Bilingual

Leanstral 1.5: Proof Abundance for All
18 days ago
- Leanstral 1.5 is a free Apache-2.0 licensed model with 6B active parameters, offering a significant performance upgrade in formal verification.
- It achieves state-of-the-art results on benchmarks: 100% on miniF2F, 587/672 on PutnamBench, 87% on FATE-H, and 34% on FATE-X.
- The model undergoes a three-stage training process: mid-training, supervised fine-tuning, and reinforcement learning with CISPO, using multiturn and code agent environments.
- It demonstrates strong test-time scaling on PutnamBench, solving more problems as token budget increases, from 44 at 50k tokens to 587 at 4M tokens.
- Leanstral 1.5 excels in code verification, proving time complexity for AVL trees and uncovering 5 previously unknown bugs across 57 repositories tested.
- The model is fully open-sourced and available via Hugging Face and a free API, with recommendations for setup using Mistral Vibe and Lean LSP MCP.
Anthropic's Method to Losing Goodwill in a Few Easy Steps
15 days ago
- Anthropic's Claude subscription and API face issues including enshittification, vendor lock-in, consumer malpractice, and price gouging, while open-source models become more competitive.
- Claude's API is notoriously unreliable, and Claude subscriptions force users to rely on Anthropic's servers, with no compatibility with third-party harnesses without extra costs.
- Anthropic introduced 'extra usage' billing for third-party tools and split subscription billing into separate pools for first-party and third-party usage, significantly increasing costs for heavy agent users.
- The author criticizes 'vibecoding' and agent-driven development for causing skill degradation, advocating instead for agent-assisted development using AI as an autocomplete tool.
- Open-source models like Qwen, GLM, and Deepseek offer viable alternatives to proprietary models, with AI gateways enabling flexible, cost-effective, and private model usage.
- The article concludes by condemning Anthropic's unethical practices, such as locking customers into closed systems and changing terms without notice, and encourages moving to open-source alternatives.
6 months to live for open models
9 days ago
- The most serious test to date for open source AI viability is happening now due to policy actions that could marginalize open models.
- White House discussions may lead to new executive orders impacting open models, potentially starting with Chinese-origin models and government uses.
- Open models lack a central economic champion to defend against regulatory actions, unlike closed models with strong lobbying.
- Likely policy action includes banning or delaying open-weights models above capability levels like GPT 5.5 within the next 6 months.
- Two key policy discussions affecting open models are distillation (regulatory capture by companies like Anthropic) and frontier capabilities.
- Anthropic's campaign against Chinese models is seen as regulatory capture, aiming to ban competitors and secure economic advantages.
- Distillation debates focus on risks like Chinese labs distilling cybersecurity capabilities, but APIs are not magically secure and have vulnerabilities.
- A ban on open models could demolish the emerging open model economy in the US, including inference and fine-tuning companies.
- Handling frontier open-weight models like Mythos requires careful policy; a flat ban may be ineffective if not globally enforced and could isolate the US.
- Open models increase safety through broad access and understanding; stopping open-source progress requires global agreement, which is currently lacking.
- Short-term solutions include a US company releasing a capable open model to shift focus from China or building a coalition to lobby for safe open model rollout.
Satya Nadella has issued a warning to companies using AI
7 days ago
- Microsoft CEO Satya Nadella warns that companies using proprietary AI models pay twice: with money and by handing over valuable proprietary business data.
- Nadella argues that enterprises are unknowingly training AI models with sensitive business knowledge, which competitors could never buy, potentially turning model makers into competitors.
- He criticizes the hypocrisy of AI labs freely using public data for training while restricting others from studying or 'distilling' their models to create cheaper alternatives.
- Nadella suggests companies retain ownership of their data and build proprietary learning environments in the cloud, along with 'orchestration layers' to switch between AI models easily.
- Enterprises are increasingly shifting to on-premise open source AI models for cost savings, control, and to avoid data leakage, a trend supported by companies like Solo.io, Vercel, and OpenRouter.
Big Tech's top executives warn enterprises are giving away too much to AI labs
3 days ago
- Satya Nadella warns enterprises are 'paying twice' for AI by purchasing tokens and inadvertently handing proprietary data to AI labs, enabling potential future competition.
- Executives like Alex Karp and David Sacks criticize frontier AI labs for extracting valuable business data while charging high prices, advocating for more control over data and infrastructure.
- Open-weight models are gaining traction as alternatives to closed-source AI, offering cost savings, transparency, and reducing vendor lock-in, with U.S. government actions prompting diversification.
- AI industry news includes xAI suing a user over deepfake misuse, OpenAI employees funding a pro-regulation PAC, and Google DeepMind researcher resigning over military deals.
- CEOs of major AI labs (Hassabis, Altman, Amodei) propose regulatory frameworks for frontier AI, though approaches vary, with concerns about entrenching incumbents over startups.
- Dario Amodei donated $1 million to a PAC pushing AI regulation, highlighting the growing political influence and public backlash against AI's rapid expansion and costs.
Kimi K3: The open-weights escalation
a day ago
- Moonshot AI released Kimi K3, a 2.8T parameter MoE model, with open weights scheduled for release on July 27th.
- Kimi K3 is the strongest open model ever, ranking high on multiple benchmarks and competing closely with top closed models like Claude Fable and GPT-5.6 Sol.
- The release reduces the performance gap between open/closed and U.S./Chinese models to an estimated 3-5 months.
- China's commitment to open-source AI was reinforced by Xi Jinping's keynote at WAIC, indicating high risk tolerance and a strategy to boost AI adoption.
- Open-weight models are economically decelerationist for frontier labs, reducing profit margins and investment but accelerating AI diffusion and customization across industries.
- Chinese labs, like Moonshot AI, demonstrate higher capital efficiency, achieving strong results with fewer resources due to optimized training and architectural innovations like Kimi Delta Attention.
- The open-model ecosystem is expanding, with Alibaba announcing Qwen 3.8 with open weights, potentially pushing Google lower on leaderboards.
- Policy debates in the U.S. about restricting open-weight models may backfire, creating asymmetrical security risks and hindering AI diffusion and safety.
- A balanced open-closed model dynamic is crucial for mitigating risks, as open models will inevitably advance and cross capability thresholds regardless of regulation.
- Kimi K3 marks a watershed moment, making frontier open-weight models a reality and shifting the AI landscape from a unipolar, U.S.-dominated sphere to a more competitive global ecosystem.

Hasty Briefsbeta

#open source ai

Leanstral 1.5: Proof Abundance for All

Anthropic's Method to Losing Goodwill in a Few Easy Steps

6 months to live for open models

Satya Nadella has issued a warning to companies using AI

Big Tech's top executives warn enterprises are giving away too much to AI labs

Kimi K3: The open-weights escalation