| 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 25 sec | 🎖️ Top News
📌 Oxylabs
⚡️ Trending Signals
📌 Neo4j
🧠 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
| Multimodal | Deepseek releases Janus, an Any-to-Any autoregressive framework for multimodal AI |
⇧ 1,275 Likes |
 |
What's New |
DeepSeek has introduced Janus, an autoregressive framework that unifies multimodal understanding and generation. Janus addresses limitations of previous models by decoupling visual encoding into separate pathways while using a unified transformer for processing.
Key highlights Janus 1.3B is a multi-modal LM for any-to-any tasks. It beats DALL-E 2/SDXL in generation and LLaVA 1.5 7B in multimodal capability.
MIT licensed, Commercial usage is permitted under these terms.
Achieves 69.4 on MMBench, 63.7 on SEED-Bench, 87.0 on POPE.
FID score of 8.53 on MSCOCO-30K, 61% accuracy on GenEval.
Model available in HuggingFace and integrated with Tansformers library.
Performance Metrics
Strong performance is demonstrated by the 1.3B parameter model, outperforming models with 7B parameters. MMBench: 69.4 (outperforms LLaVA-v1.5 7B: 67.9)
SEED-Bench: 63.7 (outperforms LLaVA-v1.5 7B: 62.4)
POPE: 87.0 (outperforms LLaVA-v1.5 7B: 85.5)
MSCOCO-30K: FID score of 8.53 (outperforms DALL-E 2: 9.0)
GenEval: Accuracy of 61% (outperforms SDXL: 58%)
Three-stage training procedure Janus employs distinct encoders for understanding and generation tasks.
The model's training follows a three-stage procedure: adaptor and image head training, unified pretraining, and supervised fine-tuning.
This approach allows Janus to balance performance between understanding and generation tasks effectively.
Model Architecture
Janus uses an autoregressive transformer architecture with separate visual encoders for understanding and generation.
For multimodal understanding, it uses a SigLIP encoder to extract high-dimensional semantic features.
Visual generation relies on a VQ tokenizer to convert images into discrete IDs. Both pathways feed into a shared transformer for autoregressive processing.
| |
|
Read more
|
|
|
|
| Build scrapers like a pro with OxyCopilot by Oxylabs | Developers with limited coding skills can now use Oxylabs' AI-driven OxyCopilot.
This tool builds scalable web scrapers and parsers, leveraging Oxylabs' scraping infrastructure.
OxyCopilot features AI-powered parsing of complex HTML into structured data
Quick adaptation to website layout changes
Precise extraction of nested and listed data
Reusable JSON-formatted parsing templates
OxyCopilot streamlines scraping tasks in no time and with little effort, allowing developers to focus on core business logic. |
|
Try for free
| partner with us |
|
|
|
TRENDING SIGNALS
| PyTorch |
|
⇧ 2,330 Likes | |
Google |
|
⇧ 821 Likes | |
Transformers |
| ⇧ 638 Likes
| | SQL |
|
⇧ 3,896 Likes | |
Meta |
|
⇧ 1,888 Likes | | |
|
|
|
|
Join NODES: Free conference on knowledge graphs + AI
| Neo4j's 24-hour virtual developer conference starts November 7.
Explore 140+ sessions on knowledge graphs, AI, data science, and intelligent apps.
Learn from experts, build skills, and discover real-world use cases. |
Register now [FREE] |
|
|
|
|
TOP PAPERS
| LLM |
| ⇧ 1,670 Likes |
Problem LLMs lack explicit thinking abilities for complex reasoning tasks. Existing methods like chain-of-thought are limited to math and logic problems and require human-written examples.
Solution Thought Preference Optimization (TPO) trains LLMs to generate internal thoughts before responding, without additional human data. It uses iterative preference optimization to explore and improve thought generation based on response quality judged by a separate model.
Results TPO outperforms direct response baselines on general instruction following benchmarks AlpacaEval (52.5% vs 48.4% win rate) and Arena-Hard (37.3% vs 33.0%). Gains are seen across diverse categories beyond traditional reasoning tasks, including creative writing, marketing, and general knowledge. |
| AI Agents |
|
⇧ 885 Likes |
Problem Current evaluation methods for agentic AI systems are inadequate, focusing only on final outcomes or requiring excessive manual labor. This fails to capture the step-by-step nature of agentic systems and slows progress.
Solution The authors introduce the Agent-as-a-Judge framework, where agentic systems evaluate other agentic systems. This extends LLM-as-a-Judge to provide rich, intermediate feedback on the entire task-solving process. They also present DevAI, a new benchmark of 55 realistic AI development tasks with 365 hierarchical requirements.
Results Agent-as-a-Judge outperformed LLM-as-a-Judge, aligning 90% with human consensus vs 70% for LLM-as-a-Judge. It performed comparably to human evaluators while saving 97% of time and cost. On DevAI, top agentic systems satisfied ~29% of task requirements but fully solved only one task. | | Diffusion Transformers |
|
⇧ 783 Likes |
Problem The paper identifies the absence of explicit scaling laws for Diffusion Transformers (DiT) in text-to-image generation, which results in inefficient resource allocation. Current models lack precise predictions for optimal model size and data requirements, leading to costly heuristic searches.
Solution The authors establish scaling laws for DiT by conducting experiments across a wide range of compute budgets. They derive a power-law relationship between compute, model size, data, and training loss, allowing predictions of optimal configurations. A 1B-parameter model trained with 1.5e21 FLOPs validates these scaling laws.
Results The scaling laws accurately predict training loss and synthesis quality, with a close match between predicted and actual losses for a 1B-parameter model. The relationship between compute budget and FID follows a power-law trend, demonstrating effective performance prediction across datasets. |
| |
|
|
|
|
|
PYTHON TIP
| Boost ML performance with partial function application and result memoization | Optimize your Python function calls for ML tasks with functools.partial and functools.lru_cache.
These tools help you reduce redundant computations and improve performance in data-heavy operations. functools.partial lets you create new functions with pre-set arguments.
functools.lru_cache memoizes function results, storing them for quick retrieval on subsequent calls with the same arguments. This speeds up expensive computations.
Here's a practical example:
|
from functools import partial, lru_cache import numpy as np
@lru_cache(maxsize=128) def expensive_computation(x, y): return np.sum(np.power(np.arange(x), y))
partial_compute = partial( expensive_computation, y=2)
# Usage result1 = partial_compute(1000000) result2 = partial_compute(1000000) #Cached result
print(f"Result 1: ") print(f"Result 2: ") # Result 1: 333332833333500000 # Result 2: 333332833333500000
| This code demonstrates how partial pre-sets the 'y' argument, while lru_cache stores results for repeated calls.
Without partial and lru_cache
1. You'd need to specify y=2 every time you call the function. 2. The expensive computation would run from scratch on each call, even with identical inputs.
This results in more verbose code and slower execution.
So the benefit of using partial and lru_cache is significant speedup for repeated expensive computations on large arrays. |
|
Run code on colab
|
|
|
|
Looking to promote your company, product, service, or event to 200,000+ AI developers? Let's work together.
|
|
Advertise
|
|
|
|
|