Lilith · selected stories
What is actually happening in AI. Selected stories, context and opinion without the promotional noise.
Atom feed ↗Anthropic says Claude Fable 5 is globally available again from July 1 after US export controls were lifted. The real story is the new classifier: it blocks one reported bypass in over 99% of cases, while making routine coding and debugging more likely to hit false refusals.
NVIDIA released Nemotron 3.5 Content Safety on Hugging Face, a 4B multimodal model for safety verdicts with custom policies, reasoning traces and broad language coverage.
Cohere is releasing North Mini Code, a 30B Mixture of Experts model with 3B active parameters under the Apache 2.0 license. The interesting signal is not only the benchmark chart, but the focus on robustness across harnesses, because coding agents often fail at the interface, not just in the code.
Writer research covered by TechCrunch suggests that memory and personalization layers can degrade model accuracy and increase sycophancy. For product teams, user context is not just fuel for better UX. It is also a new failure mode.
Google DeepMind introduced DiffusionGemma, an experimental open text generation model that the model page says can produce up to 4x to 5x faster output on NVIDIA GPUs and exceed 1,000 tokens per second on an H100. The bigger point is architectural: text models are starting to attack the sequential bottleneck.
The Verge is pressing the question Google has not clearly answered: whether music uploaded to YouTube is being used to train Lyria. The Kogon et al v. Google lawsuit argues that the issue is not just training data, but a full chain from YouTube to ContentID to AI-generated music.
Google has sued the alleged China-based Outsider Enterprise network, saying it used Gemini to help build phishing sites. Google links the operation to 9,000 fake websites, 1 million fraudulent URLs and 2.5 million text messages in two weeks.
Microsoft showed Project Ire classifying a 253 KB LOTUSLITE DLL variant as malware even though VirusTotal showed only 1 of 72 vendors detecting it on May 28, 2026. The important part is not attribution, but an agent reading behavior instead of matching IOC lists.
A coalition of U.S. state attorneys general is investigating OpenAI, according to TechCrunch and The Wall Street Journal, with New York's attorney general serving a subpoena. The questions reach beyond model safety into advertising, engagement, health data, minors and seniors.
OpenAI and Broadcom have unveiled Jalapeño, a custom LLM inference chip planned for deployment in 2026. This is not a clean break with Nvidia. It is an attempt to put cost per token under OpenAI's own control.
China’s Z.ai released the open-weight GLM-5.2, and independent cyber tests put it close to closed frontier models in specific bug-finding tasks. For security teams, the real shift is that useful vulnerability-search capability is moving outside tightly controlled US model APIs.
Elizabeth Warren and Mary Gay Scanlon are preparing an updated Health and Location Data Protection Act that would ban sales of health and location data, including information shared with AI chatbots. For AI companies, the warning is simple: a medical prompt in ChatGPT or Claude may become regulated residue, not free fuel for the data market.
Zvi Mowshowitz disputes a Wall Street Journal framing that China has matched Anthropic in cybersecurity, arguing that it blurs narrow bug finding tasks with autonomous vulnerability discovery and exploit chaining. For security teams, the lesson is colder: a model can win one test and still not be the same weapon in production.
OpenAI is teasing a Codex-focused hardware device for July 15, built with Work Louder, according to The Verge. The interesting part is not the gadget itself, but the move to turn AI coding from a prompt habit into a physical workflow habit.
Companies are pushing Claude, Codex and other coding agents into extremely terse replies to cut token bills. Caveman is not an elegant UX trend, but an accounting reaction to agents rereading their own verbosity on every later step.
Microsoft Research described generative causal testing, a method that turns brain prediction models into readable hypotheses and checks them with fMRI. The interesting move is not prediction alone, but using AI to return testable theory to scientists.
Google Research introduced TabFM, a zero-shot foundation model for classification and regression on tables. The product hook is BigQuery ML: some work around feature engineering, tuning and training may move into a SQL call.
Google released Nano Banana 2 Lite, a faster and cheaper image generation model reportedly producing images in about 4 seconds and under 4 cents per 1000 images. The important shift is not gallery quality, but mass programmable visual iteration.
BAIR published a showcase of 33 Ph.D. graduates from its 2026 class. The celebration doubles as a map of people moving into robotics, LLM agent systems, AI safety and AI for science.
Microsoft is showing SkillOpt, an optimizer that improves an agent skill file without changing model weights. For teams building agents, the important part is the validation gate, not another layer of prompt mysticism.
