to 40 MHz with ps jitter and 36-bit frequency resolution. It's cool to play with.

John Larkin Highland Tech Glen Canyon Design Center Lunatic Fringe Electronics