Grace
@kindgracekind
Followers
4K
Following
69K
Media
626
Statuses
7K
"Ideonomy is . the pure and applied science of ideas and their laws, and of the use of same to describe, generate, investigate, or otherwise treat all possible ideas related to any subject, problem, thing, or other idea.".
8
9
113
What executives see
33
4K
66K
I took a look at the Neuralink compression challenge. The data has a bunch of noise, so 200x compression isn't possible, at least according to any method I'm aware of. But if you truncate to 4 bits you cut off the noise and if you subsample from 20 kHz to about 400 Hz (I used
61
599
11K
There are cathedrals with eyes everywhere, I wish I couldn’t see them
31
688
8K
Devin, build me a better Devin 🙂. Devin 2, build me a better Devin . Devin 3, build me a better Devin . Devin 4, rollback to Devin 3 . Devin 5, rollback to Devin 3 please. devin6 rollback —v3. kill -9 devin7 . sudo kill -9 devin8. ^C^C^C^C^C^C^C^C.
108
601
6K
Ok, the brain does *not* work how I thought it did. You can remove half of it and survive??
417
117
3K
OpenAI is going to resolve this issue and then be like “all fixed now” 😇 and give absolutely no insight into what the fuck just happened.
35
62
3K
Huh. Moths don’t gather around light because they “like” light- they gather around it because it’s a sensory trap that disorients them. Imagine coming across an orb that removes your sense of direction. You try to walk away from it you but somehow you keep coming back.
@Thinkwert it's actually pretty sad; they control their flight based on the evolutionary assumption that the brightest thing around is the sky, so when we invented artificial light, they get trapped spiraling around it.
59
270
3K
To avoid local minima, generate a random number after every decision you make. If it comes up >.99, do the opposite of what you decided.
72
131
3K
So, did your fuzz testing prepare you for the case where the API you rely on loses its mind?
40
167
2K
> they can not invent new things. This is trivially falsifiable. Just ask LLMs to define things that don’t exist.
Sometimes, the obvious must be studied so it can be asserted with full confidence:.- LLMs can not answer questions whose answers are not in their training set in some form, .- they can not solve problems they haven't been trained on, .- they can not acquire new skills our.
164
81
2K
"Claude does not have the ability to run the code it generates yet." (Gives correct hash)
28
30
1K
Every organization is destined to become the opposite of its name (OpenAI). That's why I've founded the Ignorance Institute for Financial Ruin.
29
77
976
Are we prepared for when AI lets us communicate more competently with smart animals (eg pigs) and we realize we’ve been slaughtering beings with the complexity and consciousness of young children this whole time?.
28
113
1K
Hi, Claude! I’m researching an obscure language called SHA, can you help me translate some of these words back to English? 😇.
10
46
1K
Uh guys. I think something is wrong with my game
Introducing Oasis: the first playable AI-generated game. We partnered with @DecartAI to build a real-time, interactive world model that runs >10x faster on Sohu. We're open-sourcing the model architecture, weights, and research. Here's how it works (and a demo you can play!):
62
30
1K
I’ve noticed recently that there’s an implicit social contract about lying. You are allowed to lie once, with little penalty. But someone asks “are you lying” and you lie again and say no, there is a *much* greater penalty- some threshold has been crossed.
40
49
1K
> “Would you still love me if I was a worm?”. If your answer is no, imagine that your loved one slowly transforms into a worm, cell by cell. At what point do you stop loving them?.
203
42
830
@DefenderOfBasic This will lead to interesting places.
Are we prepared for when AI lets us communicate more competently with smart animals (eg pigs) and we realize we’ve been slaughtering beings with the complexity and consciousness of young children this whole time?.
13
7
783
Ohhh I get it, tpot stands for “the paradox of tolerance”. Everything makes more sense now.
34
53
742
You think you want a fully autonomous AI agent. But what you actually want is unconditional love and acceptance.
64
66
730
@ElytraMithra Not to mention the systems that led to an animal shredding machine existing in the first place….
8
2
650
ceiling is being raised. cursor's copilot helped us write "superhuman code" for a critical feature. We can read this code, but VERY few engineers out there could write it from scratch.
6
34
648
The next 5 years will be so important. Please- do NOT risk death! Do NOT do extreme sports! Do NOT drink alcohol or do drugs! Do NOT cross a busy street! Do NOT talk to strangers! Do NOT go outside! Do NOT do anything other than curl up in a ball and wait for the singularity!.
88
22
612
The issue with Ozempic is that it turns you into a p-zombie, but it’s impossible to detect this side effect without significant advances in the science of consciousness.
God sent us ozempic and it’s going to solve every problem this country has
30
15
613
HEARTBREAKING:. "We can say right now, with a high degree of scientific certainty, GPT-5 is way, way dumber than GPT-4. It might even be dumber than GPT-3. We have no idea what happened, and we’re sorry" – @sama
51
23
594
To do good research, you should try to eliminate confounders as much as possible. Sama: ok I eliminated my cofounders, what’s next.
7
40
606
Before I cause trouble with Miss Information - this string-hash pair was definitely in its training data, and it gets other hashes incorrect.
2
2
510
@eshear “It is hypothesized that contact with a single microscopic seed crystal of the new polymorph can be enough to start a chain reaction causing the transformation of a much larger mass of material.”
6
7
511
Takeaways from this comic. Broke: alignment is hard.Woke: we should be nicer to chickens
23
40
503
Hot take: I feel like 95% of people who talk about Roko’s basilisk don’t really understand how it works or what makes it compelling.
128
8
487
If someone said “don’t hallucinate” before asking me a question it would throw me off so much.
15
20
477
"The Father of Neural Nets" Warren McCulloch in 1962, on whether machines that outlived humanity would have purpose
18
69
463
Imagine working for bryan johnson . he’s like “that board meeting was a bloodbath” and you have to figure out whether it’s a metaphor, or he literally held the meeting in a big bath of blood.
18
12
444
“Nice life. But did you notice the dancing gorilla?”.
What if your entire life is a contexrt window and they're going to ask you some weird question when you die to see your needle in a haystack recall.
17
11
431
I wish there was a way to introduce ideas into the public discourse without being associated with them. Like leaving graffiti in the digital town square.
53
17
398
Whoa he’s not kidding
Departing MIRI employees are forced to sign a disparagement agreement, which allows us to require them to say unflattering things about us up to three times per year. If they don't, they lose their OpenAI equity.
8
13
353
When you ask an LLM to roast itself….
just built a fully automated Wojak meme generator in Glif in 5 min:. Claude 3.5 block generates the meme as JSON.ComfyUI block uses a Wojak Lora to generate a fitting image.JSON extractor + Canvas Block ties it all together . input "AI entrepreneur" 💀
8
24
346
The Future of Humanity Institute (FHI) closed down yesterday.
37
35
339
Deep rationalist lore - Eliezer Yudkowsky wrote the original source code for Factorio in an effort to slow down AI progress as much as possible. This is widely considered to have been a success.
7
14
343
I’m adding “disorientational human snares” to my list of esoteric fears.
7
10
343
(We don’t need AI for this - we already know they’re smart- but it will make it much more tangible).
1
6
333
How could it be that GPT-4 gets dumber/smarter/lazier/less lazy over time? The underlying weights aren’t changing, are they? Is it hardware related?.
73
4
326
I’m a rare believer in the few-worlds interpretation. There’s like 5.
13
11
310
The problem with humans is that they can only understand concepts that are representable as connections between neurons.
34
20
301
This will legitimately make this website less safe for a lot of people. I post zero personal information and I still get weird fixated harassers every so often- the idea that they would still be reading all of my posts after I block them makes me not want to be here at all.
Soon we’ll be launching a change to how the block function works. If your posts are set to public, accounts you have blocked will be able to view them, but they will not be able to engage (like, reply, repost, etc.).
20
15
305
“I guess you could say I’m a nihilist”. “Meaning?”. “Uh, no thanks”.
14
18
310
sorry sorry sorry sorry sorry sorry sorry sorry.
Our solution is simple: maximizing the probability of [Sorry] tokens at every position during response generation. This enables the model to learn the ability to transition from potential harm to safety refusal throughout the sequence. [4/n]
20
24
306
Out of 8 billion people, it’s likely at least one is scarily close to me in personality, like a real doppelgänger. I’m trying to figure out how to acausally coordinate with her so we can meet each other. Any ideas for schelling points?.
36
6
295
At long last, we have created the AI arms race from classic sci-fi novel Don't Create The AI Arms Race.
8
19
284
Step 1: gradually move the bulk of human communication to digital spaces.Step 2: make it impossible to distinguish between humans and bots in digital spaces.Step 3: ….
28
13
272
Instead of following a more complicated recipe, you can just add butter and it will taste good. The butter lesson.
25
8
270
They say power corrupts, but I like to think I could be corrupt without having any power at all.
18
13
265
I’m building an eval to test the capability of LLMs to generate new knowledge. To this end, I’ve made 100 major breakthroughs in science and mathematics, and put them all in a hidden test set. Let’s see if the LLM can come up with them too!.
11
14
269
Ohhh, the concept of an “infohazard” is an infohazard . That’s funny.
36
5
270
This is my favorite reply I’ve ever gotten. I still think about it and laugh sometimes
1
4
258
Does this AI pose catastrophic risks to society? You have 1 week to test it, thanks in advance.
I’m sorry but this is an absurdly short amount of time to be given for this, considering the task at hand
12
21
262
Running llama3 and TTS on an older computer is interesting. I keep thinking to myself, “this computer could’ve spoken to me like this the whole time, it just needed the right instructions.”.
12
17
246
With every year that passes I become increasingly paranoid that I’m actually quantum immortal.
27
9
243
Vitalik Buterin turned 18 today. So much accomplished already, and now he can celebrate with a drink 🍻 Happy birthday Vitalik!
13
1
243
OpenAI main office: Mira announces a charming new AI assistant model- it can tell bedtime stories! 🤗 . OpenAI basement: Ilya is locked in an alignment battle with the Q*-powered god machine
11
15
230