Yipeng Sun

Introducing Aurora, a new optimizer for training frontier-scale models. We train Aurora-1.1B, which achieves 100x data efficiency on open-source internet data. Despite having 25% fewer parameters, 2 orders of magnitude fewer training tokens, and using fully open-source internet-only data, Aurora matches Qwen3-1.7B on several benchmarks. Aurora was developed after identifying a major failure mode that can occur under Muon, an increasingly popular optimizer that has shown strong gains over Adam(W). We find that Muon can cause a huge percentage of neurons to effectively die early in training, reducing effective network capacity so that many parameters no longer meaningfully contribute to network outputs. By redistributing update energy more uniformly across neurons while preserving Muon’s stability properties, Aurora prevents neuron death and recovers substantial model capacity. What makes this work especially exciting is that it points toward a broader direction for ML research: better optimizers may not come purely from elegant mathematical abstractions, but from understanding and addressing the concrete dynamics and pathologies that emerge inside real training systems.

New Anthropic research: Natural Language Autoencoders. Models like Claude talk in words but think in numbers. The numbers—called activations—encode Claude’s thoughts, but not in a language we can read. Here, we train Claude to translate its activations into human-readable text.

Such a great evening to start a brand new research for NeurIPS in 3.5 days.🧘‍♂️ Day 1: planning. Night 1: running experiments and sending the abstract. Day 2: reading results fighting with Claude, and sending again. Night 2: sleep (optional). Day 3: opening Codex, and finally, write the pape in parallel. Night 3: resolving the “beef” with Claude (temporary peace) and going to sleep. Day 4: final reading, last-minute fixes, submission then some relaxation, maybe a beach walk. I’ll keep you posted on the results. This will be my only single-author paper, so I can’t hide behind other submissions if it gets rejected 😅

DocuSign Personal: $10 to $15 per month. DocuSign Standard: $25 to $45 per user per month. DocuSign Business Pro: $40 to $65 per user per month. A 10-person team on Business Pro pays $4,800 to $7,800 a year. To put signatures on PDFs. A team of 50 pays $24,000 to $39,000 a year. And there is a 100-envelopes-per-year cap on most plans. Send more contracts and you pay extra. Need SMS delivery? $0.40 per send. Need ID verification? $2.50 per attempt. Need premium support? $5,000 to $50,000 per year add-on. You are rationing digital signatures in 2026. DocuSign is a $10 billion company built entirely on this pricing model. Now meet DocuSeal. A free and open source alternative to DocuSign. Created in 2023 by a Ruby developer named Alex who was simply trying to sign one document and realised every solution online was overpriced or required a subscription. Three weeks later he had a working alternative. He pushed it to GitHub under the AGPL-3.0 license. Today it has 11,800+ stars and over 1,000 forks. Bootstrapped. No VCs. No paywalls. Here is what DocuSeal does: - Upload any PDF and turn it into a fillable, signable form - Drag and drop signature fields, dates, checkboxes, file uploads, and 13 field types - Send to multiple signers with custom signing order - Automated email reminders - Mobile signing on any device - PDF signature verification built in - Audit trail for every document - Bulk send and templates - Full API access - Self-host with one Docker command Here is what DocuSeal costs: Zero. Forever. Unlimited documents. Unlimited signers. Unlimited storage. DocuSign limits envelopes. DocuSeal doesn't. DocuSign charges per SMS. DocuSeal doesn't. DocuSign charges for ID checks. DocuSeal doesn't. DocuSign sees your contracts on their servers. DocuSeal doesn't. Here is the wildest part: The median DocuSign contract per Vendr is $17,250 per year. One Reddit thread has people saying "they want me to pay $4.80 per e-signature." Self-host DocuSeal on a $5 cloud server and a 50-person team can sign as many contracts as they want without paying a single dollar. Your contracts never leave your server. Your client lists. Your NDAs. Your employment agreements. None of it touches a third-party company. For individuals who only sign a few contracts a year, you save $180. For small teams of 10, you save up to $7,800 a year. For a 50-person company, you save up to $39,000 a year. Your documents. Your signatures. Your server. 100% Open Source. (Link in the comments)

new mistral model: 128B dense with an arch from 3 years ago (llama 2), very low context (128k), priced higher than deepseek v4 pro (1.6T total params, 1M context) and every other oss model that outperforms it this is very sad

🔥DeepSeek Input Cache Price Drop! Effective immediately, the price for input cache hits across the ENTIRE DeepSeek API series is reduced to just 1/10th of the original price! Build more efficiently for less. 📌Reminder: The DeepSeek-V4-Pro 75% OFF promotion is still active until May 5th, 2026, 15:59 (UTC Time).