Igor Babuschkin
@ibab
Followers
44,828
Following
695
Media
32
Statuses
395
Maybe the real AGI was the friends we made along the way. @xAI
San Francisco, CA
Joined February 2020
Don't wanna be here?
Send us removal request.
Explore trending content on Musk Viewer
Flamengo
• 164654 Tweets
Celtics
• 70638 Tweets
Yasuke
• 64735 Tweets
Bolívar
• 49837 Tweets
Chargers
• 42784 Tweets
#AEWDynamite
• 31726 Tweets
Alianza
• 30333 Tweets
Cavs
• 29783 Tweets
Chivas
• 26708 Tweets
Fortaleza
• 24229 Tweets
De La Cruz
• 23321 Tweets
Garland
• 22394 Tweets
Nature Campaign
• 17888 Tweets
Grabois
• 17776 Tweets
Allan
• 15623 Tweets
Marra
• 15452 Tweets
Cebolinha
• 15275 Tweets
Apolinho
• 14521 Tweets
Vidal
• 12809 Tweets
優三さん
• 11785 Tweets
Gabigol
• 11768 Tweets
Cavani
• 11385 Tweets
Al Horford
• 10871 Tweets
#DesafioXX
• 10861 Tweets
Twitter really needs to add the ability to write math formulas in tweets with LaTeX syntax.
273
207
3K
@JaxWinterbourne
The issue here is that the web is full of ChatGPT outputs, so we accidentally picked up some of them when we trained Grok on a large amount of web data. This was a huge surprise to us when we first noticed it. For what it’s worth, the issue is very rare and now that we’re aware
275
209
3K
161
393
3K
xAI is moving at light speed and there’s a lot of exciting research and engineering work ahead of us. Apply to join the team at
352
270
2K
We've released our first progress update at xAI.
78
120
1K
86
183
958
I’ve realized that the reason I started to work on AI a long time ago is because I wanted to create miracles.
To create wonderful things that nobody had ever seen before.
But during all the years I’ve worked on AI at DeepMind and OpenAI I’ve learned again and again that nothing
98
113
1K
Grok just became 3x faster. More improvements coming soon.
119
168
1K
Join the Grok community here on X if you want to help us improve Grok, or to see how others are using Grok.
57
162
615
😌 Settings > Privacy and Safety > Mute and Block > Muted words
117
72
1K
If you want to move past the AI hype and learn some real fundamental basics behind today's learning algorithms there's no better choice than MacKay's "Information Theory, Inference and Learning Algorithms". You can read the book for free on the official website:
54
118
1K
Grok should be rolled out to every Premium+ subscriber in the US now. If you're on mobile and don't see it, make sure to update and restart the app. Happy Grokking!
109
124
724
The Grok-1 repo is getting pretty popular. I will be responding to pull requests and issues. Feel free to contribute!
62
75
886
We will dramatically improve Grok over the next few months. This is just the beginning.
Grok has the potential to completely change the way we use X. It should be considered a foundational element of the platform, used to make this social network *aware of what people are saying* for the first time ever
22
21
291
75
48
877
We will soon have video games that are fully generated by AI. All that's needed is a multimodal model that generates video and audio, takes in keyboard and mouse controls and runs at 30 FPS. This is almost feasible with today's hardware.
Introducing General World Models.
We believe the next major advancement in AI will come from systems that understand the visual world and its dynamics, which is why we’re starting a new long-term research effort around general world models.
Learn more:
116
392
2K
59
69
865
Grok-1 314B running on M2 Ultra 🚀
45
78
702
We're releasing PromptIDE, one the internal tools we've built to accelerate our work on Grok. It allows you to develop and run complex prompts in the browser using an async based Python library.
Announcing the xAI PromptIDE
The xAI PromptIDE is an integrated development environment for prompt engineering and interpretability research.
It accelerates prompt engineering through an SDK that allows implementing complex prompting techniques and rich analytics that
visualize
1K
2K
10K
23
80
504
Grok is now in the X app. We’re rolling it out to all Premium+ users in the US.
Make sure to update the app and restart it if you don’t see Grok in the bottom nav bar.
Grok is rolling out to 𝕏 Premium+ subscribers with the latest app release.
Have fun!
5K
4K
37K
132
60
651
We're starting a new DeepMind research team in California to work on large-scale deep learning models. Apply if you want to contribute to our mission of building AGI!
Interested in exploring the limits of large-scale models? Consider joining our brand new Scalable Deep Learning team in California as a Research Engineer or Research Scientist, see job descriptions below!
16
64
554
7
54
519
If you had access to an infinitely intelligent AI, what's the first question you would ask?
436
37
467
I keep revisiting this great paper from
@andy_l_jones
: “Scaling scaling laws with board games”. It shows how training compute and inference compute of MCTS can be traded off against each other. 10x more MCTS steps is almost the same as training 10x more.
14
68
453
μP allows you to keep the same hyperparameters as you scale up your transformer model. No more hyperparameter tuning at large size! 🪄
It saves millions of $ for very large models.
It’s easier to implement than it seems:
You have to
1. Keep the initialization and learning rate
If you are doing LLM (>1B) training runs, you ought to do these 3 things:
1. Use SwiGLU
2. Use ALiBi
3. Use µP
Why? Your training will be almost 3X faster!
You can do 3 runs for the price of 1.
You can go for much bigger model or train longer.
There is no excuse.
1/3
10
27
241
10
62
421
@bindureddy
Elon is the only one that still has a chance of creating powerful AI that benefits all of humanity
48
40
405
Weird insight of the day: If you are setting up a tokenizer for LLM training and you choose a string like <|endoftext|> that does not occur on the internet for your special token, eventually it will occur (e.g. in the GitHub issues of your LLM project) and your model will not be
11
6
80
@xDaily
@xai
.
@BrianRoemmele
does not have access to Grok and the information he has posted is almost all incorrect. We are working on some new and exciting features though.
24
17
329
AI alignment research lead to RLHF, which lead to ChatGPT, which lead to the current AI capabilities race. In this way, perhaps AI alignment researchers have done more than anyone else to accelerate AI capabilities.
A year ago I thought AGI was probably going to destroy the world, so I spent most of 2023 working on AI alignment.
Now I'm wondering if AI alignment is going *too fast*:
24
15
222
31
22
305
Grok gives the most creative explanation of the uncertainty principle I've heard so far.
26
16
245
This is a good idea. There’s way too much training on the test set going on with LLMs right now. Difficult to avoid if you are training on the whole internet.
Someone with money should hire annotators to create a new version of an eval suite every 3-12 months. This way, we can get guaranteed uncontaminated evals by only downloading training data from before the latest version of the eval suite was released.
2
0
10
1
2
33
@marilynika
@xai
@elonmusk
Send us your referrals! We want to find the best people. Not just AI scientists but also engineers without an AI background that can demonstrate excellent prior work.
17
5
181
Turns out humans are just as susceptible to the Waluigi effect.
All humans are capable of both good and evil. And most who do evil believe they are doing good. History shows this over and over again.
1K
1K
8K
12
14
162
This is very impressive. Together AI is worth following.
Access Mixtral with the fastest inference performance anywhere! Up to 100 token/s for $0.0006/1K tokens — to our knowledge the fastest performance at the lowest price!
Mixtral-8x7b-32kseqlen
@MistralAI
& DiscoLM-mixtral-8x7b-v2 are live on Together API!
48
148
1K
4
8
160
@JaxWinterbourne
The issue here is that the web is full of ChatGPT outputs, so we accidentally picked up some of them when we trained Grok on a large amount of web data. This was a huge surprise to us when we first noticed it. For what it’s worth, the issue is very rare and now that we’re aware
275
209
3K
9
5
165
Found this amazing rendition of the Bhagavad Gita in Sanskrit. The Gita has helped me get through some dark times. Can’t recommend it highly enough.
17
19
146
@Simeon_Cps
I personally believe we need to keep our hearts and minds open at this point. We don’t have enough evidence to know which actions will lead to the best outcome for humanity. We need to avoid the danger of making things worse by taking rash actions or by forming dogmatic beliefs
8
7
129
@AravSrinivas
This is a good failure case. The reason it fails here is because Grok decided not to search on X (it makes that call on each message). We’ll fix it in the next version, but in the meantime it’s enough to ask it to search to get a better answer:
6
7
116
A nice technique for making LLM evaluation with human raters more efficient. Now I feel bad for spending hours evaluating LLM outputs manually 😂
Very excited to share our latest work:
🤔 Which Prompts Make The Difference? Data Prioritization For Efficient Human LLM Evaluation
w/
@eddotman
,
@beyzaermis
,
@mziizm
,
@sarahookr
🔗:
4
34
175
5
16
106
@AravSrinivas
This is absolutely true. It’s a real challenge for everyone building LMs right now. The ChatGPT conversations need to be filtered out of the training data.
11
3
98
@edvegaviolante
Being able to talk about mathematics. Would be especially useful for long form posts.
6
1
95
@DavieLofi
Full steam ahead. If people with good intentions quit the field, then those that remain might not have humanity’s interests as their top priority.
7
4
94
Stan did some amazing work on theorem proving with LLMs at OpenAI. I’m really looking forward to what he and the Dust team will come up with next.
4
1
85
Yet poets will turn up, to blaze
Your glory forth to after-days,
And set by folly fools a-craze.
- Mephistopheles
8
6
73
Right now I believe the biggest risk to humanity does not come from accidentally misaligned AIs, but from superintelligent AIs that are specifically trained to control humans. Whoever first decides to develop such a terrifying system will start off an international arms race.
14
7
76
@bindureddy
Thank you for the feedback. We’re working on it. The problem here is that the X post search we built didn’t find relevant posts for your query, which can cause the LM to hallucinate right now. Sometimes it can help to rephrase the question. We’ll ensure this will be more robust
5
2
78
@hbou
This doesn’t exclude the possibility of us living in a simulation, since the simulation could also contain evolution and the origin of life. Personally my view on this is that whether or not we are living in a simulation, we must act as if what we experience is real. To give up
10
0
70
I suspect someone in Balenciaga’s viral marketing team is getting a big raise this year 🤔
10
4
66
I’m so grateful to the
#wellsfargo
employee behind me in line this morning who paid for my breakfast when my card didn’t work.
3
2
62
Wer im Plutonium-Haus sitzt sollte nicht mit Neutronen werfen ☢️
7
4
46
Once, there was an explosion.
A bang which gave birth to space and time.
Once, there was an explosion.
A bang which gave birth to our planet.
Once, there was an explosion.
A bang which gave birth to life as we know it.
And then, there came the next explosion...
Death Stranding
9
6
52
@joeldangelone
@WholeMarsBlog
Thank you! We're working hard to make it available in more countries.
11
3
50
It’s a good lesson that progress in AI is difficult to predict. The road to AGI doom is paved with good intentions.
3
2
48
@Joybeanns
@xai
I really appreciate your feedback David! These are great ideas and I've sent your post to the team. Grok is still in its early days and we are working hard to make it more capable and useful.
5
1
49
Time to get all those underpaid high school teachers involved. The pay is better and neural networks tend to be much friendlier students too 🧑🏫
In the age of ChatGPT/Galactica, the value of unverified 10 page essays drops to zero. What matters now, more than ever, is fact checking. To remain grounded in reality, Effective Altruists must work with actual domain experts. Stop paying novices, and start paying experts.
5
8
81
3
4
46
@elg
@grok
@Scobleizer
@elonmusk
@SullyCNBC
Hey, don't worry. Grok does not have access to anyone's draft posts. If you ask it for them it might try to imagine what they look like. Grok does point out that they are not real at the end, but this could be more visible.
4
3
37
@EdKrassen
This is likely a bug. Regular mode should not give a funny answer like that. We’re looking into it.
6
1
37
@davinder0110v
Yeah, this is a great beginner textbook to learn about information theory. You need some basic mathematical skills like calculus and linear algebra to get the most out of it.
4
0
34
@bindureddy
Really appreciate your feedback! All of these features are planned and I hope we can ship them soon
5
2
35
@ID_AA_Carmack
There are also datacenter and operations cost: With the H100 you'll need to install fewer GPUs, and you might be able to deploy fewer racks. There will be fewer failures to respond to. The power usage for the total cluster will also be lower, although this is not a major cost
4
1
32
@calwoodford
We want Grok to be a maximally truth-seeking AI assistant. We have found multiple bias issues with our current training data and how Grok handles retrieved posts, which we'll address.
6
3
31
@ter_pieter
We will be surpassed soon. Hopefully we can use it to enhance and enlarge the human experience.
2
0
30
3
1
29
@Jeethu_tweets
@stephenbalaban
@xai
Yeah, we use Rust as our main low level language. No Haskell so far :)
4
0
26
@BasedBeffJezos
@ethanCaballero
That’s not true, but I do like using it. Very simple design that makes it very flexible in practice.
1
0
25
@rpoo
Euler-Mascheroni constant.
First appeared in 1734 and we still don’t know whether it is irrational!
0
0
23
