AI destroying economic system

[SiliconWizard]

Elysium-This is the "default path"—HuXley's Brave New World is an overly optimistic utopia. Elysium, but without the slightest possibility of rebellion, a complete separation of "superhumans" and "human resources" on a decaying planet. This is already slowly happening—the super-rich and the rest of the impoverished middle class being pushed into the muck—Musk's Mars Project as a prelude.

Yes. Brave New World is somewhere between utopia and dystopia. But it depicts a world in which the goal is making society ultra stable and making everyone happy. Even if through completely artificial means.

What is currently unfolding looks nothing like that. Yes, we have the technology, but nobody cares about people's happiness or even stability. Many studies actually show that average "happiness" is on the decline almost everywhere except for a very few countries. As to stability, I don't think I need to elaborate. Nothing we currently do even remotely looks like an utopia. It only exacerbates competition, loss of freedom and anger.

[brucehoult]

so too it seems drivers as a job may also fall by the wayside eventually

Major difference between knockerup and lamplighter, and drivers is that most drivers want to drive. Sure, there are driving jobs and no doubt those drivers could we well sick of it, but I would suggest that most cars are bought by people who want to drive. If their car comes with AI actual full self-driving, it would spend a fair amount of time disabled.

I have a relatively sporty car, an STI-spec turbocharged 2008 JDM Subaru. I enjoy giving it some welly on the numerous hilly and twisty secondary roads in New Zealand, including a fair bit of gravel.

But there is an awful lot of driving-as-transport that is simply not a lot of fun — or at least fun is incompatible with keeping your license for very long.

Which is why I very much appreciate that while it doesn't (of course) have Full Self Driving, my 2008 model car does have camera-based full speed range dynamic cruise control, so all I have to do a lot of the time is steer. I have at least one time travelled 200km, including passing through central Auckland on the motorway at a time of day with considerable traffic, without touching either accelerator or brakes.

In fact traffic makes that easier, as when there is a sharp corner or roundabout etc the car in front of you slows down appropriately, which causes your car to also slow down.

[paulca]

But there is an awful lot of driving-as-transport that is simply not a lot of fun — or at least fun is incompatible with keeping your license for very long.

The major hazard developing on the roads is because most, ie the vast majority of cars are designed to remove all possible feel of driving and all possible labour of it from the driver.  Complete isolation from the world outside, including the road - as far as they can.

Automating things to unload the pilot was discovered to be a very serious issue in aviation for exactly the same reasons it is dangerous in cars too.  Inattention grows.

On the more technical side of it however... a Hyundia I40 I had as a rental... on a wet roundabout I squeezed the accelerator gently at 20mph.  By the time it got to about 28-30mph the front end started to wash out into understeer and the car drift wide in lane.   As expected.  Safety understeer/front wheel drive.  However.  There was ZERO indication, no scrub feeling, no lightened steering feel, no vibration, nothing.  Felt like driving a "Play Station" game.

This is dangerous, but it is exactly be design.

BTW:  I own a 2017 GT86 Pro.  Stock.  Except I put 225s on the front and left the 215s on the rear, which nicely removes all the safety understeer!  In it, could tell you the name of the person who scratched their name on the cat's eyes.  More literally and honestly, you can feel the white painted lines on the road through the wheel.

With 215s on the front if you "threw" the steering into a corner it would break away at the front instantly and plough.  WIth 225s on the front if you chuck it in hard, it will bite and you better be ready when the rear doesn't.

[brucehoult]

But there is an awful lot of driving-as-transport that is simply not a lot of fun — or at least fun is incompatible with keeping your license for very long.

The major hazard developing on the roads is because most, ie the vast majority of cars are designed to remove all possible feel of driving and all possible labour of it from the driver.  Complete isolation from the world outside, including the road - as far as they can.

Most, but not all.

As you point out, the GT86/BRZ is an exception, as indeed are Mr2, MX-5, Impreza/WRX and indeed many Subaru and Mazda models. And many European cars, but they have other issues when bought outside of Europe.

Automating things to unload the pilot was discovered to be a very serious issue in aviation for exactly the same reasons it is dangerous in cars too.  Inattention grows.

I find that the car automatically maintaining following distance, but fully manual steering, is a pretty good sweet spot for reducing fatigue but not causing inattention.

With 215s on the front if you "threw" the steering into a corner it would break away at the front instantly and plough.  With 225s on the front if you chuck it in hard, it will bite and you better be ready when the rear doesn't.

I discovered by myself a technique to keep the stock tyres while completely eliminating understeer, at least on my Subarus, but I think it should work everywhere. I make a small initial turn-in, pause for a fraction of a second (I don't know how long but it's short) to wait for body roll and weight transfer to the outside wheels, and then smoothly apply the rest of the needed steering angle, which you can actually do quite aggressively. I've been doing it for 20 or 25 years and just don't even think about it, it's just automatic on any significant turn. It lets you corner far harder, at higher speeds, than "chuck it in hard".

Looking on the internet this seems to be a standard thing taught in high performance driving courses for police, racing drivers etc but almost no civilians seem to be aware of it. And there doesn't seem to be a standard name for it. Some call it "progressive steering", some "two-stage steering".  In a way it's related to the "Scandinavian flick" in rally driving, but that uses an initial small weight-transfer turn in the OPPOSITE direction to the final desired sliding turn.

[paulca]

Absolutely, progressive steering with feel is the correct technique.

On a track however it will be plain as day and night to you.  It took me to the second corner to immediately sigh and then just cope with nothing but understeer.

The technique that worked was brake in a straight line, ease off the brakes and start progressively steering in, keep the car stable.  Run in a little hot and wait and wait and wait on the front end biting, walking it gently to the apex and (GT86 specific) do NOT get on the throttle hard.  Go neutral, then a squeeze until you feel the weight on the back wheels and then put it down.  The exit was just the entry backwards.  Feel the angle required, feel the scrub, stop accelerating or you'll run wide.  Wait... wait some more... run up the exit kerb and finally get the foot down.

How the journalists produced all those 50mph 2nd gear power slides?  Surprisingly most tracks are very low grip, but also mostly clutch kicks and/or handbrake entry.  ie.  staged.

The 86 peak power is in the 5k to 7k band.  On a track that was about my hairpin apex speed, maybe a little lower, but I was right at the bottom of the power band and if you stamped on the gas pedal while the car was still unsettled (front biased) it will "snap" and spin out.  However, once settled on the rear you can apply full throttle and it will not drift or break away at the rear.

Watch any "Learning to drift" video where they give people GT86s and a sprinklered track to play on... the technique they train is clutch, handbrake revs, clutch.  No true power slide entry.

Power slide entry in 1st is available nearly all times, but power slide in 2nd is a wet weather thing without a clutch kick.  On the track I got to it would be a warning for noise abatement.  "Drifting" is specifcally denied by the "neighour hood commitee".  On the road... power slide entry in 2nd gear, in the power band is at least 50mph.  You can do it on a roundabout, but a 50mph power slide on a roundabout will land you in jail or hospital.

EDIT: Note... understeer is common everywhere even in racing cars.  However, a more 50/50 balanced car with the safety understeer tuned out is a different experience.  At the limit point on grip the car rotation becomes basically uncoupled.  You can .. and do... deliberately switch back and forth between understeer and oversteer.  "pointing" the car on the brakes and on the throttle.  The ability to overrotate it into a spin available on tap.  This is why they don't come from factor that way.

On weight transfer.  Feint motion etc.  I'm sorry but I bought a sports car, not a family saloon with an attitude   Seriously.  I can exit a roundabout on full throttle and not delay an instant in the steering from hard right to hard left.  Do I fear a lash back?  Yes.  Does it happen... no.  The weight transfer due to the low CoG and chassis/suspension design happens in step with the steering.  The time it takes you to turn the wheel is how long the car takes to unload and reload without any delayed momentum from chassis roll.  (there is a little bit of roll and momentum, but it's tiny and fast to resolve.  something I do for fun is change lanes at 60mph as quickly as possible.  It's a good accident avoidance technique.  A hard tug left, wait and a hard tug right and the car is not upset at all.  Try and do it without the pause and you will get the stability warning light flashing if not an intervention of the nannies.  Like rocking the car left to right, left to right rapidly, agresively will bring it loose.

[TizianoHV]

A bit off track from the initial topic, but some crappy tires on the rear will help  .
We have a daihatsu sirion (FWD econo box) which is understeery.
After a flat the tire shop mounted two old winter tires on the rear. When it was wet I was drifting roundabouts   . Just lifting the throttle mid turn was enough for the rear to loose grip. For me it was fun, for my sister not much. We went bak to boring tires.

I discovered by myself a technique to keep the stock tyres while completely eliminating understeer,... make a small initial turn-in, pause for a fraction of a second (I don't know how long but it's short) to wait for body roll and weight transfer to the outside wheels, and then smoothly apply the rest of the needed steering angle,

I noticed the same thing! it has lots of body roll too, and applying a first stage of turning helps at "high" speed. At low speed understeer or handbrake

[paulca]

After a flat the tire shop mounted two old winter tires on the rear. When it was wet I was drifting roundabouts   . Just lifting the throttle mid turn was enough for the rear to loose grip. For me it was fun, for my sister not much. We went bak to boring tires.

That counter effect on FWD only gets worse as the performance goes up and the weight comes down.  They tend to have such a forward bias in weight  that even lifting off the throttle on the entry to a corner basically lifts the rear end off the ground.  "Clio Cup" race cars literally "cock" the rear inside wheel in the air on turn in.  In a FWD race car the back end is literally only there to stop it's arse dragging on the road.

I have only driven that level of FWD in sims and it usually results in quite a few awkward spins to realise the "Throttle steering" is backwards in FWD cars.  "Stamp it for understeer", "Lift it for oversteer".  Same as the spin correction response.  FWD... bury the accelerator and break the front end away too.  Completely backwards.

4WD?  Press everything and hope.

Back into safe, on the road driving tips.  To avoid that "lift off oversteer" in poor grip conditions... you have to be very sensitive around the transition point... the neutral point on the throttle.  Do not "lift" fully.  Lift the pedal back to the neutral point, pause just beyond it, let the drive train slack "slap" and then lift it more fully.  The transition through "positive" and "negaitve" force on the drive train causing gear slap is what causes spins in domestic FWD cars.  If you do spin, keep your eyes pointed where you want to go and steer there.  If you want, depending on how much space you have, bury the throttle to spin the front wheels, this should reverse the spin (Nigel Mansel technique)

Bringing into topic.

What do self driving cars do when something goes wrong and they unexpectedly loose traction?  Do they have recovery algos or do they just "FAULT, take control driver, your fault."

[Rick Law]

It hurts to look at how AI eats up our energy resources, sure.

That'll be short term.

A lot of companies are going to go broke when they run out of investors and they find that people aren't willing to pay anywhere near as much for their services as they cost to provide.

I think the quoted post hits the nail on it's head.

The impact to society is an academic discussion.  I do not believe the current generation of of LLM based AI can be effectively deployed wide-scale for general use.  There may be a few ran and used by government, but none that an average Joe can afford.  Reasoning follows:

(The current generation of LLM based AI) is financially unsustainable -- no one will use a chat bot if they have to pay the real cost.  For commercial use, the AI supplier's cost is by the tokens, but token usage is hard to predict on the buying side.  The AI-using company cannot realistically budget unless it is sold by subscription instead of sold by number of million-token usage.  This mismatch needs to be resolved profitably for the AI-selling side and affordably for the AI-buying side before commercial usage growth can be expected.  My guess is, the resolution will result in much reduced usage: The cost will be high reducing users which in turn drive the cost even higher.

Costs aside, I am not even convinced that we could run enough AI data centers for general public use because of environmental constrains.  The heat generated, the water requirement, so forth.  Look at the ones in the pipe line, Oracle's Stargate for example is designed to house up to 10 giga watt of processing.  Each step of the creation 10 giga watt will take its efficiency lost.  Whether your energy source is nuclear, natural gas, or whatever, it would require high efficiency at every step to pump in say 20gW to get 10gW to use.  20gW worth of heat per hour is a lot to dissipate, every hour.   For comparison, Google's AI summary has the Hiroshima bomb at 17.5 gWh of energy.  Even if Stargate generate the power elsewhere, consuming that 10gWh power will need to dissipate more than the 1/2 the bomb's energy every hour at the using site.  Even the plants and birds within miles will need to adapt to the increased ambient temperature.

Thus, few AI data centers will be build, users will be limited to governments and a few very big and very rich companies.

----
Reference/Interesting reading:
Forbes article: Uber Burns Its 2026 AI Budget In Four Months On Claude Code
^{https://www.forbes.com/sites/janakirammsv/2026/05/17/uber-burns-its-2026-ai-budget-in-four-months-on-claude-code/}

[Simmed]

i saw an interesting bit of news today
the title reads, the ministry of something is sending the minister *name bla bla* to USA to study the effect of job lost from AI
wut ?! ?
 

in my head, im going like
a) they dont know what is ai? why do you need to go USA to "look at it" ? ai is only inside country of USA ?
b) they are clueless, so they sent another clueless dude
c) they have no data on job loss ? its year 2026 already, oh my gawd !
d) they dont know how to parse useful data from labor stats ?

geezuz F $#$^&*
waste of tax money
waste of time
i doubt they will give any meaningful answer after this

[Siwastaja]

(The current generation of LLM based AI) is financially unsustainable -- no one will use a chat bot if they have to pay the real cost. 

Nah, for the "Generic Joes", the 2022-era ChatGPT quality is sufficient, and it's possible to provide with sensible amount of resource use.

Not everyone can afford running the flagship model of the day if the real cost is exposed, that's for sure, because frontier models are always tuned, by definition, to use as much compute as is financially possible for a significant enough user base for it to make sense to pay for the R&D and training costs of said model; it's a balance: it's not financially viable to do a model so heavy only a handful of people would pay for it, because then the R&D/training cost makes it even more expensive to them when divided over a small group; but there is demand for models that can be ran for some hundreds up to some thousands per month of typical payload. (So 2x compute per inference-time token is not just 2x more expensive; if it reduces the user base size, then the non-recurring costs per user start increasing.)

The most difficult place to be in is to offer frontier model quality for $20/month, exactly what OpenAI is trying to do with endless flow of investor money. This I believe will stop, or start having very much heavier usage limits, short-term special offers only forcing to upgrade, or ads injected, or other ways to monetize the use. Or, they will revert to simpler models with less compute cost, and offer that for $20/month.

[PlainName]

(The current generation of LLM based AI) is financially unsustainable -- no one will use a chat bot if they have to pay the real cost.

They won't realise that they are paying the full cost with their data. Tracking website visits and search history would be like tap water compared to the champagne of having their AI chatbot interactions soaked up.

[brucehoult]

(The current generation of LLM based AI) is financially unsustainable -- no one will use a chat bot if they have to pay the real cost. 

Nah, for the "Generic Joes", the 2022-era ChatGPT quality is sufficient, and it's possible to provide with sensible amount of resource use.

I've been playing around with leaving llama open on a RISC-V "Raspberry Pi" equivalent (SpacemiT K3, kind of about like a Pi 5 / Rock 5 / Orange Pi 5), since I'm not using it for much else most of the time. I've got (for no particular reason) Qwen3-Coder-30B-A3B-Instruct-Q5_K_M.gguf loaded, using 26G of the 32G of RAM.

Just at random I just tried:

> What size pine timber do I need to support a 100 kg load in the middle of a 3m long gap between supports?

To determine the required pine timber size for a 100 kg load over a 3m span, I need to consider several factors:

**Key considerations:**
- **Load type**: 100 kg (approximately 980 N) concentrated load at center
- **Span**: 3m (3000mm)
- **Support conditions**: Simply supported at both ends
- **Material**: Pine (assuming structural grade)

**Calculation approach:**
For a simply supported beam with central point load:
- Maximum bending moment: M = PL/4 = 980N × 3m/4 = 735 Nm
- Maximum deflection: δ = PL³/48EI

**Recommended timber size:**
For structural pine (e.g., Douglas fir or equivalent), you'd typically need:

**Minimum dimensions:**
- **200mm × 100mm** (200mm depth) - This is conservative and should work
- **250mm × 100mm** - More comfortable with safety margin

**Important considerations:**
1. **Grade of pine** - Structural grade is essential
2. **Moisture content** - Should be <19% for structural use
3. **Deflection limits** - Typically L/250 for floor applications
4. **Safety factor** - Consider 2-3x safety margin

**Recommendation:**
Use **250mm × 100mm** structural pine timber for reliable performance. This provides adequate strength and deflection control for your application.

*Consult a structural engineer for final approval, especially if this is for a permanent installation.*

[ Prompt: 32.7 t/s | Generation: 6.3 t/s ]

That took 58 seconds.

Grok does some internet lookups, takes 15 seconds, gets the same bending moment (well, 736) but says you can get away with smalles sizes 150x50, or even 100x50 won't break if it's temporary.

ChatGPT also takes about 15 seconds, comes up with similar to Grok but recommends a bit deeper, 150 minimum, 200 better. It agress with Grok that 50mm (45mm dressed) is thick enough.

The local machine can operate without an internet connection, and uses 15W while it's doing the token generation, so around 15*60 = 900 Joules or 0.00025 kWh or 10,000 such queries for $1.

Another classic example:

> How many Rs are in Strawberry? Be brief.

3 Rs in Strawberry.

[ Prompt: 20.4 t/s | Generation: 7.6 t/s ]

That was two seconds. Without the "be brief" it takes 20 seconds to generate a large explanation.

> How many Rs are in Strawberry?

To count the number of R's in "Strawberry", I'll examine each letter:

S-t-r-a-w-b-e-r-r-y

Looking at each letter:
- S: not an R
- t: not an R
- r: this is an R (1st R)
- a: not an R
- w: not an R
- b: not an R
- e: not an R
- r: this is an R (2nd R)
- r: this is an R (3rd R)
- y: not an R

There are **3** R's in Strawberry.

[ Prompt: 21.1 t/s | Generation: 6.4 t/s ]

So that one is not a problem now. The "walk to the car wash" one gets the classic stupid LLM response though :-(

Here's another.

> What happened at Ruapekapeka?

Ruapekapeka was a significant battle in New Zealand's land wars, fought on June 21, 1845, near the Bay of Islands in the North Island. Here's what happened:

**The Context:**
- The battle was part of the broader conflict between Māori tribes and British colonial forces
- It occurred during the Northern War (also called the Bay of Islands War)
- The British were trying to assert control over the region and enforce their authority

**The Battle:**
- British forces, led by Captain William Colenso, attacked the Māori pā (fortified settlement) at Ruapekapeka
- The Māori defenders were led by Te Kani Ake, a prominent Māori leader
- The British forces included about 150 soldiers and some Māori allies
- The Māori defenders were significantly outnumbered

**Outcome:**
- The British successfully captured the pā after a prolonged siege
- The Māori defenders were forced to retreat
- This victory was part of the British campaign to establish their authority in the region

**Significance:**
- It was one of the last major battles of the New Zealand land wars
- It demonstrated the British military advantage in the region
- The battle marked the beginning of the end of Māori resistance in the Bay of Islands area

The battle is remembered as a pivotal moment in New Zealand's colonial history, showing the military effectiveness of British forces in the region.

[ Prompt: 24.4 t/s | Generation: 6.3 t/s ]

That took 50s.

It's a pretty comfortable reading speed, to be honest.

Oh, here's a very weird response:

> How far have people driven on the Moon in total?

I think there might be some confusion in your question. People haven't driven on the Moon at all - no humans have ever driven vehicles on the lunar surface.

However, I can clarify what has actually happened:

**Apollo Missions:**
- Apollo 15, 16, and 17 missions (1971-1972) included lunar rovers (called "Lunar Roving Vehicles" or LRVs)
- These were electric vehicles that astronauts drove on the Moon's surface
- The total distance driven by these rovers was approximately **36 kilometers (22 miles)** across all three missions

**The Apollo 15 rover** drove about 28 kilometers (17 miles)
**The Apollo 16 rover** drove about 27 kilometers (17 miles) 
**The Apollo 17 rover** drove about 35 kilometers (22 miles)

**Important note:** These were not "driven" by people in the traditional sense - the astronauts had to manually operate the rovers, and the vehicles were designed to be operated by humans, not driven by autonomous systems.

So while humans have driven vehicles on the Moon, the total distance is quite modest compared to what we might expect from a "driving" experience on Earth. The vehicles were primarily used to extend the range of exploration beyond the immediate vicinity of the lunar modules.

[ Prompt: 22.3 t/s | Generation: 6.1 t/s ]

That took 45 seconds.

The individual distances are correct. The total and the "corrections" about the definitions of driving are whack.

Another...

> How can I convert a number such as 123.456 to IEEE single float by hand?

To convert 123.456 to IEEE single-precision float by hand, follow these steps:

## Step 1: Convert to Binary
First, convert the decimal number to binary:

**Integer part (123):**
123 ÷ 2 = 61 remainder 1
61 ÷ 2 = 30 remainder 1
30 ÷ 2 = 15 remainder 0
15 ÷ 2 = 7 remainder 1
7 ÷ 2 = 3 remainder 1
3 ÷ 2 = 1 remainder 1
1 ÷ 2 = 0 remainder 1

So 123₁₀ = 1111011₂

**Fractional part (0.456):**
0.456 × 2 = 0.912 → 0
0.912 × 2 = 1.824 → 1
0.824 × 2 = 1.648 → 1
0.648 × 2 = 1.296 → 1
0.296 × 2 = 0.592 → 0
0.592 × 2 = 1.184 → 1
0.184 × 2 = 0.368 → 0
0.368 × 2 = 0.736 → 0
0.736 × 2 = 1.472 → 1
0.472 × 2 = 0.944 → 0

So 0.456₁₀ ≈ 0.0111010010₂

**Combined:** 123.456₁₀ ≈ 1111011.0111010010₂

## Step 2: Normalize
Move the binary point to get the normalized form:
1111011.0111010010₂ = 1.1110110111010010₂ × 2⁶

## Step 3: Calculate Exponent
IEEE single precision uses 8-bit exponent with bias of 127:
Exponent = 6 + 127 = 133₁₀ = 10000101₂

## Step 4: Extract Significand
Remove the leading 1 and keep 23 bits:
Significand = 11101101110100100000000₂

## Step 5: Combine All Parts
- Sign bit: 0 (positive)
- Exponent: 10000101₂
- Significand: 11101101110100100000000₂

## Final Result
The IEEE single-precision representation is:
0 10000101 11101101110100100000000

In hexadecimal: 0x4276D900

**Verification:** This converts back to approximately 123.456001, which matches our original number.

[ Prompt: 14.2 t/s | Generation: 5.8 t/s ]

That took 2m10s. Other than needing to do another 7 multiply by 2 steps to complete the fractional bits it's got everything right all the way to the binary representation, but mucks up the conversion to hex.

SO CLOSE.

Grok takes 15 seconds uses the same approach, also stops the fractional bits too early (but adds a note to continue for higher precision) and similarly mucks up the conversion to hex. Hmmm.

As a general knowledge encyclopaedia kind of thing that you can carry in your pocket and use without internet on 14W of power ... this seems not bad. For most people it's better than asking your parent, colleague, friend etc.

Is there a better model I should run on my 32 GB RAM (RISC-V) Raspberry Pi?

[betocool]

Hey, which actual PC is the one you have with the SpacemiT K3 chip and where did you get it?

Cheers,

Alberto

[brucehoult]

Hey, which actual PC is the one you have with the SpacemiT K3 chip and where did you get it?

I have one directly sent to me by SpacemiT for free :-)  The same as this guy:





For some reason they sent him one with 16G RAM while mine has 32G.

This guy just did a pretty deep dive on the Milk-V Jupiter 2 (which Radxa sent him, again a review unit):

https://taoofmac.com/space/reviews/2026/06/11/1830


There are a number of companies selling two reference designs from SpacemiT. I expect that later in the year they will have their own, probably cheaper, board designs.

Deep Computing have a custom board design with the K3 for the Frame 13 laptop, which they say will ship at the end of this month. You can also use that in the usual Frame 13 main board desktop case from Coolermaster.

https://deepcomputing.io/product/dc-roma-risc-v-mainboard-iii/

If I was buying one, my default choice of vendor would be Sipeed, to go with my Lichee RV, Lichee Pi 3A, Lichee Pi 4A. I like their products in general.

https://sipeed.com/k3

Unfortunately RAM prices are currently crazy. The price of these boards has already gone up a lot from the first production batch in early May.

There is a list of vendors here:

https://old.reddit.com/r/spacemit_riscv/comments/1tcrx4o/collection_of_k3_ecosystem_hardware_partner_store/

[electrodacus]

Globally it seems around 25% of labor is subsistence farmers.
I do not think they will have big issues.

In western world apparently around 80% of jobs are in service sector most of it easy to automate in the next 2 to 5 years.
Close to 20% of these are governmental jobs. The choice could be to provide even more made up jobs or maybe it will be cheaper to offer some sort of minimal basic income or negative income tax.

There is no doubt for me that humans will become irrelevant and this will include those 1% or 0.01% mentioned by many above. Control will be lost just not sure how that will happen.

[paulca]

There is no doubt for me that humans will become irrelevant and this will include those 1% or 0.01% mentioned by many above. Control will be lost just not sure how that will happen.

I think humans will break first.  In some spectacular way.  There are a lot of things currently going into the "Ignore it for now" pile.  Most of those things have long shelf lives and will just stack up.  There is a lot of elasticity in it for it to be ignored untill it's too late.

Removing people's "purpose", "craft", "mastery", "agency".  It has already started to show up in industry.  I watch a lot of people still headed directly down the rabbit hole of "AI is a person", without even realising it.

As to how that will manifest and what will happen, shurg.  But I do not predict it will go well in the next few years.

I myself had a bit of a wobble in work.  Did some tickets with claude and got ... a dump of cortisol and no dopamine as reward.  Felt like I'd been "burnt".  Couldn't get any motivation at all to do another ticket.  Crashed and burned.  Into my 3rd week off.

[Electrodynamic]

Globally it seems around 25% of labor is subsistence farmers.
I do not think they will have big issues.

In western world apparently around 80% of jobs are in service sector most of it easy to automate in the next 2 to 5 years.
Close to 20% of these are governmental jobs. The choice could be to provide even more made up jobs or maybe it will be cheaper to offer some sort of minimal basic income or negative income tax.

There is no doubt for me that humans will become irrelevant and this will include those 1% or 0.01% mentioned by many above. Control will be lost just not sure how that will happen.

I think most of it is bs and hype.

Like the world fair or expo where technology would create a utopia and we would all be using flying cars next year. They keep saying everything is going to change and yet it seldom does. For example, they can build a huge data center to operate an advanced AI but the data center is still powered by burning the residue of dead plants/animals we call fossil fuels. Why not just hire a bunch of guys out back shoveling coal into boilers?, lol. So we can pretend to be technologically advanced but reality tends to prove otherwise imo.

[SiliconWizard]

Why not just hire a bunch of guys out back shoveling coal into boilers?, lol. So we can pretend to be technologically advanced but reality tends to prove otherwise imo.

Well, I guess people will remember the hilarious case of this "AI" company actually paying people in India to reply to prompts.
Of course, that will never happen again, ever.

[Siwastaja]

Why not just hire a bunch of guys out back shoveling coal into boilers?, lol. So we can pretend to be technologically advanced but reality tends to prove otherwise imo.

Well, I guess people will remember the hilarious case of this "AI" company actually paying people in India to reply to prompts.
Of course, that will never happen again, ever.

It's no different from autonomous robot companies driving their demo with remote controls and not disclosing that. Or medical test companies making up test results. Or a startup claiming they have a new breakthrough solid state battery.

This will keep happening indefinitely because it's human nature, and quite often someone goes to jail. Investors don't like to be scammed.

[brucehoult]

Why not just hire a bunch of guys out back shoveling coal into boilers?, lol. So we can pretend to be technologically advanced but reality tends to prove otherwise imo.

Well, I guess people will remember the hilarious case of this "AI" company actually paying people in India to reply to prompts.
Of course, that will never happen again, ever.

I don't know how you could do that unless those Indians were all at least PhDs and using at least Google and possibly an actual LLM themselves.

I've already demonstrated in this thread how a 30B model running on a cheap RISC-V SBC ("cheap" is relative for anything with 32GB RAM today, but it's $600) answers general knowledge questions, with zero internet access, better than almost anyone's parents, teachers, or colleagues would. And in a minute or two. Seconds in many cases. Not perfectly in every case, but as good as you're going to get from any one normal human in fields that are not their speciality.

[Siwastaja]

Are larger and better models really going to waste more energy?

I just had Opus 4.8 hunt a subtle data discrepancy, which could have been an indicator of a more serious bug, for 500k tokens. It wanted to give up and stop a few times because the difference was small enough not to matter, but I wanted to know what causes it. I switched it to Fable (now available again) mid-session and it figured it out in seconds, some negligible amount of tokens. It was a PRNG not being reseeded and as such leaking. In hindsight, an obvious thing from just looking at the code, yet neither I nor Opus figured it out, but Fable did. Had I used it from the start, it would have maybe saved 500k tokens by getting out of that rabbit hole without ever entering it. Being smart means being efficient and not wasting time in unnecessary rabbit holes, so maybe more advanced models do not mean more serious energy crises - unless we are giving them progressively more and more unnecessary and harmful workload to do?

[mobsr]

I can not disagree! Just read a news, Amazon is going to increase its hardware price, like echo. You may already know, Apple's iPhone 18 price is already up. AI is taking all resource away from consumer electronics. 

[Siwastaja]

AI is taking all resource away from consumer electronics.

Not all, but RAM and FLASH.

There have been positive changes, too: RAM and storage have been too cheap. Now many software vendors (including Microsoft) have had to do base minimum amount of clean-up of their bloated mess to reduce RAM waste, and that's starting to work. There is so much bloat. 99.9% of average computer users never actually need 32GB RAM, it's a requirement only because software has generally bloated. Some of that bloat is relatively easy to cut out, and ironically, it's AI which enables cutting that bloat in absence of really good human team (which is a scarcity).

I'm doing it myself all the time. I'm fitting more and more advanced features on a 220KB firmware image and 128KB of RAM. Many "software experts" would easily estimate I need 10GB of storage and 16GB of RAM to do the same things, like, "you need at least Raspberry Pi 5 to do this much". And it's AI which is reading through the codebase, finding waste to be cleaned, figuring out how the changes can be done without breaking anything, implementing the changes and testing them. Before AI I had to do the same but it took 100x more time from me and that was off from the actual algorithm work.

On home computers, people are doing things that inherently require not much memory or storage at all. They do the same word processing like things which were doable on 4MB of RAM and 200MB of storage in 1990's. They are browsing web pages that render text and some 1000-pixel wide images, that was doable in early 2000's with 64MB of RAM. They are video conferencing at HD resolution, but the RAM footprint of that is still in hundreds of MB, and the video compression is heavy on CPU / GPU processing, but still not that much RAM. So where is the 32GB minimum requirement to run Windows coming from?

It's coming from nowhere, that's exactly why it's down to 16GB almost overnight from MS internally asking Claude "plz find some low hanging fruits to reduce memory consumption". With progressively a bit more work, it could easily go down to 8GB. At that point we are starting to approach modern use cases of significant minority who really needs 8GB of memory for their actual payloads.

So it's easy to blame the AI industry hogging up all the memory but that is wrong, we are just blaming the one who came to the party last from eating all the food, when those at the party earlier hoarded it all with no real need. Insane RAM spending has been industry practice before AI, and the unwillingness to stop hoarding and wasting the finite resource is equally the fault of existing RAM users - Microsoft is one, but I guess phone manufacturers are the worst. It is absolutely ridiculous we are putting 8-16GB of RAM in everyone's pocket so that they can send text messages and photos and videos that are looked at a tiny postcard size screen, listening to music, looking at videos of cats doing cat things again in postcard size. All of that can be done with 1GB, and it's not long ago smartphones did basically the same things as today with 10x less memory. Some flagship model might be a bit faster now, but even that is thanks to the CPU/GPU, not amount of memory.

Memory has been just too cheap.

AI companies are doing their share (not from goodwill, but because they have to) - improving the models without making them any larger than they already are (and they are huge). A perfect published example would be GLM-5.3 performing significantly better than GLM-5.2, at least on benchmarks, with the same architecture/same model size, just training improvements. That's the LLM development equivalency of me putting more features on the fixed 220K firmware image size.

And the software professionals are slowly waking up to this reality, too. For hardware-oriented software amateurs like me, it was obvious. For AI researchers, it was obvious. Only for software folks it wasn't. Sensible people, mostly advanced end-users and hobbyists, have talked about dangers of software bloat for decades(!), wondered why some simple software package is gigabytes to download (already in early 2000's when that was truly huge!) but all you get from Software Professional is arrogancy "no, bloat is not a problem, we have resources". And they always overshoot the resources. This is why computer use always more or less sucks, it's being regulated on the threshold of barely not sucking, first build too bloaty, then fix just barely enough that UX complaints reduce to manageable level. That's called "avoid premature optimization" in professional speak. Now these poor folks have a new* variable in the system they never anticipated: RAM/disk cost explosion which means users complain about RAM price too and suddenly ask everything to be done with half the RAM.

*) not completely new, it has happened before, but it seems this time it's even worse

Then again, maybe we can thank all the bloat. It's exactly the constant, unchanging bloat trend that has necessitated optimized RAM manufacturing at low cost, and that cheap RAM was the enabler for AI which truly needs a lot of RAM. So now we can have AI. But it's killing the feeding hand that made it possible. Now the RAM industry lives and scales up from the AI demand alone.

But classic software developers don't get a free pass from "being first at the party". The scarcity is shared equally and everyone's paying fair price. If anything - AI is harder to make smaller, traditional software is easier to optimize for memory and AI is helping with that, so really we should not be complaining, we should be solving the problem by killing it exactly where it is: in software bloat.

[PlainName]

So where is the 32GB minimum requirement to run Windows coming from?

What? Is that real?

Hmm. Just checked and it's 16GB for 'domestic' and 32GB for heavy games and stuff. Blimey - my memory (sorry!) is still thinking 4GB for office kit.

and it's not long ago smartphones did basically the same things as today with 10x less memory

I think phones get a small pass - copying from flash to RAM would speed up access, so there is a real benefit to having lots of RAM other than simple bloat.

[Siwastaja]

So where is the 32GB minimum requirement to run Windows coming from?

What? Is that real?

Hmm. Just checked

That was the whole point - MS halved the recommendation overnight because RAM is so expensive. That made big headlines.