Your curated collection of saved posts and media
CEO of Microsoft getting ratioed by an anime pfp account... https://t.co/4KY9Vz5MDM
SD 3.5 Large from @StabilityAI is now available! The @Gradio app is live now on @huggingface Spaces space: https://t.co/xleFJm47kj https://t.co/uvTiV1kyLQ
Guess who's back? Back again! π΅ @StabilityAI is back, tell a friend π€ Stable Diffusion 3.5 Large is here π₯ - ποΈ 8B parameters - Full πͺ and ποΈπ¨ 4-step Turbo variant - π§Ύ π€ commercial use (for orgs below 1M year/rev) - 𧨠day-0 LoRA fine-tuning support https://t.co/6RLQA31a6H
Successfully got Claude to order me lunch all by himself! Notes after 8 hours of using the new model: β’ Anthropic really does not want you to do this - anything involving logging into accounts and especially making purchases is RLHF'd away more intensely than usual. In fact my agents worked better on the previous model (not because the model was better, but because it cared much less when I wanted it to purchase items). I'm likely the first non-Anthropic employee to have had Sonnet-3.5 (new) autonomously purchase me food due to the difficulty. These posttraining changes have many interesting effects on the model in other areas. β’ If you use their demo repository you will hit rate limits very quickly. Even on a tier 2 or 3 API account I'd hit >2.5M tokens in ~15 minutes of agent usage. This is primarily due to a large amount of images in the context window. β’ Anthropic's demo worked instantly for me (which is impressive!), but re-implementing proper tool usage independently is cumbersome and there's few examples and only one (longer) page of documentation. β’ I don't think Anthropic intends for this to actually be used yet. The likely reasons for the release are a combination of competitive factors, financial factors, red-teaming factors, and a few others. β’ Although the restrictions can be frustrating, one has to keep in mind the scale that these companies operate at to garner sympathy; If they release a web agent that just does things it could easily delete all of your files, charge thousands to your credit card, tweet your passwords, etc. β’ A litigious milieu is the enemy of personal autonomy and freedom.
What a repo / release SOTA text2speech, MIT license Paper, Model Card, NotebookLM podcast Inference code in Keras, so jax, torch, tf ready Uses uv for installation https://t.co/E1QfRyvnpV
Online research surveys may be a dying methodology as techniques for verifying respondents as humans are failing: βAutomatic AI detection systems are currently completely unusableβ¦ Individual attention checks will no longer be a sufficient tool to ensure good data quality.β https://t.co/i6PtKgA4Uu
Generative Reward Models Presents GenRM, an iterative algorithm that trains an LLM on self-generated reasoning traces, leading to synthetic preference labels matching human preference judgments https://t.co/f7oS828PwS https://t.co/gnchZQZ8L7
Every so often, math offends our sensibilities. https://t.co/qXGIsHqOZN
This is brilliant. Thermodynamic modelling to optimise the shape of a beer glass so as to keep it cold (the author's Brazilian) for the longest time. Sadly, the resulting shape is absolutely not one I'd want to drink beer from. https://t.co/7jKwkhGDcl
Learn how to leverage LLMs for efficient report generation! ππ Report generation is a major use-case for our users, so we built a demo using Arxiv data. The key principles to get here are: β‘οΈ Techniques for extracting key information from complex documents, such as Pydantic data structures β‘οΈ Using Workflows to incorporate combinations of RAG and other extraction techniques β‘οΈ Using templates to show the LLM how to structure and present data Check out the full example in this notebook: https://t.co/0Pxvgj2kru
I wish people would stop repeating these as if they are facts that AI is plateauing. AI might hit a roadblock, we donβt know, but every one of these issues has multiple studies stating the opposite: synthetic data works, scaling is fine, etc. We need more nuance on the AI future https://t.co/Vusi7p4HXb
Unfortunately this prompt strategy does not work https://t.co/dBaFZvKXIl
IBM presents Granite 3.0, lightweight foundation models ranging from 400 million to 8B parameters. Supports coding, RAG, reasoning, and function calling, focusing on enterprise use cases, including on-premise and on-device settings. The technical report contains detailed discussions on how they collect synthetic datasets for code, reasoning, RAG, tool use, and more. "Granite 3.0 language models demonstrate strong performance across a battery of academic benchmarks for language understanding, reasoning, coding, function calling, and safety" The release includes pre-trained and post-trained versions of all Granite 3.0 models under a permissive Apache 2.0 license. Very strong release by the Granite Team at IBM.
Agentic Information Retrieval This paper provides a good introduction to agentic information retrieval, which is shaped by the capabilities of LLM agents. I've been developing with this paradigm recently and it does offer lots of interesting ways to optimize retrieval systems. https://t.co/knY20adgXW
"What Matters In Transformers?" is an interesting paper (https://t.co/2O6TxZK5Mx) that finds you can actually remove half of the attention layers in LLMs like Llama without noticeably reducing modeling performance. The concept is relatively simple. The authors delete attention layers, MLP layers, or entire transformer blocks: - Removing entire transformer blocks leads to significant performance degradation. - Removing MLP layers results in significant performance degradation. - Removing attention layers causes almost no performance degradation! In Llama 2 70B, even if half of the attention layers are deleted (which results in a 48% speed-up), there's only a 2.4% decrease in the model benchmarks. The author also recently added Llama 3 results to the paper, which are similar. The attention layers were not removed randomly but based on a cosine-based similarity score: If the input and output are very similar, the layer is redundant and can be removed. This is a super intriguing result and could potentially be combined with various model compression techniques (like pruning and quantization) for compounding effects. Furthermore, the layers are removed in a one-shot fashion (versus iterative fashion), and no (re)training is required after the removal. However, retraining the model after the removal could potentially even recover some of the lost performance. Overall, a very simple but very interesting study. It appears there might be lots of computational redundancy in larger architectures. One big caveat of this study, though, is that the focus is mostly on academic benchmarks (HellaSwag, MMLU, etc.). It's unclear how well the models perform on benchmarks measuring conversational performance.
The Case for Muon 1) We can descend 'faster' in non-Euclidean spaces 2) Adam/Shampoo/SOAP/etc. dynamically learn the preconditioner and, equivalently, the norm & space to descend in 3) Muon saves a lot of compute by simply letting the norm to vary within a fixed range https://t.co/PKpXrKSYpT
Perplexity for fact checks in action https://t.co/0toYAIuVAQ
Setup an Advanced Multimodal RAG Pipeline in Minutes β‘οΈ If you have a lot of complex documents (slide decks, product manuals), @fahdmirzaβs tutorial shows you how to index this data in minutes and it βjust worksβ. This frees up your time to build the agentic/application logic instead of getting stuck in RAG technical complexity. Itβs a great introduction to LlamaCloud, our easy-to-use, accurate, secure platform for your βRAG over a million PDFβ needs. Video: https://t.co/wa6f2e0VJv Signup: https://t.co/rzaFHLeFH1 If youβre interested in using it in an enterprise setting, come talk to us: https://t.co/ek65cohGkX
i transferred @chichengcc's state based simplified diffusion policy notebook for both Rectified Flows and Shortcut Models by @kvfrans. both work quite well, and shortcut can solve the push-T task in 1-2 diffusion steps (video below)! here are some of my takeaways: 𧡠https://t.co/W314GkS79m
DeepMind just trained a 270M transformer that can play like a grandmaster without searching through moves (MCTS), It was trained on Stockfishβs assessment of ~10M games so it didnβt really learn the game from scratch. Impressive that transformers generalize to a βlogicβ task https://t.co/l27RyTmsIt
Meta AI's hidden prompt! Quite interesting to see it uses known facts and information to make the response more personalised "If the user provides information that contradicts their data, prioritize the information that the user has provided in the conversation. Do NOT address or highlight any discrepancies between the data and the information they provided."
@fchollet's really nice presentation about why's and how's of @arcprize , a lot of new idea-directions to think about and a grounding framework to think about Abstractions, Intelligence and eventually AGI https://t.co/UnkiwRLRXS
Case Study: An Agentic System for Automated Sales Outreach π§βπΌ Research prospects and manually crafting personalized emails takes away a tremendous amount of time from the BD team. This blog by @Calsoft_Data is a great reference guide of a constrained agentic architecture that can progress from prospect research to email generation, with human-in-the-loop: 1. Ingest prospect data and enrich it 2. Query knowledge base with @llama_index to look up information that aligns with prospect needs 3. Generate emails 4. Send it Blog: https://t.co/IO8etXwKOw
Lightning Fast Advanced RAG Workflow with @GroqInc β‘οΈ This is a fantastic tutorial by Plaban Nayak on setting up a fully async RAG workflow that has lightning fast reranking + synthesis, thanks to @GroqInc. By using @llama_index workflows, you get the benefits of fully async orchestration that can handle multiple user requests, along with step-by-step debuggability + visualizations. Check it out: https://t.co/XFSmYMxJld For more details on workflows: https://t.co/EbUH0QJzLl
my article is published now π₯³ I covered my llm resources and some tips I learned along the way. I hope this helps those who are afraid to take the first step or are overwhelmed by what resources to study. this is for you ;) https://t.co/43ZhKBJ6Nq
Alright actual serious post. Lingua := super simple codebase + torch.compile for speed --> clean, hackable, but still efficient *It can train a 7B >llama2 in 24h*. Crazy. If you got the gpus, not only can you train a good 7B, you can *iterate* on it. You can do *research* https://t.co/QKC66QqHaN
π¨ RELEASE ALERT βΌοΈ https://t.co/t1A5llaLOD THIS CHANGES EVERYTHING $META just dropped a game-changing codebase! Now everyone can do LLM research! π± π§΅10 best things people are already building with lingua π₯π
launching our open source OCR tool today! try it out with some terrible pdfs and let me know how it goes: https://t.co/G18jbunPqo https://t.co/TOeQtV6sPG
Round and Round we Go! π Rotary Positional Encodings (RoPE) are a common staple of frontier LLMs. _Why_ do they work so well, and _how_ do LLMs make advantage of them? The results might surprise you, as they challenge commonly-held wisdom! Read on β©οΈ Work led by @fedzbar! https://t.co/C61UvK5zOb
π The Chat Completions API supports audio now. Pass text or audio inputs, then receive responses in text, audio, or both. https://t.co/468QclBSBU https://t.co/uUFrJa9kZH
Microsoft is living with the possibility of OpenAI cutting them off at any time of their choosing, as well as not being in control of a technology which is becoming increasingly important, and they are feeling it. Microsoft is trying to build their own SOTA model internally, and there is every reason to expect they will succeed - they have the GPUs, money, desire, paranoia and talent required to do it
ITS HAPPENING https://t.co/GzLj3NXspp
Microsoft is living with the possibility of OpenAI cutting them off at any time of their choosing, as well as not being in control of a technology which is becoming increasingly important, and they are feeling it. Microsoft is trying to build their own SOTA model internally, and
How does quantization impact the performance of LLMs? Only minimal! π€―Β @neuralmagic ran 500,000 different evaluations on @AIatMeta Llama using different quantization strategies. The impact is <1%, but the benefits are up to 2.4 faster inference and 3.5 model size reduction! π₯ TL;DR; π― Quantized models achieve 99% accuracy recovery compared to full-precision π Up to 2.4x speedup and 3.5x model size reduction with quantization. π Tested Llama 3.1 8B, 70B, and 405B models on OpenLLM Leaderboard, ArenaHard, HumanEval, and text similarity metrics. π₯W8A8-FP8 dynamic yields the best results π€Β Quantized models available on @huggingface.
"Claude, create an Excel formula that would delight me. seriously make me giggle with joy" Claude: "How about a formula that turns boring numbers into a little ASCII art cat? .... In cell A1, enter a number. This number will determine the "chonkiness" of the cat." https://t.co/Sl63zAzYT5
A few weeks ago we introduced LlamaParse Premium, our best parsing technology yet, and users have been loving it. In this in-depth LinkedIn post Hanane puts Premium through its paces and comes away impressed: https://t.co/aUWDONV1Kr Check out the original post introducing Premium here: https://t.co/dL32VlCTgq
