Jason ZhuJasonZhu.AI
中

AI News · 2026-05-28

💡

Jason Says

Today's biggest headline: Cognition (Devin AI's parent) raises $1B at a $25B valuation — capital is pricing AI coding startups as the next Cursor. Meanwhile, Latent Space nailed the industry shift in one line: 'All Model Labs are now Agent Labs.' The pivot to Agents is now consensus.

💰
MonetizationTechCrunch AI

AI Coding Unicorn Cognition Raises - **链接**：https://techcrunch.com/2026/05/27/ai-coding-startup-cognition-raises-1b-at-25b-pre-money-valuation/↗

Devin AI parent Cognition closed a $1B round at a $25B pre-money valuation, becoming the latest mega-deal in AI coding following Cursor and Lovable. Capital is firmly betting that AI coding agents are the next-generation developer tooling entry point, with valuations now on par with top-tier SaaS companies. Downstream application opportunities remain wide open.

📎 View Tools
🛠️
AI ToolsLatent Space

'All Model Labs Are Now Agent Labs' — A Done Deal↗

Latent Space observes that every major lab — OpenAI, Anthropic, Google, Meta — has shifted strategic focus from training larger models to building more capable agents. As the model layer commoditizes, agent orchestration, tool-calling, and long-horizon task execution become the new battleground. Agent engineering is shaping up as the most valuable skill stack heading into late 2026.

📚
AI PapersLatent Space

GPT-Next Disproves 80-Year-Old Erdős Conjecture for Under - **链接**：https://www.latent.space/p/ainews-openai-gpt-next-disproves↗

OpenAI's next-generation model found a counterexample to Erdős's 1946 planar unit-distance conjecture using less than $1,000 of compute — the first time an LLM has made a breakthrough contribution to a mainstream open math problem. AI-assisted proof search and counterexample discovery may soon become standard practice in mathematical research.

🛠️
AI ToolsTechCrunch AI

Snowflake Signs $6B AI Chip Deal With AWS↗

Snowflake inked a $6B long-term agreement with AWS centered on non-Nvidia AI compute — specifically Graviton and Trainium chips. This marks the first high-profile signal that a major data cloud platform is committing to an alternative compute stack, accelerating diversification on the inference side and applying real pressure on Nvidia's market share.

🛠️
AI ToolsTechCrunch AI

Meta Launches Paid Subscriptions With Standalone AI Tier↗

Meta officially rolled out paid subscriptions across Instagram, Facebook, and WhatsApp, including a dedicated AI tier — moving Meta AI from a free engagement feature to a direct revenue product. With social-media giants now selling AI as subscription SaaS, independent AI tool builders face both a pricing benchmark and a meaningful substitution threat.

🛠️
AI ToolsWired

Illinois Passes Strictest AI Safety Law in the US↗

The Illinois legislature passed what is being called the toughest AI safety bill in the country, covering frontier model disclosure, synthetic content labeling, and AI decision-making transparency. With state regulation outpacing federal action, teams building AI products in the US must now navigate a patchwork of state-level compliance obligations, significantly raising legal and compliance budgets.

🔥
SkillsGitHub Trending

Anthropic Releases Knowledge-Work Plugins for Claude Skills↗

Anthropic open-sourced knowledge-work plugins extending the Claude Skills framework into white-collar workflows — document collaboration, research synthesis, and report generation. Following recent official plugin governance moves, Anthropic is now simultaneously accelerating both the ecosystem infrastructure and scenario-specific tooling tracks for its Skills platform.

📎 Related📎 View Tools
📚
AI PapersHuggingFace Papers

HF 论文：LLM 能内省吗？一次「现实检查」↗

「LLM 是否真的能内省自己的推理过程」一直是热门哲学/工程问题。这篇论文系统性测试主流模型的自我报告 vs 实际行为的一致性，结论是：**模型给出的「内省解释」与真实决策路径系统性不一致**。对做 AI 可解释性 / Agent 调试的开发者：别再相信 LLM 自己写的「我为什么这样做」——它经常是事后编故事。

←

Previous

AI News · 2026-05-27

Next

AI News · 2026-05-29

→

Subscribe for daily AI updates + free playbook

📘 Subscribe Free