AI Guru

Helping humans understand AI, #Stanford, AI mole, Researcher, ex-@Google, ex-@Meta. Talking #AI, #GenerativeAI, #DataScience, #MachineLearning.

[AI Guru @AIGuruAIGuru]

AI will not replace you. A person using AI will.

[AI Guru @AIGuruAIGuru]

Well #AI, #chatbots

[AI Guru @AIGuruAIGuru]

“AI will take over the world”.

Meanwhile AI:

#technology #ai #ArtificialIntelligence

Follow @AIGuruAIGuru for more.

[AI Guru @AIGuruAIGuru]

#PyTorch just released an awesome tool to visualize matrices and what's happening inside them. Matrix multiplications (matmuls) are the building blocks of today’s #ML models.
It can run in browser: https://t.co/Budpu0YVQ0

#AI #datascience #machinelearning #GenerativeAI

[AI Guru @AIGuruAIGuru]

If you are thinking about joining in this #ML frenzy, stop trying to think how to best do it; just do it! Just pick a book or write some code, make mistakes, and move from there! See you on the other side!
#AI #datascience #GenerativeAI #MachineLearning

[AI Guru @AIGuruAIGuru]

To prove this, researchers tested GPT-3 and LLaMA on made-up facts in one direction and then test them on the reverse. Result: "near 0% accuracy on reversals"

[AI Guru @AIGuruAIGuru]

New paper is shaking the world of LLMs. The "Reversal Curse".
Language models trained on “A is B” cannot generalize to “B is A”, a complete failure of logical deduction.
Code: https://t.co/ga6og2gw4B
Paper: https://t.co/Pjwbq52kpi

#AI #ML #Datascience #GenAI #machinelearning

[AI Guru @AIGuruAIGuru]

If a model was trained on: “Jack Dorsey was the first president of twitter" It will not automatically be able to answer the question: “Who was the first twitter president?”

[AI Guru @AIGuruAIGuru]

Better and more efficient domain adaption would be something highly welcomed by the industry.

[AI Guru @AIGuruAIGuru]

At @Bloomberg they paid >1 million $ in training costs for something that is outperformed by a 100$ model only six months later.

This gives you an idea about the pace that LLM training and development has been moving at this year.
https://t.co/eIxgIMD3Um
https://t.co/gDs2M7wRzK

[AI Guru @AIGuruAIGuru]

The researchers from Microsoft find that this leads to a much broader skill set learned by the model than simply continuing pre-training on domain specific data. Despite the financial domain, they've also evaluated their technique for Law and Medicine.

[AI Guru @AIGuruAIGuru]

Nougat: OCRing scientific papers with a deep net trained end to end. From @Meta
Website: https://t.co/XdAaFjpaSS
Paper: https://t.co/XxAlHmzzdn
Code: https://t.co/40MTlfTsn1

[AI Guru @AIGuruAIGuru]

Trained on: * a whooping 3.5 trillion params * ~4096 GPUs * 7M GPU hours (4x more compute than LLama 2) License: Can be commercially used but under restrictive terms apparently

[AI Guru @AIGuruAIGuru]

Falcon 180B model has just been released! This makes it the largest (BLOOM 176B, OPT 175B) and the most capable (comparable with PaLM 2 Large) "openly" available language model!

TII is doing some serious engineering feats here!

Check the model here:
https://t.co/D7mZJyAPZD

#AI

[AI Guru @AIGuruAIGuru]

pushes the drone to its limits while flying through the predefined racetrack.

[AI Guru @AIGuruAIGuru]

"Champion-Level Drone Racing using Deep Reinforcement Learning,"
Paper open access: https://t.co/7116T8Azlm
Video: https://t.co/kHaaxNa1Vw

#AI #DeepLearning #visionAI #genAI #ML #Datascience

Follow @AIGuruAIGuru for more.

[AI Guru @AIGuruAIGuru]

This information is then combined with inertial sensors to estimate the drone’s position, orientation, and speed on the race track. Then, a neural network controller trained with deep reinforcement learning in simulation, combined with data collected in the physical world,