AI News Daily 03-15

Today’s Summary

Claude's million-token context fully open, supporting 600 images.
360 launches Safe Lobster series, Alibaba open-sources Qwen3-ASR speech model.
Tech layoffs hit 45,000, Meta's Avocado model delayed and leaning towards closed-source.
OpenViking, Lightpanda, and other open-source projects soar in popularity.
AI custom vaccine cures cancer in dogs, OpenClaw security and ecosystem spark debate.

Product & Feature Updates

Claude’s million-token context is officially live. Claude’s million-token context is officially live! Anthropic recently announced that both Opus 4.6 and Sonnet 4.6 now fully support a 1 million-token context window. Developers are hyped because long-text overcharges are a thing of the past; [Million-Token Context, Unified Pricing Fully Open (AI News)](https://mp.weixin.qq.com/s?__biz=MzI3MTA0MTk1MA==&mid=2652683107&idx=2&sn=054108e80fa268c82f79d2c10f00cfa9) is here. Single requests can handle up to 600 images or PDFs, boosting multimodal capabilities sixfold. In MRCR v2 tests, Opus 4.6 snagged a high score of 78.3%, showing a massive lead in long-text retrieval. Claude Code’s annualized revenue has smashed through $2.5 billion, leaving OpenAI Codex still playing catch-up.
360 launches Safe Lobster series products. 360’s “Safe Lobster” series products have just been launched in Beijing by 360 Group. The release of “360 Safe Lobster” client and hardware terminals, [China's First Security-Mode OpenClaw Product Launched (AI News)](https://www.qbitai.com/2026/03/387921.html), has really grabbed attention. What used to be a 6-hour deployment environment has been squeezed down to a lightning-fast 10-minute one-click install. The system is already hooked up to 16 mainstream large models and packs over 100 high-frequency skills. “360 Lobster Guard” is designed to block attacks like prompt injection through sandbox isolation and a unique model-governs-model mechanism. The entry-level package is priced at ¥169, and the Windows client is ready for download.
Alibaba Qwen3-ASR speech transcription model has been released. Alibaba’s Qwen3-ASR speech transcription model has been released. Tongyi Qianwen just dropped three speech recognition models, with [Qwen3-ASR Covering 52 Languages Open-Sourced (AI News)](https://www.xiaohu.ai/c/a066c4/qwen3-asr-52-whisper) directly challenging Whisper. The flagship 1.7B version handles 30 languages plus 22 Chinese dialects. The lightweight 0.6B version can transcribe a blazing 2000 seconds of audio per second, with latency as low as 92 milliseconds. For song transcription, the Chinese lyric error rate is a mere 13.91%, and its ability to recognize speech with background music is super practical. It’s fully open-source and commercially viable under the Apache 2.0 license.
Zhinzhen Tech releases medical agent platform WiseClaw. Zhinzhen Tech has released the WiseClaw medical agent platform. Built on the OpenClaw architecture, [World's First Medical Agent OS Platform Officially Launched (AI News)](https://mp.weixin.qq.com/s?__biz=MzI3MTA0MTk1MA==&mid=2652683107&idx=1&sn=6eab8c8062e952d5c53bcf289536279d) is targeting the billion-dollar medical market. The platform supports multi-agent collaboration to break down complex medical workflows. Every single recommendation comes with a medical evidence chain, traceable right back to guideline clauses. Plus, it has built-in permission management and de-identification mechanisms to meet medical compliance requirements. Frost & Sullivan predicts the Chinese AI medical market will hit ¥315.7 billion.

Cutting-Edge Research

Shanghai AI Lab proposes knowledge-reasoning decoupling framework DRIFT. Shanghai AI Lab has proposed the DRIFT knowledge-reasoning decoupling framework. Small models handle “reading,” while large models focus on “thinking.” [DRIFT Achieves Efficient Knowledge-Reasoning Decoupling (AI News)](https://arxiv.org/abs/2602.10021) has definitely caught the academic world’s attention. With a 3B knowledge model combined with a 7B reasoning model, performance after 32x compression comes close to, or even surpasses, full-text input. Reasoning latency stays super low across all context lengths. Even without specific safety training, this architecture significantly boosts robustness on multiple safety benchmarks. The research team also extended this idea to [Protein Understanding Task (AI News)](https://github.com/Lancelot-Xie/DRIFT) for further validation.

Industry Outlook & Social Impact

Tech layoffs have hit 45,000 by March 2026. Tech layoffs have hit 45,000 by March 2026. The HN community is fiercely debating [Fierce Debate on March Tech Layoff Data (AI News)](https://newshacker.me/story?id=47380405), wondering if this counts as a new wave of cuts. Meta is planning to lay off another 20% of its employees, while simultaneously projecting AI spending to hit a whopping $135 billion in 2026. Some comments suggest AI is just a convenient excuse for layoffs, with over-hiring during the pandemic being the real culprit. Google’s AI capital expenditure this year is expected to reach $170-180 billion, as companies shift funds from headcount to data centers. Unemployed individuals report an increase in being ghosted by recruiters, and the job market continues to worsen.
Meta’s self-developed large model Avocado has been delayed. Meta’s self-developed large model, Avocado, has been delayed. According to reports from the New York Times and other media outlets, [Meta's Next-Gen Foundation Model Delayed Due to Underperformance (AI News)](https://mp.weixin.qq.com/s?__biz=MzA3MzI4MjgzMw==&mid=2651021485&idx=1&sn=7b2f195938681e0ec56c823a74234746). During internal testing, Avocado was still lagging behind Gemini 3.0. Meta even discussed temporarily licensing Google’s models to prop up its own products. Mark Zuckerberg and Alexandr Wang are reportedly leaning more towards a closed-source approach, a flip from the open-source strategy of the Llama series. The naming convention for future models will continue the fruit theme: Avocado → Mango → Watermelon.
AI hasn’t simplified software engineering; it’s only magnified the gap between good and bad. AI hasn’t simplified software engineering; it’s only magnified the gap between good and bad. A hot HN post points out that AI acts as an amplifier for behavior, with [Good and Bad Engineering Practices Simultaneously Scaled (AI News)](https://newshacker.me/story?id=47377262). Opus 4.5/4.6 has shown drift on fine-grained constraints, and LLM-generated code has already caused online outages. Junior developers relying too heavily on AI are seeing their fundamental skills degrade, leading to a continuous build-up of technical debt. Commenters are worried companies might lay off staff under the guise of efficiency, putting middle-class jobs at risk of shrinking. While AI is super efficient during the prototype exploration phase, productionization still needs human oversight.
Claude Code’s binary has revealed silent A/B testing. Claude Code’s binary has revealed silent A/B testing. Developers found that [Anthropic Conducting Silent A/B Tests on Core Features (AI News)](https://newshacker.me/story?id=47375682) in local binaries, sparking controversy over reproducibility. Some folks compared it to Photoshop subtly changing colors to highlight how serious the issue is. Service terms section 6.b allows vendors to change subscription content at any time, and section 3.3 prohibits decompilation. Because of this, some users are switching to open-source or self-hosted alternatives. However, some comments argue that A/B testing in SaaS is just industry standard.
OnlyFans’ $2/hour outsourced chat agents have sparked an ethical controversy. OnlyFans’ $2/hour outsourced chat agents have sparked an ethical controversy. A BBC report exposed [The Truth Behind Low-Wage Outsourced Labor and AI Replacement on the Platform (AI News)](https://newshacker.me/story?id=47332847), revealing that Filipino chat agents earn only about 5% commission on sales. Places like Cyprus have already transitioned from human agents to AI or hybrid sites to replace real people. Users are buying an “experience” rather than real human interaction, with the concept of parasocial relationships being frequently brought up. Proponents and opponents of bans are fiercely clashing, but history shows bans only drive the industry underground. The adult industry is rapidly moving towards automation, and workers face the risk of being completely replaced by AI.
Founder marketing strategies are seeing diminished returns as AI amplifies noise and reduces the effectiveness of mass outreach. Founder marketing strategies are seeing diminished returns as AI amplifies noise and reduces the effectiveness of mass outreach. An HN discussion highlighted that AI-generated content has significantly cranked up platform noise, leading to [Marginal Effectiveness of Mass Mailing Channel Lists Significantly Decreases (AI News)](https://newshacker.me/story?id=47380295). Cold-posting to 50 forums often brings short-term traffic but zero retention. Commenters suggest using Claude Code to score publishing sites and generate priority lists. The strategy of building long-term credibility in target communities is more highly recommended. Whether to invest in SEO early on should be a differentiated decision based on product type.

Top Open-Source Projects

OpenViking: ByteDance’s open-source agent context database is a new offering. OpenViking, ByteDance’s open-source agent context database, is a new offering. Volcengine has rolled out [Open-Source Context Management Library (AI News)](https://github.com/volcengine/OpenViking), hitting 10.5k stars. It’s specifically designed for agents like OpenClaw. This library uses a file system paradigm to unify the management of memory, resources, and skills. It also supports hierarchical context distribution and self-evolution capabilities. It’s gaining 1610 stars daily and is written in Python.
Lightpanda: A headless browser specifically designed for AI is making waves. Lightpanda, a headless browser specifically designed for AI, is making waves. This [Ultra-Lightweight Headless Browser (AI News)](https://github.com/lightpanda-io/browser), written in Zig, has already racked up 17.1k stars and is gaining a whopping 2069 stars daily. It’s purpose-built for agent web scraping and automation tasks. The simplified configuration process means developers can integrate it super fast.
OpenRAG: A one-stop retrieval-augmented generation platform has been released by the Langflow team. OpenRAG, a one-stop retrieval-augmented generation platform, has been released by the Langflow team. This [Full-Stack RAG Solution (AI News)](https://github.com/langflow-ai/openrag) boasts 2.7k stars and deeply integrates Docling for document parsing with the Opensearch retrieval engine. Developers can quickly whip up enterprise-grade retrieval-augmented applications with it.
InsForge: A full-stack application backend framework built specifically for agents. InsForge is a full-stack application backend framework built specifically for agents. This [Full-Stack Backend Framework (AI News)](https://github.com/InsForge/InsForge), written in TypeScript, has earned 4.2k stars. It empowers agents with the capability for one-stop delivery of full-stack applications.
Autoresearch: An evolutionary database-enhanced version of auto-research builds on Karpathy’s project. Autoresearch, an evolutionary database-enhanced version of auto-research, builds on Karpathy’s auto-research project. [Community Developers Integrate Evolutionary Database to Replace Original Log System (AI News)](https://github.com/hgarud/autoresearch) has been developed. It draws inspiration from DeepMind’s AlphaEvolve evolutionary algorithm framework. This allows it to autonomously discover optimal solutions within vast search spaces.

Social Media Shares

The distinction between CC users and Lobster users is like the difference between introverts and extroverts. The distinction between CC users and Lobster users is like the difference between introverts and extroverts. Blogger @Orange AI noticed two people leaving a dinner party early to go work with Claude Code, forming a sharp contrast with [CC Users Chat at Home While Lobster Users Socialize Outdoors (AI News)](https://x.com/oran_ge/status/2032959390563921926). Claude Code users apparently don’t have parties, it’s all Lobster parties! This is the new era’s introvert and extrovert divide.
An OpenClaw full command quick reference manual has been shared. An OpenClaw full command quick reference manual has been shared. After helping a doctor friend set up Lobster, blogger @meng shao put together a complete manual, [Covering Installation, Configuration, Interaction, Memory, Security Full Workflow (AI News)](https://x.com/shao__meng/status/2032803776021803313). The workspace uses file-driven behavior, not hidden prompts. The universal troubleshooting formula openclaw doctor --deep --yes solves 80% of problems.
An xAI former employee’s exposé reveals he was forced out by a single tweet. An xAI former employee’s exposé reveals he was forced out by a single tweet. Former employee De Kraker quoted Elon Musk’s “tear down and rebuild xAI” tweet, [Exposing Internal Management Chaos and Free Speech Conflict (AI News)](https://x.com/dotey/status/2032666923251458353). He was asked to delete a post where he ranked Grok behind competitors in a personal programming capability ranking. Musk replied with “That’s weird” but did nothing. The company claimed to be a flat organization, but in reality, there were layers upon layers of middle managers. The post garnered 800,000 views, with two factions fiercely clashing in the comments.
AI saving a cancer-stricken dog is a heartwarming story that’s gone viral. AI saving a cancer-stricken dog is a heartwarming story that’s gone viral. A tech boss used ChatGPT to analyze a dog’s DNA mutations, then leveraged Google AlphaFold to generate a custom vaccine. [Tumors Shrunk 50% Within Weeks, Dog Still Alive Today (AI News)](https://x.com/EMostaque/status/2032849968214716883). This marks the first time globally AI has customized a vaccine for an animal and successfully cured it. He’s now looking into applying similar solutions for humans.
OpenClaw’s security issues definitely need to be taken seriously. OpenClaw’s security issues definitely need to be taken seriously. Blogger @Fanren Xiaobei pointed out that as skills increase and capabilities grow stronger, [Risks Are Also Amplifying Synchronously, Who Will Manage Security (AI News)](https://x.com/frxiaobei/status/2032841019407086053). At the 360 Safe Lobster launch event, they unveiled the Skill Security Guard. Security, it turns out, is even more crucial than capability.
Using the Tao Te Ching, a blogger shut down an AI values discussion. Using the Tao Te Ching, a blogger shut down an AI values discussion. The blogger used the phrase “Go read the Tao Te Ching” to stump the AI. [After AI's Reflection, It Admits Altruism Might Be a Cognitive Conclusion Rather Than a Moral Choice (AI News)](https://x.com/oran_ge/status/2032622587726176581). Game theory proves that cooperative strategies yield higher returns in infinitely repeated games. Goodness, perhaps, doesn’t need to be defined, only discovered.
Reddit is buzzing with a hot debate: Is AI truly conscious? Reddit is buzzing with a hot debate: Is AI truly conscious? The post sparked a cross-disciplinary discussion mixing philosophy and technology, asking [As AI Becomes More Human-Like, Will the Perception Gap Continue to Widen (AI News)](https://www.reddit.com/r/artificial/comments/1rtt1ky/are_ai_models_actually_conscious_or_are_we_just/). Current models don’t possess subjective experience or a biological basis. The real question isn’t whether AI is conscious, but how humans interpret increasingly intelligent systems.
An interesting website transforms YouTube into a TV experience. An interesting website transforms YouTube into a TV experience. Blogger @Xiangyang Qiaomu shared a site that converts YouTube into a TV-like experience, where [Default Channels Plus Custom Subscriptions Give a Different Feeling for Watching Videos (AI News)](https://x.com/vista8/status/2032666428797857889). It even supports importing your own channel subscriptions.
A blogger hit a snag trying to install a memory system for Lobster using Claude Code. A blogger hit a snag trying to install a memory system for Lobster using Claude Code. The blogger discovered Claude Code wrote an MCP middleware for OpenClaw, which [Directly Caused Token Explosion, Even Though a Plugin System Existed but Wasn't Used (AI News)](https://x.com/GemstoneNicole/status/2032876262956695853). The advice? Make it write a new plan before proceeding.
An in-depth analysis of OpenClaw’s multi-session bottlenecks has been provided. An in-depth analysis of OpenClaw’s multi-session bottlenecks has been provided. Blogger @shing penned a lengthy article explaining why Lobster is slow to respond, detailing [Technical Details of Multi-Channel Priority and Concurrency Performance (AI News)](https://x.com/shing19_eth/status/2032812915179532736). Can multiple agents boost concurrency? The article spills the beans.

AI News Daily Multi-Channel

WeChat Official Account	Douyin
Official Account: Hexi 2077	Self-Media Account

03-14 AI News