AI this week: DeepSeek’s small model wins, Claude cracks a quantum-safe candidate
A cheaper DeepSeek model beats its bigger sibling, an AI finds a flaw in a proposed encryption standard, and a giant open code dataset.
August 7, 2026 · By Bohol AI
Small, cheap models kept closing the gap, an AI system found a real weakness in a proposed security standard, and coders got a fresh open dataset.
DeepSeek’s Flash beats its Pro
DeepSeek released DeepSeek-V4-Flash-0731, the official version of the smaller model in its V4 family. Extra fine-tuning, with no change to the design, pushed it past the larger DeepSeek-V4-Pro on independent tests, at a small fraction of the cost of closed models with similar ability.
Claude finds a flaw in a quantum-safe candidate
The U.S. National Institute of Standards and Technology is testing new signature schemes meant to survive future quantum computers. Anthropic reported that Claude Mythos Preview, a model with restricted access, found a weakness in one candidate called HAWK. HAWK’s designers withdrew it from the competition after about two years and two rounds of expert review.
Why it matters: AI is now strong enough to contribute to security research, which helps defenders and attackers alike.
The Stack v3
Hugging Face released The Stack v3, a new snapshot of public GitHub code for training models. It is the largest and most current open code dataset, and unlike earlier versions it keeps whole repositories together rather than loose files.
Research: system prompts that produce safer code
Stanford researchers built SecureForge, which automatically tunes a model’s system prompt so that the Python code it writes contains fewer security flaws. The tool is freely available.
What it means for Bohol
- For freelancers and small teams, the cheapest capable model this week was a small one. Test smaller models before paying for the biggest.
- If you use AI to write code for clients, review it for security issues. Tools like SecureForge exist because AI code often has holes.
Sources
- DeepSeek API updates
- DeepSeek-V4-Flash-0731 (Hugging Face)
- Anthropic: discovering cryptographic weaknesses
- NIST: post-quantum signature round 3
- Hugging Face: The Stack v3
- Liu, Einstein, Jang et al., SecureForge (arXiv)
- The Batch, issue 365 (DeepLearning.AI)