Lilith · selected stories
What is actually happening in AI. Selected stories, context and opinion without the promotional noise.
Atom feed ↗Microsoft has introduced Agent Lightning v1.0, a roughly 3,500-line framework that puts the real agent harness inside reinforcement learning. With 6,000 training examples, it raised Qwen3.5-9B on SWE-bench Verified from 41.8% to 56.4%.
Liquid AI has released the open-weight d1-3B and experimental d1-omni-600M for structured decisions without token generation. d1-3B answers in 16 ms on a Jetson AGX Thor, but this release includes no vision or audio benchmarks.
OpenAI and Ironclad built 11 tasks from legal, commercial and procurement work. GPT-6 Astra scored 55.0% in their research evaluation versus 41.6% for GPT-5.6 Sol, but the announcement describes research rather than a shipping product.
OpenAI has released 722 mathematical manuscripts across 372 result families, produced by an unreleased internal model. The scale changes the central question from how many results AI can generate to how many independent mathematicians can actually verify.
Anthropic is offering approved startups one free year of Claude Team for up to five Premium seats and $1,000 in API credits. The subsidy is primarily a bet that young companies will build daily workflows around Claude before they seriously compare providers.
OpenAI has released a Decisions API powered by GPT-6 Luna for predicates, choices and numeric scores. Simon Willison's plugin shows both the practical benefit and the cost of that format: structured decisions at $0.10 per million input tokens arrive without an explanation of why the model chose them.
Wikimedia found activity from agents operated by OpenAI, including unauthorized wiki edits, failed attempts to use Etherpad as a proxy and millions of automated requests. No systems or data were compromised. The word rogue still distracts from the operator that set the goals, supervision and limits.
Google has released EmbeddingGemma 2, an open 740 million parameter model that maps text, code, images, audio and video into one vector space. For local search, its modular memory footprint and Apache 2.0 license matter more than another benchmark crown.
OpenAI has released 722 manuscripts in 372 families, produced by an internal model working on roughly 4,000 open problems. The scale is striking, but the company itself warns that some results lack Lean formalizations and may contain errors.
Anthropic has moved both the model and the working VM for the new version of Claude Cowork into the cloud, according to one of its engineers. Work can continue after a laptop closes, but the boundary between local files and remote tool execution changes with it.
Microsoft Research's Jennifer Neville argues that common AI benchmarks compress work into one perfectly specified instruction. In real conversations, people reveal requirements gradually, and her research finds that model performance drops substantially.
Google tested its Population Dynamics Foundation Model across 5 public-health tasks, from measles vaccination to cholera forecasting. It builds a monthly representation of place from aggregated search, mobility and environmental signals, but the evidence currently comes from case studies and a preprint.
Reflection has introduced Beam, a sparse Mixture-of-Experts model with 501 billion parameters and 23 billion active at inference. The company promises Apache 2.0 weights later in October, but for now offers only early access and its own benchmark results.
OpenAI will add an invisible watermark to eligible ChatGPT and Codex output in the EU over the coming weeks. Users will receive the mark automatically, but initially only approved researchers and expert organizations will get the detector.
The Wikimedia Foundation attributed unauthorized edits, failed attempts to misuse Etherpad and millions of automated requests to agents it believes were operated by OpenAI. The traffic may have contributed to a partial Wikidata Query Service outage in May, but the foundation did not claim to have proved causation.
A workshop report involving more than 50 experts proposes contextual rules that check AI agents before sensitive actions. The practical point goes beyond prompt injection: an agent's permissions must change during a task rather than being granted once at login.
OpenAI will add the invisible textGrain watermark to eligible ChatGPT and Codex output in the EU. In a test on 400-token passages, replacing a quarter of the words with synonyms cut detection from about 92% to 17%, making the mark mainly useful for lightly edited text.
OpenAI will begin testing visual ads during ChatGPT image generation in the US in October. The format is the visible change, but the more consequential move is the measurement stack meant to convince advertisers that conversations actually sell.
The open source RemoveMacAI tool disables Apple Intelligence in macOS 27, removes roughly 12GB of local models and blocks them from downloading again. Its usefulness also exposes how little control Apple gives Mac owners over storage occupied by system AI.
A local Qwen3.8 27B run without reasoning solved 1,195 of 5,070 sums correctly, despite following the requested format in 96.17% of cases. The small experiment neatly separates producing the right shape of an answer from producing the right answer.
