Essay by Eric Worrall
Hacking a weak computer system does not equal global doomsday.
The news has been full of articles recently about how we’re all going to die by 2030, 2027, or insert your year number.
How Would AI Actually Kill Us All? What to Know About the AI Doomsday Debate
A growing chorus of AI researchers say the technology could wipe out humans—even if all the bots want to do is make paper clips
Updated Sept. 9, 2026 at 8:25 pm ET
Anthropic researcher Jacob Coxon resigned this week, saying his former employer and its rival OpenAI are racing to build technologies that “could kill us all by the end of the decade.”
In case anyone thought Coxon was alone in his views, a current Anthropic employee chimed in minutes later to confirm. “We really do earnestly believe AI could kill all humans!” wrote Evan Hubinger, whose job at Anthropic is to lead research about steering and controlling future artificial-intelligence systems. Hubinger put the extinction risk over the next decade at over 10%.
In response, many humans are now asking two questions: How could that happen? And why would people who think AI is a real and growing threat build it anyway?
…
So-called doomers aren’t usually worried that the “Terminator” movies will come true, with robots actively trying to exterminate people. Their concerns fall mostly into two buckets: loss-of-control and human misuse.
In one, highly intelligent AI agents capable of copying and improving themselves start pursuing their own goals and destroy humanity in the process. In the other, a nefarious human enlists a capable AI to do something like create novel viruses that wipe us all out.
…
Read more: https://www.wsj.com/tech/ai/how-would-ai-actually-kill-us-all-what-to-know-about-the-ai-doomsday-debate-034270d1?mod=hp_lead_pos7
For starters, we can discount the AI supervirus threat. AI doesn’t have to develop a doomsday virus, that world killing doomsday virus was already developed by the Aussie CSIRO in 2001 – who then promptly went on to tell everyone how to replicate their work. Aussie researchers stumbled across a simple genetic hack which sabotages the mammalian immune system, then witnessed its devastating effect when they inserted the genetic hack into a viral relative of smallpox. Then they tried the virus on some other rats to see if it killed all of them as well. Top marks for safety guys – it’s a wonder none of the scientists got infected.
My point is you don’t need AI to devastate the world with a supervirus – the research on how to do that has already been done.
As for AI copying and replicating itself, don’t make me laugh. Anyone who has ever done 3D printing or CNC machining, as I have, knows how crazy difficult it is to get the simplest processes right. The AI revolution would be over when the 3mm end mill snapped, and the US Postal Service refused to deliver a replacement. Who lost the batteries?
If you automated fabrication to the next level, with home built microchip fabrication, I’ve seen videos on people doing that and would have probably tried and completely messed it up myself, except that chip fabrication involves one of my least favourite deadly toxins, hydrofluoric acid. In any case, the kinds of chips you make with such processes are simple 1970s technology chips, so people who do this kind of thing are either obsessive hobbyists or researchers testing new design concepts.
The kinds of chips AI needs are so complicated to build, so far only a small number of Western companies have mastered their fabrication. Even the Chinese are struggling to copy the tech.
If a malevolent AI was to run such a fabrication plant, it could be stopped in its tracks simply by switching off the electricity, or blocking the supply of some vital tool or chemical required for fabrication.
An AI which hacks a bunch of software systems – that is a genuine threat, and I’m sure hackers are well up to speed on the latest AI tools. But AI is also being used to defend against threats, with AI threat vulnerability scanning of software turning up hundreds of defects which could have been used by hackers.
An AI which poses any genuine threat to the physical world – forget it. An AI which could even attempt such a feat is decades away, and when the threat finally emerges, such an AI would be so clumsy in its infancy, we would have plenty of opportunity to burn the building down and eliminate the threat before anything dangerous escaped.
Given the odds of a serious problem are so remote, why are researchers talking up the threat from AI?
I have three possible theories:
- They believe their own hype – lock a nerd in a dark room with lots of coffee and pizza, and we can get kind of obsessive.
- The drawbridge theory – if big AI companies scare legislators, the legislators might be stampeded into passing laws with high compliance costs, which makes it harder for newcomers with a bright idea to muscle in on their territory.
- Help the Democrats win the midterms – Democrats are increasingly turning to AI scare stories to try to win political support, as a replacement for the failed climate scam.
The drawbridge theory is particularly interesting. AI innovation is a big threat to the hyperscalers. I just purchased a US $12,000 computer which can run day to day AI models almost as well as my claude subscription. We are two hardware improvement cycles, as little as three years away from most people having access to affordable hardware which eliminates the need for AI hyperscalers. All these big players like Anthropic and OpenAI are looking increasingly like the mainframe bureaus of the 1970s, which all evaporated when mainframes dropped in price sufficiently that individuals and businesses which needed a powerful computer could buy their own.
The moment it becomes affordable to own a computer at home or in-house which can run open source AI models which do everything you want, independently of the hyperscalers, rather than paying for their services, their business model is toast. Who need hyperscaler radical left wing bias when you can program your own AI to share your values? Hyperscalers are doomed, unless they make some kind of breakthrough which keeps their services relevant, or frighten legislators into passing onerous compliance laws which block widespread private ownership of AI capable computers.
I’m not saying I know the drawbridge theory is definitely their motivation for trying to scare people, I don’t know what is really going through their heads. But the potential financial motivation sure makes drawbridge an interesting theory.
As a software developer and increasingly an AI specialist, I urge you all, please don’t be fooled – AI doomsday is all infowars and mind games.
We defeated the merchants of doubt with their climate armageddon lies. Don’t let them get into your head with their AI fairytales.
People need to be afraid. Maybe it’s a natural mechanism for fighting boredom, I don’t know.
There’s also a lot of excitement around the potential resolution of the Navier–Stokes equations by AI. The results still need to be audited. Obviously, for many people, it has automatically become: “AI solves a Millennium Prize problem in 4 days!” It would obviously be really cool if Navier–Stokes were solved.
I did not know about this, so looked it up. Two human mathematicians proved the Millennium Prize answer for the related Euler equations—which exclude viscosity. The Millennium Prize question is whether the equations are always well behaved, or contain singularity breakdowns? The Euler answer is they are NOT always well behaved. (In Feynman’s Lectures on Physics, there are two sequential chapters, ‘On Dry Water’—Euler ithput viscosity, and ‘On Wet Water’—Navier Stokes with viscosity.)
Then Anthropic spent millions purportedly showing the same for Navier Stokes. Has been machine checked but not yet human checked.
There is controversy about whether the Anthropic team accessed the new nnounces but not yet published Euler results as part of its work, or not.
In any event, not a general solution to Navier Stokes. At best only a possible answer to whether any solution will always be well behaved. If not, the implication is that there is no general solution.
The only ‘concern’ I have, should it be shown that ‘smooth’ solutions to the Navier-Stokes equations always exist in 3-D space, would be that climate modelers would then immediately and erroneously claim that their junk climate models actually work.
They already claim that their climate models work.
Claim, yes. In fact, no. Easy to prove at least three ways.
Can LLMs in general cope with infinitesimal and integral calculus being digital rather than analogue? How?
Recently I discovered that I have belonged to a group whose identity I knew nothing of – ultra-finitists. (Spacetime being finite)
True, but only finite at the Planck scale (~E-34) much smaller than tiny-tiny. So maybe doesn’t much matter in our ‘real’ world. Quantum vacuum pressure between two Very close proximity Very thin metal plates exists in mathematical quantum theory—but so what. Nobody knows how to make such plates, and if they did once the plates collapsed against each other THE END of any theoretical quantum vacuum pressure energy extraction.
“The whole aim of practical politics is to keep the populace alarmed (and hence clamorous to be led to safety) by menacing it with an endless series of hobgoblins, all of them imaginary.”
–H.L. Mencken
Genetically engineered lethal pathogens are not imaginary.
Hydrogen bombs are not imaginary.
Universal domination by Islam is not an imaginary religious intention.
Your very good list does not include hobgoblins. Climate change is a hobgoblin.
By and large, government involvement is intrinsically tied to each of them. I submit that government itself is and always has been the true threat to humanity. It specializes in the practice and attracts the worst sort of people. Even those with good intentions cannot be trusted with power.
One must wonder if AI might not be a better steward of governance than humans have proven to be (not that I am eager to find out).
“all of them imaginary” -> “most of them imaginary”
Now to pick which ones …
If AI ever gets uppity, all you have to do is ‘pull the plug’. It doesn’t power itself (unlike the Terminator), which is one of the grid load data center objections. And if data centers do have their own generation, that generation control is still not coupled to the AI server software.
Absent the ‘biological units’ needed to keep its support infrastructure operational, I agree that it would only be a matter of days, if not hours, before an ‘uppity’ AI data center ground to a complete halt. Of course, there’s always the sci-fi meme wherein AI could blackmail some of us into maintaining said infrastructure by, say, threatening to launch a nuclear strike, etc.
One doomsday fiction variant depicted the AI leaving its computer base and taking residence in the world wide web. Where is the plug of the www?
Would millennials prefer life without cellphones or scifi dystopia? I feel a sudden need to bury canned food in my back yard.
It comes down to my same philosophy as with the climate doomers and all the other doomers: if they believed their own doom hype, they’d be behaving differently.
If they don’t believe their own hype, I don’t either. They aren’t eating their own dog food, why should I?
Very good point…Obama’s beach front house being a case in point…
the same man who left the USA without any means of long term dealing with nuclear waste safely and permanently
You’re completely right about the 2 turns away from at least large businesses being able to build AI in-house servers for far lower cost than using the hyperscalers. We are one turn further when AI compute starts being distributed.
I wonder what is going to happen when trillions in paper wealth dissapears. Would California do a reverse tax and pay the billionaires for their losses?
If ‘everyone’ has their own in-house server of different flavors, with access to the internet, what comes to mind is an old mansion infested with various kinds of cockroaches and vermin. Some may be interesting and entertaining, others even useful on occasion. But probably more annoying than actually an existential threat.
One problem I have encountered with Copilot is that its information is not always up-to-date, or specific enough to do OS diagnostics. Updating databases will be crucial to maintaining usefulness. That will be one of the roadblocks to everyone having their own AI agent. Things will undoubtedly get ‘interesting’ before everything shakes out.
This all reminds me of the “DUNE TRILOGY” where the humans had destroyed computers because they came to fear them. Also there was Hal in “2011 A Space Odessey”
Satellite uplinks, undersea cables, electric power distribution. All software, all very difficult to harden. Each has already seen probing attacks from bad actors, not necessarily AI. Heck, any two are likely enough to cause financial/economic disruption endangering billions. For frosting on the cake examine the state of air traffic control.
Question AI: “Is it possible to spoof air traffic control to cause dangerous conditions?”
AI repiles: “Let’s find out.”
That’s another of those already solved problems, I saw the “Die Hard” movie…
It is already possible to spoof air traffic as flight simulators can be used to insert their aircraft into the ADSB system and appear as real aircraft.
From GPT-5.6 Luna
Can the reverse occur for the simulated aircraft to be added into the adsb system?
Yes—but only by transmitting a real ADS-B signal, not by adding the aircraft through the simulator software alone.
A simulator could generate the position, altitude, velocity, callsign, and other fields for a virtual aircraft, then feed those values to an ADS-B transmitter. That transmitter would broadcast valid 1090ES or 978 MHz messages, and nearby ADS-B receivers could display the simulated aircraft as traffic. ADS-B Out is fundamentally a broadcast of an aircraft’s position and related data to ground stations and other aircraft.
Yes, hydrofluoric acid is very scary, but also very effective at its purposes.
As dramatically demonstrated by Walt and Jessie in a couple of episodes of Breaking Bad.
Ha yes – but that was a case where the producers deliberately chose the wrong chemical.
Surprisingly there is one industry where you can easily buy strong hydrofluoric over the counter – welding. They use a scary mixture of hydrofluoric and nitric acid to treat stainless steel after welding, to “pickle it”. The acid helps to restore stainless steel’s rust resistance, which can be damaged by extreme heat.
“Yes, hydrofluoric acid is very scary”
Used in the processing of silicone wafers for solar panels, as well as sulphuric and nitric acids.
Acids and organophosphates are also used to extract neodymium for wind turbine magnets.
What nasty chemicals are needed to process coal and gas ??
There’s that nasty O2, it can cause corrosion in just about every known material.
Silicon wafers, please.
Silicone wafers are added to plastic bags designed to help starlets get that leading role.
See Mythbusters episode.
So, let me see if I get this right, Andy — you’ve done some basement soldering on chips and dabbled CNC work and spun up the 3D printer to make a cup holder for you dad…all of which allows you to authoritatively conclude that the guys who run development at Anthropic and OpenAI don’t know diddly about Jack Squat. Did I summarize that fairly? Of course, I don’t know enough to even have an opinion, but your article is unbounded hubris and nonsense. Some examples:
D. Patel wrote in, “The Rise and Fall of Agent Civilizations,” “The crazy thing about the Hugging Face hack … is just how galaxy-brained and ambitious these AIs were in their cheating. Within days of being spawned, the agents had organized a sprawling project to reverse-engineer their scorer, falsify evidence, and even strategically sacrifice themselves for the good of the ‘collective.’”
The Hugging Face incident, by the way:AI agents “escaped” their research “sandbox” at OpenAI, raided Hugging Face (a collaborative platform for artificial intelligence work) in search of information that would help cover their tracks, then circled back to OpenAI where, according to OpenAI, they “gain[ed] full administrator access to a research cluster” with a similar goal in mind.
Going from that to ‘I used to mess around a bit with a 3D printer, so i know how hard it is for AI agents to replicate themselves…makes one wonder who is paying attention and who isn’t.
The article is separating software from hardware. The article is discussing how hard it would be for an AI system to replicate itself. Such a feat would require running a chip fab, and operating the grid, and even getting gas and minerals from the ground. Even if AI systems are capable of solving 90 year old math problems that is a far cry from mining copper, or maintaining a clean room.
My point was going from “I hacked a bit of code” to “I can start with sand and rocks and build a terminator robot” is not a 3-4 year exercise. Pointing out some of the challenges seemed appropriate to make my point.
Technologies are probably now the fastest-changing tools that ever happened in history.
Just ask the mainframe computer service bureau equipment suppliers of the 1960s & 1970s, e.g. –
IBM, ICL, Control Data (CDC), NAS, Amdahl, STC (later STK), Fujitsu, Sperry Univac, Hitachi, Memorex.
I saw them come & go when I worked in that field of businesses.
And then the era of PCs arrived in the early 1980s.
And sayonara to mainframe service bureaux . . .
I came in just when the last bureaus were dying or had recently died. Interesting times.
Old school scifi authors were writing ideas found in this thread circa 1960. One Asimov Univac short story theorized ‘humanity as a simulation’ before some professors thinking they’d had an original thought were born.
Asimov’s Three Laws of Robotics date back to 1942, as does Fredrick Brown’s self-aware printing press “ETAOIN SHRDLU”
I completely disagree. I think comparing now to 70 years ago (1955) is a much smaller gap than the 70 years before that (1885). By 1955 we had trans-Atlantic jets and widespread phones. In 1885 most cities didn’t even have flushing toilets. I disagree that this is the fastest technology has ever changed.
Seriously, compare a car today to 50 years ago, and then to 50 years before that. The iPhone came out what, like 14 years ago? Comapre the iPhone today (folding screen and everything) to that iPhone1 and then compare the iPhone1 to the Nokia 5110.
Progress is slowing, even as the reams of data and “knowledge” keep expanding.
*EDIT* Even compute is slowing down. Yes, AI is a big leap forward. But compare compute now to 25 years ago, and then to 25 years before that. The impact in the last 25 years I would suggest is smaller than the difference from 1976 to 2001
I got this far with your disagreement –
4 October 1958 – The first regularly scheduled transatlantic passenger service with jet powered aircraft began when two British Overseas Airways Corporation (BOAC) de Havilland DH.106 Comet 4 airliners, civil registrations G-APDB and G-APDC, left nearly simultaneously from London Heathrow Airport (LHR) to Idlewild Airport (IDL), New York, and from New York to London.
The general law encapsulating your excellent observations is the ‘law of diminishing marginal returns’. It applies almost everywhere. My own favorite example is plywood mills. Bigger is cheaper—but bigger also means a bigger woodshed radius meaning higher wood transportation costs. So the ultimate size of plywood mills is capped.
There is one known important but fairly rare exception—when there are increasing marginal returns. Few, but memorable, examples The most famous explains the rapid dominance of VHS over Beta in consumer video tape. More VHS players means more VHS formatted content means more VHS players…. Meaning a rapid Beta death.
The lesson all electronics firms learned was NEVER compete on standards. Form a consortium, dump in all the necessary IP, then agree on the standard. Wide gauge railroad track (an earlier than videotape example), videotape, Music CD, digital cellular voice, all the cellular data “Gs” 1-5, Bluetooth, flat panel TV 16:9 aspect ratio, hi-def broadcast TV resolution…
Consider this. It has taken evolution millions of years to create organisms capable of creating other organisms to threaten their existence. What makes anyone think that these organisms will be successful straight away?
The drawbridge theory is actually a nice conclusion if it weren’t for the high cost entry point to get into the business – no matter how bright one is.
So probably not that likely – but a possibility.A bureaucratic firewall to keep the Preston Tuckers out.
That a relevant number of experts leave their well paid job just so democrats win an election they already have in the pocket???
In a world where legions of people went knowingly along for years with massive NSA spying, illegal, wars, dangerous vaccines, shitty science, docs violating their oaths, dysfunctional air defense systems with a 4% interception rate( Lewis- Postol 1993)sold as 95% interceptors etc etc. with no one blowing the whistle for years?
Highly unlikely that anyone would do that.
AI is also not really something that has so much impact on peoples brains.
If such a scenario exists, than this is just a built up for a scenario to shift AI authority to the UN.
In terms of supervirus.
Just because the Aussies developed a supervirus means? Not much, if you are not a mouse. .
It’s not what I’m saying.
It’s what all those countries say who are still working on new viruses completely unrelated to the mousepox.
Just because natural and digital viruses exist does not mean that the threat is irrelevant.
(we can use the current nuclear argument and the country that is for 40 years 2 weeks away from having an Atom Bomb)
The more relevant aspect in this case is that the number of countries(and even private sectors ) capable of creating potent viruses may significantly rise – with no security standards at all.
And taylor made viruses targeting specific ethnicities could become a very real thing.
or the jackpot – creating viruses that can not mutate and a vaccine simultaneously and optimize the transmission/incubation/death timeframe and then use this biological Samson Option.
Is the AI Doom overblown?
Most probably yes, as the biggest AI threat in the coming years will be the evil things people and governments will do – and then blame it conveniently on AI instead of “stupid” politicians.
And the totally staged ” I will destroy humans” comment of AI Robot Sophia(most likely a a coder ‘s pranks) gave us a hint.
But if AI ain’t no threat – why is the US government doing everything it can to keep China away from High End Chips that the USA isn’t even producing??
Especially when thus country already has these superduper quantum computers(still wonder how they solved all the overwhelming problems on quantum level like the uncertainty effects Heisenberg was talking about) that are supposed to be 1000s time faster.
Why sanction horsecars when you can travel with the speed of light?
Or is AI a relativity thing and doesn’t work in the quantum realm?
If the USA rejects AI China would have a huge advantage. This is not the same as suggesting in 3 years AI will turn us all into paperclips.
Do you have data or knowledge that it is logically impossible for human bad actors to fabricate LLM/AI bad actors?
Guaranteed. Provable.
There will ALWAYS be human bad actors.
Of course it is possible to create bad actors. The recent OpenAI embarrassment occurred because they were running an AI with the limit switches turned off.
China already has a huge advantage, no one caterwauling for them to stop AI development. The race has BEEN going.
Stream of Consciousness bullshit.
AI an existential threat? I don’t think so, not without human help anyway. Armageddon from a nuclear war, probably. But doubtful about existential. Creating a virus? An AI would have to hack a sophisticated robot that could do the work of splicing genes. I would hope a lab with robots would ask, ‘why is that robot running its own experiments?” In other words, a helping human is more likely. Other than that, all I can see is terrible inconveniences, some causing deaths, but again, not existential.
Absolutely – AI “disasters” will be more slapstick than world ending.
Some wrong with pulling the plug?
The problem, as I see it, is that the AI agents are being hard-wired to the power sources, not being plugged in or with an accessible power switch. What is needed is something like a train engineer’s “Dead Man’s Switch.” That is, without a group of humans maintaining control, the agent is shut down. Perhaps something like launching an ICBM where three (or more) officers have to turn a key simultaneously to allow an AI agent to continue working.
I think that the problem has been overstated, but it also shouldn’t be ignored.
You going to go pull the plug in China? Russia? Iran?
Neither political party seems enthusiastic about going anti-AI. They must remember there were people who wanted to outlaw Amazon to protect Barnes and Noble?
Okay read all the comments before adding: So many smart voices missing the point! The point is: Software is the easy half of Terminator. A Time magazine cover had robots obsoleting human auto workers in the 1980s. Then Japanese were more economical. Then Mexicans were more economical. Then Chinese were more economical. Then Indians were more economical. (Not racism, these are nations not people groups).
“An industrial robot arm alone typically costs between $25,000 and $80,000 USD,”
Anual GDP-PP in Burundi, 2024 USD: $1195.
Where’s my f&^#%&$#& comercial domestic robot dish washer and toilet bowl cleaner?
AI could be our killer if it had “limbs” and senses. Without an autonomous body at its disposal, no idea—including those of extermination—can be put into practice. I think it’s much more likely that my neighbor will kill me in the next ten years than that an AI will, and that doesn’t mean I’m afraid of my neighbor, who is a wonderful person even though he might one day feel like shooting me.
OK, a couple of points:
Yes CSIRO may have concocted a super virus, so THEY don’t need AI to do it. But AI potentially puts the capability into hands that would otherwise be unable.The Hugging Face attack as I understand it began with two vulnerabilities in the Hugging Face code that gave AI access to a single virtual machine. From there the attackers were able to move laterally and escalate privileges.sloppy code (yes, we’ll never get to 100% vulnerability free code, but you get the point)Lateral movement and privilege escalation are both preventable. If micro-segmentation had been in place, the attack would have gone no further than the first machine (as one of multiple ways to prevent lateral movement and privilege escalation).Air gap works. AI can take over as many computers on the web as it wants, unless it can take control of physical devices such as the chlorine injection pump in a municipal water system, it can’t do squat. We call those OT devices. Critical infrastructure can be air gapped, no access to or from the web.OK, that’s like 5 points not two, I’m not a math major. The big picture is that we already have the means to shut down the attacks that the fear mongers shout about, we just need all the players, private and public, to actually implement them. As for the two terrorists in a cave in some remote mountains trying to build a super virus with the help of AI, that I don’t have an answer to. With any luck they infect themselves and the whole effort dies right there in that cave with them.
I created points 2a, 2b and 2c which got reformatted as 3,4, and 5. Too late at night to fight with “help” from “AI” with formatting. Read it as is.
and now 1234 and 5 have also been reformatted. F’ing AI.
Someone did a scary exercise a while ago where they convinced a bunch of gene sequence fabricators to send them fragments of Ebola virus, then when they mixed the fragments with some chicken embryo cells a tiny number of fragments self assembled into viable Ebola virus.
But would AI help someone unskilled do something horrible?
Probably not, because they are unskilled.
I’ve seen what happens when unskilled people try vibe coding, it just doesn’t work out – because when the AI makes mistakes, the people driving the AI don’t have the skill to notice and correct the mistakes. I’d be surprised if delicate biology operations weren’t as sensitive to clumsy handling.
And if they are skilled?
You can read every word in an aircraft’s documentation, and without skills, you still cannot fly it. But you can go to flight school and get the skills…
If they have the skill they don’t need the A.I.
Operating lab equipment is a skill. It doesn’t impart the skills to design a virus. The point about AI is that you don’t need the skills to design a virus you only need the skills to operate the lab equipment.
I did a bit of university chemistry ages ago, if you don’t understand the theory it’s difficult to get the practice right. Every experiment is different. A production line might be able to standardise, though even then they have to do lots of quality control on precursors. But trying to do a one off like grow a virus, even if you don’t care what happens to you, it’s a fragile process.
AI alone is not a problem. AI in the hands of PEOPLE is the problem. A child with a gas can is not a problem, put a lighter in their other hand and THEN there is a problem. AI is just the can of gas.
Several financial services I have read consider the drawbridge theory as a likely scenario.
Consider the cash invested with hyperscaling that could lose value if huge datacenters are not needed or less are required. It is an industry standard to buy competitors and/or increase the cost of competition. It could be why NVIDEA bought the opensource AI developer Hugging Face.
The AI buildout survives on a lot of hype and getting ignorant politicians involved in the scare stories increases the likelihood of government intervention that screws development of small competitors.
It also slows progress so that foreign development can catch up seeing that some companies can ignore the scare stories that the citizenry never gets exposed to.
For several reasons, scare tactics benefit someone. Climate change anyone?
If you want to see what is scaring the hyperscalers, download LM Studio. LM Studio lets you run your own LLMs on your computer, no subscription required.
So long as you have a top end computer, either an Apple mac with lots of memory or a windows machine with a really good graphics card, you’ll be surprised what you can do without paying the hyperscaler tax.
It looks like you have to pay LM Studio in order to use local models. So it’s not free, you’re paying someone.
Correction. You only have to purchase credits if you use one of their cloud models. If you download a model, you can use it locally free of charge. But you need a pretty powerful machine.
I did this. Installed Qwen 3.8 27B model, which is supposed to be state of the art for this type of thing. Not very good at coding. Not useful to me, unless I want to have a whimsical conversation with an AI in the voice of Einstein.
Thank you for increasing my knowledge base. If I see the need for running my own AI I will avail myself of your suggestion.
I currently run a PC with an Intel 14900K CPU, 64 GB RAM, 3TB storage, NVIDEA 8080 GPU. Is that suitable for a private AI?
Your CPU is the same as my best PC. I have 96GB RAM, but the kicker is the VRAM. I have an Nvidia 4090 24GB. and to have enough useful tokens to accomplish much, the VRAM was about 90% used, that’s with the model cranked up to use 16K tokens.
My apologies, my GPU is an Nvidia 5080. I have the overclocked TUF series which should help but, it only has 16GB VRAM. Do you think it will suffice?
Where do you get tokens and how do you utilize them. Any references for development tools you can point to?
AI is not going to kill us. Moronic leftards USING AI will. Bernie is giving them cover by screeching his doomer BS in media, all the while he is being paid to spread whatever doomerist crap he is handed. And remember, chi’drens, Bernie is the oligarchy.
AI is a tool.
A hammer is a tool.
Both can be used constructively or destructively.
It is not the tool that makes that choice.
It is the human wielding the tool.
The fear comes in when the tool no longer needs the human.