Lilith · selected stories
What is actually happening in AI. Selected stories, context and opinion without the promotional noise.
Atom feed ↗BAIR argues that GPT-4-class inference has fallen from roughly $30 per million tokens to under $1. For data teams, the bottleneck moves from the model itself to memory, coordination and control around agents.
Lilian Weng’s 28 minute analysis frames harness engineering as the practical path to self-improvement: not a model secretly rewriting its weights, but the surrounding system that manages tools, context, memory and evals.
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.
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.
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.
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.
Sebastian Raschka shows a local coding-agent stack: an open-weight model in Ollama, a harness that edits code and runs commands, and your own machine instead of a Claude Code or Codex subscription. For teams, the interesting part is not nostalgia for localhost, but protection against pricing, limits and model changes they do not control.
Pew Research Center found that 49% of U.S. adults use AI chatbots and 24% use them daily. The same survey shows the adoption ceiling: 63% say AI is advancing too quickly and 70% expect weaker personal data security.
Microsoft Research and partners described Talos, an open-source tool for automated genomic reanalysis in rare disease. In a prospective cohort of almost 5,000 patients, it added diagnoses in 5.1 % of cases.
Google Research examines why chain-of-thought helps LLMs answer simple factual questions. The study on Gemini 2.5 and Qwen3-32B points to two mechanisms: extra computation in generated tokens and factual priming.
Lilian Weng revisits scaling laws and shows why they are most useful as a tool for allocating compute, data and parameters. The practical lesson is sober: extrapolation helps only while you remember how small the underlying experiments were.
Google has rolled out frozen Multi-Token Prediction for Gemini Nano v3 on Pixel 9 and Pixel 10 phones. The practical point is simple: faster local AI without replacing the base model or adding a separate drafter into memory.
Google introduced Gemma 4 12B as a unified, encoder-free multimodal model designed for high performance directly on a laptop. The practical question is whether a 12B model can deliver enough quality for local or edge use without heavy cloud infrastructure.
Research described by 404 Media says a 13 word snippet of retrieved text from sites such as Reddit, Wikipedia, Quora or Facebook can push AI agents toward spam or scam output. For AI search, that turns SEO into a prompt injection and user-generated content moderation problem.
Google introduced an agentic RAG system for Gemini Enterprise Agent Platform that checks whether it has enough context before answering. For companies, that brake matters more than another polished retrieval layer.
Sebastian Raschka published a curated list of LLM papers from January to May 2026. It is a useful filter for teams trying to separate the research feed from topics that matter for architecture, agents and inference.
Google is launching Gemini 3.5 Live Translate for near real-time speech-to-speech translation across more than 70 languages. Users will see convenience first, but companies will care about latency, audit and trust in a voice speaking for someone else.
Perplexity describes Search as Code: an architecture where an agent does not call one monolithic search engine, but assembles a retrieval pipeline as code. The point is not a nicer search API. It is control over how evidence is found, filtered and verified.
OpenAI updated GPT-Rosalind for life sciences and is offering it in research preview to selected organizations globally. The more important move is not the scorecard, but the attempt to connect a model, Codex and bioinformatics tools into an auditable workflow.
