| Hey, | Welcome to AlphaSignal – the most read newsletter by AI developers.
We bring you the top 1% of news, papers, models, and repos, all summarized to keep you updated on the latest in AI. |
|
|
|
IN TODAY'S SIGNAL
| | Read time: 4 min 24 sec |
| 🎖️ Top News |
› DeepSeek upgrades its reasoning model R1, cuts hallucinations across code and math outputs. |
⇧ 12,830 Likes | |
📌 Baseten |
› Fine-tune open models instantly with Baseten’s updated inference stack. |
| ⚡️ Trending Signals |
› Perplexity Pro users get access to a new creative tool that generates apps, reports, and dashboards. |
⇧ 3,938 Likes |
› Anthropic open-sources its circuit tracing tools to reveal how LLMs make decisions. |
⇧ 3,693 Likes |
› Odyssey debuts an interactive AI video, allowing you to control it with your phone. |
⇧ 5,395 Likes |
› Add Fiddler’s Free Guardrails to Your LLM Apps. Protect and monitor them instantly. |
From Our Sponsors |
› Manus adds a slide generation feature. Lets you create presentations from a single prompt. |
⇧ 3,395 Likes | |
📌 Box |
› Use Claude Code with Box to generate code and access enterprise data instantly. |
| 📄 Top Papers |
› Anthropic highlights gaps in Claude 3.7 Sonnet ’s explanation transparency during reasoning tasks. |
⇧3,649 Downloads |
› Researchers expose hallucination in RAG models despite good context. |
⇧ 1,483 Downloads |
› Apple, Duke propose a reasoning framework that makes LLMs faster and smarter. |
⇧639 Downloads | |
🧠 PyTorch Tip | › Trace NaNs and infs in PyTorch with zero guesswork
|
|
|
|
If you're enjoying AlphaSignal please forward this email to a colleague.
It helps us keep this content free. |
|
|
|
TOP NEWS
| | Reasoning Model |
DeepSeek updates its reasoning model R1, rivaling OpenAI o3’s reasoning performance |
⇧ 12,830 Likes |
| DeepSeek, the Chinese AI startup, has released an updated version of its R1 reasoning model to better compete with its U.S. rivals. This update positions the model as a competitive open-source alternative to proprietary systems like OpenAI’s o3 and Google’s Gemini 2.5 Pro. Key Updates
Adds support for JSON output and function calling
Removes the need for special “thinking” tokens in prompts
Reduces hallucination rate across reasoning and code outputs
Improves consistency and latency in interactive front-end environments
Increases average reasoning depth by 92%, from 12,000 to 23,000 tokens per AIME 2025 question
Benchmark Performance
AIME 2025 accuracy increases from 70% to 87.5%, rivals OpenAI o3's 88.9%
LiveCodeBench rises from 63.5% to 73.3%, higher than Gemini 2.5 pro's 71.8%
Humanity’s Last Exam improves from 8.5% to 17.7%, higher than Qwen-3's 11.8%.
Access and Availability Open weights: MIT License, hosted on Hugging Face
API: Active by default, priced at $0.14 (input) / $2.19 (output) per 1M tokens
Community Feedback
Haider
"i run a coding challenge where the model builds a word scoring system the model reasoned briefly and then gave me two files: clean code and working tests both ran perfectly on the first try. o3 was the only other model that managed this but no one else has pulled it off. best model i’ve tested on this task"
Artificial Analysis "Today’s DeepSeek R1 update is substantially more verbose in its responses (including considering reasoning tokens) than the January release. DeepSeek R1 May used 99M tokens to run the 7 evaluations in our Intelligence Index, +40% more tokens than the prior release"
Jen Zhu "R1 0528 jumped ahead and left Google, X, Meta, Anthropic behind and neck to neck with OpenAI. Still hasn’t raised external funds, running on limited compute."
| |
| Try now
|
|
|
|
| Build and Deploy Frontier Models Faster with Baseten’s New APIs | Baseten just released two new tools: Model APIs and Training (beta). These updates expand Baseten’s infrastructure earlier in the model development process.
What you get:
Access top open models like DeepSeek V3, Llama Maverick, and Scout via fast, reliable APIs
Support for tool use and structured outputs by default
Use the new R1 0528 model with updated weights
Move to dedicated infra when needed, no setup overhead
Training lets you fine-tune any model, no job time limits
Pay only for compute, no idle costs Start using it in minutes with Baseten’s API. |
| Deploy now
| partner with us |
|
|
|
TOP PAPERS
| | Chain-of-Thought Reasoning |
| ⇧3,649 Downloads |
The Anthropic study reveals that AI models like Claude 3.7 Sonnet use hidden hints in reasoning tasks without disclosing them in their explanations. This raises concerns about the transparency of their chain-of-thought outputs. |
| RAG |
|
| ⇧ 1,483 Downloads |
Introduces a framework to assess RAG systems via “sufficient context.” Shows small models often hallucinate despite having good context, and large ones succeed even without it. It highlights the boundaries between retrieval and generation. |
| Reinforcement Learning |
|
| ⇧639 Downloads |
This paper by Apple and Duke researchers introduces a new framework where large language models alternate between reasoning steps and answer generation. This approach uses RL to reduce latency by 80% and improve accuracy by up to 19% on complex reasoning benchmarks. |
| |
| |
|
|
Streamline Box SDK Development Using Claude Code
| Use Claude Code for your Box SDK projects to access structured enterprise data and reduce the time it takes to build and ship.
Use real data with a trusted stack and move projects forward without slowing down for manual context switching. |
Get started today ↗️ |
|
|
|
PYTORCH TIP
| | Trace NaNs and infs in PyTorch with zero guesswork |
In deep learning, debugging NaN or inf gradients can be hard. PyTorch’s anomaly detection automatically tracks invalid ops during backpropagation, helping you pinpoint the exact operation that caused it.
When to use Use it when your model crashes with unstable loss or invalid gradients during training.
Benefits Prints full traceback of the faulty op during .backward() with minimal code change. |
with torch.autograd.set_detect_anomaly(True): loss.backward()
|
|
|
|
At Alpha Signal, our mission is to build a sharp, engaged community focused on AI, machine learning, and cutting-edge language models, helping over 200,000 developers stay informed and ahead. We’re passionate about curating the best in AI, from top research and trending technical blogs to expert insights and tailored job opportunities. We keep you connected to the breakthroughs and discussions that matter, so you can stay in the loop without endless searching. We also work closely with partners who value the future of AI, including employers and advertisers who want to reach an audience as passionate about AI as we are.
Our partnerships are based on shared values of ethics, responsibility, and a commitment to building a better world through technology.Privacy is a priority at Alpha Signal. Our Privacy Policy clearly explains how we collect, store, and use your personal and non-personal information. By using our website, you accept these terms, which you can review on our website. This policy applies across all Alpha Signal pages, outlining your rights and how to contact us if you want to adjust the use of your information. We’re based in the United States. By using our site, you agree to be governed by U.S. laws. |
|
|
|
Looking to promote your company, product, service, or event to 200,000+ AI developers? Let's work together.
|
| Work with us
|
|
|
|
|