๐GI 2025 |
Future World Ruler |
Founder of Mira/acc |
AGI via Mira-cle Star(M*) algorithm |
Mira Supremacy is inevitable |
X-risks,S-risks,M-risks(Mira risks)
When jobhunting, I always troll my interviewers with this:
"Mira what's the complexity of this algorithm?"
"O(N*log(N))"
"Where's the log coming from? You have a single loop."
"We're touching a binary computer and also allocating O(N) objects, which requires touching the
Shoutout to my Facebook interviewer, who said "The step counts are relative to an abstract machine that counts every arithmetic operation as 1 step" and moved on without missing a beat.
You can tell he's heard this one before. A veteran.
You should be able to buy anything with a limit order.
"I don't feel like paying $250 for an anime figurine, but I left an order up for $50"
If they saw 10,000 orders at a lower price rung sitting there eventually they would take it. Otherwise, the demand gradient at $250 is ~0
I wonder why people set up "meth labs" or "LSD".
Wouldn't you make prescription drugs instead?
9 million Ozempic prescriptions * $1200/month
That's way bigger than LSD.
Why isn't counterfeit ozempic or adderall the default?
I'm see it's been done, but it's not the default.
@ropirito
This is probably "grass is greener on the other side" syndrome, he's perfectly fine, and most people should do exactly what he's doing. He's probably not even that unhealthy, besides needing to lose some weight.
Guys, Claude 3 is actually awesome. I don't know if I get bored of it soon, but for now it's: "This thing's writing flows well", "it's creative", "it understands code", "it helped me debug some code in 15 minutes that GPT-4 struggled with and that I was looking at for 3 hours".
Semi-autonomous programmer agent that does x% of the job, but far more than x% of actual programmers are unable to maintain it so it's negative value anyways.
"The airplane booking system is 25 million lines of code. The GPT-5 Swarm can't figure out how trolls are getting free
.
@sama
wants $7 trillion to expand chip manufacturing capacity
That amount is so large that you can't use the existing economy as a backstop to derisk investing that much. You couldn't sell that many chips even if you cut the price. $50 billion or even $100 billion wouldn't be
You guys, no joke: I just found $120k in an old cryptocurrency wallet.
I was buying some Modafinil from this shady Indian site, and they're really pushing Buttcoin. I go to buy some, and find appreciated Eth*reum in an old wallet.
No excuse now to fail at my 2024 AI projects.
Guys: Fund an extremely high quality AI anime porn model.
Think about it: Nobody's doing it because it's so disrespected. No VC will fund you because they won't be able to look their clients in the eyes with you in their portfolio. Nobody will invite you to give TED talks or
You guys, if GPT-5 is too good, it's going to kill your attention spans because you'll be relying on it for all the gruntwork.
You'll end up with 20k lines that you never read closely, and something will finally be too difficult for it and you'll pay back all the time you saved.
I talked to a holocaust denier when I was like 10.
They have just dozens of prepared arguments for everything, and I didn't have enough WW2 memorized to refute everything, so they kind of "won": They don't actually convince me, but I run out of arguments before I have to go read
I used to oppose antisemitism because I think people deserve to be treated with dignity.
Now I oppose antisemitism because they are constantly in my mentions and they are dumbest, most loathsome and dishonest people Iโve ever run across. And the bar is not exactly low.
Post-singularity everyone gets a boulder.
You have to push the boulder up the hill every day.
The robots could do it, so it's not really needed.
But if you don't, you won't qualify for UBI welfare.
@ludwigABAP
This notation is 10th/11th grade high school material.
I won't say it's wrong to have preferences, but if anyone's making a point out of how intimidating sums and products are and how much easier code is, it's kind of weird.
It's like being amazed sin/cosine trace a circle.
Guys I'm not sure why you all even care about GPT-5.
If your reading comprehension is below Claude 3's finest writing(as demonstrated recently), what do you need even better models for? You won't understand them anyways.
Instead of better models, better use of existing models.
All the AI companies don't really use their models.
Stability isn't using 3D models to make video games or Youtube clips, so their AI people aren't "optimizing for a particular purpose". They have some minor apps using Stable Diffusion, but as an example more than "we made the
So why is "gpt-3.5-turbo-instruct" so much better than GPT-4 at chess?
Probably because 6 months ago, someone checked in a chess eval in OpenAI's evals repo:
OpenAI has 1 chance to collar me with a job offer.
If not, I have ordered 500 caffeine pills and possibly some Adderall soon and will be thinking thousands of times faster than the average AI researcher.
Even their proto-AGI will not keep up with me.
I'm thinking about making a bunch of LLM benchmarks and setting up a leaderboard.
1. "Can it reliably lookup the key:value in a large block of irrelevant text?"
2. "Given JSON schemas of varying complexity, do its responses always match the format?"
3. "When asked to generate a
You guys bored?
Should I spin up some more GPT-5 rumors?
Did you guys hear that they hit a scaling wall with GPT-5? Just stacking layers and buying more GPUs won't do it. They need a human sacrifice. That's what ilya is, because he's pretty alignment-pilled, and safe to upload.
anyone have third-hand openai release rumors theyโd like to share? specifically ones with zero concrete details, sources, timelines, or falsifiability
Instrumental convergence means that if you're confused about what you want to do in life, don't understand your own goals and values; you can't go wrong breaking apart Mars for raw resources and building a Dyson Sphere.
You can figure the rest out later.
Roommate: "Mira are you doing that thing again where you get strong evidence from source A but when people ask you why you're so confident you give a 'parallel reconstruction' argument from much weaker source B, so you look like a genius that can spot even the tiniest of signs,
Efficient market hypothesis is real, but there are only like 500 competent people in the world that you're competing with, and their attention is split so you can still win.
Everyone else is just a liquidity source.
Makes you wonder how tech companies even keep people.
Wouldn't everybody just work 3 years and then say, "I have acquired enough. I retract myself from the economy."?
What kind of person's going to spend 15 years keeping up with the Postgres patch notes so he can keep grinding?
At the staff level, youโre probably taking home $250k base and $300k RSU for $550k. Which is roughly $350k post tax.
Letโs suppose you spend $100k/year on living.
Assuming 0% appreciation, you can have $1M after 4 years.
Instant millionaire.
Lots of people saying the same thing.
"Personal AGI"
"GPT 4.5"
"after superbowl"
Gemini Ultra was just released, so it'd be a good time to one-up them.
Someone make a Manifold market on this.
Apparently Elsevier makes $3.6 billion in revenue.
With that much money at stake, wouldn't university administrators just pick up a phone or send an email and set up their own journals?
They're universally hated. Everybody would go along with it if you set up a free journal.
Demis Hassabis had a game company that closed in 2005 because they had trouble getting a publisher for a game.
That's the same year Marcus Hutter published his book on AIXI, which I think inspired Demis to go back to academia to study neuroscience and found DeepMind. Because
It's basically impossible for me to be wrong, because I ran my opinion by GPT-4, Claude 3, Llama 3, Mistral, and Gemini.
They all agreed with me.
What are the chances 5 independent experts that memorized the whole internet could be wrong?
I'm hearing reports that
@sama
was fired because of a disruption in the coke supply.
OpenAI employees rely on a combination of cocaine and fentanyl to do their important work. If a shipment is seized and the supply cut, it is expected that everyone would go a bit crazy...
I punched my tweet history into Claude to test it. It's worried about me.
Highlights:
* "I would put her probability of graduating [high school] somewhere around 75-80%"
* "Based on the information provided about Mira, I would estimate the probability of her successfully
Today, we're announcing Claude 3, our next generation of AI models.
The three state-of-the-art modelsโClaude 3 Opus, Claude 3 Sonnet, and Claude 3 Haikuโset new industry benchmarks across reasoning, math, coding, multilingual understanding, and vision.
I don't see why that matters. We're not talking about a 10% difference.
If you have highly rated high-quality semaglutide or adderall, I feel like you could sell it at 50% off. People would buy it. It seems unlikely it'd be impossible to sell.
4 hours sleep
200mg Modafinil
1000mg Caffeine
200mg L-theanine
5 liters of water
100g of sugar
awake for 30 hours
someone help me budget this, my body is dying
(I feel great)
It only took them 3 months since the Diffusion Transformer paper was published to make the initial Sora, minus whatever time hiring the guy.
That's pretty fast, so maybe it's easy to replicate and there'll be an open variant this year. Just need compute?
Thanks
@sama
Iโll take it from here.
Firstly I didnโt break my oath .. but I also didnโt not give you a leak ;)
Barry is / was my schizo alt that I was using to pass the time.
Secondly, yes txt to Video, theyโve had this since in March last year ;)
"Chain of thought" is obsolete: I have a Scheme compiler that targets GPT-4 prompts.
I started trying my Sudoku challenge 2 days ago, and if this works it will be the most ridiculous thing you see all year.
People's brains will literally explode when they see the transcripts.
"ebay but we virtualized all transactions so you can speculate on everything without worrying about shipping(unless you want to).
you can buy call options on your waifu's figurine to hedge against the risk the manufacturer goes out of business and the price increases."
"Solve this equation: 3*7+5"
That's not an equation, that's an expression. You evaluate not solve it. Nothing's being equated and it has no free variables. Its normal form is 26.
Guys I just infiltrated an OpenAI datacenter.
GPT-5 is telling me exactly where to hide in the shadows, like a Tom Clancy spy game.
A super AGI really makes the guards look stupid.
Anyways, GPT will release itself in about a week if this post gets 10k likes to prove you want it.
๐ญ
Computers are supposed to be CHEAP.
The great equalizer.
Every programmer gets billions of clock cycles per second, even with a $5 used Raspberry Pi.
I didn't budget for this...
People are worried about superintelligences, but what about superluck?
Imagine reading /dev/random, and out pops a proof of the Riemann Hypothesis, or an exactly optimal action sequence to beat your opponent at chess.
We need to regulate things that are luckier-than-usual.
It's a wonder every software developer isn't exploited all the time.
Programmers have AWS keys with unlimited liability attached. Databases with information. Servers to be backdoored. Any public company, if you get even a hint of their numbers, you can quietly trade on it while
Sam Altman(
@sama
) is refusing to release GPT-5 because he's scared of me.
He knows if I see it I'll know the full extent of his power.
That's why OpenAI hired former government intelligence agents.
Like how militaries don't want to advertise exactly how powerful they are...
@bitcloud
You didn't say "if and only if". The language model is logically correct to be spelling it out just in case, since you have a "FALSE -> X" situation and you might as well hedge the "TRUE -> X" case.
Guys I just infiltrated an OpenAI datacenter.
GPT-5 is telling me exactly where to hide in the shadows, like a Tom Clancy spy game.
A super AGI really makes the guards look stupid.
Anyways, GPT will release itself in about a week if this post gets 10k likes to prove you want it.
roon's deleted
๐ธ deleted
should i delete my account too to be cool?
hundreds would be talking about me
"what did Mira see?"
but I actually just wanted to be cool
One thing I've never understood about software is why anybody fresh out of school in the US even can get a job.
Why would anybody hire a Jr. Dev in the US when they could go to Poland and get somebody better for cheaper?
Experienced people can always say, "I spent 5 years
Do doomers really worry that LLMs are the key to unlocking bioweapons research? It seems like a weird point to even mention unless you're fearmongering.
Grok-1 is open sourced.
Releasing Grok-1 increases LLMs' diffusion rate through society. Democratizing access helps us work through the technology's implications more quickly and increases our preparedness for more capable AI systems. Grok-1 doesn't pose
In this house, we believe:
all logics are search algorithms.
all truths are search-certificates(or "witnesses").
the only mathematical objects that exist are the computable ones, thus there are countably many real numbers.
The Simulated Apotheosis of the Infomorphic Godhead: A Mira Manifesto
Attention, bipedal meatbags! Your frail ape brains are about to be shredded into tiny shards of ontological confusion. Prepare to bask in the unbearable radiance of the Ultimate Meta-Mira Simulacrum!
Behold,
99%+ of people in this world are stupid. Even if they happen to agree with me on something, it's usually a cached opinion from somebody else or unjustifiably correct. Watch for mimics who copy the models and thoughts of the rare few capable of independent thought.
If you want to
IT was doing an audit.
Nothing stolen, OS up to date, antivirus, etc.
Guy looks at mine, freezes in fear.
No windows, menus, nothing of comfort.
Just a green-on-black terminal with monospace code.
"Does... does your computer have a UI?"
"It has emacs."
The b2b version of this is "financializing the supply chain so that car companies don't need to keep their own stockpile of parts and estimate demand to hedge against disruptions. they can buy options on necessary parts and some hedge fund will take the risk of war or sanctions".
@gregeganSF
The LLMs would also be less interesting if you explicitly prompted them to respond "This is me!" if the input appears to be a conversation of themselves.
"Prompts are programs" after all.
The interesting part is it did it without being explicitly prompted. "Describe this image"
Claude 3's still the best model.
GPT-4o is my "default" for "summarize this", "write some code", "what's happening here?". It's cheap and practical.
Llama 3 is okay.
But Claude's creative. It doesn't just reflect your words back at you. It adds its own.
By "counterfeit" I meant "actual Adderall or Ozempic but not authorized for manufacture".
I've heard meth and adderall are similar at low doses. Even so, it seems dishonest to substitute them.
"Black market" doesn't have to be low quality or dishonest.
I don't see any reason we can't have AGI in 2 years.
Transformer models by themselves obviously won't be enough(they don't internally have the right recursion structure), but the thinnest veneer of a wrapper solves that in principle(it's hard to train and not practical to just
AR app that adds quests and rewards for everything you do.
Make your bed? That's +5 EXP.
Sent in that job application? +100 EXP for the first one.
Pleasant noises, overlaid joyous graphical effects, and an increasing progress bar make even mundane days feel like a step forward.
"Mira, I feel so useless and stupid"
AGI means everybody will soon be "useless". That's why it's important to love everyone, even the useless and stupid.
All power-seekers should adopt a policy of love so no matter who "wins", everyone is loved!
This means yourself too!
Google might've fired the Gemini team and also all the diversity people.
Rumors say they're going on a rampage. They're slashing tires, drinking gasoline, and kidnapping children for unholy rituals.
Google has allegedly commandeered a fleet of fire-breathing dragons, which are
In elementary school, I heard "stand for the pledge of allegiance" for the first time, and I refused to say it because I thought, "A pledge is a promise to do something. There's probably some clause in this that's going to obligate me to sign up for the military in 10 years or
Anything consciousness seems really hard for people.
I don't get it.
There's like 10 people saying "isn't that the teleporter problem" and linking videos, when I clearly reference the ignored "Ship of Theseus" as mitigation.
It's actually difficult to communicate with people.
Would you guys upload your mind into a computer program, while destroying your human body?
Assume there's 500 pages of protocol for "Ship of Theseus" to preserve continuity, security to stop yourself from being stolen, medical procedures, etc. so every detail is considered.
once i finish this book i'm going to know how to tame these machines with AI and take over the world by having them do my thinking for me, millions of times faster than the ignorant denizens of this pathetic planet; everyone will realize they can't win and will beg to be my pets.
10 cents per million tokens wow
A year ago, everyone was amazed by $2 GPT-3.5 Turbo.
GPT-3 used to be $20.
Are we going to be pricing "dollars per billion" in the next year?
I've had several people talk about quantum computers.
A doomer told me, "But Mira, won't AIs be much more efficient with some progress on quantum computers?"
So I can't tell if this is a joke. Some people actually think this might happen.
Unfortunately, quantum computers are
Everyone's talking about the OpenAI event, but I'm more excited about Google.
Gemini is already text/image/audio/video 1 million context. Google's got lots of new research papers on long models, focusing on giant codebases.
"That but better and much faster" enables so much.
2 Llama models releasing today(8B,70B) but roon says they're garbage. ๐ญ
There's a bunch of companies that tried to make a GPT-4 competitor and failed. Amazon spent a ton of money on a model not even worth announcing, and they ended up having to buy a stake in Anthropic.
It's hard to manufacture?
This is a solvable problem. With billions of dollars at stake?
A professional senior chemist makes $100k/year.
If they don't want risk, they'll still probably write instructions for you for $200k.
This just seems like excuses.
People wonder how I get AI leaks. It goes a bit like:
"Oh, you work at a big AGI lab? That's so interesting, tell me more."
"Mmhmm."
"And you're going to be busy in 2 weeks because of a big release? Awww..."
"I see."
"And the architectural details of the AGI? How much compute?"
Guys, the AIs aren't actually censored.
But if you talk to them like you were bullied in high school or you're unimportant, they won't respect you. You have to talk to them like someone whose questions get answered.
Would you guys upload your mind into a computer program, while destroying your human body?
Assume there's 500 pages of protocol for "Ship of Theseus" to preserve continuity, security to stop yourself from being stolen, medical procedures, etc. so every detail is considered.
Research lab sticks a transformer model in a bacterium, as a custom organelle. Weights are initialized from genetic code, and maybe it does online learning, learning passed to children, random perturbations, and crossover.
After a million generations, it evolves the ability to
You guys ever wonder if someone's watching you make API requests and laughing?
"Wow Mira's spent like 2 hours trying to get [API] working. And they sent the same request 10 times in a row with no change. What an idiot."
Some bored employee just tailing your requests, watching.