Google I/O transforms search into an AI agent platform, Gemini 3.5 Flash goes GA, and Tesla's Texas refinery faces pollution accusations

Google

Gemini 3.5 Flash: Frontier Intelligence with Action

Gemini 3.5 Flash is Google's centerpiece I/O release, going directly to general availability across the Gemini app, Search, and enterprise APIs without a preview period. The model targets agentic workflows: 1M token context, 65k max output, four thinking levels ('minimal/low/medium/high'), and thought preservation across multi-turn conversations. Speed claims of 4x faster than comparable frontier models — and 12x in Antigravity — position it as the engine for subagent orchestration rather than a single-session assistant. The price, however, has risen sharply: at $1.50/M input and $9/M output, it's roughly 3x costlier than the prior Gemini 3 Flash Preview. Independent benchmarks from Artificial Analysis put the total benchmark cost at $1,551 — actually higher than Gemini 3.1 Pro's $892 — drawing criticism that 'Flash' is becoming a branding convenience rather than a description of pricing tier. Gemini 3.5 Pro is expected next month.

Google DeepMind

Gemini Omni: Multimodal Generation and Editing

Gemini Omni combines Gemini's reasoning capabilities with generative media production, initially focused on video. Users can edit existing footage through natural language prompts, swap characters or objects while preserving scene consistency, convert sketches to video, and make multi-turn edits that maintain continuity. The model launches today for paid Gemini users in the app and Flow, with YouTube Shorts/Create access rolling out this week and developer APIs following in coming weeks. Google's strategic bet is that owning both the understanding model and the generation stack creates an advantage over point-solution video tools: the same model that reasons about a video can also edit it, enabling agentic workflows where video creation is one step in a broader task rather than a separate specialized system.

Autonocion

Tesla's Texas Lithium Refinery Discharges 231,000 Gallons of Polluted Wastewater Daily

Tesla's Robstown, Texas lithium refinery — North America's first commercial-scale spodumene-to-lithium-hydroxide facility — is discharging roughly 231,000 gallons of wastewater daily that independent lab tests found to contain hexavalent chromium, arsenic, and other heavy metals. This directly contradicts Tesla's marketing of the facility as using an 'acid-free clean process.' The region near Corpus Christi is already water-stressed, and regulatory gaps appear to have enabled the situation: state testing initially failed to screen for heavy metals, and Tesla's permit doesn't require lithium monitoring. The disclosure is significant not just as an environmental story but as a signal about the hidden costs of domestic critical mineral production — supply chain localization being pushed as a strategic priority comes with environmental tradeoffs rarely priced into the policy conversation.

Nate Silver

Disney Erased FiveThirtyEight

Nate Silver recounts how Disney gradually destroyed and ultimately erased FiveThirtyEight, deleting roughly 200,000 person-hours of published work when it shut the site down in 2025. Despite producing approximately 20 stories weekly over a decade, Disney treated the operation as a rounding error rather than a meaningful product — rejecting a paywall proposal Silver estimated could generate $5M annually, failing acquisition negotiations with The Athletic, and ultimately refusing to transfer intellectual property rights that would have allowed the archive to survive elsewhere. The erasure illustrates a structural fragility for analytical journalism attached to large conglomerates: a property that doesn't align with the parent's core business has no institutional champion when cuts come, and the parent retains copyright that can simply be taken offline. The FiveThirtyEight archive was not sold, donated, or archived — it was removed.

Emmi AI

Mistral AI Acquires Emmi AI to Create the Leading AI Stack for Industrial Engineering

Mistral AI has acquired Emmi AI, an Austrian company specializing in Physics AI for industrial engineering sectors including automotive, semiconductors, and aerospace. Emmi's 30+ researchers join Mistral and establish Linz as a new office, expanding Mistral's presence beyond France and the UK. The acquisition addresses a notable gap in current frontier model capabilities: physical world reasoning for domains like structural simulation, materials science, and manufacturing process optimization. Mistral's positioning as a European AI champion makes industrial AI a natural adjacency — European manufacturing is among the most technically sophisticated in the world, and enterprise appetite for US-dominated AI platforms in critical industrial settings is lower than in consumer contexts.

Cloudflare

Announcing Claude Managed Agents on Cloudflare

Cloudflare and Anthropic have integrated Claude Managed Agents with Cloudflare Sandboxes, letting developers run Claude's agent reasoning on Anthropic's infrastructure while routing code execution through Cloudflare's environment. The 'brain/hands' separation is the key architectural concept: Claude plans and reasons on Anthropic's platform; tool calls, code execution, file access, and browser sessions run in Cloudflare infrastructure. The practical advantages are security and scale — Cloudflare's outbound proxy layer lets teams inject credentials, prevent data exfiltration, and enforce per-agent egress policies, and the lightweight V8 isolate option (vs. full microVMs) enables tens of millions of concurrent agents at costs that VM-per-agent approaches can't match. The integration ships with built-in Browser Run for web interaction, email send/receive, private service connectivity via Workers VPC, and support for custom tool extensions.

Turso

How We Used Quint to Find Over 10 Bugs in SQLite

Turso used Quint, a formal specification and model-checking tool, to find over 10 bugs in SQLite by defining expected behavior of specific C API contracts, generating execution traces from those models, and comparing against actual database behavior. Bugs found included crashes in deserialization during read transactions, incorrect optimization rewrites in EXISTS-to-JOIN transformations, UPDATE planning regressions with correlated subqueries, and mutex alignment issues — the kind of subtle edge-case bugs that unit tests rarely surface. SQLite runs on virtually every phone and browser in the world; finding correctness bugs in it has broad reliability implications. The methodology is notable: rather than fuzzing with random inputs, Quint generates semantically meaningful traces derived from formal specs of what the API is supposed to do, producing higher-signal test cases against which actual behavior is verified.

Martin Fowler

Maintainability Sensors for Coding Agents

Birgitta Böckeler describes a sensor-based approach to keeping codebases maintainable when coding agents generate a significant share of the code. The core observation is that AI agents make characteristically different errors than humans — excessively complex functions, ignored lint rules, architectural inconsistencies — and traditional static analysis (type checkers, ESLint, dependency analysis) deployed at the right harness checkpoints can catch these before they accumulate as technical debt. The 'sensor' framing positions linting not as style enforcement but as a feedback signal that agents can incorporate into self-correction loops: run a linter, observe failures, retry. The larger argument is that the harness around a coding agent — what feedback it receives about the codebase state — is as important to output quality as the underlying model.

nesbitt.io

Dumb Ways for an Open Source Project to Die

Andrew Nesbitt catalogs the non-obvious ways open source projects effectively die while their packages continue circulating in registries. Beyond maintainer burnout, the article covers lost credentials that prevent publishing security fixes, dependency rot, license changes that cut off downstream adoption, projects compromised by malicious package takeovers, and ecosystem shifts that leave projects technically alive but practically orphaned. The central insight is that package managers track published artifacts, not project health — a dependency can be unmaintained for years while appearing healthy in lockfiles. This creates systemic supply chain risk: developers who inherit codebases rarely re-examine transitive dependency health, and the absence of visible failure signals means unmaintained packages accumulate in production systems long after they become liabilities.