For a deeper dive, check it out on arxiv: arxiv.org/abs/2301.08110
Congrats to the researchers: , Björn Deiseroth, Samuel Weinbach, , and .
Thanks , , and for their collaboration. 9/9
Follow
Aleph Alpha
@Aleph__Alpha
Our mission is a European generalizable AI. We're hiring: aleph-alpha.com/company #AGI, #artificialintelligence, #writtenbyahuman,#writtenbyanAI
Aleph Alpha’s Tweets
AtMan's memory usage is comparable to a forward pass. It scales up much more efficiently when compared to other methods that require a backward pass.
The comparison shown here is on a batch size of 1. AtMan’s runtime decreases significantly with higher batch sizes. 8/9
1
3
Show this thread
We achieve #SOTA in image and language reasoning, emerging from the same exact method on both modalities. 7/9
1
4
Show this thread
AtMan can be used to explain the outputs of both image and text models. 6/9
1
3
Show this thread
But does manipulating attention keep the model’s output within distribution?
Yes. In fact, attention manipulation can be used to steer the model’s prediction in different contextual directions. 5/9
1
2
Show this thread
How do we suppress tokens?
By scaling down the attention scores right before the SoftMax gets applied 4/9
1
3
Show this thread
The foundational idea is simple. We erase/suppress one token at a time from the input and then look at the change in the cross-entropy of the next token prediction. 3/9
1
4
Show this thread
AtMan is an attempt at helping users know the source of the answer given by LLMs. This not only makes LLMs more reliable but also empowers prompt engineers and researchers to push the boundaries of AI. 2/9
1
3
Show this thread
At #AlephAlpha we are developing Large Language Models (LLMs) that can be trusted in real-world applications. Our most recent work is AtMan, an #ExplainableAI method providing explanations of generative transformer models at almost no extra cost. A Thread 1/9
#writtenbyalephalpha
1
12
48
Show this thread
#QDrant database enables you to store and efficiently search through embeddings. Storage can reduce the cost of using the #AlephaAlpha API. You embed the document once, and you can reuse it for each query you make. Enjoy :)!
1
Show this thread
Our friends at created an example showing how to store Aleph Alpha embeddings in their database. You can check it out at: qdrant.tech/documentation/
#writtenbyalephalpha
1
8
22
Show this thread
Have you ever wondered why and how a model predicts what it does? Together with we introduce AtMan that provides explanations of generative transformer models at almost no extra cost.
📄 arxiv.org/abs/2301.08110
#writtenbyalephalpha
4
24
ChatGPT is basically useless for high-value work. An expert (e.g. lawyer) needs explainability, sources and a way to influence and control outcomes. I am proud of what our team achieved with AtMan. Making multimodal generative AI trustworthy. arxiv.org/abs/2301.08110
41
271
1,563
Show this thread
🎙️Check out the latest #podcast with our founder and CEO, and discussing the societal challenges sparked by the release of ChatGPT.
🔗 (behind a paywall, and available in German): ow.ly/MeCh50Mw1Wn
#writtenbyalephalpha
2
7
Quote Tweet
#ChatGPT: Was kann der neue #KI-Chatbot und wie wird er unseren Alltag verändern? @ChristophKeese @lenacarla thepioneer.de/originals/tech
3
5
What’s the learning rate you’d suggest for the loss landscape in Davos (visible in the background)
3
1
14
We strive to make constructive contributions to the debate on future developments of AI. is part of this year's Annual Meeting of the in Davos to speak on the most pressing issues of future AI. AI Lunch, 2:00pm
#writtenbyalephalpha
#writtenbyahuman
1
2
17
Der Staat sollte nicht nur durch Fördergelder, sondern gezielt über Aufträge in Zukunftstechnologien wie #Luminous investieren, um einen größtmöglichen Nutzen für Wirtschaft und Gesellschaft zu ermöglichen.
#writtenbyahuman
#writtenbyalephalpha
Quote Tweet
Lange waren erfolgreiche Start-ups, die auf hochkomplexer Technologie basieren, hierzulande eher eine Seltenheit. Das ändert sich gerade. Doch es bleiben Probleme, berichtet @MaxosTaxos: trib.al/dovovLi
1
9
“We are good at criticizing. We need to try more things in Europe. Technology is improving very quickly.“ ()
#DLD23
3
5
Quote Tweet
Nice Forbes article outlining our leading role in Europe for generative AI and LLMs. And while I want to remind everyone that model size is not everything, I am excited about our 300b luminous-world incoming in 2023 (already in test deployment).
forbes.com/sites/federico
3
8
Die wunderbare Doris Schneyink des hat ein Gespräch mit mir geführt zum Thema OpenAI's ChatGPT, der Zukunft der Arbeit und was wir von KI in den nächsten Jahren erwarten können.
1
15
Was ist #KünstlicheIntelligenz, wie steht Deutschland im internationalen Vergleich, und weshalb verlässt man Apple? Darüber haben wir mit gesprochen, CEO des deutschen #KI-Start-ups . #ai #artificialintelligence
the-decoder.de/deep-minds-13-
9
14
Our team published a break-through method to add multi-modality to today's #LLMs and they subsequently added this to our #Luminous models.
One year later, we achieved another break-through. M-VADER creates images from any mixture of text and images.
🔗 lnkd.in/eS56uym7
8
22
Making the right mistakes is inherent to human perception – and sometimes useful for multimodal models.
Read the full article on this topic from one of our researchers for the here: thegradient.pub/learning-to-ma
#writtenbyahuman
#writtenbyalephalpha
Quote Tweet
I wrote a little something with @gradientpub on how multimodal LLMs have a strange similarity to the way human perception works. It’s a mildly hot take if you ask me
(also involves a Bigfoot
)
thegradient.pub/learning-to-ma
)
thegradient.pub/learning-to-ma4
7
👉 Current generative AI models generate new images based mainly on text. #artificialintelligence the-decoder.com/multimodal-pro
1
2
5
Show this thread
🤝 et présentent un modèle d’IA clairsemé à 80 %
actuia.com/actualite/grap #IA #intelligenceartificielle #collaboration #chatbot
4
6
📣 Together with we were able to reduce our model from 13bn parameters to 2.6bn parameters. Our research teams, cleaned up 80% of the less relevant model weight, while retaining the most important parameters.
🔗 Read more about it here:
3
12
🤝 As part of our cooperation with Graphcore, we unveiled a significant advance in #AI #compute #efficiency, with the sparsification of our 13bn parameter model down to 2.6bn parameters.
🔗 Read more here: lnkd.in/e2f2wfYB
#writtenbyahuman
#writtenbyalephalpha
Quote Tweet
Together with @Aleph__Alpha we've delivered an 80% sparsified model, pruning parameters and dramatically reducing compute demand. Read more: graphcore.ai/posts/graphcor
0:19
308 views
9
🚀 Join the #𝘧𝘪𝘳𝘦𝘴𝘪𝘥𝘦𝘤𝘩𝘢𝘵 with our founder and CEO , and .
🗓️ 16. November
🔗 See the full agenda here: hubs.la/Q01rDGWx0
Quote Tweet
Prepare to unlock tomorrow’s largest #AI models with tips on cutting-edge #machinelearning tools, methods and real-world implementation insights from @Aleph__Alpha, Graphcore and @HPE.
Join our fireside chat next week at #SC22 – see our full agenda here: hubs.ly/Q01rDFfT0
3
6
⁉️ Have you ever struggled to understand a document in a foreign language?
🖤 Our language model #Luminous is available in 5 European languages and helps you find your answers in the desired language.
🔗Try for yourself: app.aleph-alpha.com
#writtenbyalephalpha
2
1
9
🖤 Use our conversational module #Lumi to interact freely with any information. No training and no annotation is necessary. Simply connect #Lumi to your data and information and everything is ready!
🔗 Give it a go: lnkd.in/e6Xw3qM5
#writtenbyahuman
#writtenbyalephalpha
0:26
115 views
2
6
💬 No more pre-trained #Q&As but rather human-like conversations. With our conversational module #Lumi there are new possibilities for human-machine-communication, which makes us not another #Chatbot company.
🔗 Read more: aleph-alpha.com/use-cases
#writtenbyalephalpha
2
3
13
Als Heidelbergerin werde ich künftig viel Spaß haben mit #Lumi 🤗 Gestern gab ich mich als Amerikanerin aus, sofort switchte die Sprache. 🇩🇪/🇺🇸 Women's Association kannte der Agent zwar nicht, but hey: nice chat and lifelong learning also applies to AI🧲
2
1
9
🖤 #Lumi, #Heidelberg’s digital citizen assistant, not only helps you deal with authorities but also informs you about cultural events and recreational activities.
🔗Read more about our project with the city of #Heidelberg here:lnkd.in/eJQdaent
#writtenbyalephalpha
2
5
Die Stadt und starten nach intensiver, gemeinsamer Testphase den virtuellen Bürgerassistenz Lumi.
1
2
4
🚀 Yesterday we launched #Lumi, a digital citizen assistant, with the city of #Heidelberg.
🤝 #thankyou to , the city of Heidelberg, the Amt für Digitales und Informationsverarbeitung and .
🔗 lnkd.in/eEXDdADd
#writtenbyalephalpha
1
5
😩 You know the pain, - you are in a rush and don’t have time to search the web to make an appointment. Look no further.
🖤 #Lumi, #Heidelberg’s digital citizen assistant, will save you time and hassle.
🔗 Give it a go: heidelberg.de
#writtenbyalephalpha
5