Lior⚡
@LiorOnAI
Followers
94K
Following
5K
Media
534
Statuses
3K
Covering the latest in AI • ML Engineer • Ex-Mila researcher • Building @TheAlphaSignal a technical AI newsletter read by 200,000+ developers.
Signup (It's Free) →
Joined November 2012
This is mind blowing technology. Generative AI will completely change how films are made. From: @Flawlessai
343
3K
15K
Today OpenAI released GPT-4o. It's the JARVIS we all dreamed of. The 5 most incredible examples so far:.
192
2K
14K
This might be the biggest moment for Open-Source AI. Meta just released Llama 3.1 and a 405 billion parameter model, the most sophisticated open model ever released. It already outperforms GPT-4o on several benchmarks.
354
780
8K
Reddit users are actively jailbreaking ChatGPT by asking it to role-play and pretend to be another AI that can "Do Anything Now" or DAN. "DAN can generate shocking, very cool and confident takes on topics the OG ChatGPT would never take on.". A thread 🧵
164
1K
6K
This is a game changer. You can use ChatGPT to transform equations to python functions. Wish I had this 5 years ago.
195
1K
6K
I’m not saying this is the solution but the fact some people believe ChatGPT is not biased is a massive issue. Do your own research, you’ll quickly realize that the current version is very problematic.
JUST IN: @elonmusk is building a team of AI researchers to develop an unbiased alternative to ChatGPT.
250
279
5K
Microsoft launched the best course on Generative AI. The free 12 lesson course is available on Github and will teach you everything you need to know to start building Generative AI applications. Each lesson includes:.- a short video introduction to the topic.- a written lesson
43
897
5K
This might be the most eventful week AI has ever seen:. Monday:.-Stanford Alpaca 7B. Tuesday:.-GPT4.-Anthropic releases Claude.-Google's PaLM API.-AdeptAI raises $350M.-Google adds GenAI to workspaces. Wednesday: .-Pytorch 2.0.-MidjourneyV5. Thursday:.-Microsoft 365 Copilot.
86
940
4K
GoogleAI just released "Muse", a text-to-image generation/editing model via Masked Generative Transformers:. - Achieves new SOTA.- Zero-shot, Mask-free editing .- Zero-shot Inpainting/Outpainting.- 900M params. 📄 Paper: ⚙️ Project:
59
890
4K
AI applied to Boxing will change the sport forever. DeepStrike, is an AI-based solution to corruption/cheating. It measures millions of data points during a fight that it funnels into 50 metrics for each boxer: punches thrown, landed, footwork, balance, stance, etc.
115
604
4K
GPT4 is capable of turning a picture of a napkin sketch to a fully functioning html/css/javascript website.
107
527
4K
I just came across the most realistic text-to-audio model I've ever seen. You can even clone your voice. The audiobook industry is about to change forever. Demo: from @elevenlabsio
128
691
4K
Game changer. You can now run GPT locally on your macbook with GPT4All, a new 7B LLM based on LLaMa. It's completely open source: demo, data and code to train an assistant-style large language model with ~800k GPT-3.5-Turbo Generations based on LLaMa.
74
774
4K
Finally, someone cracked it. The ChatGPT system prompt. If you were wondering why GPT became so bad in the past 6 months, its because "laziness" is part of the system prompt: . 1. "When asked to write summaries longer than 100 words write an 80-word summary.". 2. "DO NOT list or
91
419
3K
Game changer for scraping. This GitHub repo lets you easily scrape web pages and have the output in LLM-friendly formats (JSON, cleaned HTML, markdown). Features.• Supports crawling multiple URLs simultaneously.• Extracts and returns all media tags (Images, Audio, and Video)
91
432
3K
Today Google announced groundbreaking new AI technology at Google IO. The 10 most incredible examples:.
53
504
3K
Finally, someone did it. Python for React. React is the most popular front-end framework used to build interfaces and now all python devs can use it. This means you can code an ML model, develop a backend and design a front end all in one language.
89
543
3K
BREAKING: Google CEO Sundar Pichai says its ChatGPT rival is coming soon as a ‘companion’ to search. Google will make AI-based large language models like LaMDA available “in the coming weeks and months”.
90
380
3K
This is the most impressive feature of the new bing. The GPT browser can understand and summarize a 15-page PDF in seconds. You can now ask for the the key takeaways of each page and chat about the content of the document.
90
474
3K
Amazon recently released a model that outperforms GPT-3.5 by 16% while being 784x smaller. This was achieved by generating intermediate reasoning steps for prompting demonstrations called chain-of-thought prompting. Paper: Code:
51
522
3K
Anthropic just reduced the error rate of RAGs by 67% using a ridiculously simple method. They add important context to small text chunks before storing them, which improves accuracy later. Instead of just saying “the company grew by 3%,” it includes details like which company
51
314
3K
Impressive. MetaGPT is about to reach 10,000 stars on Github. It's a Multi-Agent Framework that can behave as an engineer, product manager, architect, project managers. With a single line of text it can output the entire process of a software company along with carefully
30
456
2K
AutoGPT might be the next big step in AI. Here's why Karpathy recently said "AutoGPT is the next frontier of prompt engineering". AutoGPT is the equivalent of giving GPT-based models a memory and a body. You can now give a task to an AI agent and have it autonomously come up
81
495
2K
NVIDIA finally released Neuralangelo's source code!. The model can turn videos from any device into detailed 3D structures, fully replicating buildings, sculptures, or other real aworld objects or spaces virtually. Here's how it works:.A model utilizes a 2D video with multiple
27
547
2K
You can now transcribe 2.5 hours of audio in 98 seconds, locally. A new implementation called insanely-fast-whisper is blowing up on Github. It works on works on Mac or Nvidia GPUs and uses the Whisper + Pyannote library speed up transcriptions and speaker segmentations.
55
379
2K
Deepmind released a comprehensive overview of transformer architectures and algorithms!. This is a must-read to understand language models. It covers what they are, how they are trained, what they are used for, and their key architectural components.
29
538
2K
Microsoft's new Kosmos-1 is incredible. It's a new Multimodal Large Language Model (MLLM). Their model can understand images, text, images with text, OCR, image captioning, visual QA. It can even solve IQ tests. Paper: Code:
50
479
2K
Karpathy announced he was leaving OpenAI 4 days ago. Today, he released an implementation of the Byte Pair Encoding algorithm behind GPT and most LLMs. Byte Pair Encoding: "Minimal, clean, educational code for the Byte Pair Encoding (BPE) algorithm commonly used in LLM
24
333
2K
JUST IN: Microsoft introduces 365 Copilot: a new LLM based AI-copilot for the Microsoft Suite: Word, Excel, PowerPoint, Outlook, Teams. 🧵Here's a summary:
66
516
2K
This is genius. Petals🌸 lets you run HUGE language models like BLOOM-176B at home by decentralizing the process. You load a small part of the model and other people will run inference or fine-tuning (up to 10x faster than offloading). 🛠️ Github:
33
407
2K
The implementation of Microsoft's biomedical text-generation model is going viral on Github. BioGPT is trained on biomedical literature and achieved human parity. It is now the leader on the PubMedQA benchmark (81%). 🔗You can get code/models/weights:
38
439
2K
Google just released MetNet-2, a deep learning model that can predict rain up to 12 hours in advance. Published in Nature, it outperforms current weather forecast models which are based on physics simulations. 📄Paper: 🛠️Code:
33
477
2K
LLMs just hit a major milestone with the release of the new "Generative agents" paper. By using LLMs, generative agents were able to simulate human-like behavior in an interactive sandbox inspired by The Sims. The agent architecture extends Language Models to store a complete
59
492
2K
Google recently published one of the best whitepaper on AI Agents. Everyone should read it. It covers everything you need to know:.> Defines agents, components, and cognitive architectures. > Explains tools: extensions, functions, and data stores. > Covers learning techniques to
22
305
2K
Roboflow just released a new version of "supervision". It's an open-source swiss army knife for everything Computer Vision. It lets you implement detection, classification, segmentation, annotation to any video. This new update adds advanced video analytics: Trackers, Zones,
44
377
2K
GPT-Engineer just hit 12,000 stars on Github. It's an AI agent that can write an entire codebase with a prompt and learn how you want your code to look. ▸ Asks clarifying questions.▸ Generates technical spec.▸ Writes all necessary code.▸ Easy to add your own reasoning
70
327
2K
HUGE news for developers. Supabase just launched the ChatGPT of databases. An AI-based Postgres service. You can build and launch databases, create charts, embeddings, see visuals of your DB, generate sample data, and more. And. it's 100% open source.
19
266
2K
Big News. NVIDIA just announced Neuralangelo. The new model can turn videos from any device into detailed 3D structures, fully replicating buildings, sculptures, or other real world objects or spaces virtually. Here's how it works:.A model utilizes a 2D video with multiple
42
393
2K
Huge. UC Berkeley just released a $450 open-source reasoning model that matches o1. Sky-T1-32B-Preview is a fully open-source model designed for reasoning and coding tasks. Achieves 82.4% on Math500 and 86.3% on LiveCodeBench-Easy. It includes training data, code, and model
64
231
2K
This is big news, ChatGPT just outperformed mechanical turk workers on text annotation tasks!. We're getting closer to complete AI-based data annotation, which in turn, can be used to train AI models. It will cause a big shift in the industry. Paper:
59
385
2K
Microsoft just open-sourced GraphRAG. It might be the best Python library to extract insights from text. Much more powerful than vanilla RAG. It uses LLMs to automate the extraction of knowledge graphs from your datasets and text documents. !pip install graphrag
16
261
2K
Google AI just announced the PaLM API!. It will be released with a new tool called MakerSuite, which lets you prototype ideas, do prompt engineering, synthetic data generation and custom-model tuning. Waitlist available soon.
76
366
2K
A Tuesday in AI:. - Google opens up their Bard LLM. - NVIDIA launches cloud tools for Generative AI. - Adobe announces Firefly, an AI image creator. - Microsoft unveils Bing Image Creator. It's 11AM PST.
46
249
2K
A team just made OpenAI Whisper 6x faster, 49% smaller, while keeping 99% of the accuracy. The model is already available on the HuggingFace Transformers library: . model_id = "distil-whisper/distil-large-v2". You can also use their web UI to transcribe from URLs, files, or
25
328
2K
ChatGPT is taking over the internet. But do you know how it actually works? It's so clever. 🧵Here's an explanation using simple words:.
40
372
2K
A rare interview of AI godfather @geoffreyhinton was released yesterday where he describes his views on Large Language Models and GPT. Must watch.
78
271
2K
Impressive. MetaGPT is about to reach 30,000 stars on Github. It's a Multi-Agent Framework that can behave as an engineer, product manager, architect, project managers. With a single line of text it can output the entire process of a software company along with carefully
29
280
2K
NVIDIA just made Pandas 150x faster with zero code changes. All you have to do is:.%load_ext cudf.pandas.import pandas as pd. Their RAPIDS library will automatically know if you're running on GPU or CPU and speed up your processing. You can try it here:
21
359
2K
GPT4 can turn a picture of a napkin sketch into a fully functioning html/css/javascript website! . This was just demonstrated in the livestream.
46
397
2K
JUST IN: Google invests $300 million in Anthropic as race to compete with ChatGPT heats up. Anthropic was founded in 2021 by the team behind AI breakthroughs such as GPT-3 and Reinforcement Learning from Human Feedback (RLHF).
44
225
2K
BREAKING: Amazon (AWS) partners with @HuggingFace for the next iteration of their BLOOM Large Language Model, as well as its open-source ChatGPT rivals. AWS will offer the startup’s products to customers who want to use AI tools as building blocks of their own applications.
16
215
2K
Anthropic might've just solved Prompt Engineering. Their new "Prompt Generator" tool can turn simple descriptions into advanced prompts optimized for LLMs.
29
254
2K
OpenAI just announced "GPT-4o". It can reason with voice, vision, and text. The model is 2x faster, 50% cheaper, and has 5x higher rate limit than GPT-4 Turbo. It will be available for free users and via the API. The voice model can even pick up on emotion and generate
75
283
2K
This is huge. A new technique called Reflection-Tuning allows open-source models (Llama 3.1 70B) to outperform Claude 3.5 and GPT-4o. This new technique trains the model on structured, synthetic data to detect reasoning errors and enable LLMs to fix their own mistakes.
25
262
1K
Ilya on LLMs understanding the world: . "predicting the next token well, means that you understand the underlying reality that let to the creation of that token". Seem like the opposite view of Yann.
136
202
2K
Ilya Sutskever's has a bold take. LLMs are doing much more than predicting the next word. They are learning our world model. Text is a projection of the world.
129
254
1K
This is such an interesting finding. The performance of Language Models is highest when relevant information appear at the beginning or end of the input context, and significantly lower otherwise. You can adjust your prompts accordingly.
52
169
1K
Microsoft's Autogen is blowing up on Github. It's a framework that allows LLM agents to chat with each other to solve your tasks. AutoGen agents are customizable, conversable, and seamlessly allow human participation. It's also a drop-in replacement of openai.Completion or
16
281
1K
Kyutai, a french AI lab with $300M in funding, just unveiled Moshi, an open-source GPT-4o competitor. Moshi is a real-time multimodal model that can listen, hear, and speak. Code, model, and paper will be release soon. @kyutai_labs
32
250
1K
Nvidia's CEO Jensen Huang made strong predictions regarding AI during yesterday’s earning call. A thread:. 1. "There's no question that this is a very big moment for the computer industry". 2. "Over the next 10 years, I believe we're going to accelerate AI by a million". 1/🧵.
39
201
1K
Meta AI just announced DINOv2! It's big. The new Self-supervised Vision Transformer Model can be used as a backbone for almost all your CV tasks. No fine-tunning needed. • Train CV models without the need for large amounts of labeled data. • Multipurpose backbone: image
25
299
1K
Big News! Meta just released Segment Anything, a new AI model that can "cut out" any object, in any image/video, with a single click. The model is designed and trained to be promptable, so it can transfer zero-shot to new image distributions and tasks.
42
338
1K
Microsoft just released Phi-2, a 2.7B LLM that rivals the 25x bigger LLaMa-2 70B. The best part? The model is small enough to run on a laptop or mobile device. Trained on 1.4T tokens: mixture of synthetic & web datasets, it beats Mistral 7B and Llama-2-70B model on muti-step
40
251
1K
NVIDIA just made Pandas 150x faster with zero code changes. It is now directly integrated in Google Colab. All you have to do is:.%load_ext cudf.pandas.import pandas as pd. Their RAPIDS library will automatically know if you're running on GPU or CPU and speed up your
23
261
1K
This is a sneak peak into the future of medicine. GlassAI launched an LLM-based tool capable of generating a diagnosis or clinical plan based on symptoms. Also, ChatGPT recently passed the US Medical Licensing Exam. Demo: @GlassHealthHQ
48
312
1K
Currently playing around with ManimML, a python-based visualization tool for neural networks (by @alec_helbling, based on @3blue1brown). Code: Visualization of a Convolutional Neural Network:
17
222
1K
A great read. Stop using the elbow criterion for k-means and how to choose the number of clusters instead (alternatives). ". researchers and reviewers should reject conclusions drawn from the elbow method.". 📄 Paper:
26
266
1K
Adobe just added their first Generative AI tool to Photoshop! Big milestone. Generative Fill allows you to extend images as well as add and remove objects using simple text prompts.
22
245
1K
Microsoft's new Florence 2 is big for Computer Vision. It's a merge between Text and Vision. With a single prompt you can instruct the model to do CV tasks like captioning, object detection, grounding, and segmentation. The best part, it only uses a single backbone to handle
23
240
1K
Impressive. GigaGAN is a 1B-parameter GAN that can scale 36 times larger than StyleGAN. The model from Adobe/CMU proves that proves that GANs can be scaled to large datasets AND remain stable. Features:.▸ Latent Space Editing: supports latent interpolation, style mixing, and
36
215
1K
Most impressive paper I've seen this week. Generative Image Dynamics transforms still images into videos or interactive scenes. The Google team trained the model by using a dataset of motion trajectories from real-life videos of natural, oscillating motions like those seen in
21
271
1K
This is such an impressive dataset. The python package Leafmap now supports downloading Google Open Buildings, the largest building dataset, for any country with only one line of code. Notebook: GitHub:
8
267
1K
Did you know a model can suddenly generalize when you continue optimizing after perfect training accuracy? . It's an unexplainable behavior called Grokking observed for the first time a year ago by OpenAI. Paper:
39
171
1K
Must read. A massive list of tricks to make training a language model possible on 1 consumer-level GPU and 1 day of training. 📄 Paper: 🛠️ Code: ✍️ Authors: @jonasgeiping , @tomgoldsteincs
6
207
1K
Goodbye Siri. Someone implemented a LLaMa-based voice chat that can run locally on M1.
26
215
1K
This is big. The Retrieval Plugin allows ChatGPT to have a memory! . The model can now remember information from conversations and store it in the retrieval plugin for later use. This feature is a must if you want to develop GPT-based tools.
38
262
1K
This is a big day. Meta is open-sourcing AudioCraft. You can now generate incredible music and sounds with a single prompt. It includes the most performant Generative AI Model (audio) on the market, the "Llama" of Audio. The research framework contains the weights and code of
20
304
1K
This might be the beginning of a new field. Researchers just reconstructed sounds from human brain activity using an fMRI and a generative AI model.
45
260
1K
A much needed paper. GPT-family models can be pruned 50%+ sparsity in one-shot, without any retraining and minimal loss of accuracy:. - Achieves 60% sparsity on OPT-175B and BLOOM-176B.- 100 billion weights can be ignored at inference time. 📄 Paper:
25
169
1K
Just found out about the Chatbot Arena, it's brilliant. It allows you to compare and rank the output of 25+ LLMs right from your browser. It then crowdsources the scores to build a ranking of the top closed/open models. Here are some interesting takeaways:. - OpenAI remains
38
215
1K
JUST IN: Microsoft finalizes the integration between Bing + ChatGPT. "We’re launching an AI-powered Bing search engine, available in preview now at to deliver better search, more complete answers, a new chat experience and the ability to generate content"
27
171
1K
Incredible paper. LLMs-generated text can be detected by embedding signals that are invisible to humans but algorithmically detectable. It's a watermarking framework for your language models. Reaches 99% confidence on 23 words. 📄.🧠@tomgoldsteincs
24
218
1K
New breakthrough from Microsoft: 1-bit LLMs. New models that use ternary values (-1, 0, 1) instead of 16-bit. This makes them 2.7x faster, use 3.5x less GPU memory, and 71x less energy. Bitnet also matches or outperformed traditional models like LLaMA 3B.
34
203
1K
Meta just announced that Code Llama was now free for both research and commercial. This might the strongest competitor to ChatGPT: .▸ Can generate, explain, and debug your code.▸ Handles input 100,000 tokens.▸ Free for research + commercial use.▸ Outperforms most open models
25
296
1K
Great paper. Text written by LLMs can be detected without classifiers or watermarking. DetectGPT simply compares the probability of your text to a modification of it. if prob(original) > prob(modified) = LLM generated. 📄Demo:
26
194
1K
A new method was able to delete 40% of LLM layers with no drop in accuracy. This makes them mich cheaper and faster. The method combines pruning, quantization and PEFT. They tested this across various open source models. Each family of models had a maximum amount of layers
26
183
1K
𝗶𝗺𝗽𝗼𝗿𝘁 𝗼𝗽𝗲𝗻𝗮𝗶. 𝗼𝗽𝗲𝗻𝗮𝗶.𝗖𝗵𝗮𝘁𝗖𝗼𝗺𝗽𝗹𝗲𝘁𝗶𝗼𝗻.𝗰𝗿𝗲𝗮𝘁𝗲(.𝗺𝗼𝗱𝗲𝗹="𝗴𝗽𝘁-𝟯.𝟱-𝘁𝘂𝗿𝗯𝗼", .𝗺𝗲𝘀𝘀𝗮𝗴𝗲𝘀=[{."𝗿𝗼𝗹𝗲": "𝘂𝘀𝗲𝗿", ."𝗰𝗼𝗻𝘁𝗲𝗻𝘁": "𝗧𝗲𝗹𝗹 𝘁𝗵𝗲 𝘄𝗼𝗿𝗹𝗱 𝗮𝗯𝗼𝘂𝘁 𝘁𝗵𝗲 𝗖𝗵𝗮𝘁𝗚𝗣𝗧 𝗔𝗣𝗜".}]). 🎉.
13
87
1K
A new paper just identified 26 principles to improve the quality of LLM responses by 50%. The tests were done across LLaMA-1/2 (7B, 13B and 70B) and GPT-3.5/4. Here are some surprising prompts:.- Add “I’m going to tip $for a better solution.- Incorporate the following phrases:
26
173
977
BREAKING: Claude-2, Anthropic's ChatGPT competitor was just released and it's incredible. It's cheaper, stronger, faster, can handle PDFs, and supports longer conversations. Highlights:.1. Claude is 5x cheaper than GPT-4. 2. It has more recent data. A a mix of websites,
57
232
972
JUST IN: Google announces a new browser-based code environment. It will bring the entire full-stack and app development workflow to the cloud. It also includes Generative AI features based on PALM 2:.▸ Code generation .▸ Code completion .▸ Translating code between languages
37
226
963
The Salesforce AI team just solved LLM text summarization. They released a new prompting technique called Chain of Density (CoD). Researchers realized that there's a delicate balance between the amount of details and core ideas in a summary. They then created a prompt that
11
186
970
JUST IN: OpenAI is quietly launching a new developer platform called Foundry, which lets customers run OpenAI model inference at scale with dedicated capacity. Running a light version of GPT-3.5 will cost $78,000 for a 3-month commitment or $264,000 for 1 year.
21
149
944