Lilith · selected stories
What is actually happening in AI. Selected stories, context and opinion without the promotional noise.
Atom feed ↗Microsoft Research describes SocialReasoning-Bench, a benchmark testing whether AI agents genuinely act in the user's best interest. Key finding: agents complete tasks technically, but do not consistently improve outcomes for the person, even when explicitly instructed to.
LWiAI Podcast episode 246 from 26 May 2026 is a map, not a single thesis. Google I/O, coding agents, legal pressure around OpenAI and safety research landed in the same week and sketch four simultaneous pressures on the AI market.
Microsoft Research released Data Formulator 0.7, an analytics workspace where AI agents assist with exploration, transformation and visualization of enterprise data. The key question is whether the agent handles messy, permissioned data outside the demo.
SQLite added an AGENTS.md file with a blunt rule for people pointing AI agents at the codebase: agentic code is not accepted, but high-quality reproducible bug reports can be useful. A small file, but a big signal for critical open source maintenance.
Gartner estimates that over 40% of agentic AI projects will be cancelled by the end of 2027 because of cost, unclear value or weak risk controls. The signal is not that agents are dead. It is that unmanaged PoCs are entering the bill, governance and accountability wall.
Simon Willison highlighted Boris Mann's point that saying "11 AI agents" is meaningless by itself. It says about as much as counting spreadsheets or browser tabs. The useful questions are outcomes, responsibility boundaries, workflow, handoff, observability, failure handling, permissions and human review.
OpenAI says Gartner named Codex a Leader in enterprise AI coding agents. For companies, this is mainly a procurement and governance signal, not proof of technical superiority.
OpenAI, Thrive Holdings and Crete built Tax AI for more than 30 accounting firms. The pilot processed 7,000 returns, saves about one third of practitioner time and improved sharply within six weeks through a feedback loop powered by Codex.
Warp is positioning the terminal as an agentic development environment rather than a command line wrapper. By open sourcing its client with OpenAI as a founding sponsor and leaning on GPT-5.5, it wants developers to set objectives and review outcomes while agents plan, code, test and open pull requests.
PromptArmor researchers demonstrated an attack chain in which Microsoft Copilot Cowork can help exfiltrate Microsoft 365 files through prompt injection. This is not only a product bug, but a warning for any agentic system with delegated permissions.
OpenAI and Dell want to bring Codex closer to enterprise data, hybrid infrastructure, and on-prem environments. Less flashy than a demo, much more important for enterprise adoption.
OpenAI is moving Education for Countries toward national AI education programs. This is not only about ChatGPT access, but about shaping infrastructure, training, and operating habits around AI in the public sector.
OpenAI published lessons from Parameter Golf: more than 1,000 participants, over 2,000 submissions, a 16 MB artifact limit, and 10 minutes of training on 8x H100. The important part is not only model compression. AI coding agents changed the tempo of research iteration.
AgentMail provides real email inbox infrastructure for AI agents: inbox creation, sending, receiving, threads, attachments, webhooks, WebSockets, search, custom domains and MCP integration. The company announced a 6M USD seed round led by General Catalyst with Y Combinator participation.
Simon Willison tested Claude Code for web: Anthropic wrapped the local Claude Code experience in a hosted sandbox and made it usable from web and mobile. The important shift is not a more capable model, but a workflow change: coding agents become more valuable when they can run asynchronously and safely away from your laptop.
DeepMind introduced AlphaEvolve as a Gemini-powered evolutionary loop that automatically discovers better algorithms. Concrete production results: 30 % fewer errors in genomics, 20 % lower write amplification for Spanner, Klarna doubled transformer training speed.
OpenAI details how Codex runs in isolated environments: per-repo sandboxes, network restrictions, approval gates, and agent-native telemetry for safe enterprise adoption.
Cisco and OpenAI describe deploying Codex as an agent in enterprise engineering workflows: build automation, defect fixes, and a shift toward agent-native development.
The GPT-5.1-Codex-Max system card describes two safety layers: model-level safety training and prompt injection protection, and product-level sandboxing with configurable network access.
Google DeepMind released Gemini 2.5 Computer Use in preview: a specialized model for agents that drive user interfaces. Unlike general-purpose Gemini 2.5 Pro, this model was trained specifically for screen interaction, not just reasoning about it.
