Piotr Mazurek
@tugot17
Followers
550
Following
460
Media
188
Statuses
834
enjoying the late pre-agi; making llms go brrr @Aleph__Alpha
Berlin
Joined September 2021
Don't wanna be here?
Send us removal request.
Explore trending content on Musk Viewer
#EleccionesMéxico2024
• 415231 Tweets
casillas
• 338179 Tweets
Flamengo
• 321191 Tweets
緊急速報
• 188304 Tweets
緊急地震速報
• 174478 Tweets
O Vasco
• 158129 Tweets
Whindersson
• 112157 Tweets
地震大丈夫
• 74801 Tweets
震度5強
• 67924 Tweets
Gabigol
• 59322 Tweets
PREP
• 47840 Tweets
能登半島
• 43810 Tweets
#FileninSultanları
• 39241 Tweets
地震警報
• 38773 Tweets
Cruzeiro
• 35320 Tweets
Soto
• 30220 Tweets
アラート
• 23618 Tweets
Natalie
• 23191 Tweets
Tigre
• 22962 Tweets
Valle
• 22481 Tweets
LEALTAD FURIOSA
• 22140 Tweets
Mario Delgado
• 18370 Tweets
Pereira
• 16702 Tweets
目覚まし
• 16007 Tweets
Homer
• 14373 Tweets
Bucaramanga
• 13165 Tweets
Borja
• 13086 Tweets
Marlon
• 12199 Tweets
#MissUniverseColombia
• 11614 Tweets
Fonseca
• 11524 Tweets
Calleri
• 11266 Tweets
Barranquilla
• 11119 Tweets
Pinned Tweet
As predicted in the llama-author-scaling-law, 328 people on the most recent llama paper 🦙
1/3 One of the intriguing aspects of the new
#Llama
paper is the considerable increase in the project complexity and the number of authors. The leap from 14 authors in v1 to 68 in v2 is impressive. Should this trend persist, we might see over 300 contributors in version 3 :)
2
3
34
1
19
185
Future is now
You can "de-code" a prompt for a given picture. This + Cross Attention image editing + some human feedback RL in the future and I already see the end of the prompt engineering paradigm as we know it today
22
64
451
@weirddalle
img-to-audio from the image:
Thanks to the CLIP investigator, we can now produce music based on the style of an image
Pic: Lofi nuclear war to relax and study
*turn on the sound🔈
9
28
175
4
9
196
Thanks to the CLIP investigator, we can now produce music based on the style of an image
Pic: Lofi nuclear war to relax and study
*turn on the sound🔈
9
28
175
@weirddalle
@weirddalle
Finally I found the right prompt `hybridized` is the word I have been looking for, thanks :)
2
1
82
@the_transit_guy
To this day I am traumatized by the memory of crossing the streets in Orlando 🇺🇸
1
1
68
@HarryStebbings
Definitely a bigger problem than literally 90% of the space occupied by cars 😏
2
1
61
The fact that all the major LLM benchmarks contain a significant number of wrong answers, and yet we use a fraction of a percent difference between models to determine which one is the best is quite astonishing🧐
1
5
46
@_akhaliq
@huggingface
@MokadyRon
@hertz_amir
Here you can find an unofficial implementation adjusted for Stable Diffusion:
E.g. my attempt with changing a dog into a teddy bear
Cross-attention control is such a great tool. It lets you edit the image directly with prompts. You can change the target of the image or change the whole style, preserving the content.
1/4
1
5
34
0
2
45
@nathanbenaich
To be fair, simply considering the CPI does not give a full picture of the scale of these projects. Manhattan Project accounted for a much larger portion of GDP, in addition, a larger portion of that GDP had to be spent on sustaining the population leaving a smaller surplus
2
0
36
Cross-attention control is such a great tool. It lets you edit the image directly with prompts. You can change the target of the image or change the whole style, preserving the content.
1/4
1
5
34
@MichaelTrazzi
Google at least have some LLM experience to strike back in the near future. Think what must happening now at the 10B a year team at Alexa xd
0
0
31
@josephdviviano
@ethanCaballero
@GOFAI_
Isn't Salesforce AI Research a rather cool institution? They produce all this great papers like the "AI Economist" and others. Seems pretty cool to me xd
1
1
26
@Noahpinion
The funny thing is that Russia literally has a GDP the size of Spain, and yet we 🇪🇺, the continent 10x wealthier, are unable to help our neighbor beat it.
2
0
25
1
2
21
@LangrJakub
@lexfridman
@ZelenskyyUa
@Ukraine
@APUkraine
@DmytroKuleba
@MFA_Ukraine
@Klitschko
@Kyiv
After watching Lex's interview with Oliver Stone, I have absolutely zero confidence that he will do an honest job in presenting the war
4
2
19
@nathanbenaich
If anything it just gave a bunch of startups tons of moat. They are not "just a GPT wrapper" anymore
1
0
20
The method was originally designed for Image Gen (), but was adjusted and open-sourced and works quite well with Stable Diffusion. Here you can find the open-source version adjusted for
#stablediffusion
:
4/4
0
1
16
@amasad
It is quite the opposite. In 2022, finally, the "western decline" narrative failed completely. The fall of the Russian army, the covid disaster in China, and dozens of other failures of autocracies, while "the west" despite some problems, seems stronger than ever
0
0
16
So now there are 4 Falcon models, the worse result is hardly different from LLAMA2, and even the best is not statistically significantly better than LLAMA2
Kind of strange for 4x the compute used for training 🫤
2
2
13
@gerstenzang
Google founders were literally doing PhD in graph theory and ranking, so yeah you can assume they were at least interesting in the topic
0
0
12
@armchairexp
@YIMBYPoland
Lol what? Turks in Poland are awesome, what cultural integration issues you are talking about?
3
0
12
@dimelovitito
@AlexLoftus19
@Appyg99
Pure copium, social mobility is about moving across different income groups, how easy it is to join the top quantile, aka not how to become Jeff Bezos, but how easy it is to move from a trailer park to become a layer or doctor, and this is easier in Europe.
6
0
10
@ylecun
@BlancheMinerva
Ok but how does it matter if the parameters are released under a non-commercial license? Like who cares under what the code is released?
1
0
10
@tdudzik_
@chrisalbon
In practice there are only two topics, unlimited intelligence and unlimited energy, everything else will come as a result once we solve this
2
1
10
@AlexLoftus19
@Appyg99
This is not what this graph is showing. It shows how easy it is to move up from the bottom quarter
2
0
9
@woj_zaremba
How did you make it aware that it doesn't know something and prevent it from hallucinating?
0
0
8
@EMostaque
@AMD
The M1 version would be great. It would open a lot of possibilities for showing off the model to all non-tech people using macs
0
0
9
@dylan522p
@abundantclimber
@sailaunderscore
This is clearly something that American mind can’t comprehend. Skling in Germany or Austria is clearly w working class activity, enjoyed by a wide spectrum of people
1
0
8
@DavidSacks
This is by far the most disrespectful thing I read this month, and it is very sad someone with such huge reach is spreading a propaganda piece like this
0
0
9
@realGeorgeHotz
Literally all deep learning technology was invented after 2010. Just compare how awful Theano was compared to modern deep learning frameworks
2
0
8
@tszzl
Since when prompt engineering is dead?
At this very moment, a whole new industry of tools for prompt engineering is being built. With tools such as
@LangChainAI
or
@dust4ai
we are yet to reach the golden age of prompting
1
0
8
@BlinkDL_AI
@StabilityAI
@AiEleuther
@EMostaque
What is the effective context length? After how many tokens it starts to forget the previous informations?
1
0
8
@multimodalart
Theoretically, this result should be easily achieved if the new LAION5B dataset is used. So there is some hope of creating an open source version in the near future
0
0
8
@EMostaque
At the time of buying it Nuance was alrady present in like every single radiology department in America. MSFT didn't buy the shity speeach-to-text tech, they bought enterence to every PACS system in the country, which is waiting to be put on a cloud and integrated with AI models
0
0
7
@ChrisJBakke
It is pretty bad explanation, obviously FTT, aka the magic beans, were worthless and should not be kept as collateral, but the question is where the real money went. Did Alameda lost it somehow or maybe the clients all refused to pay. It is the main unanswered question
1
1
8
@culturaltutor
It is also the basis for the greatest musical the German-speaking world has ever produced 🇦🇹
0
0
8
Perplexity AI - combination of Bing search and ChatGPT. Ok, I can buy that this might be a Google killer. I'm wondering how long it will take to put it in the basic version of Bing.
2
1
7
@hardmaru
@craiyonAI
@StableDiffusion
Why do you think it is the case since both models are trained using the LAION dataset?
3
0
7
@SamuelAinsworth
From Bloom, we know that multi-lingual training severely decreases performance on English benchmarks. Does it mean that we can just stop doing multu-language training, train separate models for different languages, merge them and keep the original performance on every language?
1
0
6
@DanImmergluck
You may not know this, but high quality mass transport is basically equivalent of apartaid
0
0
6
@metaphorsystems
Do you have any ideas on how such queries, which require some context about the source, could be better handled? The answer is nowhere to be found in the results, while Google somehow, as a first result, returns me a website discussing the answer I'm looking for
1
0
6
a fantasy landscape with a maple forest
a watercolor painting of..., a van gogh painting of..., a charcoal pencil sketch of...
3/4
1
2
6
@armchairexp
@YIMBYPoland
Just have the nice ones 😅
To teraz już wiecie jak głosowały te tłumy młodych ludzi w Warszawie. Kilicdaroglu 85%, Erdogan 8%.
13
37
553
1
0
5
0
0
6
Unpopular opinion
In a way,
@OpenAI
's
#Dale2
is inferior when compared with
@craiyonAI
's. It comprehends contemporary culture far less well than
#Dallemini
. It fails to recognize public figures and generally struggles with more abstract concepts.
1
1
5
@LangrJakub
@jeremyphoward
@lexfridman
@kanyewest
I had a very similar feeling. The Stone interview was the first one where I had to pause the podcast several times because I couldn't stand the amount of lies that were told and in no way straightened out by Lex
1
0
5
Prediction
By the end of April, FAIR will open-source the LLAMa under the Apache License
The incentives to do so are just too strong, and there is no real downside to doing this. This will 10x the speed of the development of the Llama ecosystem
ChatGPT-like model by July 1st
1
2
5
This is a very interesting take on why
@GrzegorzRutko14
seems so "over-represented" in the AI Art space .
TL; DR: The problematic part might be the CLIP model, not the
@laion_ai
dataset.
1
0
5
@__tinygrad__
Memory bandwidth is at 5.2Tb/s, flops also better than H100.
It is a genuine question; Why it is more optimal to use 6 consumer GPUs rather than a single Mi300?
2
1
5
@TerribleMaps
local market with colorful townhouses is missing, there is one in every city from Antwerp to Talinn
0
0
5
0
0
4
I'm absolutely bullish on Taiwan 🇹🇼. A random newspaper in the random store—GTC summary on the front page. Taiwan is sooo back 🚀
0
0
4
@irinarish
The biggest problem with Paris is the language barrier. Speaking English means you are a second class citizen
2
0
5
A photo of a Corgi dog riding a bike in Times Square wearing sunglasses and beach hat, cinestill, 800t, 35mm, full-HD
A photo of a teddy bear riding a bike in Times Square wearing sunglasses and beach hat, cinestill, 800t, 35mm, full-HD*
*
@LexicaArt
2/4
1
2
5
This is not an investment advice but you should put all your net worth in hands of this men 🫡
0
0
5
And people say that ChatGPT is useless ... try finding this with google 😄
0
0
5
1
0
5
Next year
In Untitled23.ipynb you can find the working example of AGI, in scripts/control.py you can find how to fine tune the model to manage and maximize the profits of a pharmaceutical company
0
0
5
@pfau
GATO seems like a great foundation for future research to me. It has been said for some time that the LLM of the future will need to be trained in multi-modal tasks to achieve more impressive results, and now we have the first working indication of how impressive they might get
0
1
4
1/3 One of the intriguing aspects of the new
#Llama
paper is the considerable increase in the project complexity and the number of authors. The leap from 14 authors in v1 to 68 in v2 is impressive. Should this trend persist, we might see over 300 contributors in version 3 :)
2
3
34
@gordic_aleksa
@DeepMind
Good luck; I don't think there has been anyone building AI-bio tech in public yet, looking forward to reading about this
1
0
4
Just realized that a small object tracking project I created 2 years ago surpassed 150 stars a few days ago 🥳🎉
1
0
4
@Noahpinion
Fun fact; in Poland 🇵🇱 Girkin is quite a popular online figure, many people like to hear him say how much Russia is screwed
0
0
4
@marktenenholtz
There is basically one algorithm: SGD, and everything else just builds on top of that
1
0
4
