| Hey, | Welcome to today's edition of AlphaSignal, a newsletter for developers by developers.
We identify and summarize the top 1% news, papers, models, and repos in the AI industry. |
|
|
|
IN TODAY'S SIGNAL
| Read time: 6 min 34 sec | 🎖️ Top News
📌 Oxylab
⚡️ Trending Signals
📌 Tilores
🧠 Top Papers
🧠 Python Tip
|
|
|
|
If you're enjoying AlphaSignal please forward this email to a colleague.
It helps us keep this content free. |
|
|
|
TOP NEWS
| Google | Google Releases Gemini 1.5 Flash-8B: A Faster, Cheaper AI Model for High-Volume Tasks |
⇧ 2,911 Likes |
 |
What's New | Google has launched Gemini 1.5 Flash-8B, a production-ready variant of its lightweight language model.
This release marks a significant advancement in efficient AI, designed for high-volume, multimodal applications and long context summarization tasks.
Key highlights 50% lower price compared to 1.5 Flash
2x higher rate limits (4,000 requests per minute)
Lower latency on small prompts
$0.01 per million tokens for cached prompts <128K
Available through Google AI Studio and Gemini API
Gemini 1.5 Flash-8B builds on the original 1.5 Flash model released in May 2024.
Performance Outperforms 1.5 Flash (May 2024) in math tasks, scoring 58.7% vs 54.9% on MATH benchmark
Significant improvement in HiddenMath, scoring 32.8% compared to 20.3% for 1.5 Flash (May 2024)
Better performance in audio tasks, with 13.6% on FLEURS compared to 9.8% for 1.5 Flash (May 2024)
Improved safety performance, scoring 92.6% on XSTest versus 86.9% for 1.5 Flash (May 2024)
Maintains comparable performance to 1.5 Flash (May 2024) in most other benchmarks, with slight differences in various categories Pricing structure
The pricing structure positions Gemini 1.5 Flash-8B as the most cost-effective model in the Gemini lineup.
Input tokens cost $0.0375 per million, while output tokens are priced at $0.15 per million for prompts under 128K tokens.
This pricing applies to the paid tier, with billing set to begin on October 14, 2024.
The increased rate limits further support its use in high-throughput scenarios.
The emphasis on multimodal capabilities implies the model can process various input types, though exact details on modalities are not specified.
the announcement doesn't specify the maximum context length. | |
|
Read more
|
|
|
|
| Introducing OxyCopilot: The first AI-powered assistant for scalable web scraping | Oxylabs just released OxyCopilot, an AI-powered assistant for web scraping. It handles diverse website structures without manual coding.
OxyCopilot generates scraping requests and parsing instructions, identifying complex patterns using advanced language models.
Developers can access OxyCopilot through Oxylabs' Scraper API Playground. Input a prompt and receive ready-to-use code in seconds.
This tool aims to enhance data collection efficiency for various projects. It's designed to streamline the scraping process, potentially saving time on manual code writing. |
|
Learn more
| partner with us |
|
|
|
TRENDING SIGNALS
| OpenAI |
|
⇧ 19,714 Likes | |
Meta |
|
⇧ 4,458 Likes | |
Medical AI |
|
⇧ 950 Likes | |
Andrej Karpathy |
|
⇧ 7,300 Likes | |
Image Generation |
|
⇧ 1,100 Likes | | |
|
|
|
|
|
Provide unified customer data to your LLMs
| Tilores just released a LangChain integration, enabling developers to use their identity resolution technology as an "identity RAG" source for customer data in LLM applications.
This new feature is particularly valuable for regulated companies dealing with scattered customer data across various silos, where precise retrieval is crucial. |
Try it now ↗️ |
|
|
|
| TOP PAPERS |
RNN | |
⇧ 2,489 Likes | Problem Transformers have scalability limitations for long sequences. Recent recurrent models like S4, Mamba, and Aaren achieve comparable performance but are complex.
Solution The authors revisit traditional LSTMs and GRUs, simplifying them by removing hidden state dependencies from gates. This results in minimal versions (minLSTMs and minGRUs) that are fully parallelizable during training and use fewer parameters.
Results
minLSTMs/GRUs are 175x faster to train than traditional LSTMs/GRUs for sequences of length 512. They match performance of recent sequence models on tasks like selective copying, reinforcement learning, and language modeling. minGRU outperforms Decision S4 and matches Decision Transformer and Decision Mamba on D4RL benchmark. |
| LLM |
|
| ⇧ 771 Likes |
Problem LLMs struggle with complex logical reasoning tasks despite their broad capabilities. Existing prompting methods like Chain-of-Thought can improve reasoning but suffer from unfaithful conclusions. Some neuro-symbolic approaches aim to address this but face information loss when extracting logical expressions.
Solution The paper proposes Logic-of-Thought (LoT) prompting. LoT extracts propositions and logical expressions from input, expands them using logical reasoning laws, translates expanded expressions back to natural language, and uses this as additional augmentation to input prompts. This preserves original context while adding logical guidance.
Results LoT boosts performance across multiple reasoning tasks and prompting methods. It improves Chain-of-Thought accuracy on ReClor by +4.35%, Self-Consistency on ReClor by +6.52%, Chain-of-Thought with Self-Consistency on LogiQA by +5%, and Tree-of-Thoughts on ProofWriter by +8%. | | RLHF |
| ⇧ 731 Likes
| Problem LLMs face challenges in multi-task learning due to reward hacking and conflicting objectives. Current RLHF methods struggle to balance multiple tasks effectively. Reward models have limitations in fine-grained criteria alignment and are prone to misspecification.
Solution The paper introduces Constrained Generative Policy Optimization (CGPO), a novel RLHF framework. It uses a Mixture of Judges (MoJ) with rule-based and LLM-based constraints to mitigate reward hacking. CGPO applies task-specific optimization strategies, including customized reward models, MoJs, and hyperparameters. Three new constrained RLHF optimizers are proposed: CRPG, CODPO, and CRRAFT.
Results CGPO consistently outperforms PPO and DPO baselines across various benchmarks. It improves performance by 7.4% in AlpacaEval-2, 12.5% in Arena-Hard, 2% in IFEval, 2% in MATH and GSM8K, 5% in HumanEval, and 2% in ARC challenge. CGPO effectively addresses reward hacking issues observed in PPO, particularly in coding benchmarks. |
| |
| |
|
|
PYTHON TIP
| Create your own audiobooks from PDFs with Python |
You can transform any PDF into an audiobook using just a few lines of Python code. This technique helps you convert technical papers, research articles, or even entire books into audio format for on-the-go learning.
Here's how to do it Extract text from PDF
Convert extracted text to speech
First, install the required libraries:
pip install pdfplumber gTTS |
import pdfplumber as pp from gtts import gTTS
pdf_text = '' with pp.open('attention_is_all_you_need.pdf') as pdf: for page in pdf.pages: pdf_text += page.extract_text()
tts = gTTS(text=pdf_text, lang='en') tts.save('audio_book.mp3')
| The pdfplumber library is for plumbing a PDF for detailed information about each char, rectangle, line, et cetera — and easily extract text and tables.
And the gTTS is a python library and CLI tool to interface with Google Translate's text-to-speech API.
This script reads all pages of your PDF, extracts the text, and converts it to an MP3 file using Google's Text-to-Speech engine.
You can customize the language, voice speed, and output format. The gTTS library supports multiple languages and accents.
This method works well for most PDFs, but complex layouts or scanned documents might require additional preprocessing. |
|
Code in github
|
|
|
|
Looking to promote your company, product, service, or event to 200,000+ AI developers? Let's work together.
|
|
Advertise
|
|
|
|
|