Anthropic Surpasses OpenAI with $1 Trillion Valuation — The AI Industry's Employment Divide Comes Into Sharp Focus

According to Latent Space's AINews report, Anthropic's valuation has reached $1–1.2 trillion, overtaking OpenAI to become the 11th to 15th most valuable company in the world.

The drivers are an 80x annual growth rate in what has been called a 'miraculous Q1' and a $15 billion increase in ARR (annual recurring revenue) in a single month. Secondary market activity and media reports from outlets such as the FT back up this valuation.

Meanwhile, a stark employment divide lurks behind this growth. While Block has cut 40%, Coinbase 14%, and Cloudflare 20% of their workforce, the 'polarization' — with Anthropic alone continuing to grow 10x — stands out. The tech industry as a whole is in the middle of an AI transition, and the gap between companies investing in AI and those that aren't is widening fast.

'People Hate AI Art' — Backlash Against AI Art Sparks Debate

The article 'People Hate AI Art,' which racked up 74 points on Hacker News, has laid bare how people feel about AI-generated content.

Even as the technical quality of AI art improves, many people still feel it is 'not real.' The piece argues that this backlash is not mere prejudice, but is rooted in the value of human intent and experience in the act of creation.

With AI now surpassing human skill in image generation, the fundamental question of what makes art 'valuable' is resurfacing. As AI-generated works spread, how to position human creativity looks set to become a challenge for the entire creative industry.

AI's 'Circular Psychosis' — A Self-Referential Trap

'AI's Circular Psychosis' from Where's Your Ed At poses a self-referential challenge to the AI industry.

The article, which sparked lively discussion on Hacker News, points out that the 'loop' of AI talking about AI, using AI to develop AI, and evaluating AI's output with AI may be dulling the industry's judgment as a whole. The observation that much of AI's economic value is sustained by use cases that consume AI itself raises an important question about the durability of the current boom.

Qwen3.6 35B A3B Uncensored Version Released with Full MTP Preservation

An uncensored version of Qwen3.6-35B-A3B called 'heretic' has been released in Reddit's LocalLLaMA community.

It fully preserves all 19 MTP (Multi-Token Prediction) heads and achieves an extremely low KLD (Kullback-Leibler Divergence) of 0.0015. Across 100 tests, refusals occurred only 10 times. It is available in multiple formats including Safetensors, GGUF, NVFP4, and GPTQ-Int4.

Models that strip out filtering of sensitive content while keeping MTP functionality intact are in high demand in the local LLM community, making this a highly practical option.

Gemma 4 in Practice — The Community Is Split

Also on LocalLLaMA, real-world user feedback on Google's Gemma 4 series was shared.

A user who had been getting good results with Qwen3.6 models tried Gemma 4, testing a Q5-quantized 31B model and a Q8-quantized 27B model, but reported they were 'disappointing.' Even with the KV cache kept in FP16, a performance gap versus Qwen3.6 was observed in coding and data extraction tasks.

That said, some users do prefer Gemma 4, confirming once again that the best model depends on the use case.

Claude Code Security Checklist — Sharing Practical Lessons

A 'Claude Code Security Checklist' was published on Qiita. Drawing on firsthand experience — losing files twice to 'rm -rf' while actually using Claude Code, nearly git-pushing a .env file, and publishing a broken package via npm publish — the author compiled a 10-item security checklist.

As AI coding tools become widespread, the risk of unintended destructive operations has become very real. Sharing practical know-how like this should be a useful reference for any team adopting AI development tools.

Getting AI Agents to Explore 'Uncertainty'

On Zenn, an article examined how to handle highly uncertain development situations with AI agents — cases where you simply cannot set a clear goal.

When you instruct an AI agent to 'improve the code quality,' it will find and fix problems, but that only solves 'the problems the agent found' — it doesn't mean the code has reached the quality level you expect. The piece tackles the practical challenge of how to collaborate with AI agents in development settings where the goal cannot be clearly defined.

Claude Code v2.1.133–136 — MCP Reauthentication Issue Resolved

Also on Qiita, the changelogs for Claude Code v2.1.133 through v2.1.136 were translated and explained.

v2.1.136 fixes the daily reauthentication issue in environments using multiple MCP servers. It also includes improvements such as a better --workspace option.

---

**References:**

[Anthropic growing 10x/year while everyone else is laying off >10% of their workforce](https://www.latent.space/p/ainews-anthropic-growing-10xyear) — Latent Space

[People Hate AI Art](https://mccue.dev/pages/5-8-26-ai-art) — mccue.dev

[AI's Circular Psychosis](https://www.wheresyoured.at/premium-ais-circular-psychosis/) — Where's Your Ed At

[Qwen3.6 35B A3B uncensored heretic Native MTP Preserved](https://www.reddit.com/r/LocalLLaMA/comments/1t7qfaq/qwen36_35b_a3b_uncensored_heretic_native_mtp/) — Reddit LocalLLaMA

[Those of you who like Gemma4 models - how are you guys using them?](https://www.reddit.com/r/LocalLLaMA/comments/1t7rco2/those_of_you_who_like_gemma4_models_how_are_you/) — Reddit LocalLLaMA

[Claude CodeのSecurityチェックリスト——10項目でAIの安全性を確認する](https://qiita.com/yurukusa/items/3b44d09a34e5ca0f86a5) — Qiita

[AIエージェントに不確実性を探索させるためにやること](https://zenn.dev/kinzal/articles/6e789454354cfc) — Zenn

[Claude Code v2.1.133〜v2.1.136 リリース｜毎日Changelog解説](https://qiita.com/moha0918_/items/f55caa82337f0894a793) — Qiita