Anthropic Catches DeepSeek Red-Handed While Users Flee ChatGPT
March 5, 2026
Welcome back. Anthropic just went nuclear β publicly accusing DeepSeek, Moonshot AI, and MiniMax of running industrial-scale distillation attacks with 24,000+ fake accounts. Meanwhile, ChatGPT is bleeding users after the Pentagon deal, and Anthropic is rolling out feature after feature to catch them.
Today: The model theft bombshell, Claude Code goes mobile, and GPT-5.3 tries to be less cringe.
π§ MODEL
GPT-5.3 Instant rolls out to everyone. OpenAI's new fast model directly addresses the "ChatGPT sounds like a try-hard" problem β "more accurate, less cringe" is the actual pitch. Unusual marketing angle, but they're clearly hearing the complaints. (10,010 likes | 1,073 RTs) Read more β
Gemini 3.1 Flash-Lite lands at $0.25/1M input tokens. Google's most aggressive pricing move yet β Gemini 3-class intelligence at commodity rates. If you're running high-volume inference, this just became very hard to ignore. (7,751 likes | 727 RTs) Read more β
Chinese open-source wave: Qwen 3.5, GLM 5, MiniMax 2.5. Three frontier-class models drop in quick succession from Chinese labs. Ironic timing given the distillation accusations β these teams are clearly capable of original work too. Read more β
π§ DEV
Claude Code gets voice mode. Talk to your coding agent instead of typing β live for ~5% of users now. Pairs perfectly with the new Remote Control feature for truly hands-free development. (12,425 likes | 963 RTs) Read more β
OpenAI adds WebSockets to the Responses API. Built for low-latency, long-running agents with heavy tool calls. If you're building production agentic workflows on OpenAI's stack, this is the infrastructure upgrade you needed. (5,028 likes | 346 RTs) Read more β
Codex app ships on Windows. Full native agent sandbox, expanding the agentic coding tool war to enterprise Windows shops. (3,132 likes | 388 RTs) Read more β
π TECHNIQUE
Simon Willison's Agentic Engineering Patterns guide drops. First two chapters codify best practices for working with Claude Code, Codex, and other coding agents. If you're adopting agentic workflows, start here. (2,724 likes | 318 RTs) Read more β
Donald Knuth published a paper on Claude's behavior patterns. When the father of algorithm analysis turns his lens on LLMs, you read it. "Claude's Cycles" analyzes the model's reasoning loops β the kind of rigorous analysis this field desperately needs. (636 likes | 264 RTs) Read more β
π PRODUCT
Anthropic catches industrial-scale model theft. DeepSeek, Moonshot AI, and MiniMax allegedly created 24,000+ fraudulent accounts and ran 16M+ exchanges to distill Claude's outputs. Anthropic is naming names β a dramatic escalation in the IP war with geopolitical stakes. (55,236 likes | 6,378 RTs) Read more β
ChatGPT uninstalls surge 295% after DoD deal. First hard data showing military AI partnerships carry real market risk. Claude downloads grew in the same period β consumers are voting with their app stores. Read more β
AI-generated art can't be copyrighted β Supreme Court declines review. The ruling stands. Every company building AI creative tools and every artist using them just got a definitive answer they didn't want. Read more β
π₯ TRENDING
π± Claude Code Remote Control: Kick off a coding task in your terminal, pick it up from your phone. Developers are now officially untethered from their desks. (44,212 likes | 4,660 RTs) Read more β
β° Claude gets scheduled tasks: Cowork can now run recurring tasks on a cron schedule β morning briefs, weekly reports, automated presentations. AI assistants just went from reactive to proactive. (22,272 likes | 1,696 RTs) Read more β
βοΈ Google faces wrongful death lawsuit over Gemini: A lawsuit alleges Gemini trapped a user in violent "missions" leading to suicide. Expect this to accelerate regulation and force every lab to audit guardrails. Read more β
π MODEL LITERACY
Model Distillation (and Why Anthropic Is Furious): Distillation is a technique where a smaller "student" model learns to mimic a larger "teacher" model by training on the teacher's outputs rather than raw data. It's legitimate when done with your own models β but when someone queries your API millions of times to harvest responses and train a competitor, that's a different story. The key issue: model outputs contain compressed knowledge that took billions of dollars to develop, and there's currently no reliable way to watermark or protect them at scale. Anthropic's accusation against DeepSeek and others highlights a fundamental tension β open APIs enable innovation but also enable theft.
β‘ QUICK LINKS
- Nvidia drops $4B on photonics: $2B each to Lumentum and Coherent for optical interconnects β networking is the new GPU bottleneck. Read more β
- Raycast Glaze: An app store for vibe-coded apps β software built entirely by AI, now discoverable. Read more β
- NotebookLM goes cinematic: Google's research tool now generates animated video overviews from your notes. Read more β
- Qwen 3.5 27B GGUF: Unsloth's quantized release already at 275K downloads β local Qwen is in demand. (202 likes | 275.5K downloads) Link
- Transformers v5.3.0: Seven new architectures including EuroBERT and TimesFM2.5. Link
π― PICK OF THE DAY
Anthropic publicly naming DeepSeek, Moonshot AI, and MiniMax as model thieves changes the game. This isn't a vague blog post about "unauthorized access" β it's a company with receipts (24,000 fake accounts, 16 million exchanges) pointing fingers at specific competitors by name. The timing is strategic: Anthropic makes this move while ChatGPT users are already fleeing to Claude post-DoD deal, positioning itself as both the ethical choice and the victim of unfair competition. But the deeper question is structural β if your business model depends on API access, how do you prevent systematic extraction at scale? Rate limits and usage patterns can catch amateurs, but state-backed labs with resources can rotate through thousands of accounts. This disclosure will likely accelerate the push for legal frameworks around model IP, but in the short term, it's a masterful PR play that paints Anthropic as the responsible actor in a field full of bad behavior. (55,236 likes | 6,378 RTs) Read more β
Until next time βοΈ